Introduction
Cluster-randomized trials, in which groups of individuals rather than individual participants are randomized to treatment conditions, are widely used to evaluate interventions in health, biomedical, and social science research. The primary analysis of a cluster-randomized trial typically assesses whether an intervention is effective by estimating its total effect on a final outcome (Hayes and Moulton, 2017; Li et al., 2025). The total effect alone, however, does not reveal the mechanisms through which the intervention operates. Causal mediation analysis complements this overall assessment by examining pathways through post-treatment intermediate variables, referred to as mediators [8 ] . By distinguishing indirect effects operating through the mediator from direct effects operating through other pathways, causal mediation analysis can deepen our understanding of how an intervention works and inform the design and refinement of future interventions.
A growing literature has investigated causal mechanisms in cluster-randomized trials. Relative to simpler individually randomized trials, mediation analysis in cluster-randomized trials is further complicated by within-cluster interference, under which one individual’s mediator may affect not only that individual’s own outcome but also the outcomes of other members of the same cluster. In the single-mediator setting, VanderWeele et al. [17 ] decomposed the natural indirect effect into an individual mediation effect, capturing pathways through an individual’s own mediator, and a spillover mediation effect, capturing pathways through the mediators of other individuals in the same cluster. Cheng and Li [2 ] established semiparametric efficiency theory for inferring these effect estimands and developed efficient estimators that accommodate data-adaptive machine-learning methods for nuisance modeling. Subsequent work has extended this literature to settings with multiple mediators. For two unordered mediators, Ohnishi and Li [12 ] decomposed the mediator-specific exit indirect effect [20 ] into exit individual and exit spillover mediation effects, thereby providing a finer characterization of individual and spillover pathways, and developed a Bayesian nonparametric approach for estimation and inference. Tong et al. [16 ] generalized these exit mediation effects to an arbitrary number of mediators without requiring a known causal ordering among them and constructed efficient debiased machine-learning estimators. Together, these developments have substantially expanded the range of causal mechanisms that can be studied in cluster-randomized trials. Existing methods, however, mostly focus on settings without treatment-induced confounding of the mediator–outcome relationship.
Treatment-induced mediator-outcome confounding arises when the treatment affects an intermediate post-treatment variable that, in turn, influences both the mediator and the outcome. In this setting, natural mediation effects are generally not identified under sequential ignorability assumptions, and identifying them requires additional unverifiable restrictions on counterfactual quantities [1 , 21 ] (Tchetgen Tchetgen and VanderWeele 2014). One approach to this problem is to target the natural mediation effects and assess their sensitivity to the unidentified cross-world dependence between the potential mediator and the potential outcome. For example, Ohnishi et al. [13 ] developed a Bayesian sensitivity analysis framework for natural mediation effects in cluster-randomized trials subject to post-treatment confounding, using a Gaussian copula model to parameterize the residual dependence between counterfactual mediator and outcome under different treatment conditions. As a complementary approach, we target interventional analogues of natural mediation effects, which replace individual-specific natural mediator values with stochastic draws from treatment-specific mediator distributions [14 , 18 , 19 ] . In the independent data setting, these effects can be identified in the presence of observed treatment-induced mediator-outcome confounding under appropriate exchangeability and positivity conditions, without specifying the unidentified cross-world dependence required for natural effects [5 , 7 ] . In cluster-randomized trials, an individual’s outcome may depend on both their own mediator and those of other cluster members, motivating interventions on the joint mediator distribution to assess how changes in marginal distributions and their dependence affect the outcomes. The resulting contrasts quantifies the potential value of hypothetical mediator interventions by comparing mean outcomes under treatment and control mediator distributions while holding treatment fixed, and can differ from the treatment effects carried through individual causal pathways that natural effects explain [11 ] . Interventional effects provide a complementary, intervention-oriented perspective on the role of mediators.
We seek to develop novel semiparametric methods for interventional mediation analysis in cluster-randomized trials with treatment-induced mediator-outcome confounding. Our contributions are threefold. First, we define interventional direct and indirect effects by drawing the within-cluster mediator vector from its treatment-specific distribution conditional on baseline information, and further decompose the interventional indirect effect into individual and spillover components operating through an individual’s own mediator and the mediators of other cluster members, respectively. The proposed estimands allow arbitrary within-cluster interference and accommodate both cluster-average and individual-average target populations through general cluster-level weighting [9 ] . To our knowledge, this is the first effort to address interventional effect estimands in the presence of mediator spillover. Second, we derive the efficient influence functions and construct cluster-level cross-fitted estimators that accommodate flexible nuisance estimation [3 ] . Exact bias decompositions characterize their multiple robustness properties. The estimators are consistent under the required multiple robustness conditions for the nuisance functions. And, when all nuisance estimators are consistent and the required product rate conditions hold, the estimators are asymptotically normal and semiparametrically efficient. Third, to facilitate practical implementation, we further develop a lower-dimensional reparameterization of the efficient influence function that replaces high-dimensional post-treatment confounder densities and integration over the full mediator vector with density ratios, treatment classifiers, univariate mediator densities, and conditional mean regressions. In particular, we adapt the augmented-classification approach of Rudolph et al. [15 ] to estimate the key density ratio without directly modeling the corresponding high-dimensional conditional densities. We establish corresponding robustness and efficiency results under compatibility conditions on the auxiliary regressions and examine both implementations in simulation studies. We illustrate the proposed methods in the PPACT cluster-randomized trial of cognitive behavioral therapy for chronic pain.
Notation, estimands and assumptions
Consider a cluster-randomized trial with K K K clusters, indexed by i ∈ { 1 , … , K } i \in\{1,\ldots,K\} i ∈ { 1 , … , K } . Let N i N_i N i denote the size of cluster i i i , and let A i ∈ { 0 , 1 } A_i \in\{0,1\} A i ∈ { 0 , 1 } denote its treatment assignment, with A i = 1 A_i=1 A i = 1 indicating treatment and A i = 0 A_i=0 A i = 0 indicating control. All individuals in a cluster receive the same assignment. Baseline information consists of cluster-level covariates V i ∈ R d V × 1 V_i \in\mathbb{R}^{d_V \times1} V i ∈ R d V × 1 and individual-level covariates X i j ∈ R d X × 1 X_{ij} \in\mathbb{R}^{d_X \times1} X ij ∈ R d X × 1 , for j ∈ { 1 , … , N i } j \in\{1,\ldots,N_i\} j ∈ { 1 , … , N i } . We also write X i = { V i , X i 1 , … , X i N i } X_i = \{V_i,X_{i1},\ldots,X_{iN_i}\} X i = { V i , X i 1 , … , X i N i } for the full baseline covariate information in the cluster, with X i = [ X i 1 , … , X i N i ] T ∈ R N i d X × 1 X_i = [X_{i1},\ldots,X_{iN_i}]^T \in\mathbb{R}^{N_i d_X \times1} X i = [ X i 1 , … , X i N i ] T ∈ R N i d X × 1 . After assignment, we observe post-treatment confounders Z i j ∈ R d Z × 1 Z_{ij} \in\mathbb{R}^{d_Z \times1} Z ij ∈ R d Z × 1 , followed by a mediator M i j ∈ R M_{ij} \in\mathbb{R} M ij ∈ R and an outcome Y i j ∈ R Y_{ij} \in\mathbb{R} Y ij ∈ R . Z i j Z_{ij} Z ij may be affected by treatment and may, in turn, affect both the mediators and the outcomes. Denote Z i = [ Z i 1 , … , Z i N i ] T ∈ R N i × d Z Z_i = [Z_{i1},\ldots,Z_{iN_i}]^T \in\mathbb{R}^{N_i \times d_Z} Z i = [ Z i 1 , … , Z i N i ] T ∈ R N i × d Z , M i = [ M i 1 , … , M i N i ] T ∈ R N i × 1 M_i = [M_{i1},\ldots,M_{iN_i}]^T \in\mathbb{R}^{N_i \times1} M i = [ M i 1 , … , M i N i ] T ∈ R N i × 1 , and Y i = [ Y i 1 , … , Y i N i ] T ∈ R N i × 1 Y_i = [Y_{i1},\ldots,Y_{iN_i}]^T \in\mathbb{R}^{N_i \times1} Y i = [ Y i 1 , … , Y i N i ] T ∈ R N i × 1 . For an individual j j j , M i ( − j ) ∈ R ( N i − 1 ) × 1 M_{i(-j)} \in\mathbb{R}^{(N_i-1) \times1} M i ( − j ) ∈ R ( N i − 1 ) × 1 denotes the mediators of all other individuals in the same cluster. The observed data for cluster i i i are O i = { N i , C i , A i , Z i , M i , Y i } O_i = \{N_i,C_i,A_i,Z_i,M_i,Y_i\} O i = { N i , C i , A i , Z i , M i , Y i } . We suppress the cluster index “i i i ” when referring to a generic cluster, using a “∙ \bullet ∙ ” in its place for individual-level variables; for example, we write M M M and M ∙ j M_{\bullet j} M ∙ j for M i M_i M i and M i j M_{ij} M ij .
Let Z ∙ j ( a ) Z_{\bullet j}(a) Z ∙ j ( a ) denote the post-treatment confounders for individual j j j under cluster treatment a ∈ { 0 , 1 } a \in\{0,1\} a ∈ { 0 , 1 } , and let M ∙ j ( a , z ) M_{\bullet j}(a,z) M ∙ j ( a , z ) denote the potential mediator when treatment is set to a a a and the full cluster confounder vector is set to z = { z ∙ j , z ∙ ( − j ) } z = \{z_{\bullet j},z_{\bullet(-j)}\} z = { z ∙ j , z ∙ ( − j ) } . Similarly, Y ∙ j ( a , z , m ) = Y ∙ j ( a , z ∙ j , z ∙ ( − j ) , m ∙ j , m ∙ ( − j ) ) Y_{\bullet j}(a,z,m) = Y_{\bullet j}(a,z_{\bullet j},z_{\bullet(-j)},m_{\bullet j},m_{\bullet(-j)}) Y ∙ j ( a , z , m ) = Y ∙ j ( a , z ∙ j , z ∙ ( − j ) , m ∙ j , m ∙ ( − j ) ) is the potential outcome when treatment, confounders, and mediators are set to a a a , z z z , and m = { m ∙ j , m ∙ ( − j ) } m = \{\mathbf{m}_{\bullet j}, \mathbf{m}_{\bullet(-j)}\} m = { m ∙ j , m ∙ ( − j ) } , respectively. Dependence on the full vectors z z z and m m m permits an individual’s outcome to depend on both their own and other cluster members’ post-treatment variables. The potential variables are indexed only by interventions within the same cluster, thereby excluding interference across clusters.
We assume composition: Y ∙ j ( a , m ) = Y ∙ j ( a , Z ( a ) , m ) Y_{\bullet j}(a,m) = Y_{\bullet j}(a,Z(a),m) Y ∙ j ( a , m ) = Y ∙ j ( a , Z ( a ) , m ) and M ∙ j ( a ) = M ∙ j ( a , Z ( a ) ) M_{\bullet j}(a) = M_{\bullet j}(a,Z(a)) M ∙ j ( a ) = M ∙ j ( a , Z ( a )) , where Z ( a ) Z(a) Z ( a ) and M ( a ) M(a) M ( a ) collect the corresponding potential variables over cluster members. Thus, an intervention on the mediators leaves the post-treatment confounders at their natural values under treatment a a a . We also assume Y ∙ j ( a ) = Y ∙ j ( a , Z ( a ) , M ( a ) ) = Y ∙ j ( a , Z ( a ) , M ( a , Z ( a ) ) ) Y_{\bullet j}(a) = Y_{\bullet j}(a,Z(a),M(a)) = Y_{\bullet j}(a,Z(a),M(a,Z(a))) Y ∙ j ( a ) = Y ∙ j ( a , Z ( a ) , M ( a )) = Y ∙ j ( a , Z ( a ) , M ( a , Z ( a ))) , i.e., the potential outcome under treatment alone equals the potential outcome obtained by setting both post-treatment vectors to their natural values under that treatment.
To define interventional mediation effects, let G ( a ) G(a) G ( a ) be a random draw from the conditional distribution of M ( a ) M(a) M ( a ) given ( C , N ) (C,N) ( C , N ) . Conditional on ( C , N ) (C,N) ( C , N ) , we construct G ( 0 ) G(0) G ( 0 ) and G ( 1 ) G(1) G ( 1 ) to be independent of one another and of all potential variables. Each G ( a ) G(a) G ( a ) is drawn as a joint cluster mediator vector , allowing the dependence among its components under treatment a a a . This construction extends stochastic mediator interventions for intermediate confounding to clustered data setting [5 , 15 ] . Specifically, we define the interventional direct effect (IDE) and interventional indirect effect (IIE) by
I D E = h ( θ ( 1 , 0 ) , θ ( 0 , 0 ) ) , I I E = h ( θ ( 1 , 1 ) , θ ( 1 , 0 ) ) . \mathrm{IDE} = h(\theta(1,0),\theta(0,0)), \qquad\mathrm{IIE} = h(\theta(1,1),\theta(1,0)). IDE = h ( θ ( 1 , 0 ) , θ ( 0 , 0 )) , IIE = h ( θ ( 1 , 1 ) , θ ( 1 , 0 )) .
where h h h is a known contrast function specifying the effect scale, and
θ ( a , a ′ ) = E [ W N ∑ j = 1 N Y ∙ j ( a , G ∙ j ( a ′ ) , G ∙ ( − j ) ( a ′ ) ) ] E [ W ] , for a , a ′ ∈ { 0 , 1 } . (1) \theta(a,a') = \frac{\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}Y_{\bullet j}\left(a,G_{\bullet j}(a'),G_{\bullet(-j)}(a')\right)\right]}{\mathbb{E}[W]}, \qquad\text{for } a,a' \in\{0,1\}. \tag*{(1)} θ ( a , a ′ ) = E [ W ] E [ N W ∑ j = 1 N Y ∙ j ( a , G ∙ j ( a ′ ) , G ∙ ( − j ) ( a ′ ) ) ] , for a , a ′ ∈ { 0 , 1 } . (1)
The weight W W W is a user-specified function of { N , V } \{N,V\} { N , V } . As illustrated in Li et al. [10 ] , W W W specifies how cluster-specific mean outcomes contribute to the target population mean. In particular, W = 1 W=1 W = 1 gives equal weight to clusters, whereas W = N W=N W = N gives weight proportional to cluster size, yielding cluster-average and individual-average interventional effect estimands, respectively. Possible scale functions include h ( x , y ) = x / y h(x,y)=x/y h ( x , y ) = x / y and h ( x , y ) = ( x / ( 1 − x ) ) / ( y / ( 1 − y ) ) h(x,y)=(x/(1-x))/(y/(1-y)) h ( x , y ) = ( x / ( 1 − x )) / ( y / ( 1 − y )) , indicating risk ratio and odds ratio scales. In this paper, we focus on the mean difference scale with h ( x , y ) = x − y h(x,y)=x-y h ( x , y ) = x − y .
In θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) , treatment is set to a a a , the confounders take their natural values Z ( a ) Z(a) Z ( a ) , and the entire mediator vector is independently drawn from its distribution under a ′ a' a ′ among clusters with the same ( C , N ) (C,N) ( C , N ) . The IDE therefore changes treatment while holding the mediator intervention at the control distribution; it allows treatment to affect the outcome through the post-treatment confounders with the mediator draw held fixed in distribution. The IIE holds treatment at 1 and changes the joint mediator distribution from control to treatment. These effects sum to θ ( 1 , 1 ) − θ ( 0 , 0 ) \theta(1,1)-\theta(0,0) θ ( 1 , 1 ) − θ ( 0 , 0 ) . This sum need not equal the total effect of treatment alone. That is, even when a = a ′ a=a' a = a ′ , the stochastic intervention breaks the natural dependence between Z ( a ) Z(a) Z ( a ) and M ( a ) M(a) M ( a ) conditional on ( C , N ) (C,N) ( C , N ) . This distinction between interventional and total treatment effects is discussed by Díaz et al. [5 ] and Rudolph et al. [15 ] .
To distinguish changes involving an individual’s own mediator from those involving the mediators of other cluster members, we further define the interventional individual mediation effect (IIME) and the interventional spillover mediation effect (ISME) as
I I M E = h ( τ , θ ( 1 , 0 ) ) , I S M E = h ( θ ( 1 , 1 ) , τ ) , \mathrm{IIME} = h(\tau,\theta(1,0)), \qquad\mathrm{ISME} = h(\theta(1,1),\tau), IIME = h ( τ , θ ( 1 , 0 )) , ISME = h ( θ ( 1 , 1 ) , τ ) ,
where
τ = E [ W N ∑ j = 1 N Y ∙ j ( 1 , G ∙ j ( 1 ) , G ∙ ( − j ) ( 0 ) ) ] E [ W ] . (2) \tau= \frac{\mathbb{E}\left[ \frac{W}{N}\sum_{j=1}^{N}Y_{\bullet j}\left(1,G_{\bullet j}(1),G_{\bullet(-j)}(0)\right)\right]}{\mathbb{E}[W]}. \tag*{(2)} τ = E [ W ] E [ N W ∑ j = 1 N Y ∙ j ( 1 , G ∙ j ( 1 ) , G ∙ ( − j ) ( 0 ) ) ] . (2)
For each individual j j j , the intervention in τ \tau τ draws their own mediator from the treatment distribution and the vector of other members’ mediators from the control distribution, independently conditional on ( C , N ) (C,N) ( C , N ) . The parameter averages outcomes under these individual-specific interventions; it does not require a single mediator assignment to implement all of them simultaneously. Note that the IIME and ISME sum exactly to the IIE. Their interpretation also depends on the joint mediator distributions under treatment and control. Moving from θ ( 1 , 0 ) \theta(1,0) θ ( 1 , 0 ) to τ \tau τ changes the own-mediator marginal distribution and removes its dependence on the peer mediator vector under control. Moving from τ \tau τ to θ ( 1 , 1 ) \theta(1,1) θ ( 1 , 1 ) changes the peer marginal distribution and restores the dependence between own and peer mediators under treatment. Consequently, these contrasts can reflect changes in mediator dependence as well as changes in the respective marginal distributions.
Assumption A. We require the following identifying assumptions:
(A1) If A i = a A_i=a A i = a , then Z i = Z i ( a ) Z_i=Z_i(a) Z i = Z i ( a ) . If, in addition, Z i = z Z_i=z Z i = z , then M i = M i ( a , z ) M_i=M_i(a,z) M i = M i ( a , z ) . If A i = a A_i=a A i = a , Z i = z Z_i=z Z i = z , and M i = m M_i=m M i = m , then Y i j = Y i j ( a , z , m ) Y_{ij}=Y_{ij}(a,z,m) Y ij = Y ij ( a , z , m ) for every j ∈ { 1 , … , N i } j \in\{1,\ldots,N_i\} j ∈ { 1 , … , N i } .
(A2) The treatment indicators are mutually independent with P ( A i = 1 ) = π ∈ ( 0 , 1 ) \mathbb{P}(A_i=1)=\pi\in(0,1) P ( A i = 1 ) = π ∈ ( 0 , 1 ) . For ( z , m ) (z,m) ( z , m ) over their valid support, A i ⊥ ⊥ { C i , N i , Z i ( a ) , M i ( a , z ) , Y i j ( a , z , m ) : a ∈ { 0 , 1 } , j = 1 , … , N i } A_i \perp\!\!\!\perp\{C_i,N_i,Z_i(a),M_i(a,z),Y_{ij}(a,z,m):a \in\{0,1\},\ j=1,\ldots,N_i\} A i ⊥ ⊥ { C i , N i , Z i ( a ) , M i ( a , z ) , Y ij ( a , z , m ) : a ∈ { 0 , 1 } , j = 1 , … , N i } .
(A3) For every j j j , a a a , z z z and m m m over their valid support, M i ( a , z ) ⊥ ⊥ Y i j ( a , z , m ) ∣ Z i ( a ) = z , C i , N i M_i(a,z) \perp\!\!\!\perp Y_{ij}(a,z,m)\mid Z_i(a)=z,C_i,N_i M i ( a , z ) ⊥ ⊥ Y ij ( a , z , m ) ∣ Z i ( a ) = z , C i , N i .
(A4) The observed clustered data O 1 , … , O K O_1,\ldots,O_K O 1 , … , O K are mutually independent, with N N N having a finite support in N + \mathbb{N}^{+} N + . Conditional on N N N , the observed-data distribution admits P Y ∣ M , Z , A , C , N × P M ∣ Z , A , C , N × P Z ∣ A , C , N × P A × P C ∣ N P_{Y\mid M,Z,A,C,N}\times P_{M\mid Z,A,C,N}\times P_{Z\mid A,C,N}\times P_A\times P_{C\mid N} P Y ∣ M , Z , A , C , N × P M ∣ Z , A , C , N × P Z ∣ A , C , N × P A × P C ∣ N , with E ( Y ∘ j 2 ) < ∞ \mathbb{E}(Y_{\circ j}^{2})<\infty E ( Y ∘ j 2 ) < ∞ across j ≤ N j\leq N j ≤ N . For every a , a ′ ∈ { 0 , 1 } a,a' \in\{0,1\} a , a ′ ∈ { 0 , 1 } , suppose
f M ∣ A , Z , C , N ( m ∣ a , z , c , n ) > 0 if f Z ∣ A , C , N ( z ∣ a , c , n ) f M ∣ A , C , N ( m ∣ a ′ , c , n ) > 0 , f_{M\mid A,Z,C,N}(m\mid a,z,c,n)>0\quad\text{if}\quad f_{Z\mid A,C,N}(z\mid a,c,n)f_{M\mid A,C,N}(m\mid a',c,n)>0, f M ∣ A , Z , C , N ( m ∣ a , z , c , n ) > 0 if f Z ∣ A , C , N ( z ∣ a , c , n ) f M ∣ A , C , N ( m ∣ a ′ , c , n ) > 0 ,
and f M ∣ A , Z , C , N ( m ∣ 1 , z , c , n ) > 0 \text{and}\quad f_{M\mid A,Z,C,N}(m\mid1,z,c,n)>0 and f M ∣ A , Z , C , N ( m ∣ 1 , z , c , n ) > 0
if f Z ∣ A , C , N ( z ∣ 1 , c , n ) f M ∙ j ∣ A , C , N ( m ∙ j ∣ 1 , c , n ) f M ∙ ( − j ) ∣ A , C , N ( m ∙ ( − j ) ∣ 0 , c , n ) > 0. \text{if}\quad f_{Z\mid A,C,N}(z\mid1,c,n)f_{M_{\bullet j}\mid A,C,N}(m_{\bullet j}\mid1,c,n)f_{M_{\bullet(-j)}\mid A,C,N}(m_{\bullet(-j)}\mid0,c,n)>0. if f Z ∣ A , C , N ( z ∣ 1 , c , n ) f M ∙ j ∣ A , C , N ( m ∙ j ∣ 1 , c , n ) f M ∙ ( − j ) ∣ A , C , N ( m ∙ ( − j ) ∣ 0 , c , n ) > 0.
In addition, the weight W W W is nonnegative and bounded.
Assumption (A1) links the observed post-treatment variables to their potential values, accommodating arbitrary within-cluster interference while excluding interference across clusters. Assumption (A2) specifies independent Bernoulli assignment of clusters. Assumption (A3) supplies the additional mediator–outcome exchangeability condition. Conditional on ( C i , N i ) (C_i,N_i) ( C i , N i ) and the natural confounder values Z i ( a ) = z Z_i(a)=z Z i ( a ) = z , M ( a , z ) M(a,z) M ( a , z ) is independent of each Y ∘ j ( a , z , m ) Y_{\circ j}(a,z,m) Y ∘ j ( a , z , m ) . Thus, after accounting for the measured baseline information and the full post-treatment confounder vector, there does not exist remaining confounding of the joint mediator intervention and an individual’s outcome. The condition allows Z i Z_i Z i to be affected by treatment and to confound the mediator–outcome relationship. In cluster-randomized trials, this differs from the sequential ignorability condition for identifying natural mediation effects in Cheng and Li [2 ] , which excludes treatment-induced mediator–outcome confounding. Assumption (A4) requires the clustered data are mutually independent and the cluster size is finite. It also requires positivity conditions for the mediator densities.
Before we proceed, we define several key nuisance parameters underlying the observed data distribution. Let η ∘ j ( a , z , m , c , n ) = E [ Y ∘ j ∣ A = a , Z = z , M = m , C = c , N = n ] \eta_{\circ j}(a,z,m,c,n)=\mathbb{E}[Y_{\circ j}\mid A=a,Z=z,M=m,C=c,N=n] η ∘ j ( a , z , m , c , n ) = E [ Y ∘ j ∣ A = a , Z = z , M = m , C = c , N = n ] be the conditional outcome mean. Let ξ ( m ∣ a , z , c , n ) = f M ∣ A , Z , C , N ( m ∣ a , z , c , n ) \xi(m\mid a,z,c,n)=f_{M\mid A,Z,C,N}(m\mid a,z,c,n) ξ ( m ∣ a , z , c , n ) = f M ∣ A , Z , C , N ( m ∣ a , z , c , n ) and κ ( m ∣ a , c , n ) = f M ∣ A , C , N ( m ∣ a , c , n ) \kappa(m\mid a,c,n)=f_{M\mid A,C,N}(m\mid a,c,n) κ ( m ∣ a , c , n ) = f M ∣ A , C , N ( m ∣ a , c , n ) denote the conditional distributions of the mediator vector given ( A , Z , C , N ) (A,Z,C,N) ( A , Z , C , N ) and ( A , C , N ) (A,C,N) ( A , C , N ) , respectively. Also define ϕ ( z ∣ a , c , n ) = f Z ∣ A , C , N ( z ∣ a , c , n ) \phi(z\mid a,c,n)=f_{Z\mid A,C,N}(z\mid a,c,n) ϕ ( z ∣ a , c , n ) = f Z ∣ A , C , N ( z ∣ a , c , n ) as the conditional distribution of the post-treatment confounder vector given ( A , C , N ) (A,C,N) ( A , C , N ) . We abbreviate the four nuisance parameters as h = ( η , ϕ , κ , ξ ) h=(\eta,\phi,\kappa,\xi) h = ( η , ϕ , κ , ξ ) . Estimating h h h can be nontrivial because the nuisance parameters involve multidimensional conditional distributions. Section H of the Appendix describes nuisance estimation details, including an extension of the copula approach of Cheng and Li [2 ] to estimate ( ξ , κ , ϕ ) (\xi,\kappa,\phi) ( ξ , κ , ϕ ) . Further let κ ∙ j ( m ∙ j ∣ a , c , n ) \kappa_{\bullet j}(m_{\bullet j}\mid a,c,n) κ ∙ j ( m ∙ j ∣ a , c , n ) and κ ∙ ( − j ) ( m ∙ ( − j ) ∣ a , c , n ) \kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid a,c,n) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ a , c , n ) denote the corresponding conditional densities of M ∙ j M_{\bullet j} M ∙ j and M ∙ ( − j ) M_{\bullet(-j)} M ∙ ( − j ) , which can be extracted from κ \kappa κ based on κ ∙ j ( m ∙ j ∣ a , c , n ) = ∫ κ ( m ∣ a , c , n ) d m ∙ ( − j ) \kappa_{\bullet j}(m_{\bullet j}\mid a,c,n)=\int\kappa(m\mid a,c,n)\,dm_{\bullet(-j)} κ ∙ j ( m ∙ j ∣ a , c , n ) = ∫ κ ( m ∣ a , c , n ) d m ∙ ( − j ) and κ ∙ ( − j ) ( m ∙ ( − j ) ∣ a , c , n ) = ∫ κ ( m ∣ a , c , n ) d m ∙ j \kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid a,c,n)=\int\kappa(m\mid a,c,n)\,dm_{\bullet j} κ ∙ ( − j ) ( m ∙ ( − j ) ∣ a , c , n ) = ∫ κ ( m ∣ a , c , n ) d m ∙ j .
Identification and the nonparametric efficiency bounds
We focus on identifying θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) and τ \tau τ as functionals of the observed-data distribution, based on which all interventional mediation effects are identified. The following theorem provides the core identification formulas.
Theorem 1. Under Assumptions (A1)–(A4), θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) is identified by
θ ( a , a ′ ) = E [ W N ∑ j = 1 N ∬ η ∙ j ( a , z , m , C , N ) ϕ ( z ∣ a , C , N ) κ ( m ∣ a ′ , C , N ) d z d m ] E [ W ] (3) \theta(a,a')=\frac{\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iint\eta_{\bullet j}(a,z,m,C,N)\phi(z\mid a,C,N)\kappa(m\mid a',C,N)\,dz\,dm\right]}{\mathbb{E}[W]} \tag*{(3)} θ ( a , a ′ ) = E [ W ] E [ N W ∑ j = 1 N ∬ η ∙ j ( a , z , m , C , N ) ϕ ( z ∣ a , C , N ) κ ( m ∣ a ′ , C , N ) d z d m ] (3) for any a , a ′ ∈ { 0 , 1 } a,a'\in\{0,1\} a , a ′ ∈ { 0 , 1 } . Additionally, τ \tau τ is identified by
τ = E [ W N ∑ j = 1 N ∬ η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) d z d m ] E [ W ] (4) \tau=\frac{\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iint\eta_{\bullet j}(1,z,m,C,N)\phi(z\mid1,C,N)\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\,dz\,dm\right]}{\mathbb{E}[W]} \tag*{(4)} τ = E [ W ] E [ N W ∑ j = 1 N ∬ η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) d z d m ] (4)
The proof is given in Section B of the Appendix. Based on the identification results for our causal targets, their efficient influence functions (EIFs) then characterize the efficiency bounds for estimating these functionals under the nonparametric model. To express the EIFs, we introduce the auxiliary functions defined in Table 1 (Panel A), which depend on the nuisance parameters ( η , ξ , κ , ϕ ) (\eta,\xi,\kappa,\phi) ( η , ξ , κ , ϕ ) .
The ratios ω ( 1 ) \omega^{(1)} ω ( 1 ) and ω ∙ j ( 2 ) \omega_{\bullet j}^{(2)} ω ∙ j ( 2 ) compare the mediator distribution specified by an intervention with the observed mediator distribution conditional on treatment and Z Z Z . They reweight outcome residuals to the relevant intervention distribution. The functions u ∙ j ( 1 ) − u ∙ j ( 3 ) u_{\bullet j}^{(1)}-u_{\bullet j}^{(3)} u ∙ j ( 1 ) − u ∙ j ( 3 ) and r ∙ j ( 1 ) − r ∙ j ( 4 ) r_{\bullet j}^{(1)}-r_{\bullet j}^{(4)} r ∙ j ( 1 ) − r ∙ j ( 4 ) are integrated outcome regressions used to express the EIFs of θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) and τ \tau τ , respectively. The next theorem gives the EIFs of θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) and τ \tau τ .
Theorem 2. Suppose that Assumptions (A1)–(A4) hold and that the displayed scores below are square-integrable. The EIF of θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) in the nonparametric observed-data model with known treatment probability is D θ ( a , a ′ ) = { ψ θ ( a , a ′ ; O ) − θ ( a , a ′ ) W } / E [ W ] D_{\theta(a,a')}=\{\psi_\theta(a,a';O)-\theta(a,a')W\}/\mathbb{E}[W] D θ ( a , a ′ ) = { ψ θ ( a , a ′ ; O ) − θ ( a , a ′ ) W } / E [ W ] , where
ψ θ ( a , a ′ ; O ) = W N ∑ j = 1 N { I ( A = a ) π a ( 1 − π ) 1 − a ω ( 1 ) ( a , a ′ , Z , M , C , N ) { Y ∙ j − η ∙ j ( a , Z , M , C , N ) } \psi_\theta(a,a';O)=\frac{W}{N}\sum_{j=1}^{N}\{\frac{I(A=a)}{\pi^{a}(1-\pi)^{1-a}}\omega^{(1)}(a,a',Z,M,C,N)\{Y_{\bullet j}-\eta_{\bullet j}(a,Z,M,C,N)\} ψ θ ( a , a ′ ; O ) = N W j = 1 ∑ N { π a ( 1 − π ) 1 − a I ( A = a ) ω ( 1 ) ( a , a ′ , Z , M , C , N ) { Y ∙ j − η ∙ j ( a , Z , M , C , N )} + I ( A = a ) π a ( 1 − π ) 1 − a { u ∙ j ( 1 ) ( a , a ′ , Z , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N ) } \qquad+\frac{I(A=a)}{\pi^{a}(1-\pi)^{1-a}}\left\{u_{\bullet j}^{(1)}(a,a',Z,C,N)-u_{\bullet j}^{(3)}(a,a',C,N)\right\} + π a ( 1 − π ) 1 − a I ( A = a ) { u ∙ j ( 1 ) ( a , a ′ , Z , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N ) }
Table 1. Auxiliary functions, nuisance estimation requirements, and properties of the EIF estimators.
Table 1 (PDF p. 7). Nuisance estimation requirements and properties.
Notes. The checkmarks in the nuisance rows indicate required estimators. The auxiliary functions’ means are evaluated by integration or conditional regression; e ∙ j e_{\bullet j} e ∙ j is constructed using (14) . Each alternative lists nuisance functions whose correct limits suffice for consistency under the stated regularity conditions. Indexed conditions apply to every j j j ; the low-dimensional results require compatible auxiliary limits as defined in § 5 . a ^{a} a Each alternative lists nuisance functions whose correct limits suffice for consistency under the stated regularity conditions. Indexed conditions apply to every j j j ; the low-dimensional results require compatible auxiliary limits as defined in § 5 . b ^{b} b Efficiency requires correct nuisance limits and the convergence and remainder conditions of Theorems 3 and 7 for the canonical and low-dimensional estimators, respectively. Detailed nuisance estimators appear in Tables S1–Table S3 of § 10.9 .§ 3 of the Appendix.
Table S3. Compatibility-preserving low-dimensional scenarios and separate theoretical guarantees.
+ I ( A = a ′ ) π a ′ ( 1 − π ) 1 − a ′ { u ∙ j ( 2 ) ( a , M , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N ) } + u ∙ j ( 3 ) ( a , a ′ , C , N ) (5) +\frac{I(A=a')}{\pi^{a'}(1-\pi)^{1-a'}}\left\{u_{\bullet j}^{(2)}(a,M,C,N)-u_{\bullet j}^{(3)}(a,a',C,N)\right\}
+u_{\bullet j}^{(3)}(a,a',C,N) \tag*{(5)} + π a ′ ( 1 − π ) 1 − a ′ I ( A = a ′ ) { u ∙ j ( 2 ) ( a , M , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N ) } + u ∙ j ( 3 ) ( a , a ′ , C , N ) (5)
The EIF of τ \tau τ is D τ = { ψ τ ( O ) − τ W } / E [ W ] D_\tau=\{\psi_\tau(O)-\tau W\}/\mathbb{E}[W] D τ = { ψ τ ( O ) − τ W } / E [ W ] , where
ψ τ ( O ) = W N ∑ j = 1 N { A π ω ∙ j ( 2 ) ( 1 , 1 , 0 , Z , M , C , N ) { Y ∙ j − η ∙ j ( 1 , Z , M , C , N ) } + A π { r ∙ j ( 1 ) ( 1 , 1 , 0 , Z , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } + A π { r ∙ j ( 2 ) ( 1 , 0 , M ∙ j , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } + 1 − A 1 − π { r ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } (6) \begin{aligned}
\psi_\tau(O)={}&\frac{W}{N}\sum_{j=1}^{N}\{\frac{A}{\pi}\omega_{\bullet j}^{(2)}(1,1,0,Z,M,C,N)\{Y_{\bullet j}-\eta_{\bullet j}(1,Z,M,C,N)\}\\
&+\frac{A}{\pi}\{r_{\bullet j}^{(1)}(1,1,0,Z,C,N)-r_{\bullet j}^{(4)}(1,1,0,C,N)\}\\
&+\frac{A}{\pi}\{r_{\bullet j}^{(2)}(1,0,M_{\bullet j},C,N)-r_{\bullet j}^{(4)}(1,1,0,C,N)\}\\
&+\frac{1-A}{1-\pi}\{r_{\bullet j}^{(3)}(1,1,M_{\bullet(-j)},C,N)-r_{\bullet j}^{(4)}(1,1,0,C,N)\} \tag*{(6)}
\end{aligned} ψ τ ( O ) = N W j = 1 ∑ N { π A ω ∙ j ( 2 ) ( 1 , 1 , 0 , Z , M , C , N ) { Y ∙ j − η ∙ j ( 1 , Z , M , C , N )} + π A { r ∙ j ( 1 ) ( 1 , 1 , 0 , Z , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} + π A { r ∙ j ( 2 ) ( 1 , 0 , M ∙ j , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} + 1 − π 1 − A { r ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} (6)
+ r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } . (7) +r_{\bullet j}^{(4)}(1,1,0,C,N)\}. \tag*{(7)} + r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} . (7)
The proof is provided in Section C of the Appendix. Importantly, the EIFs only depend on the observed data O O O and nuisance parameters ( η , ξ , κ , ϕ ) (\eta,\xi,\kappa,\phi) ( η , ξ , κ , ϕ ) . Under the conditions of Theorem 2, the variances of D θ ( a , a ′ ) D_{\theta(a,a')} D θ ( a , a ′ ) and D τ D_\tau D τ give the nonparametric efficiency bounds per independent cluster. In the following Section 4, we establish the corresponding robustness and asymptotic results for the canonical EIF estimators directly motivated by Theorem 2.
The canonical EIF estimators and their asymptotic behaviors
We first consider efficient estimators of θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) and τ \tau τ based on their canonical EIFs. Here, since we will use machine learning models to estimate the nuisance parameters, we leverage the cross-fitting algorithm to control the over-fitting bias associated with the nuisance estimation. Specifically, let R ≥ 2 R\ge2 R ≥ 2 be fixed and partition the K K K clusters into validation folds I 1 , … , I R \mathcal{I}_1,\ldots,\mathcal{I}_R I 1 , … , I R , with K r = ∣ I r ∣ K_r=|\mathcal{I}_r| K r = ∣ I r ∣ . For each fold r ∈ { 1 , … , R } r\in\{1,\ldots,R\} r ∈ { 1 , … , R } , we estimate the nuisance parameters h = ( η , ϕ , κ , ξ ) h=(\eta,\phi,\kappa,\xi) h = ( η , ϕ , κ , ξ ) using the clusters outside I r \mathcal{I}_r I r , yielding h ^ ( − r ) \widehat{h}^{(-r)} h ( − r ) . All members of a cluster remain in the same fold, so nuisance estimation excludes every observation in a validation cluster [2 ] . Then, the canonical EIF estimators are given by
θ ^ c f ( a , a ′ ) = K − 1 ∑ r = 1 R ∑ i ∈ I r ψ θ { a , a ′ ; O i ; h ^ ( − r ) } P K ( W ) , τ ^ c f = K − 1 ∑ r = 1 R ∑ i ∈ I r ψ τ { O i ; h ^ ( − r ) } P K ( W ) , \widehat{\theta}_{\mathrm{cf}}(a,a')=\frac{K^{-1}\sum_{r=1}^{R}\sum_{i\in\mathcal{I}_r}\psi_\theta\{a,a';O_i;\widehat{h}^{(-r)}\}}{\mathbb{P}_K(W)},\qquad
\widehat{\tau}_{\mathrm{cf}}=\frac{K^{-1}\sum_{r=1}^{R}\sum_{i\in\mathcal{I}_r}\psi_\tau\{O_i;\widehat{h}^{(-r)}\}}{\mathbb{P}_K(W)}, θ cf ( a , a ′ ) = P K ( W ) K − 1 ∑ r = 1 R ∑ i ∈ I r ψ θ { a , a ′ ; O i ; h ( − r ) } , τ cf = P K ( W ) K − 1 ∑ r = 1 R ∑ i ∈ I r ψ τ { O i ; h ( − r ) } ,
where P K ( f ) = K − 1 ∑ i = 1 K f ( O i ) \mathbb{P}_K(f)=K^{-1}\sum_{i=1}^{K}f(O_i) P K ( f ) = K − 1 ∑ i = 1 K f ( O i ) .
With a slight abuse of notations, we use the subscript 0 to denote true nuisance parameters so that h 0 = ( η 0 , ϕ 0 , κ 0 , ξ 0 ) h_0=(\eta_0,\phi_0,\kappa_0,\xi_0) h 0 = ( η 0 , ϕ 0 , κ 0 , ξ 0 ) are true values of h h h aligning with the true observed data distribution. Further, let f ∙ j f_{\bullet j} f ∙ j be a generic nuisance parameter or auxiliary function with a possibly individual-specific component j j j , and f ∙ j ⋆ f_{\bullet j}^{\star} f ∙ j ⋆ be its convergent value of the cross-fitted estimator f ^ ∙ j ( − r ) \widehat{f}_{\bullet j}^{(-r)} f ∙ j ( − r ) , which may differ from the true value f ∙ j , 0 f_{\bullet j,0} f ∙ j , 0 . Theorems 8 and 9 in Section D.1 of the Appendix establish the multiple robustness conditions for the canonical EIF estimators. To describe the multiple robustness property, we use f ∙ j ⋆ = f ∙ j , 0 f_{\bullet j}^{\star}=f_{\bullet j,0} f ∙ j ⋆ = f ∙ j , 0 to denote the correct specification of f ^ ∙ j ( − r ) \widehat{f}_{\bullet j}^{(-r)} f ∙ j ( − r ) converges to f ∙ j , 0 f_{\bullet j,0} f ∙ j , 0 in L 2 ( P ) L_2(P) L 2 ( P ) -norm for every j j j . Let ν 0 : = E ( W ) > 0 \nu_0 := \mathbb{E}(W) > 0 ν 0 := E ( W ) > 0 , and N max N_{\max} N m a x denote the largest possible cluster size. For θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) , E { ψ θ ( a , a ′ ; O ; h ) } = ν 0 θ 0 ( a , a ′ ) \mathbb{E}\{\psi_\theta(a,a';O;h)\}=\nu_0\theta_0(a,a') E { ψ θ ( a , a ′ ; O ; h )} = ν 0 θ 0 ( a , a ′ ) if either (i) ( η , ϕ ) = ( η 0 , ϕ 0 ) (\eta,\phi)=(\eta_0,\phi_0) ( η , ϕ ) = ( η 0 , ϕ 0 ) , (ii) ( η , κ ) = ( η 0 , κ 0 ) (\eta,\kappa)=(\eta_0,\kappa_0) ( η , κ ) = ( η 0 , κ 0 ) , or (iii) ( κ , ξ ) = ( κ 0 , ξ 0 ) (\kappa,\xi)=(\kappa_0,\xi_0) ( κ , ξ ) = ( κ 0 , ξ 0 ) . For τ \tau τ , E { ψ τ ( O ; h ) } = ν 0 τ 0 \mathbb{E}\{\psi_\tau(O;h)\}=\nu_0\tau_0 E { ψ τ ( O ; h )} = ν 0 τ 0 if either (i) ( κ , η ) = ( κ 0 , η 0 ) (\kappa,\eta)=(\kappa_0,\eta_0) ( κ , η ) = ( κ 0 , η 0 ) , (ii) ( κ , ξ ) = ( κ 0 , ξ 0 ) (\kappa,\xi)=(\kappa_0,\xi_0) ( κ , ξ ) = ( κ 0 , ξ 0 ) , (iii) ( η , ϕ , κ ∙ j ) = ( η 0 , ϕ 0 , κ 0 ∙ j ) (\eta,\phi,\kappa_{\bullet j})=(\eta_0,\phi_0,\kappa_{0\bullet j}) ( η , ϕ , κ ∙ j ) = ( η 0 , ϕ 0 , κ 0 ∙ j ) , or (iv) ( η , ϕ , κ ∙ ( − j ) ) = ( η 0 , ϕ 0 , κ 0 ∙ ( − j ) ) (\eta,\phi,\kappa_{\bullet(-j)})=(\eta_0,\phi_0,\kappa_{0\bullet(-j)}) ( η , ϕ , κ ∙ ( − j ) ) = ( η 0 , ϕ 0 , κ 0 ∙ ( − j ) ) . Under L 2 ( P ) L_2(P) L 2 ( P ) convergence of ψ θ { a , a ′ ; ∙ ; h ^ ( − r ) } \psi_\theta\{a,a';\bullet;\widehat{h}^{(-r)}\} ψ θ { a , a ′ ; ∙ ; h ( − r ) } and ψ τ { ∙ ; h ^ ( − r ) } \psi_\tau\{\bullet;\widehat{h}^{(-r)}\} ψ τ { ∙ ; h ( − r ) } to their respective limits uniformly over the fixed number of folds, θ ^ c f ( a , a ′ ) \widehat{\theta}_{\mathrm{cf}}(a,a') θ cf ( a , a ′ ) and τ ^ c f \widehat{\tau}_{\mathrm{cf}} τ cf are consistent whenever their corresponding conditions above hold at the limiting nuisance functions, even if the remaining nuisance functions are misspecified.
We need two regularity conditions for the asymptotic linearity and asymptotic normality of the canonical EIF estimators and we describe them below.
Assumption B. The number of folds R R R is fixed and K r / K → p r ∈ ( 0 , 1 ) K_r/K \to p_r \in(0,1) K r / K → p r ∈ ( 0 , 1 ) for every r r r .
Assumption B requires each fold to contain a nonvanishing proportion of clusters. The next assumption specifies the nuisance limits and the additional convergence conditions. To do so, we define λ f , r = max 1 ≤ j ≤ N max ∥ f ^ ∙ j ( − r ) − f ∙ j ⋆ ∥ 2 , P \lambda_{f,r}=\max_{1\le j\le N_{\max}}\|\widehat{f}^{(-r)}_{\bullet j}-f^\star_{\bullet j}\|_{2,P} λ f , r = max 1 ≤ j ≤ N m a x ∥ f ∙ j ( − r ) − f ∙ j ⋆ ∥ 2 , P as its estimation error compared to its convergent value under the L 2 ( P ) L_2(P) L 2 ( P ) -norm, where P P P denotes the observed-data distribution. We further define
δ η , r = ∥ W { η ^ ( − r ) − η 0 } ∥ 2 , P , δ ξ , r = ∥ W { ξ 0 ξ ^ ( − r ) − 1 } ∥ 2 , P , \delta_{\eta,r}=\|\sqrt{W}\{\widehat{\eta}^{(-r)}-\eta_0\}\|_{2,P}, \qquad
\delta_{\xi,r}=\left\|\sqrt{W}\left\{\frac{\xi_0}{\widehat{\xi}^{(-r)}}-1\right\}\right\|_{2,P}, δ η , r = ∥ W { η ( − r ) − η 0 } ∥ 2 , P , δ ξ , r = W { ξ ( − r ) ξ 0 − 1 } 2 , P ,
δ η ϕ , r = ∥ W { η 0 − η ^ ( − r ) ϕ ^ ( − r ) ϕ 0 } ∥ 2 , P , δ g , r = ∥ W { g ^ ( − r ) g 0 − 1 } ∥ 2 , P , \delta_{\eta\phi,r}=\left\|\sqrt{W}\left\{\eta_0-\widehat{\eta}^{(-r)}\frac{\widehat{\phi}^{(-r)}}{\phi_0}\right\}\right\|_{2,P}, \qquad
\delta_{g,r}=\left\|\sqrt{W}\left\{\frac{\widehat{g}^{(-r)}}{g_0}-1\right\}\right\|_{2,P}, δ η ϕ , r = W { η 0 − η ( − r ) ϕ 0 ϕ ( − r ) } 2 , P , δ g , r = W { g 0 g ( − r ) − 1 } 2 , P ,
as the estimation errors compared to their true values, where g ∈ { κ , κ ∙ j , κ ∙ ( − j ) } g\in\{\kappa,\kappa_{\bullet j},\kappa_{\bullet(-j)}\} g ∈ { κ , κ ∙ j , κ ∙ ( − j ) } , and δ κ ∙ j , r \delta_{\kappa_{\bullet j},r} δ κ ∙ j , r and δ κ ∙ ( − j ) , r \delta_{\kappa_{\bullet(-j)},r} δ κ ∙ ( − j ) , r denote the errors for individual j j j ’s mediator distribution and the joint mediator distribution of the other cluster members, respectively.
Assumption C. There exists a fixed measurable limit h ⋆ = ( η ⋆ , ϕ ⋆ , κ ⋆ , ξ ⋆ ) h^\star=(\eta^\star,\phi^\star,\kappa^\star,\xi^\star) h ⋆ = ( η ⋆ , ϕ ⋆ , κ ⋆ , ξ ⋆ ) such that, uniformly over the fixed number of folds, max 1 ≤ r ≤ R { λ η , r + λ ϕ , r + λ κ , r + λ ξ , r } = o p ( 1 ) \max_{1\le r\le R}\{\lambda_{\eta,r}+\lambda_{\phi,r}+\lambda_{\kappa,r}+\lambda_{\xi,r}\}=o_p(1) max 1 ≤ r ≤ R { λ η , r + λ ϕ , r + λ κ , r + λ ξ , r } = o p ( 1 ) . Assume E { ψ θ ( a , a ′ ; O ; h ⋆ ) 2 } < ∞ \mathbb{E}\{\psi_\theta(a,a';O;h^\star)^2\}<\infty E { ψ θ ( a , a ′ ; O ; h ⋆ ) 2 } < ∞ , and E { ψ τ ( O ; h ⋆ ) 2 } < ∞ \mathbb{E}\{\psi_\tau(O;h^\star)^2\}<\infty E { ψ τ ( O ; h ⋆ ) 2 } < ∞ . For each target under consideration, impose the corresponding conditions below.
(i) For θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) , assume max 1 ≤ r ≤ R { λ η , r + λ ω ( 1 ) , r + ∑ ℓ = 1 3 λ u ( ℓ ) , r } = o p ( 1 ) \max_{1\le r\le R}\{\lambda_{\eta,r}+\lambda_{\omega(1),r}+\sum_{\ell=1}^{3}\lambda_{u^{(\ell)},r}\}=o_p(1) max 1 ≤ r ≤ R { λ η , r + λ ω ( 1 ) , r + ∑ ℓ = 1 3 λ u ( ℓ ) , r } = o p ( 1 ) . The limit h ⋆ h^\star h ⋆ satisfies one of the three conditions in Theorem 8, and
max 1 ≤ r ≤ R { δ κ , r δ η ϕ , r + δ η , r δ ξ , r } = o p ( K − 1 / 2 ) . (8) \max_{1\le r\le R}\{\delta_{\kappa,r}\delta_{\eta\phi,r}+\delta_{\eta,r}\delta_{\xi,r}\}=o_p(K^{-1/2}). \tag*{(8)} 1 ≤ r ≤ R max { δ κ , r δ η ϕ , r + δ η , r δ ξ , r } = o p ( K − 1/2 ) . (8) (ii) For τ \tau τ , assume max 1 ≤ r ≤ R { λ η , r + λ ω ( 2 ) , r + ∑ ℓ = 1 4 λ r ( ℓ ) , r } = o p ( 1 ) \max_{1\le r\le R}\{\lambda_{\eta,r}+\lambda_{\omega(2),r}+\sum_{\ell=1}^{4}\lambda_{r^{(\ell)},r}\}=o_p(1) max 1 ≤ r ≤ R { λ η , r + λ ω ( 2 ) , r + ∑ ℓ = 1 4 λ r ( ℓ ) , r } = o p ( 1 ) . The limit h ⋆ h^\star h ⋆ satisfies one of the four conditions in Theorem 9, and
max 1 ≤ r ≤ R max 1 ≤ j ≤ N max { ( δ κ ∙ j , r + δ κ ∙ ( − j ) , r ) δ η ϕ , r + δ κ ∙ j , r δ κ ∙ ( − j ) , r + δ η , r δ ξ , r } = o p ( K − 1 / 2 ) . (9) \max_{1\le r\le R}\max_{1\le j\le N_{\max}}\left\{(\delta_{\kappa_{\bullet j},r}+\delta_{\kappa_{\bullet(-j)},r})\delta_{\eta\phi,r}+\delta_{\kappa_{\bullet j},r}\delta_{\kappa_{\bullet(-j)},r}+\delta_{\eta,r}\delta_{\xi,r}\right\}=o_p(K^{-1/2}). \tag*{(9)} 1 ≤ r ≤ R max 1 ≤ j ≤ N m a x max { ( δ κ ∙ j , r + δ κ ∙ ( − j ) , r ) δ η ϕ , r + δ κ ∙ j , r δ κ ∙ ( − j ) , r + δ η , r δ ξ , r } = o p ( K − 1/2 ) . (9)
Assumption C allows nuisance functions to converge to possibly misspecified limits, provided the relevant robustness condition holds. The conditions in (8) and (9), which are relevant to asymptotic linearity, concern products of errors relative to the truth and are stronger than the corresponding zero-bias conditions in Theorems 8 and 9. If one error has a nonzero limit, the paired error must converge faster for its product to be o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) .
Define ψ θ , c a n ⋆ ( O ) : = ψ θ ( a , a ′ ; O ; h ⋆ ) \psi_{\theta,\mathrm{can}}^{\star}(O) := \psi_{\theta}(a,a';O;h^{\star}) ψ θ , can ⋆ ( O ) := ψ θ ( a , a ′ ; O ; h ⋆ ) , and ψ τ , c a n ⋆ ( O ) = ψ τ ( O ; h ⋆ ) \psi_{\tau,\mathrm{can}}^{\star}(O) = \psi_{\tau}(O;h^{\star}) ψ τ , can ⋆ ( O ) = ψ τ ( O ; h ⋆ ) , where
D θ , c a n ⋆ ( O ) = ψ θ , c a n ⋆ ( O ) − θ 0 ( a , a ′ ) W ν 0 , D τ , c a n ⋆ ( O ) = ψ τ , c a n ⋆ ( O ) − τ 0 W ν 0 . D_{\theta,\mathrm{can}}^{\star}(O) = \frac{\psi_{\theta,\mathrm{can}}^{\star}(O)-\theta_{0}(a,a')W}{\nu_{0}}, \qquad D_{\tau,\mathrm{can}}^{\star}(O) = \frac{\psi_{\tau,\mathrm{can}}^{\star}(O)-\tau_{0}W}{\nu_{0}}. D θ , can ⋆ ( O ) = ν 0 ψ θ , can ⋆ ( O ) − θ 0 ( a , a ′ ) W , D τ , can ⋆ ( O ) = ν 0 ψ τ , can ⋆ ( O ) − τ 0 W .
We now state the asymptotic expansions for the two cross-fitted estimators. We assume the support and boundedness conditions stated at the beginning of Section F of the Appendix.
Theorem 3. Under Assumptions B and C(i),
K { θ ^ c f ( a , a ′ ) − θ 0 ( a , a ′ ) } = 1 K ∑ i = 1 K D θ , c a n ⋆ ( O i ) + o p ( 1 ) . (10) \sqrt{K}\{\widehat{\theta}_{\mathrm{cf}}(a,a')-\theta_{0}(a,a')\}=\frac{1}{\sqrt{K}}\sum_{i=1}^{K}D_{\theta,\mathrm{can}}^{\star}(O_i)+o_p(1). \tag*{(10)} K { θ cf ( a , a ′ ) − θ 0 ( a , a ′ )} = K 1 i = 1 ∑ K D θ , can ⋆ ( O i ) + o p ( 1 ) . (10) Under Assumptions B and C(ii),
K ( τ ^ c f − τ 0 ) = 1 K ∑ i = 1 K D τ , c a n ⋆ ( O i ) + o p ( 1 ) . (11) \sqrt{K}(\widehat{\tau}_{\mathrm{cf}}-\tau_{0})=\frac{1}{\sqrt{K}}\sum_{i=1}^{K}D_{\tau,\mathrm{can}}^{\star}(O_i)+o_p(1). \tag*{(11)} K ( τ cf − τ 0 ) = K 1 i = 1 ∑ K D τ , can ⋆ ( O i ) + o p ( 1 ) . (11) Consequently, the estimators converge to centered Gaussian limits with variances Σ θ , c a n ⋆ = E { ( D θ , c a n ⋆ ) 2 } \Sigma_{\theta,\mathrm{can}}^{\star}=\mathbb{E}\{(D_{\theta,\mathrm{can}}^{\star})^{2}\} Σ θ , can ⋆ = E {( D θ , can ⋆ ) 2 } and Σ τ , c a n ⋆ = E { ( D τ , c a n ⋆ ) 2 } \Sigma_{\tau,\mathrm{can}}^{\star}=\mathbb{E}\{(D_{\tau,\mathrm{can}}^{\star})^{2}\} Σ τ , can ⋆ = E {( D τ , can ⋆ ) 2 } , respectively. If h ⋆ = h 0 h^{\star}=h_{0} h ⋆ = h 0 , the limiting EIFs are the EIFs in Theorem 2, and the corresponding variances attain nonparametric efficiency bounds.
The proof is given in Section F of the Appendix. Define D ^ θ , c a n , i = ψ θ { a , a ′ ; O i ; h ^ ( − r ) } − θ ^ c f ( a , a ′ ) W i P K ( W ) \widehat{D}_{\theta,\mathrm{can},i}=\frac{\psi_{\theta}\{a,a';O_i;\widehat{h}^{(-r)}\}-\widehat{\theta}_{\mathrm{cf}}(a,a')W_i}{\mathbb{P}_{K}(W)} D θ , can , i = P K ( W ) ψ θ { a , a ′ ; O i ; h ( − r ) } − θ cf ( a , a ′ ) W i and D ^ τ , c a n , i = ψ τ { O i ; h ^ ( − r ) } − τ ^ c f W i P K ( W ) \widehat{D}_{\tau,\mathrm{can},i}=\frac{\psi_{\tau}\{O_i;\widehat{h}^{(-r)}\}-\widehat{\tau}_{\mathrm{cf}}W_i}{\mathbb{P}_{K}(W)} D τ , can , i = P K ( W ) ψ τ { O i ; h ( − r ) } − τ cf W i , for i ∈ I r i\in\mathcal{I}_{r} i ∈ I r . For the purpose of inference, the asymptotic variances Σ θ , c a n ⋆ \Sigma_{\theta,\mathrm{can}}^{\star} Σ θ , can ⋆ and Σ τ , c a n ⋆ \Sigma_{\tau,\mathrm{can}}^{\star} Σ τ , can ⋆ can be estimated by the cross-fitted sample variances P K ( D ^ θ , c a n 2 ) \mathbb{P}_{K}(\widehat{D}_{\theta,\mathrm{can}}^{2}) P K ( D θ , can 2 ) and P K ( D ^ τ , c a n 2 ) \mathbb{P}_{K}(\widehat{D}_{\tau,\mathrm{can}}^{2}) P K ( D τ , can 2 ) , respectively, and Wald-type confidence intervals can be constructed. Once we have θ ^ c f ( a , a ′ ) \widehat{\theta}_{\mathrm{cf}}(a,a') θ cf ( a , a ′ ) and τ ^ c f \widehat{\tau}_{\mathrm{cf}} τ cf , the interventional mediation effects can be obtained immediately.
Low-dimensional reparameterization and alternative efficient estimators
Recall that the canonical nuisance parameters ( η , ϕ , κ , ξ ) (\eta,\phi,\kappa,\xi) ( η , ϕ , κ , ξ ) determine the auxiliary functions ω ( 1 ) \omega^{(1)} ω ( 1 ) , ω ∙ j ( 2 ) \omega^{(2)}_{\bullet j} ω ∙ j ( 2 ) , u ∙ j ( 1 ) u_{\bullet j}^{(1)} u ∙ j ( 1 ) –u ∙ j ( 3 ) u_{\bullet j}^{(3)} u ∙ j ( 3 ) , and r ∙ j ( 1 ) r_{\bullet j}^{(1)} r ∙ j ( 1 ) –r ∙ j ( 4 ) r_{\bullet j}^{(4)} r ∙ j ( 4 ) , which in turn determine the EIFs in Theorem 2. A direct implementation requires estimating conditional densities of cluster-level vectors whose dimensions increase with cluster size. Forming ratios of separately estimated densities can amplify estimation errors, particularly where the denominator density is small. Evaluating the auxiliary regressions also requires multivariate integration, which can be computationally demanding and introduces numerical approximation error when analytic expressions are unavailable. To avoid separately estimating multivariate densities and evaluating their integrals, we re-express these auxiliary functions using density ratios, treatment probabilities, univariate mediator densities, and conditional-mean regressions, similar to Cheng and Li [2 ] , Díaz et al. [5 ] , and Rudolph et al. [15 ] . We denote π a = π a ( 1 − π ) 1 − a \pi_a=\pi^a(1-\pi)^{1-a} π a = π a ( 1 − π ) 1 − a . Define
κ ∙ j ⋆ ( m ∙ j ∣ m ∙ ( − j ) , a , c , n ) = f M ∙ j ∣ M ∙ ( − j ) , A , C , N ( m ∙ j ∣ m ∙ ( − j ) , a , c , n ) \kappa_{\bullet j}^{\star}(m_{\bullet j}\mid m_{\bullet(-j)},a,c,n)=f_{M_{\bullet j}\mid M_{\bullet(-j)},A,C,N}(m_{\bullet j}\mid m_{\bullet(-j)},a,c,n) κ ∙ j ⋆ ( m ∙ j ∣ m ∙ ( − j ) , a , c , n ) = f M ∙ j ∣ M ∙ ( − j ) , A , C , N ( m ∙ j ∣ m ∙ ( − j ) , a , c , n ) ,
d ( a , z , m , c , n ) = κ ( m ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) . (12) d(a,z,m,c,n)=\frac{\kappa(m\mid a,c,n)}{\xi(m\mid a,z,c,n)}. \tag*{(12)} d ( a , z , m , c , n ) = ξ ( m ∣ a , z , c , n ) κ ( m ∣ a , c , n ) . (12)
s ( a ∣ m , c , n ) = P ( A = a ∣ M = m , C = c , N = n ) . (13) s(a\mid m,c,n)=P(A=a\mid M=m,C=c,N=n). \tag*{(13)} s ( a ∣ m , c , n ) = P ( A = a ∣ M = m , C = c , N = n ) . (13)
and s − j ( a ∣ m ∙ ( − j ) , c , n ) = P { A = a ∣ M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n } s_{-j}(a \mid\mathbf{m}_{\bullet(-j)},c,n)=P\{A=a\mid\mathbf{M}_{\bullet(-j)}=\mathbf{m}_{\bullet(-j)},C=c,N=n\} s − j ( a ∣ m ∙ ( − j ) , c , n ) = P { A = a ∣ M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n } . For the spillover functional, further define
e ∙ j ( z , m , c , n ) = d ( 1 , z , m , c , n ) κ ∙ j ( m ∙ j ∣ 1 , c , n ) κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) . (14) e_{\bullet j}(z,m,c,n)=d(1,z,m,c,n)\frac{\kappa_{\bullet j}(m_{\bullet j}\mid1,c,n)}{\kappa_{\bullet j}^{*}(m_{\bullet j}\mid\mathbf{m}_{\bullet(-j)},1,c,n)}. \tag*{(14)} e ∙ j ( z , m , c , n ) = d ( 1 , z , m , c , n ) κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ ∙ j ( m ∙ j ∣ 1 , c , n ) . (14)
Proposition 1. At the observed data distribution, on the common support of the relevant conditional distributions, the auxiliary density ratios in Theorem 2 satisfy
ω ( 1 ) ( a , a ′ , z , m , c , n ) = d ( a , z , m , c , n ) π a π a ′ s ( a ′ ∣ m , c , n ) s ( a ∣ m , c , n ) . (15) \omega^{(1)}(a,a',z,m,c,n)=d(a,z,m,c,n)\frac{\pi_a}{\pi_{a'}}\frac{s(a'\mid m,c,n)}{s(a\mid m,c,n)}. \tag*{(15)} ω ( 1 ) ( a , a ′ , z , m , c , n ) = d ( a , z , m , c , n ) π a ′ π a s ( a ∣ m , c , n ) s ( a ′ ∣ m , c , n ) . (15) ω ∙ j ( 2 ) ( 1 , 1 , 0 , z , m , c , n ) = e ∙ j ( z , m , c , n ) π 1 − π s − j ( 0 ∣ m ∙ ( − j ) , c , n ) s − j ( 1 ∣ m ∙ ( − j ) , c , n ) . (16) \omega_{\bullet j}^{(2)}(1,1,0,z,m,c,n)=e_{\bullet j}(z,m,c,n)\frac{\pi}{1-\pi}\frac{s_{-j}(0\mid\mathbf{m}_{\bullet(-j)},c,n)}{s_{-j}(1\mid\mathbf{m}_{\bullet(-j)},c,n)}. \tag*{(16)} ω ∙ j ( 2 ) ( 1 , 1 , 0 , z , m , c , n ) = e ∙ j ( z , m , c , n ) 1 − π π s − j ( 1 ∣ m ∙ ( − j ) , c , n ) s − j ( 0 ∣ m ∙ ( − j ) , c , n ) . (16) For any square-integrable working outcome regression η \eta η for which the following expectations exist, the associated auxiliary regressions, with the conditional distributions evaluated under the observed-data distribution, are
u ∙ j ( 1 ) ( a , a ′ , z , c , n ) = E { η ∙ j ( a , z , M , c , n ) ∣ A = a ′ , C = c , N = n } . (17) u_{\bullet j}^{(1)}(a,a',z,c,n)=\mathbb{E}\{\eta_{\bullet j}(a,z,M,c,n)\mid A=a',C=c,N=n\}. \tag*{(17)} u ∙ j ( 1 ) ( a , a ′ , z , c , n ) = E { η ∙ j ( a , z , M , c , n ) ∣ A = a ′ , C = c , N = n } . (17) u ∙ j ( 2 ) ( a , m , c , n ) = E { η ∙ j ( a , Z , m , c , n ) ∣ A = a , C = c , N = n } . (18) u_{\bullet j}^{(2)}(a,m,c,n)=\mathbb{E}\{\eta_{\bullet j}(a,Z,m,c,n)\mid A=a,C=c,N=n\}. \tag*{(18)} u ∙ j ( 2 ) ( a , m , c , n ) = E { η ∙ j ( a , Z , m , c , n ) ∣ A = a , C = c , N = n } . (18) u ∙ j ( 3 ) ( a , a ′ , c , n ) = E { u ∙ j ( 1 ) ( a , a ′ , Z , C , N ) ∣ A = a , C = c , N = n } = E { u ∙ j ( 2 ) ( a , M , C , N ) ∣ A = a ′ , C = c , N = n } . (19) \begin{aligned}
u_{\bullet j}^{(3)}(a,a',c,n)&=\mathbb{E}\{u_{\bullet j}^{(1)}(a,a',Z,C,N)\mid A=a,C=c,N=n\}\\
&=\mathbb{E}\{u_{\bullet j}^{(2)}(a,M,C,N)\mid A=a',C=c,N=n\}. \tag*{(19)}
\end{aligned} u ∙ j ( 3 ) ( a , a ′ , c , n ) = E { u ∙ j ( 1 ) ( a , a ′ , Z , C , N ) ∣ A = a , C = c , N = n } = E { u ∙ j ( 2 ) ( a , M , C , N ) ∣ A = a ′ , C = c , N = n } . (19)
The corresponding representations of r ∙ j ( 1 ) r_{\bullet j}^{(1)} r ∙ j ( 1 ) –r ∙ j ( 4 ) r_{\bullet j}^{(4)} r ∙ j ( 4 ) use conditional regressions and the univariate densities κ ∙ j \kappa_{\bullet j} κ ∙ j and κ ∙ j ∗ \kappa_{\bullet j}^{*} κ ∙ j ∗ , as given in (20)–(23) of the Appendix. At η = η 0 \eta=\eta_0 η = η 0 , these regressions coincide with the true auxiliary functions defined in § 3 .
The proof is given in Section G.2 of the Appendix. This reparameterization reduces the dimension of the density responses and replaces multivariate integration by conditional regression. Corresponding estimation of the nuisance and auxiliary functions is described in Section H of the Appendix. In particular, the additional ratio d d d accounts for conditioning on the post-treatment confounder vector Z Z Z . We extend the augmented method proposed by Rudolph et al. [15 ] to estimate d d d from two classifiers distinguishing observed cluster records from reference copies, without separately estimating κ \kappa κ and ξ \xi ξ ; see Section G.1 of the Appendix for further details.
In our setting, a nuisance tuple is called compatible when its auxiliary regressions are obtained from the same working outcome regression η \eta η through Proposition 1, using the true conditional distributions of Z Z Z and M M M under the observed-data distribution. We construct θ ^ l d ( a , a ′ ) \widehat{\theta}_{\mathrm{ld}}(a,a') θ ld ( a , a ′ ) and τ ^ l d \widehat{\tau}_{\mathrm{ld}} τ ld using cluster-level cross-fitting, as in § 4 . Theorem 4 establishes multiple robustness for compatible tuples. For compatible fitted tuples, Assumptions B and D, together with the stated regularity conditions, yield root-K K K asymptotic normality by Theorem 5 , with the nonparametric efficiency bounds attained at the true nuisance limits. Both theorems and Assumption D are presented in Section A of the Appendix. For separately estimated auxiliary regressions converging to compatible limits, Theorem 6 identifies their additional contributions to E { ψ ^ θ , r ( O ) } − ν 0 θ 0 ( a , a ′ ) \mathbb{E}\{\widehat{\psi}_{\theta,r}(O)\}-\nu_0\theta_0(a,a') E { ψ θ , r ( O )} − ν 0 θ 0 ( a , a ′ ) and E { ψ ^ τ , r ( O ) } − ν 0 τ 0 \mathbb{E}\{\widehat{\psi}_{\tau,r}(O)\}-\nu_0\tau_0 E { ψ τ , r ( O )} − ν 0 τ 0 . Theorem 7 gives sufficient conditions for consistency and asymptotic normality. Both theorems, their assumptions, and the corresponding constructions are presented in Section A.3 of the Appendix, along with a discussion of the related compatibility considerations in Rudolph et al. [15 ] .
Panel B in Table 1 summarizes the nuisance estimation requirements and robustness and efficiency properties of the proposed EIF estimators.
Simulation studies
We conducted two simulation studies with shared data-generating process to examine the finite-sample performance of canonical and low-dimensional EIF estimators. The complete code for implementing the estimators and reproducing the simulation studies is available at https://github.com/daeyoungspam-stat/CRT-interventional-mediation . At each K ∈ { 50 , 75 , 100 , 150 } K \in\{50, 75, 100, 150\} K ∈ { 50 , 75 , 100 , 150 } , we generated 2,000 independent cluster-randomized trials. The data-generating model for O i O_i O i and true values of interventional mediation effects are given in § 10.9 .§ 1 of the Appendix. The nuisance parameters were estimated and auxiliary functions were evaluated using five-fold cluster-level cross-fitting. Both studies used the nuisance fits specified in Table S1 of § 10.9 .§ 3 of the Appendix. We also applied the stabilization motivated by Cheng and Li [2 ] ; see Appendix Section I.2 for implementation considerations including these numerical stabilization procedures and the construction of EIF-based confidence intervals. We expect the asymptotic analysis of the outcome regression stabilization described in Appendix Section I.2 (excluding the truncation discussed below) to follow arguments similar to those used to prove Propositions S1 and S2 of Cheng and Li [2 ] , and therefore omit the details here. We compared our EIF-based estimators to the g-computation-based plug-in method, which is calculated based on the identification formulas in Theorem 1 directly.
Simulation study for the canonical EIF estimators
The canonical study evaluated all 16 correct/incorrect combinations of ( η , ϕ , κ , ξ ) (\eta, \phi, \kappa, \xi) ( η , ϕ , κ , ξ ) , denoted C-η ϕ κ ξ \eta\phi\kappa\xi η ϕ κ ξ . The correctly specified models used the covariates and interactions listed for each nuisance in Table S1 in § 10.9 .§ 3 of the Appendix. The misspecification scenarios omitted baseline covariates and their interactions while considering treatment, cluster size, and the relevant mediator or post-treatment-confounder arguments. As in Appendix B of Cheng and Li [2 ] , the residual scale and exchangeable correlation were estimated in each training fold. We compared stabilized canonical EIF estimators using canonical parametric regression and flexible nuisance estimation. The latter supplied the same prespecified correct or misspecified feature sets to Super Learner with SL.library set to c("SL.glm","SL.ranger"). The parametric g-computation used the corresponding η , ϕ , κ \eta,\phi,\kappa η , ϕ , κ fits, and exploited a 200-resample cluster bootstrap for inference.
Table S2 in § 10.9 .§ 3 summarizes the guarantees for point consistency, asymptotic linearity for each estimator, in accordance with Theorems 8 and 9 (see Section D.1 of the Appendix) for point consistency, and Theorem 3 for asymptotic linearity. Among the scenarios considered here, the asymptotic linearity guarantees in Table S2 apply to the theoretical canonical EIF estimators under C-1111 and to parametric g-computation under both C-1111 and C-1110.
Table S2. Table S2: Canonical nuisance scenarios and separate theoretical guarantees.
Figure 1 and Figure 2 display the cluster-average IDE and ISME results, respectively. Under C-1111, both canonical EIF methods remained approximately unbiased, and their performance improved as K K K increased. The IDE coverage remained close to nominal, whereas the mild ISME undercoverage at smaller K K K diminished as K K K increased. The g-computation was also approximately unbiased with coverage close to nominal, reflecting the correct specification of its parametric models. When only η \eta η was misspecified (C-0111), both EIF estimators remained approximately unbiased for IDE and ISME, as expected from the point-consistency condition requiring correct ( κ , ξ ) (\kappa,\xi) ( κ , ξ ) . Their coverage was generally close to nominal at larger K K K . The g-computation exhibited little IDE bias, but its ISME bias persisted as K K K increased. When only ϕ \phi ϕ was misspecified (C-1011), both EIF estimators again remained approximately unbiased, now through correct ( η , κ ) (\eta, \kappa) ( η , κ ) .
Figure 1. Canonical-study performance comparison for the cluster-average IDE in C-1111, C-0000, and the four scenarios with exactly one wrong nuisance. Panel A shows estimation error over 2,000 replications. Panel B shows empirical coverage with Monte Carlo error bars.
Figure 2. Canonical-study performance comparison for the cluster-average ISME in the six scenarios as in Figure 1 .
The distinction between point consistency and inference was particularly evident when κ \kappa κ or ξ \xi ξ was misspecified. Under C-1101, correct ( η , ϕ ) (\eta, \phi) ( η , ϕ ) sufficed for IDE point consistency. Accordingly, both EIF estimators remained close to the IDE target and became more concentrated with K K K , but their intervals undercovered, more noticeably for the parametric implementation. For ISME, the same nuisance combination did not satisfy a point-consistency condition: bias persisted, and coverage deteriorated as K K K increased. Under C-1110, both IDE and ISME estimates remained unbiased, but coverage became conservative, especially for IDE.
Under C-0000, the sampling distributions narrowed with K K K but remained displaced from the targets. For ISME, coverage deteriorated for all three estimators, with a substantially larger undercoverage for g-computation. The IDE coverage for the EIF estimators remained comparatively close to nominal despite persistent bias and considerably greater variation than under C-1111. We note that Figure S1 –Figure S3 in § 10.9 .4.1 extend these comparisons to all 16 nuisance scenarios and include IIME. The individual-average results in Figure S7 –Figure S9 of § 10.9 .4.2 show similar patterns.
Figure S1. Complete canonical-study cluster-average IDE results for all 16 nuisance scenarios, three estimators, and four values of K K K .
Figure S3. Complete canonical-study cluster-average ISME results.
Figure S7. Complete canonical-study individual-average IDE results for all 16 nuisance scenarios, three estimators, and four values of K K K .
Figure S9. Complete canonical-study individual-average ISME results.
Simulation study for the low-dimensional EIF estimators
To estimate the density ratio d d d , we used the augmented-classification approach of Rudolph et al. [15 ] , adapted to clustered data. For each observed training cluster, we created random copies by replacing its mediator vector with a draw from an auxiliary distribution λ ~ \widetilde{\lambda} λ , while keeping its treatment, post-treatment confounders, baseline covariates, and cluster size unchanged. The distribution λ ~ \widetilde{\lambda} λ could depend on treatment, baseline covariates, and cluster size, but not on the post-treatment confounders. Two binary classifiers were then trained to distinguish the generated copies from the observed data, and their fitted probabilities were combined to estimate d d d . We considered two constructions based on fitted arm-specific mediator means: Gaussian residual draws (G) and independent empirical residual-bootstrap draws (RB). Each mechanism was combined with glmnet, random forest (RF), or Super Learner (SL), yielding six different versions. The definition of the reference distribution and the classification-based representation of d d d are provided in Section G.1. The nuisance estimation stage used the same cluster-level cross-fitting scheme as the canonical study. The working models and implementation details are provided in § 10.9 .§ 3 .
The scenario label LD-AC denotes all-correct specification. The labels LD-T2–LD-T6 vary the specification of the nuisances for θ \theta θ while holding correct specification for τ \tau τ ; LD-U2–LD-U6 reverse these roles. Table S3 in § 10.9 .§ 3 summarizes the scenario-specific guarantees for point consistency and asymptotic linearity, following Theorems 4 and 5 , together with Theorem 7 for estimated auxiliary functions. Among these simulation scenarios, asymptotic linearity with the corresponding EIFs, and hence the corresponding interval validity, is guaranteed only under LD-AC (as in the canonical study), subject to the stated rate conditions.
Figure 3 and Figure 4 display the cluster-average IDE and ISME results for G+SL and RB+SL under LD-AC and LD-T2–LD-T6. Under LD-AC, both estimators remained approximately centered, and their sampling distributions became more concentrated as K K K increased. The ISME coverage approached nominal at larger K K K , whereas IDE intervals remained mildly conservative. The two reference distributions yielded very similar centering, dispersion, and coverage throughout the displayed scenarios, indicating little sensitivity to this choice in the present Gaussian design.
Figure 3. Low-dimensional-study performance comparison for the cluster-average IDE, comparing G+SL and RB+SL in LD-AC and LD-T2–LD-T6. Plot elements are as in Figure 1 .
Figure 4. Low-dimensional-study performance comparison for the cluster-average ISME for the same estimators and scenarios as Figure 3 .
Under LD-T2 and LD-T3, both effects satisfied the point-consistency conditions. The estimates remained close to their targets. Their coverage patterns nevertheless differed: LD-T2 intervals remained more conservative than those under LD-AC, whereas LD-T3 coverage was closer to nominal at larger K K K , particularly for ISME. Thus, the two scenarios exhibited similar point-estimation
performance but different behavior of the confidence intervals.
LD-T4 distinguished the nuisance requirements for the two effects. ISME has a point-consistency guarantee and remained close to its target, with coverage approaching nominal as K K K increased. IDE also exhibited little bias and increasingly concentrated estimates, although its nuisance specification did not satisfy a point-consistency condition. Under LD-T5 and LD-T6, neither displayed effect had a point-consistency guarantee. The observed biases were modest over the examined values of K K K , but the intervals remained more conservative than under LD-AC, especially for IDE.
Additional scenarios and IIME results are reported in Figure S4 –S6 of Section I.4.1. These results further illustrate the effect-specific point-consistency guarantees. The individual-average results in Figure S10 –S12 of Section I.4.2 showed similar patterns.
Figure S4. Complete low-dimensional-study cluster-average IDE results for all 11 scenarios, six reference-distribution–classifier pairs, and four values of K K K .
Figure S10. Complete low-dimensional-study individual-average IDE results for all 11 scenarios, six reference-distribution–classifier pairs, and four values of K K K .
Across the six estimators under LD-AC, glmnet and RF had similar RMSE, and the SL implementations achieved comparable point-estimation performance. The SL implementations generally produced wider intervals than glmnet and RF, particularly for IDE and ISME at smaller K K K . The comprehensive simulation reports can be found in Section I.4.
Data application: Analysis of the PPACT cluster-randomized trial
We reanalyzed the Pain Program for Active Coping and Training (PPACT) trial [4 ] , a pragmatic cluster-randomized trial to evaluate a primary care-based cognitive behavioral therapy intervention among adults with chronic pain receiving long-term opioid therapy. The PPACT trial comprises 850 participants within 106 primary care provider clusters, with cluster sizes ranging from 3 to 13 participants. In our mediation analysis, the outcome is pain impact at 12 months, measured using the 0–10 PEGS scale, for which higher values indicate greater pain impact. We consider the Roland–Morris Disability Questionnaire (RMDQ) score at 6 months (range 0–1, with higher values indicating greater pain-related disability) as the mediator of interest. Satisfaction with primary care services at 6 months (range 1–5) was treated as the post-treatment confounder of the mediator–outcome relationship. We adjusted for cluster size and the following individual-level baseline covariates: sex, age, diagnosis of at least two chronic conditions, count of pain types, benzodiazepine receipt during the 6 months before randomization, baseline PEGS score, baseline RMDQ score, and average daily opioid dose during the 6 months before randomization.
We applied the five estimators described in the simulation study (§ 6 ) to investigate the cluster-average interventional mediation effects, including parametric g-computation, the stabilized canonical EIF estimator with parametric nuisance estimation, the stabilized canonical EIF estimator with flexible nuisance estimation, and the low-dimensional EIF estimators based on either a Gaussian reference distribution or a residual-bootstrap reference distribution. The g-computation and parametric canonical EIF estimators used generalized linear models only. For the flexible canonical estimator, the Super Learner library contained generalized linear models, elastic-net regression, and random forests; the low-dimensional estimators used the corresponding prespecified Super Learner libraries.
The mediation analysis results are shown in Figure 5 . First, all five estimators indicated that the intervention reduced pain impact at 12 months. The TEs ranged from −0.54 to −0.46 PEGS points, and every 95% confidence interval excluded zero. Second, IDEs ranged from −0.27 to −0.15, which are all statistically insignificant. The indirect effects were consistently negative. For the IIEs, the g-computation and both canonical EIF intervals excluded zero, whereas the intervals from the two low-dimensional estimators were wider and included zero. Third, the decomposition effects from IIE (i.e., the IIME and ISME) suggested that the mediated pathway operated primarily through the
Figure 5. Estimated effects of the PPACT cognitive behavioral therapy intervention on the 12-month PEGS score under cluster-average weighting (W i = 1 W_i = 1 W i = 1 ). Points are estimates and horizontal bars are 95% confidence intervals. TE denotes the total effect, IDE the interventional direct effect, IIE the interventional indirect effect, IIME the interventional individual mediation effect, and ISME the interventional spillover mediation effect. Negative values favor the intervention because higher PEGS scores indicate worse pain impact.
participant’s own RMDQ score rather than through same-cluster spillover. The IIMEs ranged from − 0.31 -0.31 − 0.31 to − 0.24 -0.24 − 0.24 ; their confidence intervals excluded zero for g-computation and the two canonical EIF estimators but not for the low-dimensional EIF estimators. By comparison, the ISMEs were smaller and less stable, ranging from − 0.12 -0.12 − 0.12 to 0.10 0.10 0.10 , and all intervals included zero. Taken together, the results provide evidence of a protective total effect of the intervention, which also suggests some evidence for existence of indirect effect. However, the results provide little evidence of spillover mediation through interactions with other participants in the same provider cluster. The wider intervals from the low-dimensional EIF estimators, particularly for the IIME and ISME, warrant cautious interpretation of the detailed decomposition. We remark that the low-dimensional EIF estimators also yielded wider intervals than their canonical EIF counterparts in the simulation studies, particularly for ISME (see § 10.9 .4.1 in the Appendix).
Based on the canonical EIF estimator with machine learning nuisance estimation, we compared interventional mediation effects, which account for satisfaction with primary care services as a post-treatment confounder, with the corresponding natural mediation effects, which do not adjust for this confounder. As shown in Supplementary Table S4 in Appendix § 10.10 , the estimated indirect effect, individual mediation effect, and spillover mediation effect were similar across the two analyses. However, the direct effect differed substantially. The interventional direct effect was − 0.163 -0.163 − 0.163 (95% CI: − 0.426 -0.426 − 0.426 , 0.101 0.101 0.101 ), whereas the natural direct effect was − 0.277 -0.277 − 0.277 (95% CI: − 0.502 -0.502 − 0.502 , − 0.052 -0.052 − 0.052 ). This suggests that accounting for post-treatment confounding has little influence on the substantive conclusion regarding mediation through RMDQ but materially changes the magnitude and significance of the direct effect. Because the two analyses target different estimands, however, this difference should be interpreted descriptively and not a formal estimate of confounding bias.
Conclusion
We develop, to our knowledge, the first semiparametric framework for mediation analysis in cluster-randomized trials that jointly accommodates treatment-induced mediator–outcome confounding and within-cluster spillover. The central barrier is that treatment can alter common causes of the mediator and outcome, thereby hindering identification of natural mediation effects. By defining stochastic interventions on the joint within-cluster mediator distribution, we identify interventional direct and indirect effects without requiring cross-world independence assumptions and decompose the indirect effect into individual and spillover components. The proposed estimands quantify the consequences of hypothetical interventions on the joint mediator distribution and can differ from treatment effects carried through individual causal pathways, thereby providing an intervention-oriented perspective on the role of mediators that complements natural effects. For these estimands, we derive efficient influence functions and construct cluster-level cross-fitted estimators with multiple robustness properties, establishing asymptotic normality and semiparametric efficiency under suitable nuisance estimation conditions. A lower-dimensional reparameterization facilitates flexible nuisance estimation while reducing the burden of multivariate density estimation and integration. Extensions to multiple mediators and sensitivity analyses for residual mediator–outcome confounding merit further investigation.
Acknowledgement
Research in this article was supported by the Patient-Centered Outcomes Research Institute® (PCORI® Award ME-2023C1-31350). The statements presented in this article are solely the responsibility of the authors and do not necessarily represent the views of PCORI® , its Board of Governors or Methodology Committee.
Supplementary Material
Additional technical details for the low-dimensional estimator
This section gives the identities and explicit score and estimator formulas underlying the low-dimensional reparameterization, its exact bias decompositions and multiple-robustness conditions, and the assumptions and asymptotic results for cross-fitted inference.
We give the spillover-regression representations stated in Proposition 1 of Section 5. For any square-integrable working regression η \eta η , the compatible partial means for τ \tau τ can be written using conditional regressions and densities with the scalar response M ∙ j M_{\bullet j} M ∙ j , with the conditional distributions evaluated under the observed-data distribution:
r ∙ j ( 1 ) ( 1 , 1 , 0 , z , c , n ) = E [ η ∙ j ( 1 , z , M , C , N ) κ ∙ j ( M ∙ j ∣ 1 , c , n ) κ ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 0 , c , n ) | A = 0 , C = c , N = n ] . (20) r_{\bullet j}^{(1)}(1,1,0,z,c,n)=\mathbb{E}\left[\eta_{\bullet j}(1,z,M,C,N)\frac{\kappa_{\bullet j}(M_{\bullet j}\mid1,c,n)}{\kappa_{\bullet j}^{*}(M_{\bullet j}\mid M_{\bullet(-j)},0,c,n)}\ \middle|\ A=0,C=c,N=n\right]. \tag*{(20)} r ∙ j ( 1 ) ( 1 , 1 , 0 , z , c , n ) = E [ η ∙ j ( 1 , z , M , C , N ) κ ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 0 , c , n ) κ ∙ j ( M ∙ j ∣ 1 , c , n ) A = 0 , C = c , N = n ] . (20)
r ∙ j ( 2 ) ( 1 , 0 , m ∙ j , c , n ) = E [ u ∙ j ( 2 ) ( 1 , M , C , N ) κ ∙ j ( M ∙ j ∣ 0 , C , N ) κ ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 0 , C , N ) | A = 0 , M ∙ j = m ∙ j , C = c , N = n ] . (21) r_{\bullet j}^{(2)}(1,0,m_{\bullet j},c,n)=\mathbb{E}\left[u_{\bullet j}^{(2)}(1,M,C,N)\frac{\kappa_{\bullet j}(M_{\bullet j}\mid0,C,N)}{\kappa_{\bullet j}^{*}(M_{\bullet j}\mid M_{\bullet(-j)},0,C,N)}\ \middle|\ A=0,M_{\bullet j}=m_{\bullet j},C=c,N=n\right]. \tag*{(21)} r ∙ j ( 2 ) ( 1 , 0 , m ∙ j , c , n ) = E [ u ∙ j ( 2 ) ( 1 , M , C , N ) κ ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 0 , C , N ) κ ∙ j ( M ∙ j ∣ 0 , C , N ) A = 0 , M ∙ j = m ∙ j , C = c , N = n ] . (21)
r ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , c , n ) = E [ u ∙ j ( 2 ) ( 1 , M , C , N ) κ ∙ j ( M ∙ j ∣ 1 , C , N ) κ ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 1 , C , N ) | A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ] . (22) r_{\bullet j}^{(3)}(1,1,m_{\bullet(-j)},c,n)=\mathbb{E}\left[u_{\bullet j}^{(2)}(1,M,C,N)\frac{\kappa_{\bullet j}(M_{\bullet j}\mid1,C,N)}{\kappa_{\bullet j}^{*}(M_{\bullet j}\mid M_{\bullet(-j)},1,C,N)}\ \middle|\ A=1,M_{\bullet(-j)}=m_{\bullet(-j)},C=c,N=n\right]. \tag*{(22)} r ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , c , n ) = E [ u ∙ j ( 2 ) ( 1 , M , C , N ) κ ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 1 , C , N ) κ ∙ j ( M ∙ j ∣ 1 , C , N ) A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ] . (22)
r ∙ j ( 4 ) ( 1 , 1 , 0 , c , n ) = ∫ r ∙ j ( 2 ) ( 1 , 0 , m ∙ j , c , n ) κ ∙ j ( m ∙ j ∣ 1 , c , n ) d m ∙ j = E { r ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 0 , C = c , N = n } . (23) \begin{aligned}
r_{\bullet j}^{(4)}(1,1,0,c,n)
&=\int r_{\bullet j}^{(2)}(1,0,m_{\bullet j},c,n)\kappa_{\bullet j}(m_{\bullet j}\mid1,c,n)\,dm_{\bullet j}\\
&=\mathbb{E}\{r_{\bullet j}^{(3)}(1,1,M_{\bullet(-j)},C,N)\mid A=0,C=c,N=n\}. \tag*{(23)}
\end{aligned} r ∙ j ( 4 ) ( 1 , 1 , 0 , c , n ) = ∫ r ∙ j ( 2 ) ( 1 , 0 , m ∙ j , c , n ) κ ∙ j ( m ∙ j ∣ 1 , c , n ) d m ∙ j = E { r ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 0 , C = c , N = n } . (23)
When η = η 0 \eta=\eta_0 η = η 0 , these functions coincide with the auxiliary functions in Section 3. The proof is given in Section G.2 of the Appendix.
Substituting the auxiliary functions introduced in Section 5 into the canonical scores, we obtain
ψ θ , l d ( a , a ′ ; O ) = W N ∑ j = 1 N [ I ( A = a ) π a d ( a , Z , M , C , N ) π a s ( a ′ ∣ M , C , N ) π a ′ s ( a ∣ M , C , N ) { Y ∙ j − η ∙ j ( a , Z , M , C , N ) } + I ( A = a ) π a { u ∙ j ( 1 ) ( a , a ′ , Z , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N ) } + I ( A = a ′ ) π a ′ { u ∙ j ( 2 ) ( a , M , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N ) } + u ∙ j ( 3 ) ( a , a ′ , C , N ) ] . (24) \begin{aligned}
\psi_{\theta,\mathrm{ld}}(a,a';O)
={}&\frac{W}{N}\sum_{j=1}^{N}\Bigg[
\frac{I(A=a)}{\pi_a}d(a,Z,M,C,N)
\frac{\pi_a\,s(a'\mid M,C,N)}{\pi_{a'}\,s(a\mid M,C,N)}
\{Y_{\bullet j}-\eta_{\bullet j}(a,Z,M,C,N)\}\\
&+\frac{I(A=a)}{\pi_a}
\{u_{\bullet j}^{(1)}(a,a',Z,C,N)-u_{\bullet j}^{(3)}(a,a',C,N)\}\\
&+\frac{I(A=a')}{\pi_{a'}}
\{u_{\bullet j}^{(2)}(a,M,C,N)-u_{\bullet j}^{(3)}(a,a',C,N)\}
+u_{\bullet j}^{(3)}(a,a',C,N)
\Bigg]. \tag*{(24)}
\end{aligned} ψ θ , ld ( a , a ′ ; O ) = N W j = 1 ∑ N [ π a I ( A = a ) d ( a , Z , M , C , N ) π a ′ s ( a ∣ M , C , N ) π a s ( a ′ ∣ M , C , N ) { Y ∙ j − η ∙ j ( a , Z , M , C , N )} + π a I ( A = a ) { u ∙ j ( 1 ) ( a , a ′ , Z , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N )} + π a ′ I ( A = a ′ ) { u ∙ j ( 2 ) ( a , M , C , N ) − u ∙ j ( 3 ) ( a , a ′ , C , N )} + u ∙ j ( 3 ) ( a , a ′ , C , N ) ] . (24)
and
ψ τ , l d ( O ) = W N ∑ j = 1 N [ A π e ∙ j ( Z , M , C , N ) π 1 − π s − j ( 0 ∣ M ∙ ( − j ) , C , N ) s − j ( 1 ∣ M ∙ ( − j ) , C , N ) { Y ∙ j − η ∙ j ( 1 , Z , M , C , N ) } + A π { r ∙ j ( 1 ) ( 1 , 1 , 0 , Z , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } + A π { r ∙ j ( 2 ) ( 1 , 0 , M ∙ j , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } + 1 − A 1 − π { r ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) } + r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) ] . (25) \psi_{\tau,\mathrm{ld}}(O)
=
\frac{W}{N}\sum_{j=1}^{N}\left[
\frac{A}{\pi}e_{\bullet j}(Z,M,C,N)\frac{\pi}{1-\pi}\frac{s_{-j}(0\mid M_{\bullet(-j)},C,N)}{s_{-j}(1\mid M_{\bullet(-j)},C,N)}\{Y_{\bullet j}-\eta_{\bullet j}(1,Z,M,C,N)\}
+\frac{A}{\pi}\{r_{\bullet j}^{(1)}(1,1,0,Z,C,N)-r_{\bullet j}^{(4)}(1,1,0,C,N)\}
+\frac{A}{\pi}\{r_{\bullet j}^{(2)}(1,0,M_{\bullet j},C,N)-r_{\bullet j}^{(4)}(1,1,0,C,N)\}
+\frac{1-A}{1-\pi}\{r_{\bullet j}^{(3)}(1,1,M_{\bullet(-j)},C,N)-r_{\bullet j}^{(4)}(1,1,0,C,N)\}+r_{\bullet j}^{(4)}(1,1,0,C,N)
\right]. \tag*{(25)} ψ τ , ld ( O ) = N W j = 1 ∑ N [ π A e ∙ j ( Z , M , C , N ) 1 − π π s − j ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( 0 ∣ M ∙ ( − j ) , C , N ) { Y ∙ j − η ∙ j ( 1 , Z , M , C , N )} + π A { r ∙ j ( 1 ) ( 1 , 1 , 0 , Z , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} + π A { r ∙ j ( 2 ) ( 1 , 0 , M ∙ j , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} + 1 − π 1 − A { r ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) − r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N )} + r ∙ j ( 4 ) ( 1 , 1 , 0 , C , N ) ] . (25)
The following lemma shows that the reparameterized scores recover the canonical scores at the true nuisance functions.
Lemma 1. At the true nuisance functions, ψ θ , l d ( a , a ′ ; O ) \psi_{\theta,\mathrm{ld}}(a,a';O) ψ θ , ld ( a , a ′ ; O ) equals ψ θ ( a , a ′ ; O ) \psi_{\theta}(a,a';O) ψ θ ( a , a ′ ; O ) in (5) , and ψ τ , l d ( O ) \psi_{\tau,\mathrm{ld}}(O) ψ τ , ld ( O ) equals ψ τ ( O ) \psi_{\tau}(O) ψ τ ( O ) in (7) .
The proof is provided in Section G.3 of the Appendix. Thus, the reparameterization preserves the canonical gradients at the true nuisance functions.
Bias decomposition and multiple robustness
The next theorem gives exact bias decompositions for the reparameterized scores within the compatible class.
Theorem 4. For every compatible candidate nuisance tuple,
E { ψ θ , l d ( a , a ′ ; O ) } − E ( W ) θ 0 ( a , a ′ ) = E [ W N ∑ j = 1 N E [ π a π a ′ s ( a ′ ∣ M , C , N ) s ( a ∣ M , C , N ) ( d − d 0 ) ( a , Z , M , C , N ) ( η 0 ∙ j − η ∙ j ) ( a , Z , M , C , N ) + π a π a ′ { s ( a ′ ∣ M , C , N ) s ( a ∣ M , C , N ) − s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) } { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) } ( a , M , C , N ) ∣ A = a , C , N ] ] . (26) \begin{aligned}
\mathbb{E}\{\psi_{\theta,\mathrm{ld}}(a,a';O)\}-\mathbb{E}(W)\theta_0(a,a')
={}&\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\mathbb{E}\left[
\frac{\pi_a}{\pi_{a'}}\frac{s(a'\mid M,C,N)}{s(a\mid M,C,N)}(d-d_0)(a,Z,M,C,N)(\eta_{0\bullet j}-\eta_{\bullet j})(a,Z,M,C,N)\right.\right.\\
&\left.\left.\qquad+\frac{\pi_a}{\pi_{a'}}\left\{\frac{s(a'\mid M,C,N)}{s(a\mid M,C,N)}-\frac{s_0(a'\mid M,C,N)}{s_0(a\mid M,C,N)}\right\}\{u_{0\bullet j}^{(2)}-u_{\bullet j}^{(2)}\}(a,M,C,N)\mid A=a,C,N\right]\right]. \tag*{(26)}
\end{aligned} E { ψ θ , ld ( a , a ′ ; O )} − E ( W ) θ 0 ( a , a ′ ) = E [ N W j = 1 ∑ N E [ π a ′ π a s ( a ∣ M , C , N ) s ( a ′ ∣ M , C , N ) ( d − d 0 ) ( a , Z , M , C , N ) ( η 0 ∙ j − η ∙ j ) ( a , Z , M , C , N ) + π a ′ π a { s ( a ∣ M , C , N ) s ( a ′ ∣ M , C , N ) − s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) } { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) } ( a , M , C , N ) ∣ A = a , C , N ] ] . (26) Moreover,
E { ψ τ , l d ( O ) } − E ( W ) τ 0 = E [ W N ∑ j = 1 N E [ π 1 − π s − j ( 0 ∣ M ∙ ( − j ) , C , N ) s − j ( 1 ∣ M ∙ ( − j ) , C , N ) ( e ∙ j − e 0 ∙ j ) ( Z , M , C , N ) ( η 0 ∙ j − η ∙ j ) ( 1 , Z , M , C , N ) + π 1 − π { s − j ( 0 ∣ M ∙ ( − j ) , C , N ) s − j ( 1 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) } { r ∙ j ( 3 ) − r 0 ∙ j ( 3 ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 1 , C , N ] ] . (27) \begin{aligned}
\mathbb{E}\{\psi_{\tau,\mathrm{ld}}(O)\}-\mathbb{E}(W)\tau_0
={}&\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\mathbb{E}\left[
\frac{\pi}{1-\pi}\frac{s_{-j}(0\mid M_{\bullet(-j)},C,N)}{s_{-j}(1\mid M_{\bullet(-j)},C,N)}
(e_{\bullet j}-e_{0\bullet j})(Z,M,C,N)(\eta_{0\bullet j}-\eta_{\bullet j})(1,Z,M,C,N)\right.\right.\\
&\left.\left.\qquad+\frac{\pi}{1-\pi}\left\{\frac{s_{-j}(0\mid M_{\bullet(-j)},C,N)}{s_{-j}(1\mid M_{\bullet(-j)},C,N)}-\frac{s_{0,-j}(0\mid M_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid M_{\bullet(-j)},C,N)}\right\}\{r_{\bullet j}^{(3)}-r_{0\bullet j}^{(3)}\}(1,1,M_{\bullet(-j)},C,N)\mid A=1,C,N\right]\right]. \tag*{(27)}
\end{aligned} E { ψ τ , ld ( O )} − E ( W ) τ 0 = E [ N W j = 1 ∑ N E [ 1 − π π s − j ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( 0 ∣ M ∙ ( − j ) , C , N ) ( e ∙ j − e 0 ∙ j ) ( Z , M , C , N ) ( η 0 ∙ j − η ∙ j ) ( 1 , Z , M , C , N ) + 1 − π π { s − j ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( 0 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) } { r ∙ j ( 3 ) − r 0 ∙ j ( 3 ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 1 , C , N ] ] . (27) The proof is given in Section G.4 of the Appendix. Each decomposition consists of two products of nuisance errors. Theorem 4 gives the following sufficient conditions for zero population bias for
{ d = d 0 or η = η 0 } and { s ( a ′ ∣ M , C , N ) s ( a ∣ M , C , N ) = s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) or u ( 2 ) = u 0 ( 2 ) } . (28) \{d=d_0\text{ or }\eta=\eta_0\}\quad\text{and}\quad
\left\{\frac{s(a'\mid M,C,N)}{s(a\mid M,C,N)}
=\frac{s_0(a'\mid M,C,N)}{s_0(a\mid M,C,N)}
\text{ or }u^{(2)}=u_0^{(2)}\right\}. \tag*{(28)} { d = d 0 or η = η 0 } and { s ( a ∣ M , C , N ) s ( a ′ ∣ M , C , N ) = s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) or u ( 2 ) = u 0 ( 2 ) } . (28)
For τ \tau τ , the corresponding conditions are, for every j j j ,
{ e ∙ j = e 0 ∙ j or η ∙ j = η 0 ∙ j } and { s − j ( 0 ∣ M ∙ ( − j ) , C , N ) s − j ( 1 ∣ M ∙ ( − j ) , C , N ) = s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) or r ∙ j ( 3 ) = r 0 ∙ j ( 3 ) } . (29) \{e_{\bullet j}=e_{0\bullet j}\text{ or }\eta_{\bullet j}=\eta_{0\bullet j}\}\quad\text{and}\quad
\left\{\frac{s_{-j}(0\mid M_{\bullet(-j)},C,N)}{s_{-j}(1\mid M_{\bullet(-j)},C,N)}
=\frac{s_{0,-j}(0\mid M_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid M_{\bullet(-j)},C,N)}
\text{ or }r_{\bullet j}^{(3)}=r_{0\bullet j}^{(3)}\right\}. \tag*{(29)} { e ∙ j = e 0 ∙ j or η ∙ j = η 0 ∙ j } and { s − j ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( 0 ∣ M ∙ ( − j ) , C , N ) = s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) or r ∙ j ( 3 ) = r 0 ∙ j ( 3 ) } . (29)
Under compatibility, correct η \eta η also gives the correct u ( 2 ) u^{(2)} u ( 2 ) and r ∙ j ( 3 ) r_{\bullet j}^{(3)} r ∙ j ( 3 ) , making both population bias terms zero. Otherwise, correct d d d or e ∙ j e_{\bullet j} e ∙ j removes the first term, while the second term vanishes if either the corresponding classifier-odds or integrated regression is correct. For a = a ′ a=a' a = a ′ , the classifier-odds ratio is identically one, so the second term for θ ( a , a ) \theta(a,a) θ ( a , a ) is zero. When e ∙ j e_{\bullet j} e ∙ j is constructed through (14) , correctness of d ( 1 , ∙ ) d(1,\bullet) d ( 1 , ∙ ) and κ ∙ j / κ ∙ j ⋆ \kappa_{\bullet j}/\kappa_{\bullet j}^{\star} κ ∙ j / κ ∙ j ⋆ is sufficient for its correctness.
Asymptotic properties
Use the same cluster-level folds as in § 4 . For each fold I r \mathcal{I}_r I r , estimate the low-dimensional nuisance functions and their associated auxiliary regressions outside the fold and evaluate the scores on its validation clusters.
The results through Theorem 5 concern fitted tuples in the compatible class. Section A.3 gives conditions for (30) and (31) with estimated auxiliary regressions. Define
θ ^ l d ( a , a ′ ) = K − 1 ∑ r = 1 R ∑ i ∈ I r ψ ^ θ , r ( O i ) P K ( W ) . (30) \widehat{\theta}_{\mathrm{ld}}(a,a')
=
\frac{K^{-1}\sum_{r=1}^{R}\sum_{i\in\mathcal{I}_r}\widehat{\psi}_{\theta,r}(O_i)}
{\mathbb{P}_K(W)}. \tag*{(30)} θ ld ( a , a ′ ) = P K ( W ) K − 1 ∑ r = 1 R ∑ i ∈ I r ψ θ , r ( O i ) . (30)
τ ^ l d = K − 1 ∑ r = 1 R ∑ i ∈ I r ψ ^ τ , r ( O i ) P K ( W ) . (31) \widehat{\tau}_{\mathrm{ld}}
=
\frac{K^{-1}\sum_{r=1}^{R}\sum_{i\in\mathcal{I}_r}\widehat{\psi}_{\tau,r}(O_i)}
{\mathbb{P}_K(W)}. \tag*{(31)} τ ld = P K ( W ) K − 1 ∑ r = 1 R ∑ i ∈ I r ψ τ , r ( O i ) . (31)
Let h θ = ( η , d , s , u ( 1 ) , u ( 2 ) , u ( 3 ) ) h_\theta=(\eta,d,s,u^{(1)},u^{(2)},u^{(3)}) h θ = ( η , d , s , u ( 1 ) , u ( 2 ) , u ( 3 ) ) and h τ = ( η , e , s − j , r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) ) h_\tau=(\eta,e,s_{-j},r^{(1)},r^{(2)},r^{(3)},r^{(4)}) h τ = ( η , e , s − j , r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) ) . These tuples include both the nuisance functions ( η , d , s ) (\eta,d,s) ( η , d , s ) or ( η , e , s − j ) (\eta,e,s_{-j}) ( η , e , s − j ) and the associated auxiliary regressions u ( 1 ) , u ( 2 ) , u ( 3 ) u^{(1)},u^{(2)},u^{(3)} u ( 1 ) , u ( 2 ) , u ( 3 ) or r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) r^{(1)},r^{(2)},r^{(3)},r^{(4)} r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) , so that the convergence requirements below cover both. Again, for vector-indexed nuisance or auxiliary functions, every norm below is understood as the maximum over j ≤ N max j\leq N_{\max} j ≤ N m a x .
The following assumption specifies the nuisance limits and error products for the reparameterized scores. Denote ν 0 = E ( W ) ∈ ( 0 , ∞ ) \nu_0=\mathbb{E}(W)\in(0,\infty) ν 0 = E ( W ) ∈ ( 0 , ∞ ) .
Assumption D. The weight W W W is nonnegative and bounded. For each target, there exists a fixed compatible nuisance limit and the following conditions hold uniformly over the fixed number of folds.
(i) For θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) ,
∥ η ^ ( − r ) − η ⋆ ∥ 2 , P + ∥ d ^ ( − r ) − d ⋆ ∥ 2 , P + ∥ s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) − s ⋆ ( a ′ ∣ M , C , N ) s ⋆ ( a ∣ M , C , N ) ∥ 2 , P + ∑ ℓ = 1 3 ∥ u ^ ( ℓ ) , ( − r ) − u ( ℓ ) , ⋆ ∥ 2 , P = o p ( 1 ) . (32) \left\|\widehat{\eta}^{(-r)}-\eta^\star\right\|_{2,P}
+\left\|\widehat{d}^{(-r)}-d^\star\right\|_{2,P}
+\left\|
\frac{\widehat{s}^{(-r)}(a'\mid M,C,N)}
{\widehat{s}^{(-r)}(a\mid M,C,N)}
-
\frac{s^\star(a'\mid M,C,N)}
{s^\star(a\mid M,C,N)}
\right\|_{2,P}
+\sum_{\ell=1}^{3}
\left\|\widehat{u}^{(\ell),(-r)}-u^{(\ell),\star}\right\|_{2,P}
=o_p(1). \tag*{(32)} η ( − r ) − η ⋆ 2 , P + d ( − r ) − d ⋆ 2 , P + s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) − s ⋆ ( a ∣ M , C , N ) s ⋆ ( a ′ ∣ M , C , N ) 2 , P + ℓ = 1 ∑ 3 u ( ℓ ) , ( − r ) − u ( ℓ ) , ⋆ 2 , P = o p ( 1 ) . (32)
The limit satisfies (28) , and
max 1 ≤ r ≤ R [ ∥ W { d ^ ( − r ) − d 0 } ∥ 2 , P ∥ W { η ^ ( − r ) − η 0 } ∥ 2 , P + ∥ W { s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) − s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) } ∥ 2 , P ∥ W { u ^ ( 2 ) , ( − r ) − u 0 ( 2 ) } ∥ 2 , P ] = o p ( K − 1 / 2 ) . (33) \max_{1\le r\le R}\left[\left\|\sqrt{W}\{\widehat{d}^{(-r)}-d_0\}\right\|_{2,P}\left\|\sqrt{W}\{\widehat{\eta}^{(-r)}-\eta_0\}\right\|_{2,P}
+\left\|\sqrt{W}\left\{\frac{\widehat{s}^{(-r)}(a' \mid M,C,N)}{\widehat{s}^{(-r)}(a \mid M,C,N)}-\frac{s_0(a' \mid M,C,N)}{s_0(a \mid M,C,N)}\right\}\right\|_{2,P}\left\|\sqrt{W}\{\widehat{u}^{(2),(-r)}-u_0^{(2)}\}\right\|_{2,P}\right]=o_p(K^{-1/2}). \tag*{(33)} 1 ≤ r ≤ R max [ W { d ( − r ) − d 0 } 2 , P W { η ( − r ) − η 0 } 2 , P + W { s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) − s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) } 2 , P W { u ( 2 ) , ( − r ) − u 0 ( 2 ) } 2 , P ] = o p ( K − 1/2 ) . (33)
(ii) For τ \tau τ ,
∥ η ^ ( − r ) − η ⋆ ∥ 2 , P + ∥ e ^ ( − r ) − e ⋆ ∥ 2 , P + max 1 ≤ j ≤ N max ∥ s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) − s − j ⋆ ( 0 ∣ M ∙ ( − j ) , C , N ) s − j ⋆ ( 1 ∣ M ∙ ( − j ) , C , N ) ∥ 2 , P + ∑ ℓ = 1 4 ∥ r ( ℓ ) , ( − r ) − r ( ℓ ) , ⋆ ∥ 2 , P = o p ( 1 ) . (34) \left\|\widehat{\eta}^{(-r)}-\eta^\star\right\|_{2,P}+\left\|\widehat{e}^{(-r)}-e^\star\right\|_{2,P}+\max_{1\le j\le N_{\max}}\left\|\frac{\widehat{s}_{-j}^{(-r)}(0\mid M_{\bullet(-j)},C,N)}{\widehat{s}_{-j}^{(-r)}(1\mid M_{\bullet(-j)},C,N)}-\frac{s_{-j}^\star(0\mid M_{\bullet(-j)},C,N)}{s_{-j}^\star(1\mid M_{\bullet(-j)},C,N)}\right\|_{2,P}
+\sum_{\ell=1}^{4}\left\|r^{(\ell),(-r)}-r^{(\ell),\star}\right\|_{2,P}=o_p(1). \tag*{(34)} η ( − r ) − η ⋆ 2 , P + e ( − r ) − e ⋆ 2 , P + 1 ≤ j ≤ N m a x max s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) − s − j ⋆ ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ⋆ ( 0 ∣ M ∙ ( − j ) , C , N ) 2 , P + ℓ = 1 ∑ 4 r ( ℓ ) , ( − r ) − r ( ℓ ) , ⋆ 2 , P = o p ( 1 ) . (34)
The limit satisfies (29) for every j j j , and
max 1 ≤ r ≤ R max 1 ≤ j ≤ N max [ ∥ W { e ^ ∙ j ( − r ) − e 0 ∙ j } ∥ 2 , P ∥ W { η ^ ∙ j ( − r ) − η 0 ∙ j } ∥ 2 , P + ∥ W { s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) } ∥ 2 , P ∥ W { r ^ ∙ j ( 3 ) , ( − r ) − r 0 ∙ j ( 3 ) } ∥ 2 , P ] = o p ( K − 1 / 2 ) . (35) \max_{1\le r\le R}\max_{1\le j\le N_{\max}}\left[\left\|\sqrt{W}\{\widehat{e}_{\bullet j}^{(-r)}-e_{0\bullet j}\}\right\|_{2,P}\left\|\sqrt{W}\{\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{0\bullet j}\}\right\|_{2,P}
+\left\|\sqrt{W}\left\{\frac{\widehat{s}_{-j}^{(-r)}(0\mid M_{\bullet(-j)},C,N)}{\widehat{s}_{-j}^{(-r)}(1\mid M_{\bullet(-j)},C,N)}-\frac{s_{0,-j}(0\mid M_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid M_{\bullet(-j)},C,N)}\right\}\right\|_{2,P}\left\|\sqrt{W}\{\widehat{r}_{\bullet j}^{(3),(-r)}-r_{0\bullet j}^{(3)}\}\right\|_{2,P}\right]=o_p(K^{-1/2}). \tag*{(35)} 1 ≤ r ≤ R max 1 ≤ j ≤ N m a x max W { e ∙ j ( − r ) − e 0 ∙ j } 2 , P W { η ∙ j ( − r ) − η 0 ∙ j } 2 , P + W { s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) } 2 , P W { r ∙ j ( 3 ) , ( − r ) − r 0 ∙ j ( 3 ) } 2 , P = o p ( K − 1/2 ) . (35)
Assumption D includes convergence of all auxiliary regressions entering the scores. The conditions in (33) and (35), which are required for asymptotic linearity, require each paired error product to be o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) . As in Section 4, zero population bias at the nuisance limit does not by itself establish these rates.
Define
ψ θ , l d ⋆ ( O ) = ψ θ , l d ( a , a ′ ; O ; h θ ⋆ ) , D θ , l d ⋆ ( O ) = ψ θ , l d ⋆ ( O ) − θ 0 ( a , a ′ ) ν 0 W , \psi^\star_{\theta,\mathrm{ld}}(O)=\psi_{\theta,\mathrm{ld}}(a,a';O;h_\theta^\star),\qquad D^\star_{\theta,\mathrm{ld}}(O)=\frac{\psi^\star_{\theta,\mathrm{ld}}(O)-\theta_0(a,a')}{\nu_0}W, ψ θ , ld ⋆ ( O ) = ψ θ , ld ( a , a ′ ; O ; h θ ⋆ ) , D θ , ld ⋆ ( O ) = ν 0 ψ θ , ld ⋆ ( O ) − θ 0 ( a , a ′ ) W ,
ψ τ , l d ⋆ ( O ) = ψ τ , l d ( O ; h τ ⋆ ) , D τ , l d ⋆ ( O ) = ψ τ , l d ⋆ ( O ) − τ 0 ν 0 W . \psi^\star_{\tau,\mathrm{ld}}(O)=\psi_{\tau,\mathrm{ld}}(O;h_\tau^\star),\qquad D^\star_{\tau,\mathrm{ld}}(O)=\frac{\psi^\star_{\tau,\mathrm{ld}}(O)-\tau_0}{\nu_0}W. ψ τ , ld ⋆ ( O ) = ψ τ , ld ( O ; h τ ⋆ ) , D τ , ld ⋆ ( O ) = ν 0 ψ τ , ld ⋆ ( O ) − τ 0 W .
Assume the boundedness conditions stated at the beginning of Section G.5. The following theorem states the asymptotic expansions for the low-dimensional estimators.
Theorem 5. Under Assumptions B and D(i),
K { θ ^ l d ( a , a ′ ) − θ 0 ( a , a ′ ) } = 1 K ∑ i = 1 K D θ , l d ⋆ ( O i ) + o p ( 1 ) . (36) \sqrt{K}\{\widehat{\theta}_{\mathrm{ld}}(a,a')-\theta_0(a,a')\}=\frac{1}{\sqrt{K}}\sum_{i=1}^{K}D^\star_{\theta,\mathrm{ld}}(O_i)+o_p(1). \tag*{(36)} K { θ ld ( a , a ′ ) − θ 0 ( a , a ′ )} = K 1 i = 1 ∑ K D θ , ld ⋆ ( O i ) + o p ( 1 ) . (36) Under Assumptions B and D(ii),
K ( τ ^ l d − τ 0 ) = 1 K ∑ i = 1 K D τ , l d ⋆ ( O i ) + o p ( 1 ) . (37) \sqrt{K}(\widehat{\tau}_{\mathrm{ld}}-\tau_0)=\frac{1}{\sqrt{K}}\sum_{i=1}^{K}D^\star_{\tau,\mathrm{ld}}(O_i)+o_p(1). \tag*{(37)} K ( τ ld − τ 0 ) = K 1 i = 1 ∑ K D τ , ld ⋆ ( O i ) + o p ( 1 ) . (37) Consequently, the estimators converge to centered normal limits with variances Σ θ , l d ⋆ = E { ( D θ , l d ⋆ ) 2 } \Sigma^\star_{\theta,\mathrm{ld}}=\mathbb{E}\{(D^\star_{\theta,\mathrm{ld}})^2\} Σ θ , ld ⋆ = E {( D θ , ld ⋆ ) 2 } and Σ τ , l d ⋆ = E { ( D τ , l d ⋆ ) 2 } \Sigma^\star_{\tau,\mathrm{ld}}=\mathbb{E}\{(D^\star_{\tau,\mathrm{ld}})^2\} Σ τ , ld ⋆ = E {( D τ , ld ⋆ ) 2 } , respectively. If all nuisance limits are correct, the corresponding variances attain the nonparametric efficiency bounds.
The proof is deferred to Section G.5 of the Appendix. At the true nuisance limits, Lemma 1 identifies the limiting scores with the canonical scores, giving the same efficiency bounds as in § 4 .
For i ∈ I r i \in\mathcal{I}_r i ∈ I r , define D ^ θ , l d , i = ψ ^ θ , r ( O i ) − θ ^ l d ( a , a ′ ) W i P K ( W ) \widehat{D}_{\theta,\mathrm{ld},i} = \frac{\widehat{\psi}_{\theta,r}(O_i)-\widehat{\theta}_{\mathrm{ld}}(a,a^\prime)W_i}{\mathbb{P}_K(W)} D θ , ld , i = P K ( W ) ψ θ , r ( O i ) − θ ld ( a , a ′ ) W i and D ^ τ , l d , i = ψ ^ τ , r ( O i ) − τ ^ l d W i P K ( W ) \widehat{D}_{\tau,\mathrm{ld},i} = \frac{\widehat{\psi}_{\tau,r}(O_i)-\widehat{\tau}_{\mathrm{ld}}W_i}{\mathbb{P}_K(W)} D τ , ld , i = P K ( W ) ψ τ , r ( O i ) − τ ld W i . The variance estimators associated with Theorem 5 are P K ( D ^ θ , l d 2 ) \mathbb{P}_K(\widehat{D}_{\theta,\mathrm{ld}}^2) P K ( D θ , ld 2 ) and P K ( D ^ τ , l d 2 ) \mathbb{P}_K(\widehat{D}_{\tau,\mathrm{ld}}^2) P K ( D τ , ld 2 ) .
Consistency and asymptotic normality with estimated auxiliary regressions
We study (30) and (31) with every fitted function in (24) and (25) estimated outside I r \mathcal{I}_r I r . Expectations involving these functions integrate a new cluster under P P P conditional on the training observations. The sampling and moment conditions of Assumption (A4), the data-independent folds of Assumption B, and the boundedness requirement in Section G.5 apply to the asymptotic results below. We maintain h θ , l d ⋆ h^\star_{\theta,\mathrm{ld}} h θ , ld ⋆ , h τ , l d ⋆ h^\star_{\tau,\mathrm{ld}} h τ , ld ⋆ , ψ θ , l d ⋆ \psi^\star_{\theta,\mathrm{ld}} ψ θ , ld ⋆ , ψ τ , l d ⋆ \psi^\star_{\tau,\mathrm{ld}} ψ τ , ld ⋆ , D θ , l d ⋆ D^\star_{\theta,\mathrm{ld}} D θ , ld ⋆ , and D τ , l d ⋆ D^\star_{\tau,\mathrm{ld}} D τ , ld ⋆ as specified in Assumption D and Theorem 5 . The functions u ^ ( 1 ) , ( − r ) \widehat{u}^{(1),(-r)} u ( 1 ) , ( − r ) , u ^ ( 2 ) , ( − r ) \widehat{u}^{(2),(-r)} u ( 2 ) , ( − r ) , u ^ ( 3 ) , ( − r ) \widehat{u}^{(3),(-r)} u ( 3 ) , ( − r ) and r ^ ( 1 ) , ( − r ) , … , r ^ ( 4 ) , ( − r ) \widehat{r}^{(1),(-r)}, \ldots, \widehat{r}^{(4),(-r)} r ( 1 ) , ( − r ) , … , r ( 4 ) , ( − r ) may be fitted separately.
Theorem 6. Suppose the support conditions in Assumption (A4) hold, the fitted ratios below are finite on the support of P P P , and all displayed expectations and integrals are absolutely integrable. For every fold r r r , (24) gives the following bias decomposition:
E { ψ ^ θ , r ( O ) } − ν 0 θ 0 ( a , a ′ ) = E [ W N ∑ j = 1 N E [ π a π a ′ s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) { d ^ ( − r ) − d 0 } ( a , Z , M , C , N ) { η 0 ∙ j − η ^ ∙ j ( − r ) } ( a , Z , M , C , N ) + { π a π a ′ s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) − π a π a ′ s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) } { u ^ 0 ∙ j ( 2 ) − u ^ ∙ j ( 2 ) , ( − r ) } ( a , M , C , N ) ∣ A = a , C , N ] + E [ { u ^ ∙ j ( 1 ) , ( − r ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } − u ^ ∙ j ( 3 ) , ( − r ) ( a , a ′ , C , N ) + ∫ κ 0 ( m ∣ a , C , N ) π a π a ′ s ^ ( − r ) ( a ′ ∣ m , C , N ) s ^ ( − r ) ( a ∣ m , C , N ) { u ^ ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) − ∫ η ^ ∙ j ( − r ) ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) d z d m } ] ] . (38) \begin{aligned}
\mathbb{E}\{\widehat{\psi}_{\theta,r}(O)\}-\nu_0\theta_0(a,a^\prime)
={}&\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\mathbb{E}\left[\frac{\pi_a}{\pi_{a^\prime}}\frac{\widehat{s}^{(-r)}(a^\prime\mid M,C,N)}{\widehat{s}^{(-r)}(a\mid M,C,N)}
\{\widehat{d}^{(-r)}-d_0\}(a,Z,M,C,N)\{\eta_{0\bullet j}-\widehat{\eta}^{(-r)}_{\bullet j}\}(a,Z,M,C,N)\right.\right.\\
&\left.\left.\quad+\left\{\frac{\pi_a}{\pi_{a^\prime}}\frac{\widehat{s}^{(-r)}(a^\prime\mid M,C,N)}{\widehat{s}^{(-r)}(a\mid M,C,N)}
-\frac{\pi_a}{\pi_{a^\prime}}\frac{s_0(a^\prime\mid M,C,N)}{s_0(a\mid M,C,N)}\right\}
\{\widehat{u}^{(2)}_{0\bullet j}-\widehat{u}^{(2),(-r)}_{\bullet j}\}(a,M,C,N)\mid A=a,C,N\right]\right.\\
&\left.+\mathbb{E}\left[\{\widehat{u}^{(1),(-r)}_{\bullet j}(a,a^\prime,Z,C,N)\mid A=a,C,N\}
-\widehat{u}^{(3),(-r)}_{\bullet j}(a,a^\prime,C,N)\right.\right.\\
&\left.\left.\quad+\int\kappa_0(m\mid a,C,N)\frac{\pi_a}{\pi_{a^\prime}}\frac{\widehat{s}^{(-r)}(a^\prime\mid m,C,N)}{\widehat{s}^{(-r)}(a\mid m,C,N)}
\left\{\widehat{u}^{(2),(-r)}_{\bullet j}(a,m,C,N)
-\int\widehat{\eta}^{(-r)}_{\bullet j}(a,z,m,C,N)\phi_0(z\mid a,C,N)\,dz\,dm\right\}\right]\right]. \tag*{(38)}
\end{aligned} E { ψ θ , r ( O )} − ν 0 θ 0 ( a , a ′ ) = E [ N W j = 1 ∑ N E [ π a ′ π a s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) { d ( − r ) − d 0 } ( a , Z , M , C , N ) { η 0 ∙ j − η ∙ j ( − r ) } ( a , Z , M , C , N ) + { π a ′ π a s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) − π a ′ π a s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) } { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) , ( − r ) } ( a , M , C , N ) ∣ A = a , C , N ] + E [ { u ∙ j ( 1 ) , ( − r ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } − u ∙ j ( 3 ) , ( − r ) ( a , a ′ , C , N ) + ∫ κ 0 ( m ∣ a , C , N ) π a ′ π a s ( − r ) ( a ∣ m , C , N ) s ( − r ) ( a ′ ∣ m , C , N ) { u ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) − ∫ η ∙ j ( − r ) ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) d z d m } ] ] . (38)
Likewise, (25) gives
E { ψ ^ τ , r ( O ) } − ν 0 τ 0 = E [ W N ∑ j = 1 N E [ π 1 − π s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) { e ^ ∙ j ( − r ) − e 0 ∙ j } ( Z , M , C , N ) { η 0 ∙ j − η ^ ∙ j ( − r ) } ( 1 , Z , M , C , N ) + { π 1 − π s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) − π 1 − π s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) } { r ^ 0 ∙ j ( 3 ) − r ^ ∙ j ( 3 ) , ( − r ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 1 , C , N ] + E [ { r ^ ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , Z , C , N ) ∣ A = 1 , C , N } + r ^ ∙ j ( 2 ) , ( − r ) ( 1 , 0 , M ∙ j , C , N ) ∣ A = 1 , C , N − 2 r ^ ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , C , N ) ] + ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) π 1 − π s ^ − j ( − r ) ( 0 ∣ m ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ m ∙ ( − j ) , C , N ) [ r ^ ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) − ∬ η ^ ∙ j ( − r ) ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j d m ∙ ( − j ) ] ] ] . (39) \begin{aligned}
\mathbb{E}\{\widehat{\psi}_{\tau,r}(O)\}-\nu_0\tau_0
={}&\mathbb{E}[\frac{W}{N}\sum_{j=1}^{N}\mathbb{E}[\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid M_{\bullet(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid M_{\bullet(-j)},C,N)}
\{\widehat{e}^{(-r)}_{\bullet j}-e_{0\bullet j}\}(Z,M,C,N)\{\eta_{0\bullet j}-\widehat{\eta}^{(-r)}_{\bullet j}\}(1,Z,M,C,N)\\
&\quad+\{\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid M_{\bullet(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid M_{\bullet(-j)},C,N)}
-\frac{\pi}{1-\pi}\frac{s_{0,-j}(0\mid M_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid M_{\bullet(-j)},C,N)}\}
\{\widehat{r}^{(3)}_{0\bullet j}-\widehat{r}^{(3),(-r)}_{\bullet j}\}(1,1,M_{\bullet(-j)},C,N)\mid A=1,C,N]\\
&+\mathbb{E}[\{\widehat{r}^{(1),(-r)}_{\bullet j}(1,1,0,Z,C,N)\mid A=1,C,N\}\\
&\quad+\widehat{r}^{(2),(-r)}_{\bullet j}(1,0,M_{\bullet j},C,N)\mid A=1,C,N
-2\widehat{r}^{(4),(-r)}_{\bullet j}(1,1,0,C,N)]\\
&\quad+\int\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid1,C,N)
\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid m_{\bullet(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid m_{\bullet(-j)},C,N)}
[\widehat{r}^{(3),(-r)}_{\bullet j}(1,1,m_{\bullet(-j)},C,N)\\
&\quad-\iint\widehat{\eta}^{(-r)}_{\bullet j}(1,z,m,C,N)\phi_0(z\mid1,C,N)\kappa_{0\bullet j}(m_{\bullet j}\mid1,C,N)\,dz\,dm_{\bullet j}\,dm_{\bullet(-j)}]]]. \tag*{(39)}
\end{aligned} E { ψ τ , r ( O )} − ν 0 τ 0 = E [ N W j = 1 ∑ N E [ 1 − π π s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) { e ∙ j ( − r ) − e 0 ∙ j } ( Z , M , C , N ) { η 0 ∙ j − η ∙ j ( − r ) } ( 1 , Z , M , C , N ) + { 1 − π π s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) − 1 − π π s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) } { r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) , ( − r ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 1 , C , N ] + E [{ r ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , Z , C , N ) ∣ A = 1 , C , N } + r ∙ j ( 2 ) , ( − r ) ( 1 , 0 , M ∙ j , C , N ) ∣ A = 1 , C , N − 2 r ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , C , N )] + ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) 1 − π π s − j ( − r ) ( 1 ∣ m ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ m ∙ ( − j ) , C , N ) [ r ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) − ∬ η ∙ j ( − r ) ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j d m ∙ ( − j ) ]]] . (39)
The proof is given in Section G.6. Under the integral equalities defining the compatible tuples in Section 5 and (20)–(23), (38) and (39) reduce to (26) and (27). The next result allows the fitted functions to approach those equalities through (32) and (34).
Theorem 7. Under the sampling, moment, fold, and boundedness requirements stated in Section A.3, the following assertions hold. For h θ ⋆ h_{\theta}^{\star} h θ ⋆ satisfying (28), (32) implies θ ^ l d ( a , a ′ ) → p θ 0 ( a , a ′ ) \widehat{\theta}_{\mathrm{ld}}(a,a') \to_{p} \theta_{0}(a,a') θ ld ( a , a ′ ) → p θ 0 ( a , a ′ ) and
max 1 ≤ r ≤ R ∣ E [ W N ∑ j = 1 N { E { u ^ ∙ j ( 1 ) , ( − r ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } − u ^ ∙ j ( 3 ) , ( − r ) ( a , a ′ , C , N ) \max_{1\le r\le R}\left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{\mathbb{E}\left\{\widehat{u}_{\bullet j}^{(1),(-r)}(a,a',Z,C,N)\mid A=a,C,N\right\}-\widehat{u}_{\bullet j}^{(3),(-r)}(a,a',C,N)\right.\right.\right. 1 ≤ r ≤ R max E [ N W j = 1 ∑ N { E { u ∙ j ( 1 ) , ( − r ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } − u ∙ j ( 3 ) , ( − r ) ( a , a ′ , C , N ) + ∫ κ 0 ( m ∣ a , C , N ) π a ( − r ) s ^ ( − r ) ( a ′ ∣ m , C , N ) π a ′ ( − r ) s ^ ( − r ) ( a ∣ m , C , N ) { u ^ ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) − ∫ η ^ ∙ j ( − r ) ( a , z , m , C , N ) φ ^ 0 ( z ∣ a , C , N ) d z d m } } ] ∣ = o p ( 1 ) . (40) \left.\left.\left.
+\int\kappa_{0}(m\mid a,C,N)\frac{\pi_{a}^{(-r)}\widehat{s}^{(-r)}(a'\mid m,C,N)}{\pi_{a'}^{(-r)}\widehat{s}^{(-r)}(a\mid m,C,N)}
\left\{\widehat{u}_{\bullet j}^{(2),(-r)}(a,m,C,N)-\int\widehat{\eta}_{\bullet j}^{(-r)}(a,z,m,C,N)\widehat{\varphi}_{0}(z\mid a,C,N)\,dz\,dm\right\}\right\}\right]\right|=o_{p}(1). \tag*{(40)} + ∫ κ 0 ( m ∣ a , C , N ) π a ′ ( − r ) s ( − r ) ( a ∣ m , C , N ) π a ( − r ) s ( − r ) ( a ′ ∣ m , C , N ) { u ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) − ∫ η ∙ j ( − r ) ( a , z , m , C , N ) φ 0 ( z ∣ a , C , N ) d z d m } } ] = o p ( 1 ) . (40) If, in addition, (33) holds and the left-hand side of (40) is o p ( K − 1 / 2 ) o_{p}(K^{-1/2}) o p ( K − 1/2 ) , then (36) holds and
K { θ ^ l d ( a , a ′ ) − θ 0 ( a , a ′ ) } ⇝ N ( 0 , Σ θ , l d ⋆ ) , P K ( D ^ θ , l d 2 ) → p Σ θ , l d ⋆ . (41) \sqrt{K}\{\widehat{\theta}_{\mathrm{ld}}(a,a')-\theta_{0}(a,a')\}\rightsquigarrow N(0,\Sigma_{\theta,\mathrm{ld}}^{\star}),\qquad\mathbb{P}_{K}(\widehat{D}_{\theta,\mathrm{ld}}^{2})\to_{p}\Sigma_{\theta,\mathrm{ld}}^{\star}. \tag*{(41)} K { θ ld ( a , a ′ ) − θ 0 ( a , a ′ )} ⇝ N ( 0 , Σ θ , ld ⋆ ) , P K ( D θ , ld 2 ) → p Σ θ , ld ⋆ . (41) For h τ ⋆ h_{\tau}^{\star} h τ ⋆ satisfying (29) for every j j j , (34) implies τ ^ l d → p τ 0 \widehat{\tau}_{\mathrm{ld}}\to_{p}\tau_{0} τ ld → p τ 0 and
max 1 ≤ r ≤ R ∣ E [ W N ∑ j = 1 N { E [ r ^ ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , Z , C , N ) ∣ A = 1 , C , N ] + E [ r ^ ∙ j ( 2 ) , ( − r ) ( 1 , 0 , M ∙ j , C , N ) ∣ A = 1 , C , N ] − 2 r ^ ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , C , N ) \max_{1\le r\le R}\left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{\mathbb{E}\left[\widehat{r}_{\bullet j}^{(1),(-r)}(1,1,0,Z,C,N)\mid A=1,C,N\right]+\mathbb{E}\left[\widehat{r}_{\bullet j}^{(2),(-r)}(1,0,M_{\bullet j},C,N)\mid A=1,C,N\right]-2\widehat{r}_{\bullet j}^{(4),(-r)}(1,1,0,C,N)\right.\right.\right. 1 ≤ r ≤ R max E [ N W j = 1 ∑ N { E [ r ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , Z , C , N ) ∣ A = 1 , C , N ] + E [ r ∙ j ( 2 ) , ( − r ) ( 1 , 0 , M ∙ j , C , N ) ∣ A = 1 , C , N ] − 2 r ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , C , N ) + κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) π ( − r ) ( 0 ∣ m ∙ ( − j ) , C , N ) 1 − π ( − r ) ( 1 ∣ m ∙ ( − j ) , C , N ) [ r ^ ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) − ∬ η ^ ∙ j ( − r ) ( 1 , z , m , C , N ) φ ^ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j ] d m ∙ ( − j ) } ] ∣ = o p ( 1 ) . (42) \left.\left.\left.
+\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid1,C,N)\frac{\pi^{(-r)}(0\mid m_{\bullet(-j)},C,N)}{1-\pi^{(-r)}(1\mid m_{\bullet(-j)},C,N)}
\left[\widehat{r}_{\bullet j}^{(3),(-r)}(1,1,m_{\bullet(-j)},C,N)-\iint\widehat{\eta}_{\bullet j}^{(-r)}(1,z,m,C,N)\widehat{\varphi}_{0}(z\mid1,C,N)\kappa_{0\bullet j}(m_{\bullet j}\mid1,C,N)\,dz\,dm_{\bullet j}\right]dm_{\bullet(-j)}\right\}\right]\right|=o_{p}(1). \tag*{(42)} + κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) 1 − π ( − r ) ( 1 ∣ m ∙ ( − j ) , C , N ) π ( − r ) ( 0 ∣ m ∙ ( − j ) , C , N ) [ r ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) − ∬ η ∙ j ( − r ) ( 1 , z , m , C , N ) φ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j ] d m ∙ ( − j ) } ] = o p ( 1 ) . (42) If, in addition, (35) holds and the left-hand side of (42) is o p ( K − 1 / 2 ) o_{p}(K^{-1/2}) o p ( K − 1/2 ) , then (37) holds and
K ( τ ^ l d − τ 0 ) ⇝ N ( 0 , Σ τ , l d ⋆ ) , P K ( D ^ τ , l d 2 ) → p Σ τ , l d ⋆ . (43) \sqrt{K}(\widehat{\tau}_{\mathrm{ld}}-\tau_{0})\rightsquigarrow N(0,\Sigma_{\tau,\mathrm{ld}}^{\star}),\qquad\mathbb{P}_{K}(\widehat{D}_{\tau,\mathrm{ld}}^{2})\to_{p}\Sigma_{\tau,\mathrm{ld}}^{\star}. \tag*{(43)} K ( τ ld − τ 0 ) ⇝ N ( 0 , Σ τ , ld ⋆ ) , P K ( D τ , ld 2 ) → p Σ τ , ld ⋆ . (43) For Σ θ , l d ⋆ > 0 \Sigma_{\theta,\mathrm{ld}}^{\star}>0 Σ θ , ld ⋆ > 0 and Σ τ , l d ⋆ > 0 \Sigma_{\tau,\mathrm{ld}}^{\star}>0 Σ τ , ld ⋆ > 0 , respectively,
K { θ ^ l d ( a , a ′ ) − θ 0 ( a , a ′ ) } { P K ( D ^ θ , l d 2 ) } 1 / 2 ⇝ N ( 0 , 1 ) , K ( τ ^ l d − τ 0 ) { P K ( D ^ τ , l d 2 ) } 1 / 2 ⇝ N ( 0 , 1 ) . (44) \frac{\sqrt{K}\{\widehat{\theta}_{\mathrm{ld}}(a,a')-\theta_{0}(a,a')\}}{\{\mathbb{P}_{K}(\widehat{D}_{\theta,\mathrm{ld}}^{2})\}^{1/2}}\rightsquigarrow N(0,1),\qquad
\frac{\sqrt{K}(\widehat{\tau}_{\mathrm{ld}}-\tau_{0})}{\{\mathbb{P}_{K}(\widehat{D}_{\tau,\mathrm{ld}}^{2})\}^{1/2}}\rightsquigarrow N(0,1). \tag*{(44)} { P K ( D θ , ld 2 ) } 1/2 K { θ ld ( a , a ′ ) − θ 0 ( a , a ′ )} ⇝ N ( 0 , 1 ) , { P K ( D τ , ld 2 ) } 1/2 K ( τ ld − τ 0 ) ⇝ N ( 0 , 1 ) . (44) At the true nuisance limits, the variances in (41) and (43) equal the corresponding bounds in Theorem 2.
The proof is given in Section G.7. The requirements in (32) and (34) concern all components of h θ ⋆ h_{\theta}^{\star} h θ ⋆ and h τ ⋆ h_{\tau}^{\star} h τ ⋆ , including their auxiliary regressions. Together with (28) or (29), they give both consistency and (40) or (42). The additional requirement for (36) or (37) is the o p ( K − 1 / 2 ) o_{p}(K^{-1/2}) o p ( K − 1/2 ) rate in (40) or (42), together with the existing (33) or (35).
The following construction gives some sufficient conditions for (40) and (42). Choose fitted conditional densities φ ^ ( − r ) \widehat{\varphi}^{(-r)} φ ( − r ) and κ ^ ∙ j ( − r ) \widehat{\kappa}_{\bullet j}^{(-r)} κ ∙ j ( − r ) with unit integrals, satisfying φ ^ ( − r ) ( z ∣ a , c , n ) = 0 \widehat{\varphi}^{(-r)}(z\mid a,c,n)=0 φ ( − r ) ( z ∣ a , c , n ) = 0 wherever φ 0 ( z ∣ a , c , n ) = 0 \varphi_{0}(z\mid a,c,n)=0 φ 0 ( z ∣ a , c , n ) = 0 and κ ^ ∙ j ( − r ) ( m ∙ j ∣ 1 , c , n ) = 0 \widehat{\kappa}_{\bullet j}^{(-r)}(m_{\bullet j}\mid1,c,n)=0 κ ∙ j ( − r ) ( m ∙ j ∣ 1 , c , n ) = 0 wherever κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) = 0 \kappa_{0\bullet j}(m_{\bullet j}\mid1,c,n)=0 κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) = 0 , for a ∈ { 0 , 1 } a\in\{0,1\} a ∈ { 0 , 1 } and almost every ( c , n ) (c,n) ( c , n ) . Take the ratios below to be one where both numerator and denominator are zero. For θ ^ l d ( a , a ′ ) \widehat{\theta}_{\mathrm{ld}}(a,a') θ ld ( a , a ′ ) , choose u ^ ∙ j ( 2 ) , ( − r ) \widehat{u}_{\bullet j}^{(2),(-r)} u ∙ j ( 2 ) , ( − r ) and u ^ ∙ j ( 3 ) , ( − r ) \widehat{u}_{\bullet j}^{(3),(-r)} u ∙ j ( 3 ) , ( − r ) to satisfy
u ^ ∙ j ( 2 ) , ( − r ) ( a , m , c , n ) = ∫ η ^ ∙ j ( − r ) ( a , z , m , c , n ) φ ^ ( − r ) ( z ∣ a , c , n ) d z , \widehat{u}_{\bullet j}^{(2),(-r)}(a,m,c,n)=\int\widehat{\eta}_{\bullet j}^{(-r)}(a,z,m,c,n)\widehat{\varphi}^{(-r)}(z\mid a,c,n)\,dz, u ∙ j ( 2 ) , ( − r ) ( a , m , c , n ) = ∫ η ∙ j ( − r ) ( a , z , m , c , n ) φ ( − r ) ( z ∣ a , c , n ) d z , u ^ ∙ j ( 3 ) , ( − r ) ( a , a ′ , c , n ) = ∫ u ^ ∙ j ( 1 ) , ( − r ) ( a , a ′ , z , c , n ) φ ^ ( − r ) ( z ∣ a , c , n ) d z . (45) \widehat{u}_{\bullet j}^{(3),(-r)}(a,a',c,n)=\int\widehat{u}_{\bullet j}^{(1),(-r)}(a,a',z,c,n)\widehat{\varphi}^{(-r)}(z\mid a,c,n)\,dz. \tag*{(45)} u ∙ j ( 3 ) , ( − r ) ( a , a ′ , c , n ) = ∫ u ∙ j ( 1 ) , ( − r ) ( a , a ′ , z , c , n ) φ ( − r ) ( z ∣ a , c , n ) d z . (45)
The regression u ^ ( 1 ) , ( − r ) \widehat{u}^{(1),(-r)} u ( 1 ) , ( − r ) may be fitted separately. Under (45), a sufficient condition for (40) is
max 1 ≤ r ≤ R max 1 ≤ j ≤ N max ∥ W { ϕ ^ ( − r ) ( Z ∣ a , C , N ) ϕ 0 ( Z ∣ a , C , N ) − 1 } ∥ 2 , P [ ∥ W { u ^ ∙ j ( 1 ) , ( − r ) − u 0 ∙ j ( 1 ) } ( a , a ′ , Z , C , N ) ∥ 2 , P + ∥ W { η ^ ∙ j ( − r ) − η 0 ∙ j } ( a , Z , M , C , N ) ∥ 2 , P + ∥ W { s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) − s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) } ∥ 2 , P ] = o P ( 1 ) . (46) \max_{1\le r\le R}\max_{1\le j\le N_{\max}}\left\|\sqrt{W}\left\{\frac{\widehat{\phi}^{(-r)}(Z\mid a,C,N)}{\phi_{0}(Z\mid a,C,N)}-1\right\}\right\|_{2,P}
\left[
\left\|\sqrt{W}\{\widehat{u}_{\bullet j}^{(1),(-r)}-u_{0\bullet j}^{(1)}\}(a,a',Z,C,N)\right\|_{2,P}
+\left\|\sqrt{W}\{\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{0\bullet j}\}(a,Z,M,C,N)\right\|_{2,P}
+\left\|\sqrt{W}\left\{\frac{\widehat{s}^{(-r)}(a'\mid M,C,N)}{\widehat{s}^{(-r)}(a\mid M,C,N)}-\frac{s_{0}(a'\mid M,C,N)}{s_{0}(a\mid M,C,N)}\right\}\right\|_{2,P}
\right]=o_{P}(1). \tag*{(46)} 1 ≤ r ≤ R max 1 ≤ j ≤ N m a x max W { ϕ 0 ( Z ∣ a , C , N ) ϕ ( − r ) ( Z ∣ a , C , N ) − 1 } 2 , P [ W { u ∙ j ( 1 ) , ( − r ) − u 0 ∙ j ( 1 ) } ( a , a ′ , Z , C , N ) 2 , P + W { η ∙ j ( − r ) − η 0 ∙ j } ( a , Z , M , C , N ) 2 , P + W { s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) − s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) } 2 , P ] = o P ( 1 ) . (46)
Replacing o P ( 1 ) o_{P}(1) o P ( 1 ) in (46) by o P ( K − 1 / 2 ) o_{P}(K^{-1/2}) o P ( K − 1/2 ) gives the o P ( K − 1 / 2 ) o_{P}(K^{-1/2}) o P ( K − 1/2 ) rate in (40). For example, under the boundedness requirement in Section G.5,
max 1 ≤ r ≤ R ∥ W { ϕ ^ ( − r ) ( Z ∣ a , C , N ) ϕ 0 ( Z ∣ a , C , N ) − 1 } ∥ 2 , P = o P ( 1 ) (47) \max_{1\le r\le R}\left\|\sqrt{W}\left\{\frac{\widehat{\phi}^{(-r)}(Z\mid a,C,N)}{\phi_{0}(Z\mid a,C,N)}-1\right\}\right\|_{2,P}=o_{P}(1) \tag*{(47)} 1 ≤ r ≤ R max W { ϕ 0 ( Z ∣ a , C , N ) ϕ ( − r ) ( Z ∣ a , C , N ) − 1 } 2 , P = o P ( 1 ) (47)
implies (46). If (47) instead has rate O P ( K − 1 / 2 ) O_{P}(K^{-1/2}) O P ( K − 1/2 ) , and (32) holds with ( η ⋆ , s ⋆ ) = ( η 0 , s 0 ) (\eta^{\star},s^{\star})=(\eta_{0},s_{0}) ( η ⋆ , s ⋆ ) = ( η 0 , s 0 ) , then (46) has rate o P ( K − 1 / 2 ) o_{P}(K^{-1/2}) o P ( K − 1/2 ) . Alternatively, o P ( K − 1 / 4 ) o_{P}(K^{-1/4}) o P ( K − 1/4 ) for every norm in (46) gives that rate.
For τ ~ l d \widetilde{\tau}_{\mathrm{ld}} τ ld , choose r ^ ∙ j ( 3 ) , ( − r ) \widehat{r}_{\bullet j}^{(3),(-r)} r ∙ j ( 3 ) , ( − r ) and r ^ ∙ j ( 4 ) , ( − r ) \widehat{r}_{\bullet j}^{(4),(-r)} r ∙ j ( 4 ) , ( − r ) to satisfy
r ^ ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , c , n ) = ∬ η ^ ( − r ) ( 1 , z , m , c , n ) ϕ ^ ( − r ) ( z ∣ 1 , c , n ) κ ^ ∙ j ( − r ) ( m ∙ j ∣ 1 , c , n ) d z d m ∙ j , \widehat{r}_{\bullet j}^{(3),(-r)}(1,1,m_{\bullet(-j)},c,n)=\iint\widehat{\eta}^{(-r)}(1,z,m,c,n)\widehat{\phi}^{(-r)}(z\mid1,c,n)\widehat{\kappa}_{\bullet j}^{(-r)}(m_{\bullet j}\mid1,c,n)\,dz\,dm_{\bullet j}, r ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , c , n ) = ∬ η ( − r ) ( 1 , z , m , c , n ) ϕ ( − r ) ( z ∣ 1 , c , n ) κ ∙ j ( − r ) ( m ∙ j ∣ 1 , c , n ) d z d m ∙ j ,
2 r ^ ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , c , n ) = ∫ r ^ ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , z , c , n ) ϕ ^ ( − r ) ( z ∣ 1 , c , n ) d z + ∫ r ^ ∙ j ( 2 ) , ( − r ) ( 1 , 0 , m ∙ j , c , n ) κ ^ ∙ j ( − r ) ( m ∙ j ∣ 1 , c , n ) d m ∙ j . (48) 2\widehat{r}_{\bullet j}^{(4),(-r)}(1,1,0,c,n)=\int\widehat{r}_{\bullet j}^{(1),(-r)}(1,1,0,z,c,n)\widehat{\phi}^{(-r)}(z\mid1,c,n)\,dz
+\int\widehat{r}_{\bullet j}^{(2),(-r)}(1,0,m_{\bullet j},c,n)\widehat{\kappa}_{\bullet j}^{(-r)}(m_{\bullet j}\mid1,c,n)\,dm_{\bullet j}. \tag*{(48)} 2 r ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , c , n ) = ∫ r ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , z , c , n ) ϕ ( − r ) ( z ∣ 1 , c , n ) d z + ∫ r ∙ j ( 2 ) , ( − r ) ( 1 , 0 , m ∙ j , c , n ) κ ∙ j ( − r ) ( m ∙ j ∣ 1 , c , n ) d m ∙ j . (48)
The regressions r ^ ( 1 ) , ( − r ) \widehat{r}^{(1),(-r)} r ( 1 ) , ( − r ) and r ^ ( 2 ) , ( − r ) \widehat{r}^{(2),(-r)} r ( 2 ) , ( − r ) may be fitted separately. Under (48), a sufficient condition for (42) is
max 1 ≤ r ≤ R max 1 ≤ j ≤ N max [ ∥ W { ϕ ^ ( − r ) ( Z ∣ 1 , C , N ) ϕ 0 ( Z ∣ 1 , C , N ) − 1 } ∥ 2 , P × [ ∥ W { r ^ ∙ j ( 1 ) , ( − r ) − r 0 ∙ j ( 1 ) } ( 1 , 1 , 0 , Z , C , N ) ∥ 2 , P + ∥ W { η ^ ∙ j ( − r ) − η 0 ∙ j } ( 1 , Z , M , C , N ) ∥ 2 , P ] + ∥ W { s ^ ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) } ∥ 2 , P ] = o P ( 1 ) . (49) \begin{aligned}
&\max_{1\le r\le R}\max_{1\le j\le N_{\max}}
\Bigg[
\left\|\sqrt{W}\left\{\frac{\widehat{\phi}^{(-r)}(Z\mid1,C,N)}{\phi_{0}(Z\mid1,C,N)}-1\right\}\right\|_{2,P}\\
&\quad\times\Bigg[\left\|\sqrt{W}\{\widehat{r}_{\bullet j}^{(1),(-r)}-r_{0\bullet j}^{(1)}\}(1,1,0,Z,C,N)\right\|_{2,P}
+\left\|\sqrt{W}\{\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{0\bullet j}\}(1,Z,M,C,N)\right\|_{2,P}\Bigg]\\
&\quad+\left\|\sqrt{W}\left\{\frac{\widehat{s}^{(-r)}(0\mid M_{\bullet(-j)},C,N)}{\widehat{s}^{(-r)}(1\mid M_{\bullet(-j)},C,N)}-\frac{s_{0,-j}(0\mid M_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid M_{\bullet(-j)},C,N)}\right\}\right\|_{2,P}
\Bigg]=o_{P}(1). \tag*{(49)}
\end{aligned} 1 ≤ r ≤ R max 1 ≤ j ≤ N m a x max [ W { ϕ 0 ( Z ∣ 1 , C , N ) ϕ ( − r ) ( Z ∣ 1 , C , N ) − 1 } 2 , P × [ W { r ∙ j ( 1 ) , ( − r ) − r 0 ∙ j ( 1 ) } ( 1 , 1 , 0 , Z , C , N ) 2 , P + W { η ∙ j ( − r ) − η 0 ∙ j } ( 1 , Z , M , C , N ) 2 , P ] + W { s ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) } 2 , P ] = o P ( 1 ) . (49)
Replacing o P ( 1 ) o_{P}(1) o P ( 1 ) in (49) by o P ( K − 1 / 2 ) o_{P}(K^{-1/2}) o P ( K − 1/2 ) gives the o P ( K − 1 / 2 ) o_{P}(K^{-1/2}) o P ( K − 1/2 ) rate in (42). In particular, (47) with a = 1 a=1 a = 1 and
max 1 ≤ r ≤ R max 1 ≤ j ≤ N max ∥ W { κ ^ ∙ j ( − r ) ( M ∙ j ∣ 1 , C , N ) κ 0 ∙ j ( M ∙ j ∣ 1 , C , N ) − 1 } ∥ 2 , P = o P ( 1 ) (50) \max_{1\le r\le R}\max_{1\le j\le N_{\max}}
\left\|\sqrt{W}\left\{\frac{\widehat{\kappa}_{\bullet j}^{(-r)}(M_{\bullet j}\mid1,C,N)}{\kappa_{0\bullet j}(M_{\bullet j}\mid1,C,N)}-1\right\}\right\|_{2,P}=o_{P}(1) \tag*{(50)} 1 ≤ r ≤ R max 1 ≤ j ≤ N m a x max W { κ 0 ∙ j ( M ∙ j ∣ 1 , C , N ) κ ∙ j ( − r ) ( M ∙ j ∣ 1 , C , N ) − 1 } 2 , P = o P ( 1 ) (50)
imply (49) under the same boundedness requirement. If (47) with a = 1 a=1 a = 1 and (50) instead have rates O p ( K − 1 / 2 ) O_p(K^{-1/2}) O p ( K − 1/2 ) , and (34) holds with ( η ⋆ , s ⋆ ) j = ( η 0 , s 0 , − j ) (\eta^\star, s^\star)_j = (\eta_0, s_{0,-j}) ( η ⋆ , s ⋆ ) j = ( η 0 , s 0 , − j ) for every j j j , then (49) has rate o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) . Alternatively, o p ( K − 1 / 4 ) o_p(K^{-1/4}) o p ( K − 1/4 ) for every norm in (49) gives that rate. Section G.7 derives (46) and (49) from exact integral identities. Combining their o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) versions with (33) or (35) and the respective component convergence requirements gives (36) and (41), or (37) and (43).
The augmented-classification estimator of d d d in Section G.1 follows Rudolph et al. [15 ] . The condition in Rudolph et al. [15 ] controls agreement between fitted integral representations at the rates required by Rudolph et al. [15 ] . Here, (45) and (48) give explicit relationships among the estimated auxiliary regressions, and (46) and (49) give sufficient nuisance rates for (40) and (42).
Proof of Theorem 1
Proof. Fix a , a ′ ∈ { 0 , 1 } a, a' \in\{0,1\} a , a ′ ∈ { 0 , 1 } . Conditional on ( C , N ) (C,N) ( C , N ) , G ( a ′ ) G(a') G ( a ′ ) is an independent random draw from the conditional distribution of M ( a ′ ) M(a') M ( a ′ ) . Therefore, by iterated expectation, the numerator of (1) is
E [ W N ∑ j = 1 N Y ∙ j { a , G ( a ′ ) } ] = E [ W N ∑ j = 1 N ∫ E { Y ∙ j ( a , m ) ∣ C , N } κ ( m ∣ a ′ , C , N ) d m ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}Y_{\bullet j}\{a,G(a')\}\right]
=
\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\int\mathbb{E}\{Y_{\bullet j}(a,m)\mid C,N\}\kappa(m\mid a',C,N)\,dm\right]. E [ N W j = 1 ∑ N Y ∙ j { a , G ( a ′ )} ] = E [ N W j = 1 ∑ N ∫ E { Y ∙ j ( a , m ) ∣ C , N } κ ( m ∣ a ′ , C , N ) d m ] .
where randomization and consistency identify the conditional distribution of M ( a ′ ) M(a') M ( a ′ ) given ( C , N ) (C,N) ( C , N ) with κ ( ∙ ∣ a ′ , C , N ) \kappa(\mathord{\bullet}\mid a',C,N) κ ( ∙ ∣ a ′ , C , N ) . By the composition, Y ∙ j ( a , m ) = Y ∙ j { a , Z ( a ) , m } Y_{\bullet j}(a,m)=Y_{\bullet j}\{a,Z(a),m\} Y ∙ j ( a , m ) = Y ∙ j { a , Z ( a ) , m } , and hence
E { Y ∙ j ( a , m ) ∣ C , N } = ∫ E { Y ∙ j ( a , z , m ) ∣ Z ( a ) = z , C , N } ϕ ( z ∣ a , C , N ) d z . \mathbb{E}\{Y_{\bullet j}(a,m)\mid C,N\}
=
\int\mathbb{E}\{Y_{\bullet j}(a,z,m)\mid Z(a)=z,C,N\}\phi(z\mid a,C,N)\,dz. E { Y ∙ j ( a , m ) ∣ C , N } = ∫ E { Y ∙ j ( a , z , m ) ∣ Z ( a ) = z , C , N } ϕ ( z ∣ a , C , N ) d z .
For each ( z , m ) (z,m) ( z , m ) in the relevant support, randomization gives
E { Y ∙ j ( a , z , m ) ∣ Z ( a ) = z , C , N } = E { Y ∙ j ( a , z , m ) ∣ A = a , Z ( a ) = z , C , N } . \mathbb{E}\{Y_{\bullet j}(a,z,m)\mid Z(a)=z,C,N\}
=
\mathbb{E}\{Y_{\bullet j}(a,z,m)\mid A=a,Z(a)=z,C,N\}. E { Y ∙ j ( a , z , m ) ∣ Z ( a ) = z , C , N } = E { Y ∙ j ( a , z , m ) ∣ A = a , Z ( a ) = z , C , N } .
Assumption (A3) then permits conditioning on M ( a , z ) = m M(a,z)=m M ( a , z ) = m , and Assumption (A1) maps the resulting counterfactual variables to their observed counterparts:
E { Y ∙ j ( a , z , m ) ∣ A = a , Z ( a ) = z , C , N } = E { Y ∙ j ( a , z , m ) ∣ A = a , Z ( a ) = z , M ( a , z ) = m , C , N } = E ( Y ∙ j ∣ A = a , Z = z , M = m , C , N ) = η ∙ j ( a , z , m , C , N ) . \begin{aligned}
\mathbb{E}\{Y_{\bullet j}(a,z,m)\mid A=a,Z(a)=z,C,N\}
&=\mathbb{E}\{Y_{\bullet j}(a,z,m)\mid A=a,Z(a)=z,M(a,z)=m,C,N\}\\
&=\mathbb{E}(Y_{\bullet j}\mid A=a,Z=z,M=m,C,N)=\eta_{\bullet j}(a,z,m,C,N).
\end{aligned} E { Y ∙ j ( a , z , m ) ∣ A = a , Z ( a ) = z , C , N } = E { Y ∙ j ( a , z , m ) ∣ A = a , Z ( a ) = z , M ( a , z ) = m , C , N } = E ( Y ∙ j ∣ A = a , Z = z , M = m , C , N ) = η ∙ j ( a , z , m , C , N ) .
Substitution and division by E ( W ) \mathbb{E}(W) E ( W ) prove (3).
For (2), G ( 1 ) G(1) G ( 1 ) and G ( 0 ) G(0) G ( 0 ) are conditionally independent draws. Consequently, G ∙ j ( 1 ) G_{\bullet j}(1) G ∙ j ( 1 ) and G ∙ ( − j ) ( 0 ) G_{\bullet(-j)}(0) G ∙ ( − j ) ( 0 ) have the product distribution with densities κ ∙ j ( ∙ ∣ 1 , C , N ) \kappa_{\bullet j}(\mathord{\bullet}\mid1,C,N) κ ∙ j ( ∙ ∣ 1 , C , N ) and κ ∙ ( − j ) ( ∙ ∣ 0 , C , N ) \kappa_{\bullet(-j)}(\mathord{\bullet}\mid0,C,N) κ ∙ ( − j ) ( ∙ ∣ 0 , C , N ) , and
E [ W N ∑ j = 1 N Y ∙ j { 1 , G ∙ j ( 1 ) , G ∙ ( − j ) ( 0 ) } ] = E [ W N ∑ j = 1 N ∫ E { Y ∙ j ( 1 , m ) ∣ C , N } κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) d m ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}Y_{\bullet j}\{1,G_{\bullet j}(1),G_{\bullet(-j)}(0)\}\right]
=
\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\int\mathbb{E}\{Y_{\bullet j}(1,m)\mid C,N\}\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\,dm\right]. E [ N W j = 1 ∑ N Y ∙ j { 1 , G ∙ j ( 1 ) , G ∙ ( − j ) ( 0 )} ] = E [ N W j = 1 ∑ N ∫ E { Y ∙ j ( 1 , m ) ∣ C , N } κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) d m ] .
In addition, we have
E { Y ∙ j ( 1 , m ) ∣ C , N } = ∫ η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) d z . \mathbb{E}\{Y_{\bullet j}(1,m)\mid C,N\}
=
\int\eta_{\bullet j}(1,z,m,C,N)\phi(z\mid1,C,N)\,dz. E { Y ∙ j ( 1 , m ) ∣ C , N } = ∫ η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) d z .
Substitution and division by E ( W ) \mathbb{E}(W) E ( W ) prove (4) .
Proof of Theorem 2
Proof. Because the treatment mechanism is known by design, an arbitrary regular one-dimensional parametric submodel through the observed-data distribution can be written as
f t ( O ) = f t ( Y ∣ A , Z , M , C , N ) f t ( M ∣ A , Z , C , N ) f t ( Z ∣ A , C , N ) f ( A ) f t ( C , N ) , f_t(O) = f_t(Y \mid A, Z, M, C, N)f_t(M \mid A, Z, C, N)f_t(Z \mid A, C, N)f(A)f_t(C, N), f t ( O ) = f t ( Y ∣ A , Z , M , C , N ) f t ( M ∣ A , Z , C , N ) f t ( Z ∣ A , C , N ) f ( A ) f t ( C , N ) ,
with score S ( O ) S(O) S ( O ) at t = 0 t = 0 t = 0 . We first differentiate the numerator of (3) . Its value along the submodel is
∫ w n ∑ j = 1 n ∬ η ∙ j , t ( a , z , m , c , n ) ϕ t ( z ∣ a , c , n ) κ t ( m ∣ a ′ , c , n ) d z d m f t ( c , n ) d c d n , \int\frac{w}{n} \sum_{j=1}^{n} \iint\eta_{\bullet j,t}(a, z, m, c, n)\phi_t(z \mid a, c, n)\kappa_t(m \mid a', c, n)\,dz\,dm\,f_t(c,n)\,dc\,dn, ∫ n w j = 1 ∑ n ∬ η ∙ j , t ( a , z , m , c , n ) ϕ t ( z ∣ a , c , n ) κ t ( m ∣ a ′ , c , n ) d z d m f t ( c , n ) d c d n ,
where
κ t ( m ∣ a ′ , c , n ) = ∫ ξ t ( m ∣ a ′ , z , c , n ) ϕ t ( z ∣ a ′ , c , n ) d z \kappa_t(m \mid a', c, n) = \int\xi_t(m \mid a', z, c, n)\phi_t(z \mid a', c, n)\,dz κ t ( m ∣ a ′ , c , n ) = ∫ ξ t ( m ∣ a ′ , z , c , n ) ϕ t ( z ∣ a ′ , c , n ) d z
is the marginal conditional distribution of M M M in arm a ′ a' a ′ .
For the outcome regression, the standard conditional-mean derivative calculation yields
∂ ∂ t η ∙ j , t ( a , z , m , c , n ) ∣ t = 0 = E [ { Y ∙ j − η ∙ j ( a , z , m , c , n ) } S ( O ) ∣ A = a , Z = z , M = m , C = c , N = n ] . \left.\frac{\partial}{\partial t}\eta_{\bullet j,t}(a,z,m,c,n)\right|_{t=0}
= \mathbb{E}\left[\{Y_{\bullet j}-\eta_{\bullet j}(a,z,m,c,n)\}S(O)\mid A=a,Z=z,M=m,C=c,N=n\right]. ∂ t ∂ η ∙ j , t ( a , z , m , c , n ) t = 0 = E [ { Y ∙ j − η ∙ j ( a , z , m , c , n )} S ( O ) ∣ A = a , Z = z , M = m , C = c , N = n ] .
Changing measure from ϕ ( z ∣ a , c , n ) κ ( m ∣ a ′ , c , n ) \phi(z \mid a,c,n)\kappa(m \mid a',c,n) ϕ ( z ∣ a , c , n ) κ ( m ∣ a ′ , c , n ) to the observed conditional distribution ϕ ( z ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) \phi(z \mid a,c,n)\xi(m \mid a,z,c,n) ϕ ( z ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) therefore yields
E [ W N ∑ j = 1 N I ( A = a ) π a ( 1 − π ) 1 − a ω ∙ j ( 1 ) ( a , a ′ , Z , M , C , N ) { Y ∙ j − η ∙ j ( a , Z , M , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{I(A=a)}{\pi^a(1-\pi)^{1-a}}\omega_{\bullet j}^{(1)}(a,a',Z,M,C,N)\{Y_{\bullet j}-\eta_{\bullet j}(a,Z,M,C,N)\}S(O)\right]. E [ N W j = 1 ∑ N π a ( 1 − π ) 1 − a I ( A = a ) ω ∙ j ( 1 ) ( a , a ′ , Z , M , C , N ) { Y ∙ j − η ∙ j ( a , Z , M , C , N )} S ( O ) ] .
The conditional score of Z Z Z in arm a a a is represented by E { S ( O ) ∣ A = a , Z , C , N } − E { S ( O ) ∣ A = a , C , N } \mathbb{E}\{S(O)\mid A=a,Z,C,N\}-\mathbb{E}\{S(O)\mid A=a,C,N\} E { S ( O ) ∣ A = a , Z , C , N } − E { S ( O ) ∣ A = a , C , N } . Differentiating the explicit factor ϕ t ( z ∣ a , c , n ) \phi_t(z\mid a,c,n) ϕ t ( z ∣ a , c , n ) thus contributes
E [ W N ∑ j = 1 N I ( A = a ) π a ( 1 − π ) 1 − a { ω ∙ j ( 1 ) ( a , a ′ , Z , C , N ) − ω ∙ j ( 3 ) ( a , a ′ , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{I(A=a)}{\pi^a(1-\pi)^{1-a}}\{\omega_{\bullet j}^{(1)}(a,a',Z,C,N)-\omega_{\bullet j}^{(3)}(a,a',C,N)\}S(O)\right]. E [ N W j = 1 ∑ N π a ( 1 − π ) 1 − a I ( A = a ) { ω ∙ j ( 1 ) ( a , a ′ , Z , C , N ) − ω ∙ j ( 3 ) ( a , a ′ , C , N )} S ( O ) ] .
Likewise, because κ t ( ⋅ ∣ a ′ , c , n ) \kappa_t(\cdot\mid a',c,n) κ t ( ⋅ ∣ a ′ , c , n ) is the marginal distribution of M M M in arm a ′ a' a ′ , its conditional score is represented by E { S ( O ) ∣ A = a ′ , M , C , N } − E { S ( O ) ∣ A = a ′ , C , N } \mathbb{E}\{S(O)\mid A=a',M,C,N\}-\mathbb{E}\{S(O)\mid A=a',C,N\} E { S ( O ) ∣ A = a ′ , M , C , N } − E { S ( O ) ∣ A = a ′ , C , N } . Its derivative therefore contributes
E [ W N ∑ j = 1 N I ( A = a ′ ) π a ′ ( 1 − π ) 1 − a ′ { ω ∙ j ( 2 ) ( a , M , C , N ) − ω ∙ j ( 3 ) ( a , a ′ , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{I(A=a')}{\pi^{a'}(1-\pi)^{1-a'}}\{\omega_{\bullet j}^{(2)}(a,M,C,N)-\omega_{\bullet j}^{(3)}(a,a',C,N)\}S(O)\right]. E [ N W j = 1 ∑ N π a ′ ( 1 − π ) 1 − a ′ I ( A = a ′ ) { ω ∙ j ( 2 ) ( a , M , C , N ) − ω ∙ j ( 3 ) ( a , a ′ , C , N )} S ( O ) ] .
This representation remains valid when a = a ′ a=a' a = a ′ ; in that case the two factor-specific derivatives are both present and add in the same treatment arm. Finally, differentiation of the baseline distribution f t ( c , n ) f_t(c,n) f t ( c , n ) gives
E [ { W N ∑ j = 1 N ω ⋅ j ( 3 ) ( a , a ′ , C , N ) − E ( W ) θ ( a , a ′ ) } S ( O ) ] . \mathbb{E}\left[\left\{\frac{W}{N}\sum_{j=1}^{N}\omega_{\cdot j}^{(3)}(a,a',C,N)-\mathbb{E}(W)\theta(a,a')\right\}S(O)\right]. E [ { N W j = 1 ∑ N ω ⋅ j ( 3 ) ( a , a ′ , C , N ) − E ( W ) θ ( a , a ′ ) } S ( O ) ] .
Combining the four derivatives shows that the influence function of the numerator is ψ θ ( a , a ′ ; O ) − E ( W ) θ ( a , a ′ ) \psi_\theta(a,a';O)-\mathbb{E}(W)\theta(a,a') ψ θ ( a , a ′ ; O ) − E ( W ) θ ( a , a ′ ) .
For the numerator of (4), the same argument applies. Differentiation of the outcome regression and change of measure to the product intervention give
E [ W N ∑ j = 1 N A π ω ⋅ j ( 2 ) ( 1 , 1 , 0 , Z , M , C , N ) { Y ⋅ j − η ⋅ j ( 1 , Z , M , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{A}{\pi}\omega_{\cdot j}^{(2)}(1,1,0,Z,M,C,N)\{Y_{\cdot j}-\eta_{\cdot j}(1,Z,M,C,N)\}S(O)\right]. E [ N W j = 1 ∑ N π A ω ⋅ j ( 2 ) ( 1 , 1 , 0 , Z , M , C , N ) { Y ⋅ j − η ⋅ j ( 1 , Z , M , C , N )} S ( O ) ] .
Differentiation of the explicit treated-arm Z Z Z distribution contributes
E [ W N ∑ j = 1 N A π { r ⋅ j ( 1 ) ( 1 , 1 , 0 , Z , C , N ) − r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{A}{\pi}\left\{r_{\cdot j}^{(1)}(1,1,0,Z,C,N)-r_{\cdot j}^{(4)}(1,1,0,C,N)\right\}S(O)\right]. E [ N W j = 1 ∑ N π A { r ⋅ j ( 1 ) ( 1 , 1 , 0 , Z , C , N ) − r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) } S ( O ) ] .
The score of the treated-arm marginal distribution of M ⋅ j M_{\cdot j} M ⋅ j is represented by conditioning S ( O ) S(O) S ( O ) on ( A = 1 , M ⋅ j , C , N ) (A=1,M_{\cdot j},C,N) ( A = 1 , M ⋅ j , C , N ) and then centering given ( A = 1 , C , N ) (A=1,C,N) ( A = 1 , C , N ) ; its derivative contributes
E [ W N ∑ j = 1 N A π { r ⋅ j ( 2 ) ( 1 , 0 , M ⋅ j , C , N ) − r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{A}{\pi}\left\{r_{\cdot j}^{(2)}(1,0,M_{\cdot j},C,N)-r_{\cdot j}^{(4)}(1,1,0,C,N)\right\}S(O)\right]. E [ N W j = 1 ∑ N π A { r ⋅ j ( 2 ) ( 1 , 0 , M ⋅ j , C , N ) − r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) } S ( O ) ] .
Similarly, differentiating the control-arm marginal distribution of M ⋅ ( − j ) M_{\cdot(-j)} M ⋅ ( − j ) contributes
E [ W N ∑ j = 1 N 1 − A 1 − π { r ⋅ j ( 3 ) ( 1 , 1 , M ⋅ ( − j ) , C , N ) − r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) } S ( O ) ] . \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\frac{1-A}{1-\pi}\left\{r_{\cdot j}^{(3)}(1,1,M_{\cdot(-j)},C,N)-r_{\cdot j}^{(4)}(1,1,0,C,N)\right\}S(O)\right]. E [ N W j = 1 ∑ N 1 − π 1 − A { r ⋅ j ( 3 ) ( 1 , 1 , M ⋅ ( − j ) , C , N ) − r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) } S ( O ) ] .
Differentiation of f t ( c , n ) f_t(c,n) f t ( c , n ) gives
E [ { W N ∑ j = 1 N r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) − E ( W ) τ } S ( O ) ] . \mathbb{E}\left[\left\{\frac{W}{N}\sum_{j=1}^{N}r_{\cdot j}^{(4)}(1,1,0,C,N)-\mathbb{E}(W)\tau\right\}S(O)\right]. E [ { N W j = 1 ∑ N r ⋅ j ( 4 ) ( 1 , 1 , 0 , C , N ) − E ( W ) τ } S ( O ) ] .
Thus the influence function of this numerator is ψ τ ( O ) − E ( W ) τ \psi_\tau(O)-\mathbb{E}(W)\tau ψ τ ( O ) − E ( W ) τ .
The influence function of E ( W ) \mathbb{E}(W) E ( W ) is W − E ( W ) W-\mathbb{E}(W) W − E ( W ) . We get
ψ θ − E ( W ) θ − θ { W − E ( W ) } E ( W ) = ψ θ − θ W E ( W ) . \frac{\psi_\theta-\mathbb{E}(W)\theta-\theta\{W-\mathbb{E}(W)\}}{\mathbb{E}(W)}
=
\frac{\psi_\theta-\theta W}{\mathbb{E}(W)}. E ( W ) ψ θ − E ( W ) θ − θ { W − E ( W )} = E ( W ) ψ θ − θ W .
and the analogous expression for τ \tau τ .
Proofs of Theorems 8 and 9
Bias decompositions
We use the subscript 0 to denote true parameter and nuisance values. For a candidate tuple ( η , ϕ , κ , ξ ) (\eta,\phi,\kappa,\xi) ( η , ϕ , κ , ξ ) , the scores use the density ratios and integrated outcome regressions obtained from that tuple as defined in Section 3. The following theorem gives the exact bias decomposition for θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) .
Theorem 8. Fix arbitrary candidate nuisance functions ( η , ϕ , κ , ξ ) (\eta,\phi,\kappa,\xi) ( η , ϕ , κ , ξ ) satisfying the support conditions in Assumption ( A 4 ) (A4) ( A 4 ) , then
E { ψ θ ( a , a ′ ; O ; η , ϕ , κ , ξ ) } − E ( W ) θ 0 ( a , a ′ ) = E [ W N ∑ j = 1 N { ∬ { κ − κ 0 } ( m ∣ a ′ , C , N ) [ η 0 ∙ j ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) − η ∙ j ( a , z , m , C , N ) ϕ ( z ∣ a , C , N ) ] d z d m + ∬ { η 0 ∙ j − η ∙ j } ( a , z , m , C , N ) κ ( m ∣ a ′ , C , N ) ϕ 0 ( z ∣ a , C , N ) { ξ 0 ( m ∣ a , z , C , N ) ξ ( m ∣ a , z , C , N ) − 1 } d z d m } ] . (51) \begin{aligned}
\mathbb{E}\{\psi_{\theta}(a,a';O;\eta,\phi,\kappa,\xi)\}-\mathbb{E}(W)\theta_{0}(a,a') = \mathbb{E}\bigg[\frac{W}{N}\sum_{j=1}^{N}\bigg\{&\iint\{\kappa-\kappa_{0}\}(m\mid a',C,N) [\eta_{0\bullet j}(a,z,m,C,N)\phi_{0}(z\mid a,C,N)\\
&-\eta_{\bullet j}(a,z,m,C,N)\phi(z\mid a,C,N)]\,dz\,dm\\
&+\iint\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,z,m,C,N)\kappa(m\mid a',C,N)\phi_{0}(z\mid a,C,N)\\
&\qquad\left\{\frac{\xi_{0}(m\mid a,z,C,N)}{\xi(m\mid a,z,C,N)}-1\right\}\,dz\,dm\bigg\}\bigg]. \tag*{(51)}
\end{aligned} E { ψ θ ( a , a ′ ; O ; η , ϕ , κ , ξ )} − E ( W ) θ 0 ( a , a ′ ) = E [ N W j = 1 ∑ N { ∬ { κ − κ 0 } ( m ∣ a ′ , C , N ) [ η 0 ∙ j ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) − η ∙ j ( a , z , m , C , N ) ϕ ( z ∣ a , C , N )] d z d m + ∬ { η 0 ∙ j − η ∙ j } ( a , z , m , C , N ) κ ( m ∣ a ′ , C , N ) ϕ 0 ( z ∣ a , C , N ) { ξ ( m ∣ a , z , C , N ) ξ 0 ( m ∣ a , z , C , N ) − 1 } d z d m } ] . (51) Consequently, if at least one of the following conditions holds for every j j j : (i) ( η ∙ j , ϕ ) = ( η 0 ∙ j , ϕ 0 ) (\eta_{\bullet j},\phi)=(\eta_{0\bullet j},\phi_{0}) ( η ∙ j , ϕ ) = ( η 0 ∙ j , ϕ 0 ) ; (ii) ( η ∙ j , κ ) = ( η 0 ∙ j , κ 0 ) (\eta_{\bullet j},\kappa)=(\eta_{0\bullet j},\kappa_{0}) ( η ∙ j , κ ) = ( η 0 ∙ j , κ 0 ) for every j j j ; (iii) ( κ , ξ ) = ( κ 0 , ξ 0 ) (\kappa,\xi)=(\kappa_{0},\xi_{0}) ( κ , ξ ) = ( κ 0 , ξ 0 ) . then
E { ψ θ ( a , a ′ ; O ; η , ϕ , κ , ξ ) } = E ( W ) θ 0 ( a , a ′ ) . \mathbb{E}\{\psi_{\theta}(a,a';O;\eta,\phi,\kappa,\xi)\}=\mathbb{E}(W)\theta_{0}(a,a'). E { ψ θ ( a , a ′ ; O ; η , ϕ , κ , ξ )} = E ( W ) θ 0 ( a , a ′ ) . For τ \tau τ , we establish the following bias decomposition.
Theorem 9. Fix arbitrary candidate nuisance functions ( η , ϕ , κ , ξ ) (\eta,\phi,\kappa,\xi) ( η , ϕ , κ , ξ ) satisfying the support conditions in Assumption ( A 4 ) (A4) ( A 4 ) . Then
E { ψ τ ( O ; η , ϕ , κ , ξ ) } − E ( W ) τ 0 = E [ W N ∑ j = 1 N ∬ { { κ ∙ j − κ 0 ∙ j } ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) } + { κ ∙ ( − j ) − κ 0 ∙ ( − j ) } ( m ∙ ( − j ) ∣ 0 , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) } − η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) { κ ∙ j − κ 0 ∙ j } ( m ∙ j ∣ 1 , C , N ) { κ ∙ ( − j ) − κ 0 ∙ ( − j ) } ( m ∙ ( − j ) ∣ 0 , C , N ) + { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) ϕ 0 ( z ∣ 1 , C , N ) { ξ 0 ( m ∣ 1 , z , C , N ) ξ ( m ∣ 1 , z , C , N ) − 1 } } d m d z ] . (52) \begin{aligned}
\mathbb{E}\{\psi_{\tau}(O;\eta,\phi,\kappa,\xi)\}-\mathbb{E}(W)\tau_{0}
=\mathbb{E}\bigg[\frac{W}{N}\sum_{j=1}^{N}\iint\bigg\{&
\{\kappa_{\bullet j}-\kappa_{0\bullet j}\}(m_{\bullet j}\mid1,C,N)\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\\
&\quad\{\eta_{0\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)-\eta_{\bullet j}(1,z,m,C,N)\phi(z\mid1,C,N)\}\\
&+\{\kappa_{\bullet(-j)}-\kappa_{0\bullet(-j)}\}(m_{\bullet(-j)}\mid0,C,N)\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\\
&\quad\{\eta_{0\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)-\eta_{\bullet j}(1,z,m,C,N)\phi(z\mid1,C,N)\}\\
&-\eta_{0\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)
\{\kappa_{\bullet j}-\kappa_{0\bullet j}\}(m_{\bullet j}\mid1,C,N)\\
&\quad\{\kappa_{\bullet(-j)}-\kappa_{0\bullet(-j)}\}(m_{\bullet(-j)}\mid0,C,N)\\
&+\{\eta_{0\bullet j}-\eta_{\bullet j}\}(1,z,m,C,N)\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\\
&\quad\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\phi_{0}(z\mid1,C,N)
\left\{\frac{\xi_{0}(m\mid1,z,C,N)}{\xi(m\mid1,z,C,N)}-1\right\}
\bigg\}\,dm\,dz\bigg]. \tag*{(52)}
\end{aligned} E { ψ τ ( O ; η , ϕ , κ , ξ )} − E ( W ) τ 0 = E [ N W j = 1 ∑ N ∬ { { κ ∙ j − κ 0 ∙ j } ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N )} + { κ ∙ ( − j ) − κ 0 ∙ ( − j ) } ( m ∙ ( − j ) ∣ 0 , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N )} − η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) { κ ∙ j − κ 0 ∙ j } ( m ∙ j ∣ 1 , C , N ) { κ ∙ ( − j ) − κ 0 ∙ ( − j ) } ( m ∙ ( − j ) ∣ 0 , C , N ) + { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) ϕ 0 ( z ∣ 1 , C , N ) { ξ ( m ∣ 1 , z , C , N ) ξ 0 ( m ∣ 1 , z , C , N ) − 1 } } d m d z ] . (52) Consequently, if at least one of the following conditions holds for every j j j : (i) ( κ , η ∙ j ) = ( κ 0 , η 0 ∙ j ) (\kappa,\eta_{\bullet j})=(\kappa_{0},\eta_{0\bullet j}) ( κ , η ∙ j ) = ( κ 0 , η 0 ∙ j ) ; (ii) ( κ , ξ ) = ( κ 0 , ξ 0 ) (\kappa,\xi)=(\kappa_{0},\xi_{0}) ( κ , ξ ) = ( κ 0 , ξ 0 ) ; (iii) ( η ∙ j , ϕ , κ ∙ j ) = ( η 0 ∙ j , ϕ 0 , κ 0 ∙ j ) (\eta_{\bullet j},\phi,\kappa_{\bullet j})=(\eta_{0\bullet j},\phi_{0},\kappa_{0\bullet j}) ( η ∙ j , ϕ , κ ∙ j ) = ( η 0 ∙ j , ϕ 0 , κ 0 ∙ j ) ; (iv) ( η ∙ j , ϕ , κ ∙ ( − j ) ) = ( η 0 ∙ j , ϕ 0 , κ 0 ∙ ( − j ) ) (\eta_{\bullet j},\phi,\kappa_{\bullet(-j)})=(\eta_{0\bullet j},\phi_{0},\kappa_{0\bullet(-j)}) ( η ∙ j , ϕ , κ ∙ ( − j ) ) = ( η 0 ∙ j , ϕ 0 , κ 0 ∙ ( − j ) ) , then E { ψ τ ( O ; η , ϕ , κ , ξ ) } = E ( W ) τ 0 \mathbb{E}\{\psi_{\tau}(O;\eta,\phi,\kappa,\xi)\}=\mathbb{E}(W)\tau_{0} E { ψ τ ( O ; η , ϕ , κ , ξ )} = E ( W ) τ 0 .
The proofs of Theorems 8 and 9 are given in Sections D.2 and D.3 of the Appendix.
Each listed combination of correct nuisance functions makes the corresponding population bias zero. For θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) , correct η \eta η and ϕ \phi ϕ suffice even when both mediator density models are misspecified. For τ \tau τ , the product of the two mediator marginal errors remains, so correctness of η \eta η and ϕ \phi ϕ alone does not generally suffice; the last two robustness conditions additionally require one of the relevant mediator marginals to be correct.
Multiple robustness of θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ )
Proof. By iterated expectation, we have
\begin{aligned}
\mathbb{E}\{\psi_\theta(a,a';O;\eta,\phi,\kappa,\xi)\}
={}&\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{
\mathbb{E}\left[\omega^{(1)}(a,a',Z,M,C,N)\{\eta_{0\mathbin{\bullet}j}-\eta_{\mathbin{\bullet}j}\}(a,Z,M,C,N)\mid A=a,C,N\right]\right.\\
&\quad+\iint\eta_{\mathbin{\bullet}j}(a,z,m,C,N)\phi_0(z\mid a,C,N)\kappa(m\mid a',C,N)\,dz\,dm\\
&\quad+\iint\eta_{\mathbin{\bullet}j}(a,z,m,C,N)\phi(z\mid a,C,N)\kappa_0(m\mid a',C,N)\,dz\,dm\\
&\quad\left.-\iint\eta_{\mathbin{\bullet}j}(a,z,m,C,N)\phi(z\mid a,C,N)\kappa(m\mid a',C,N)\,dz\,dm
\right\}\right].
\end{aligned}
Subtracting E ( W ) θ 0 ( a , a ′ ) \mathbb{E}(W)\theta_0(a,a') E ( W ) θ 0 ( a , a ′ ) , adding and subtracting the true ratio ω 0 ( 1 ) \omega_0^{(1)} ω 0 ( 1 ) , and using
E [ ω 0 ( 1 ) ( η 0 ∙ j − η ∙ j ) ∣ A = a , C , N ] = ∬ ( η 0 ∙ j − η ∙ j ) ϕ 0 κ 0 d z d m \mathbb{E}\left[\omega_0^{(1)}(\eta_{0\mathbin{\bullet}j}-\eta_{\mathbin{\bullet}j})\mid A=a,C,N\right]
=\iint(\eta_{0\mathbin{\bullet}j}-\eta_{\mathbin{\bullet}j})\phi_0\kappa_0\,dz\,dm E [ ω 0 ( 1 ) ( η 0 ∙ j − η ∙ j ) ∣ A = a , C , N ] = ∬ ( η 0 ∙ j − η ∙ j ) ϕ 0 κ 0 d z d m
yield
\begin{aligned}
\mathbb{E}\{\psi_\theta(a,a';O;\eta,\phi,\kappa,\xi)\}-\mathbb{E}(W)\theta_0(a,a')
={}&\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{
\mathbb{E}\left[(\omega^{(1)}-\omega_0^{(1)})(\eta_{0\mathbin{\bullet}j}-\eta_{\mathbin{\bullet}j})(a,Z,M,C,N)\mid A=a,C,N\right]\right.\\
&\quad\left.+\iint\eta_{\mathbin{\bullet}j}(a,z,m,C,N)\{\phi_0-\phi\}(z\mid a,C,N)\{\kappa-\kappa_0\}(m\mid a',C,N)\,dz\,dm
\right\}\right].
\end{aligned}
Now
( ω ( 1 ) − ω 0 ( 1 ) ) ξ 0 = ( κ − κ 0 ) + κ ( ξ 0 ξ − 1 ) . (\omega^{(1)}-\omega_0^{(1)})\xi_0=(\kappa-\kappa_0)+\kappa\left(\frac{\xi_0}{\xi}-1\right). ( ω ( 1 ) − ω 0 ( 1 ) ) ξ 0 = ( κ − κ 0 ) + κ ( ξ ξ 0 − 1 ) .
Substitution and rearrangement gives the exact identity
E { ψ θ ( a , a ′ ; O ; η , ϕ , κ , ξ ) } − E ( W ) θ 0 ( a , a ′ ) = E [ W N ∑ j = 1 N { ∬ { κ − κ 0 } ( m ∣ a ′ , C , N ) [ η 0 ∙ j ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) − η ∙ j ( a , z , m , C , N ) ϕ ( z ∣ a , C , N ) ] d z d m + ∬ { η 0 ∙ j − η ∙ j } ( a , z , m , C , N ) κ ( m ∣ a ′ , C , N ) ϕ 0 ( z ∣ a , C , N ) × { ξ 0 ( m ∣ a , z , C , N ) ξ ( m ∣ a , z , C , N ) − 1 } d z d m } ] . \begin{aligned}
\mathbb{E}\{\psi_\theta(a,a';O;\eta,\phi,\kappa,\xi)\}-\mathbb{E}(W)\theta_0(a,a')
={}&\mathbb{E}[\frac{W}{N}\sum_{j=1}^{N}\{
\iint\{\kappa-\kappa_0\}(m\mid a',C,N)
[\eta_{0\mathbin{\bullet}j}(a,z,m,C,N)\phi_0(z\mid a,C,N)\\
&\qquad-\eta_{\mathbin{\bullet}j}(a,z,m,C,N)\phi(z\mid a,C,N)]\,dz\,dm\\
&\quad+\iint\{\eta_{0\mathbin{\bullet}j}-\eta_{\mathbin{\bullet}j}\}(a,z,m,C,N)\kappa(m\mid a',C,N)\phi_0(z\mid a,C,N)\\
&\qquad\times\{\frac{\xi_0(m\mid a,z,C,N)}{\xi(m\mid a,z,C,N)}-1\}\,dz\,dm
\}].
\end{aligned} E { ψ θ ( a , a ′ ; O ; η , ϕ , κ , ξ )} − E ( W ) θ 0 ( a , a ′ ) = E [ N W j = 1 ∑ N { ∬ { κ − κ 0 } ( m ∣ a ′ , C , N ) [ η 0 ∙ j ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) − η ∙ j ( a , z , m , C , N ) ϕ ( z ∣ a , C , N )] d z d m + ∬ { η 0 ∙ j − η ∙ j } ( a , z , m , C , N ) κ ( m ∣ a ′ , C , N ) ϕ 0 ( z ∣ a , C , N ) × { ξ ( m ∣ a , z , C , N ) ξ 0 ( m ∣ a , z , C , N ) − 1 } d z d m }] .
Each of the three configurations in Theorem 8 makes both terms in (51) equal to zero, proving the theorem.
Multiple robustness of τ \tau τ
Proof. Again by iterated expectation, the conditional mean of the outcome-residual term can be expressed as
∬ { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) × ϕ 0 ( z ∣ 1 , C , N ) ξ 0 ( m ∣ 1 , z , C , N ) ξ ( m ∣ 1 , z , C , N ) d m d z . \iint\left\{\eta_{0\bullet j}-\eta_{\bullet j}\right\}(1,z,m,C,N)\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\times\phi_{0}(z\mid1,C,N)\frac{\xi_{0}(m\mid1,z,C,N)}{\xi(m\mid1,z,C,N)}\,dm\,dz. ∬ { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) × ϕ 0 ( z ∣ 1 , C , N ) ξ ( m ∣ 1 , z , C , N ) ξ 0 ( m ∣ 1 , z , C , N ) d m d z .
The conditional means of the four remaining score components are, respectively,
∬ η ∙ j ϕ 0 κ ∙ j κ ∙ ( − j ) d m d z , ∬ η ∙ j ϕ κ 0 ∙ j κ ∙ ( − j ) d m d z , \iint\eta_{\bullet j}\phi_{0}\kappa_{\bullet j}\kappa_{\bullet(-j)}\,dm\,dz,\qquad\iint\eta_{\bullet j}\phi\kappa_{0\bullet j}\kappa_{\bullet(-j)}\,dm\,dz, ∬ η ∙ j ϕ 0 κ ∙ j κ ∙ ( − j ) d m d z , ∬ η ∙ j ϕ κ 0 ∙ j κ ∙ ( − j ) d m d z ,
∬ η ∙ j ϕ κ ∙ j κ 0 ∙ ( − j ) d m d z , − 2 ∬ η ∙ j ϕ κ ∙ j κ ∙ ( − j ) d m d z , \iint\eta_{\bullet j}\phi\kappa_{\bullet j}\kappa_{0\bullet(-j)}\,dm\,dz,\qquad-2\iint\eta_{\bullet j}\phi\kappa_{\bullet j}\kappa_{\bullet(-j)}\,dm\,dz, ∬ η ∙ j ϕ κ ∙ j κ 0 ∙ ( − j ) d m d z , − 2 ∬ η ∙ j ϕ κ ∙ j κ ∙ ( − j ) d m d z ,
where every function is evaluated at the arguments shown in (7): ϕ \phi ϕ is under arm 1, κ ∙ j \kappa_{\bullet j} κ ∙ j under arm 1, and κ ∙ ( − j ) \kappa_{\bullet(-j)} κ ∙ ( − j ) under arm 0. Write ξ 0 / ξ = 1 + ( ξ 0 / ξ − 1 ) \xi_{0}/\xi=1+(\xi_{0}/\xi-1) ξ 0 / ξ = 1 + ( ξ 0 / ξ − 1 ) , subtract the true numerator, and collect terms. This yields
E { ψ τ ( O ; η , ϕ , κ , ξ ) } − E ( W ) τ 0 = E [ W N ∑ j = 1 N ∬ { ( κ ∙ j − κ 0 ∙ j ) ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) × { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) } + ( κ ∙ ( − j ) − κ 0 ∙ ( − j ) ) ( m ∙ ( − j ) ∣ 0 , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) × { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N ) } − η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) ( κ ∙ j − κ 0 ∙ j ) ( m ∙ j ∣ 1 , C , N ) × ( κ ∙ ( − j ) − κ 0 ∙ ( − j ) ) ( m ∙ ( − j ) ∣ 0 , C , N ) + { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) ϕ 0 ( z ∣ 1 , C , N ) × { ξ 0 ( m ∣ 1 , z , C , N ) ξ ( m ∣ 1 , z , C , N ) − 1 } } d m d z ] . (53) \begin{aligned}
&\mathbb{E}\{\psi_{\tau}(O;\eta,\phi,\kappa,\xi)\}-\mathbb{E}(W)\tau_{0}\\
={}&\mathbb{E}[\frac{W}{N}\sum_{j=1}^{N}\iint\Biggl\{(\kappa_{\bullet j}-\kappa_{0\bullet j})(m_{\bullet j}\mid1,C,N)\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\\
&\quad\times\{\eta_{0\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)-\eta_{\bullet j}(1,z,m,C,N)\phi(z\mid1,C,N)\}\\
&\quad+(\kappa_{\bullet(-j)}-\kappa_{0\bullet(-j)})(m_{\bullet(-j)}\mid0,C,N)\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\\
&\quad\times\{\eta_{0\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)-\eta_{\bullet j}(1,z,m,C,N)\phi(z\mid1,C,N)\}\\
&\quad-\eta_{0\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)(\kappa_{\bullet j}-\kappa_{0\bullet j})(m_{\bullet j}\mid1,C,N)\\
&\quad\times(\kappa_{\bullet(-j)}-\kappa_{0\bullet(-j)})(m_{\bullet(-j)}\mid0,C,N)\\
&\quad+\{\eta_{0\bullet j}-\eta_{\bullet j}\}(1,z,m,C,N)\kappa_{\bullet j}(m_{\bullet j}\mid1,C,N)\kappa_{\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)\phi_{0}(z\mid1,C,N)\\
&\quad\times\{\frac{\xi_{0}(m\mid1,z,C,N)}{\xi(m\mid1,z,C,N)}-1\}\Biggr\}\,dm\,dz]. \tag*{(53)}
\end{aligned} = E { ψ τ ( O ; η , ϕ , κ , ξ )} − E ( W ) τ 0 E [ N W j = 1 ∑ N ∬ { ( κ ∙ j − κ 0 ∙ j ) ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) × { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N )} + ( κ ∙ ( − j ) − κ 0 ∙ ( − j ) ) ( m ∙ ( − j ) ∣ 0 , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) × { η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) − η ∙ j ( 1 , z , m , C , N ) ϕ ( z ∣ 1 , C , N )} − η 0 ∙ j ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) ( κ ∙ j − κ 0 ∙ j ) ( m ∙ j ∣ 1 , C , N ) × ( κ ∙ ( − j ) − κ 0 ∙ ( − j ) ) ( m ∙ ( − j ) ∣ 0 , C , N ) + { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) κ ∙ j ( m ∙ j ∣ 1 , C , N ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) ϕ 0 ( z ∣ 1 , C , N ) × { ξ ( m ∣ 1 , z , C , N ) ξ 0 ( m ∣ 1 , z , C , N ) − 1 } } d m d z ] . (53)
Each configuration in Theorem 9 makes every term in (52) vanish, proving the theorem.
Generic cross-fitting machinery
The following lemma isolates the argument shared by the canonical and low-dimensional estimators. It is stated at the cluster level, so the independent sample size is K K K .
Lemma 2. Let ζ 0 \zeta_{0} ζ 0 be a generic scalar estimand, let ν 0 = E ( W ) ∈ ( 0 , ∞ ) \nu_{0}=\mathbb{E}(W)\in(0,\infty) ν 0 = E ( W ) ∈ ( 0 , ∞ ) , and set μ 0 = ν 0 ζ 0 \mu_{0}=\nu_{0}\zeta_{0} μ 0 = ν 0 ζ 0 . For r = 1 , … , R r=1,\ldots,R r = 1 , … , R , let ψ ^ r \widehat{\psi}_{r} ψ r be a random function measurable with respect to the training clusters outside I r \mathcal{I}_{r} I r , and define
μ ^ = ∑ r = 1 R K r K P K , r ( ψ ^ r ) , ζ ^ = μ ^ P K ( W ) . \widehat{\mu}=\sum_{r=1}^{R}\frac{K_{r}}{K}\mathbb{P}_{K,r}(\widehat{\psi}_{r}),\qquad\widehat{\zeta}=\frac{\widehat{\mu}}{\mathbb{P}_{K}(W)}. μ = r = 1 ∑ R K K r P K , r ( ψ r ) , ζ = P K ( W ) μ . where P K , r \mathbb{P}_{K,r} P K , r is the empirical mean over I r \mathcal{I}_{r} I r . Suppose Assumption B B B holds and there is a fixed square-integrable function ψ ⋆ \psi^{\star} ψ ⋆ such that, uniformly over r r r ,
∥ ψ ^ r − ψ ⋆ ∥ 2 , P = o p ( 1 ) . (54) \left\lVert\widehat{\psi}_{r}-\psi^{\star}\right\rVert_{2,P}=o_{p}(1). \tag*{(54)} ψ r − ψ ⋆ 2 , P = o p ( 1 ) . (54) E ψ ⋆ = μ 0 . (55) \mathbb{E}\psi^{\star}=\mu_{0}. \tag*{(55)} E ψ ⋆ = μ 0 . (55) ∣ E ( ψ ^ r − ψ ⋆ ) ∣ = o p ( K − 1 / 2 ) . (56) \left|\mathbb{E}\left(\widehat{\psi}_{r}-\psi^{\star}\right)\right|=o_{p}\left(K^{-1/2}\right). \tag*{(56)} E ( ψ r − ψ ⋆ ) = o p ( K − 1/2 ) . (56)
Then
K ( ζ ^ − ζ 0 ) = 1 K ∑ i = 1 K ψ ⋆ ( O i ) − ζ 0 W i ν 0 + o p ( 1 ) . (57) \sqrt{K}(\widehat{\zeta}-\zeta_0)=\frac{1}{\sqrt{K}}\sum_{i=1}^{K}\frac{\psi^\star(O_i)-\zeta_0W_i}{\nu_0}+o_p(1). \tag*{(57)} K ( ζ − ζ 0 ) = K 1 i = 1 ∑ K ν 0 ψ ⋆ ( O i ) − ζ 0 W i + o p ( 1 ) . (57)
Consequently, if D ⋆ ( O ) = { ψ ⋆ ( O ) − ζ 0 W } / ν 0 D^\star(O)=\{\psi^\star(O)-\zeta_0W\}/\nu_0 D ⋆ ( O ) = { ψ ⋆ ( O ) − ζ 0 W } / ν 0 , then K ( ζ ^ − ζ 0 ) ⇝ N { 0 , E ( D ⋆ 2 ) } \sqrt{K}(\widehat{\zeta}-\zeta_0)\rightsquigarrow N\{0,\mathbb{E}(D^{\star2})\} K ( ζ − ζ 0 ) ⇝ N { 0 , E ( D ⋆ 2 )} .
If, for i ∈ I r i\in\mathcal{I}_r i ∈ I r ,
D ^ i = ψ ^ r ( O i ) − ζ ^ W i P K ( W ) , \widehat{D}_i=\frac{\widehat{\psi}_r(O_i)-\widehat{\zeta}W_i}{\mathbb{P}_K(W)}, D i = P K ( W ) ψ r ( O i ) − ζ W i ,
then P K ( D ^ 2 ) → p E ( D ⋆ 2 ) \mathbb{P}_K(\widehat{D}^2)\to_p\mathbb{E}(D^{\star2}) P K ( D 2 ) → p E ( D ⋆ 2 ) .
Proof. Let p r = K r / K p_r=K_r/K p r = K r / K . Adding and subtracting ψ ⋆ \psi^\star ψ ⋆ gives the exact numerator decomposition
μ ^ − μ 0 = ( P K − E ) ψ ⋆ + R 1 K + R 2 K . (58) \widehat{\mu}-\mu_0=(\mathbb{P}_K-\mathbb{E})\psi^\star+R_{1K}+R_{2K}. \tag*{(58)} μ − μ 0 = ( P K − E ) ψ ⋆ + R 1 K + R 2 K . (58)
where
R 1 K = ∑ r = 1 R p r ( P K , r − E ) ( ψ ^ r − ψ ⋆ ) , R 2 K = ∑ r = 1 R p r E ( ψ ^ r − ψ ⋆ ) . R_{1K}=\sum_{r=1}^{R}p_r(\mathbb{P}_{K,r}-\mathbb{E})(\widehat{\psi}_r-\psi^\star),\qquad R_{2K}=\sum_{r=1}^{R}p_r\mathbb{E}(\widehat{\psi}_r-\psi^\star). R 1 K = r = 1 ∑ R p r ( P K , r − E ) ( ψ r − ψ ⋆ ) , R 2 K = r = 1 ∑ R p r E ( ψ r − ψ ⋆ ) .
For a fixed fold r r r , condition on the sigma-field generated by the training clusters. The function g r = ψ ^ r − ψ ⋆ g_r=\widehat{\psi}_r-\psi^\star g r = ψ r − ψ ⋆ is then fixed and independent of the validation clusters, and
K Var { p r ( P K , r − E ) g r ∣ F − r } ≤ p r E ( g r 2 ) . K\operatorname{Var}\{p_r(\mathbb{P}_{K,r}-\mathbb{E})g_r\mid\mathcal{F}_{-r}\}\leq p_r\mathbb{E}(g_r^2). K Var { p r ( P K , r − E ) g r ∣ F − r } ≤ p r E ( g r 2 ) .
By (54) and Chebyshev’s inequality, we obtain K p r ( P K , r − E ) g r = o p ( 1 ) \sqrt{K}p_r(\mathbb{P}_{K,r}-\mathbb{E})g_r=o_p(1) K p r ( P K , r − E ) g r = o p ( 1 ) . Since R R R is fixed, we have
K R 1 K = o p ( 1 ) . (59) \sqrt{K}R_{1K}=o_p(1). \tag*{(59)} K R 1 K = o p ( 1 ) . (59)
Condition (56) similarly gives
K R 2 K = o p ( 1 ) . (60) \sqrt{K}R_{2K}=o_p(1). \tag*{(60)} K R 2 K = o p ( 1 ) . (60)
Combining (58) –(60) , we get
K ( μ ^ − μ 0 ) = K ( P K − E ) ψ ⋆ + o p ( 1 ) . (61) \sqrt{K}(\widehat{\mu}-\mu_0)=\sqrt{K}(\mathbb{P}_K-\mathbb{E})\psi^\star+o_p(1). \tag*{(61)} K ( μ − μ 0 ) = K ( P K − E ) ψ ⋆ + o p ( 1 ) . (61)
Thus μ ^ − μ 0 = O p ( K − 1 / 2 ) \widehat{\mu}-\mu_0=O_p(K^{-1/2}) μ − μ 0 = O p ( K − 1/2 ) . Also, ν ^ = P K ( W ) \widehat{\nu}=\mathbb{P}_K(W) ν = P K ( W ) satisfies ν ^ − ν 0 = O p ( K − 1 / 2 ) \widehat{\nu}-\nu_0=O_p(K^{-1/2}) ν − ν 0 = O p ( K − 1/2 ) and ν ^ → p ν 0 \widehat{\nu}\to_p\nu_0 ν → p ν 0 . The identity
ζ ^ − ζ 0 = ( μ ^ − μ 0 ) − ζ 0 ( ν ^ − ν 0 ) ν 0 + ( 1 ν ^ − 1 ν 0 ) { ( μ ^ − μ 0 ) − ζ 0 ( ν ^ − ν 0 ) } \widehat{\zeta}-\zeta_0=\frac{(\widehat{\mu}-\mu_0)-\zeta_0(\widehat{\nu}-\nu_0)}{\nu_0}+\left(\frac{1}{\widehat{\nu}}-\frac{1}{\nu_0}\right)\{(\widehat{\mu}-\mu_0)-\zeta_0(\widehat{\nu}-\nu_0)\} ζ − ζ 0 = ν 0 ( μ − μ 0 ) − ζ 0 ( ν − ν 0 ) + ( ν 1 − ν 0 1 ) {( μ − μ 0 ) − ζ 0 ( ν − ν 0 )}
shows that the second term is O p ( K − 1 ) O_p(K^{-1}) O p ( K − 1 ) . Substitution of (61) proves (57) ; the cluster-level central limit theorem concludes the Gaussian limit claim.
For variance estimation, score convergence and the same conditional Markov argument imply
1 K ∑ r = 1 R ∑ i ∈ I r { ψ ^ r ( O i ) − ψ ⋆ ( O i ) } 2 = o p ( 1 ) . \frac{1}{K}\sum_{r=1}^{R}\sum_{i\in\mathcal{I}_r}\{\widehat{\psi}_r(O_i)-\psi^\star(O_i)\}^2=o_p(1). K 1 r = 1 ∑ R i ∈ I r ∑ { ψ r ( O i ) − ψ ⋆ ( O i ) } 2 = o p ( 1 ) .
Together with ζ ^ → p ζ 0 \widehat{\zeta}\to_p\zeta_0 ζ → p ζ 0 , ν ^ → p ν 0 \widehat{\nu}\to_p\nu_0 ν → p ν 0 , and P K ( W 2 ) = O p ( 1 ) \mathbb{P}_K(W^2)=O_p(1) P K ( W 2 ) = O p ( 1 ) , this yields P K { ( D ^ − D ⋆ ) 2 } = o p ( 1 ) \mathbb{P}_K\{(\widehat{D}-D^\star)^2\}=o_p(1) P K {( D − D ⋆ ) 2 } = o p ( 1 ) . Hence, by the Cauchy–Schwarz inequality, we obtain
∣ P K ( D ^ 2 ) − P K ( D ⋆ 2 ) ∣ = o p ( 1 ) , \left|\mathbb{P}_K(\widehat{D}^2)-\mathbb{P}_K(D^{\star2})\right|=o_p(1), P K ( D 2 ) − P K ( D ⋆ 2 ) = o p ( 1 ) ,
and the distribution of large numbers yields P K ( D ⋆ 2 ) → p E ( D ⋆ 2 ) \mathbb{P}_{K}(D^{\star2}) \to_{p} \mathbb{E}(D^{\star2}) P K ( D ⋆ 2 ) → p E ( D ⋆ 2 ) .
Proof of Theorem 3
We assume that the outcome-residual density ratios in (5)–(7), the auxiliary regressions entering the applicable score, and the fitted-to-true ratios that appear in the product-rate conditions below are well-defined on their corresponding supports. Moreover, we assume that there is a finite constant B B B such that, with probability tending to one, the true, limiting, and fold-specific values of the outcome regressions, the outcome-residual density ratios, the auxiliary regressions entering the applicable score, and the fitted-to-true ratios in (8)–(9) are bounded in absolute value by B B B .
Proof. We verify the hypotheses of Lemma 2 separately for θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) and τ \tau τ . Throughout, condition on the training clusters outside a fixed validation fold I r \mathcal{I}_r I r . The fold-specific nuisance functions are then fixed, while the validation clusters remain independent cluster-level observations with distribution P P P .
We first compute the elementary weighted truncation step used to derive score convergence from the componentwise conditions in Assumption C. If G G G is nonnegative integrable, ∣ f r ∣ ≤ B |f_r| \le B ∣ f r ∣ ≤ B , and E ( f r 2 ) = o p ( 1 ) \mathbb{E}(f_r^2) = o_p(1) E ( f r 2 ) = o p ( 1 ) , then for every L > 0 L > 0 L > 0 ,
E ( G f r 2 ) ≤ L E ( f r 2 ) + B 2 E { G I ( G > L ) } . (62) \mathbb{E}(Gf_r^2) \le L\mathbb{E}(f_r^2) + B^2\mathbb{E}\{G\mathbb{I}(G > L)\}. \tag*{(62)} E ( G f r 2 ) ≤ L E ( f r 2 ) + B 2 E { G I ( G > L )} . (62)
Letting first K → ∞ K \to\infty K → ∞ and then L → ∞ L \to\infty L → ∞ gives E ( G f r 2 ) = o p ( 1 ) \mathbb{E}(Gf_r^2) = o_p(1) E ( G f r 2 ) = o p ( 1 ) . We apply this with
G = W 2 { 1 + 1 N ∑ j = 1 N Y ∙ j 2 } , G = W^2\left\{1 + \frac{1}{N}\sum_{j=1}^{N}Y_{\bullet j}^2\right\}, G = W 2 { 1 + N 1 j = 1 ∑ N Y ∙ j 2 } ,
which is integrable by Assumptions (A4) and C.
For θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) , set
ψ ^ r ( O ) = ψ θ { a , a ′ ; O ; h ^ ( − r ) } , ψ ⋆ ( O ) = ψ θ , c a n ⋆ ( O ) . \widehat{\psi}_r(O) = \psi_\theta\{a,a';O;\widehat{h}^{(-r)}\}, \qquad\psi^\star(O) = \psi^\star_{\theta,\mathrm{can}}(O). ψ r ( O ) = ψ θ { a , a ′ ; O ; h ( − r ) } , ψ ⋆ ( O ) = ψ θ , can ⋆ ( O ) .
Since h ⋆ h^\star h ⋆ satisfies one of the configurations in Theorem 8, the exact identity (51) gives E ψ ⋆ = ν 0 θ 0 ( a , a ′ ) \mathbb{E}\psi^\star= \nu_0\theta_0(a,a') E ψ ⋆ = ν 0 θ 0 ( a , a ′ ) .
We next derive score convergence. Expanding the finite cluster-level sum in (5), the difference ψ ^ r − ψ ⋆ \widehat{\psi}_r-\psi^\star ψ r − ψ ⋆ is a finite sum, over at most N max N_{\max} N m a x indices, of terms of the following forms:
W ( ω ^ ( 1 ) , ( − r ) − ω ( 1 ) , ⋆ ) Y ∙ j , W ω ^ ( 1 ) , ( − r ) ( η ^ ∙ j ( − r ) − η ∙ j ⋆ ) , W { u ^ ∙ j ( ℓ ) , ( − r ) − u ∙ j ( ℓ ) , ⋆ } , W(\widehat{\omega}^{(1),(-r)}-\omega^{(1),\star})Y_{\bullet j}, \qquad W\widehat{\omega}^{(1),(-r)}(\widehat{\eta}^{(-r)}_{\bullet j}-\eta^\star_{\bullet j}), \qquad W\{\widehat{u}^{(\ell),(-r)}_{\bullet j}-u^{(\ell),\star}_{\bullet j}\}, W ( ω ( 1 ) , ( − r ) − ω ( 1 ) , ⋆ ) Y ∙ j , W ω ( 1 ) , ( − r ) ( η ∙ j ( − r ) − η ∙ j ⋆ ) , W { u ∙ j ( ℓ ) , ( − r ) − u ∙ j ( ℓ ) , ⋆ } ,
up to fixed treatment-probability constants and signs, with ℓ = 1 , 2 , 3 \ell=1,2,3 ℓ = 1 , 2 , 3 . The second and third forms are controlled directly by boundedness and Assumption C(i). The first form contains Y ∙ j Y_{\bullet j} Y ∙ j and is controlled by (61). Since the number of terms is finite by Assumption (A4), we get
∥ ψ ^ r − ψ θ , c a n ⋆ ∥ 2 , P = o p ( 1 ) (63) \lVert\widehat{\psi}_r-\psi^\star_{\theta,\mathrm{can}}\rVert_{2,P}=o_p(1) \tag*{(63)} ∥ ψ r − ψ θ , can ⋆ ∥ 2 , P = o p ( 1 ) (63)
uniformly over folds. Thus the score-convergence condition in Lemma 2 is verified for the canonical θ \theta θ score.
It remains to verify the conditional population drift. Insert h ^ ( − r ) \widehat{h}^{(-r)} h ( − r ) into the exact remainder (51). Fix ( C , N ) (C,N) ( C , N ) and j j j . For the first product in that identity, insert the true product-intervention density and apply the Cauchy–Schwarz inequality under the ( z , m ) (z,m) ( z , m ) distribution:
∣ ∬ { κ ^ ( − r ) − κ 0 } { m ∣ a ′ , C , N } { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } d z d m ∣ = ∣ ∬ ϕ 0 ( z ∣ a , C , N ) κ 0 ( m ∣ a ′ , C , N ) { κ ^ ( − r ) ( m ∣ a ′ , C , N ) κ 0 ( m ∣ a ′ , C , N ) − 1 } { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ ^ ( − r ) ( z ∣ a , C , N ) ϕ 0 ( z ∣ a , C , N ) } d z d m ∣ ≤ [ ∬ ϕ 0 κ 0 { κ ^ ( − r ) κ 0 − 1 } 2 d z d m ] 1 / 2 [ ∬ ϕ 0 κ 0 { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ ^ ( − r ) ϕ 0 } 2 d z d m ] 1 / 2 . \begin{aligned}
\left|\iint\left\{\hat{\kappa}^{(-r)}-\kappa_0\right\}\left\{m\mid a',C,N\right\}\left\{\eta_{0\bullet j}\phi_0-\hat{\eta}^{(-r)}_{\bullet j}\hat{\phi}^{(-r)}\right\} \, dz\,dm\right| \\
&= \left|\iint\phi_0(z\mid a,C,N)\kappa_0(m\mid a',C,N)\left\{\frac{\hat{\kappa}^{(-r)}(m\mid a',C,N)}{\kappa_0(m\mid a',C,N)}-1\right\}\left\{\eta_{0\bullet j}-\hat{\eta}^{(-r)}_{\bullet j}\frac{\hat{\phi}^{(-r)}(z\mid a,C,N)}{\phi_0(z\mid a,C,N)}\right\}\,dz\,dm\right| \\
&\leq\left[\iint\phi_0\kappa_0\left\{\frac{\hat{\kappa}^{(-r)}}{\kappa_0}-1\right\}^{2}\,dz\,dm\right]^{1/2}\left[\iint\phi_0\kappa_0\left\{\eta_{0\bullet j}-\hat{\eta}^{(-r)}_{\bullet j}\frac{\hat{\phi}^{(-r)}}{\phi_0}\right\}^{2}\,dz\,dm\right]^{1/2}.
\end{aligned} ∬ { κ ^ ( − r ) − κ 0 } { m ∣ a ′ , C , N } { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } d z d m = ∬ ϕ 0 ( z ∣ a , C , N ) κ 0 ( m ∣ a ′ , C , N ) { κ 0 ( m ∣ a ′ , C , N ) κ ^ ( − r ) ( m ∣ a ′ , C , N ) − 1 } { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ 0 ( z ∣ a , C , N ) ϕ ^ ( − r ) ( z ∣ a , C , N ) } d z d m ≤ [ ∬ ϕ 0 κ 0 { κ 0 κ ^ ( − r ) − 1 } 2 d z d m ] 1/2 ∬ ϕ 0 κ 0 { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ 0 ϕ ^ ( − r ) } 2 d z d m 1/2 .
Multiplying by W W W , averaging over ( C , N ) (C,N) ( C , N ) and over j j j , and applying another Cauchy–Schwarz inequality in the outer cluster-level distribution yields
∣ E [ W N ∑ j = 1 N ∬ { κ ^ ( − r ) − κ 0 } { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } d z d m ] ∣ ≤ C ∥ W { κ ^ ( − r ) κ 0 − 1 } ∥ 2 , P ∥ W { η 0 − η ^ ( − r ) ϕ ^ ( − r ) ϕ 0 } ∥ 2 , P . (64) \left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iint\left\{\hat{\kappa}^{(-r)}-\kappa_0\right\}\left\{\eta_{0\bullet j}\phi_0-\hat{\eta}^{(-r)}_{\bullet j}\hat{\phi}^{(-r)}\right\}\,dz\,dm\right]\right|
\leq C\left\|\sqrt{W}\left\{\frac{\hat{\kappa}^{(-r)}}{\kappa_0}-1\right\}\right\|_{2,P}\left\|\sqrt{W}\left\{\eta_0-\hat{\eta}^{(-r)}\frac{\hat{\phi}^{(-r)}}{\phi_0}\right\}\right\|_{2,P}. \tag*{(64)} E [ N W j = 1 ∑ N ∬ { κ ^ ( − r ) − κ 0 } { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } d z d m ] ≤ C W { κ 0 κ ^ ( − r ) − 1 } 2 , P W { η 0 − η ^ ( − r ) ϕ 0 ϕ ^ ( − r ) } 2 , P . (64)
The constant C C C depends only on the treatment probabilities, N max N_{\max} N m a x , and the bounded nuisance ratios.
For the second product in (51) , rewrite
∬ { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ( − r ) ϕ 0 { ξ 0 ξ ^ ( − r ) − 1 } d z d m = ∬ ϕ 0 κ 0 { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ( − r ) κ 0 { ξ 0 ξ ^ ( − r ) − 1 } d z d m . \begin{aligned}
\iint\left\{\eta_{0\bullet j}-\hat{\eta}^{(-r)}_{\bullet j}\right\}\hat{\kappa}^{(-r)}\phi_0\left\{\frac{\xi_0}{\hat{\xi}^{(-r)}}-1\right\}\,dz\,dm \\
&= \iint\phi_0\kappa_0\left\{\eta_{0\bullet j}-\hat{\eta}^{(-r)}_{\bullet j}\right\}\frac{\hat{\kappa}^{(-r)}}{\kappa_0}\left\{\frac{\xi_0}{\hat{\xi}^{(-r)}}-1\right\}\,dz\,dm.
\end{aligned} ∬ { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ( − r ) ϕ 0 { ξ ^ ( − r ) ξ 0 − 1 } d z d m = ∬ ϕ 0 κ 0 { η 0 ∙ j − η ^ ∙ j ( − r ) } κ 0 κ ^ ( − r ) { ξ ^ ( − r ) ξ 0 − 1 } d z d m .
By the boundedness of κ ^ ( − r ) / κ 0 \hat{\kappa}^{(-r)}/\kappa_0 κ ^ ( − r ) / κ 0 , a similar Cauchy–Schwarz inequality argument provides
∣ E [ W N ∑ j = 1 N ∬ { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ( − r ) ϕ 0 { ξ 0 ξ ^ ( − r ) − 1 } d z d m ] ∣ ≤ C ∥ W { η ^ ( − r ) − η 0 } ∥ 2 , P ∥ W { ξ 0 ξ ^ ( − r ) − 1 } ∥ 2 , P . (65) \left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iint\left\{\eta_{0\bullet j}-\hat{\eta}^{(-r)}_{\bullet j}\right\}\hat{\kappa}^{(-r)}\phi_0\left\{\frac{\xi_0}{\hat{\xi}^{(-r)}}-1\right\}\,dz\,dm\right]\right|
\leq C\left\|\sqrt{W}\left\{\hat{\eta}^{(-r)}-\eta_0\right\}\right\|_{2,P}\left\|\sqrt{W}\left\{\frac{\xi_0}{\hat{\xi}^{(-r)}}-1\right\}\right\|_{2,P}. \tag*{(65)} E [ N W j = 1 ∑ N ∬ { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ( − r ) ϕ 0 { ξ ^ ( − r ) ξ 0 − 1 } d z d m ] ≤ C W { η ^ ( − r ) − η 0 } 2 , P W { ξ ^ ( − r ) ξ 0 − 1 } 2 , P . (65)
Adding (64) and (65) , and then using (8) , gives
∣ E ψ r − E ( W ) θ 0 ( a , a ′ ) ∣ = o P ( K − 1 / 2 ) \left|\mathbb{E}\psi_r-\mathbb{E}(W)\theta_0(a,a')\right|=o_P(K^{-1/2}) ∣ E ψ r − E ( W ) θ 0 ( a , a ′ ) ∣ = o P ( K − 1/2 )
uniformly over folds. Since E ψ ⋆ = E ( W ) θ 0 ( a , a ′ ) \mathbb{E}\psi^\star=\mathbb{E}(W)\theta_0(a,a') E ψ ⋆ = E ( W ) θ 0 ( a , a ′ ) , we have
∣ E ( ψ ^ r − ψ ⋆ ) ∣ = o P ( K − 1 / 2 ) , (66) \left|\mathbb{E}(\hat{\psi}_r-\psi^\star)\right|=o_P(K^{-1/2}), \tag*{(66)} E ( ψ ^ r − ψ ⋆ ) = o P ( K − 1/2 ) , (66)
which verifies the condition for Lemma 2 regarding θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) . Lemma 2 proves (10) and variance consistency.
For τ \tau τ , set
ψ ^ r ( O ) = ψ τ { O ; h ^ ( − r ) } , ψ ⋆ ( O ) = ψ τ , can ⋆ ( O ) . \hat{\psi}_r(O)=\psi_\tau\{O;\hat{h}^{(-r)}\},\qquad\psi^\star(O)=\psi^\star_{\tau,\operatorname{can}}(O). ψ ^ r ( O ) = ψ τ { O ; h ^ ( − r ) } , ψ ⋆ ( O ) = ψ τ , can ⋆ ( O ) .
Because h ⋆ h^{\star} h ⋆ satisfies one of the configurations in Theorem 9 , the exact identity (52) gives E ψ ⋆ = ν 0 τ 0 E\psi^{\star}=\nu_{0}\tau_{0} E ψ ⋆ = ν 0 τ 0 . Expanding (7) , the difference ψ ^ r − ψ τ , c a n ⋆ \widehat{\psi}_{r}-\psi_{\tau,\mathrm{can}}^{\star} ψ r − ψ τ , can ⋆ is a finite sum of terms of the forms
W ( ω ^ ∙ j ( 2 ) , ( − r ) − ω ∙ j ( 2 ) , ⋆ ) Y ∙ j , W ω ^ ∙ j ( 2 ) , ( − r ) ( η ^ ∙ j ( − r ) − η ∙ j ⋆ ) , W { r ^ ∙ j ( ℓ ) , ( − r ) − r ∙ j ( ℓ ) , ⋆ } , W\left(\widehat{\omega}_{\bullet j}^{(2),(-r)}-\omega_{\bullet j}^{(2),\star}\right)Y_{\bullet j},\quad W\widehat{\omega}_{\bullet j}^{(2),(-r)}\left(\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{\bullet j}^{\star}\right),\quad W\left\{\widehat{r}_{\bullet j}^{(\ell),(-r)}-r_{\bullet j}^{(\ell),\star}\right\}, W ( ω ∙ j ( 2 ) , ( − r ) − ω ∙ j ( 2 ) , ⋆ ) Y ∙ j , W ω ∙ j ( 2 ) , ( − r ) ( η ∙ j ( − r ) − η ∙ j ⋆ ) , W { r ∙ j ( ℓ ) , ( − r ) − r ∙ j ( ℓ ) , ⋆ } ,
with ℓ = 1 , 2 , 3 , 4 \ell=1,2,3,4 ℓ = 1 , 2 , 3 , 4 , up to fixed treatment-probability constants and signs. Boundedness, Assumption C(ii), the finite support of N N N , and (62) therefore imply
∥ ψ ^ r − ψ τ , c a n ⋆ ∥ 2 , P = o P ( 1 ) (67) \left\|\widehat{\psi}_{r}-\psi_{\tau,\mathrm{can}}^{\star}\right\|_{2,P}=o_{P}(1) \tag*{(67)} ψ r − ψ τ , can ⋆ 2 , P = o P ( 1 ) (67)
uniformly over folds.
It remains to verify the drift. Insert h ^ ( − r ) \widehat{h}^{(-r)} h ( − r ) into (52). For the first product, write the integrand relative to ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) \phi_{0}(z\mid1,C,N)\kappa_{0\bullet j}(m_{\bullet j}\mid1,C,N)\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) :
{ κ ^ ∙ j ( − r ) − κ 0 ∙ j } κ ^ ∙ ( − j ) ( − r ) { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } = ϕ 0 κ 0 ∙ j κ 0 ∙ ( − j ) { κ ^ ∙ j ( − r ) κ 0 ∙ j − 1 } κ ^ ∙ ( − j ) ( − r ) κ 0 ∙ ( − j ) { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ ^ ( − r ) ϕ 0 } . \begin{aligned}
\left\{\widehat{\kappa}_{\bullet j}^{(-r)}-\kappa_{0\bullet j}\right\}\widehat{\kappa}_{\bullet(-j)}^{(-r)}\left\{\eta_{0\bullet j}\phi_{0}-\widehat{\eta}_{\bullet j}^{(-r)}\widehat{\phi}^{(-r)}\right\} \\
&=\phi_{0}\kappa_{0\bullet j}\kappa_{0\bullet(-j)}
\left\{\frac{\widehat{\kappa}_{\bullet j}^{(-r)}}{\kappa_{0\bullet j}}-1\right\}
\frac{\widehat{\kappa}_{\bullet(-j)}^{(-r)}}{\kappa_{0\bullet(-j)}}
\left\{\eta_{0\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\frac{\widehat{\phi}^{(-r)}}{\phi_{0}}\right\}.
\end{aligned} { κ ∙ j ( − r ) − κ 0 ∙ j } κ ∙ ( − j ) ( − r ) { η 0 ∙ j ϕ 0 − η ∙ j ( − r ) ϕ ( − r ) } = ϕ 0 κ 0 ∙ j κ 0 ∙ ( − j ) { κ 0 ∙ j κ ∙ j ( − r ) − 1 } κ 0 ∙ ( − j ) κ ∙ ( − j ) ( − r ) { η 0 ∙ j − η ∙ j ( − r ) ϕ 0 ϕ ( − r ) } .
By boundedness of κ ^ ∙ ( − j ) ( − r ) / κ 0 ∙ ( − j ) \widehat{\kappa}_{\bullet(-j)}^{(-r)}/\kappa_{0\bullet(-j)} κ ∙ ( − j ) ( − r ) / κ 0 ∙ ( − j ) and the Cauchy–Schwarz inequality, we get
∣ E [ W N ∑ j = 1 N ∭ { κ ^ ∙ j ( − r ) − κ 0 ∙ j } κ ^ ∙ ( − j ) ( − r ) { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } d m d z ] ∣ ≤ C max 1 ≤ j ≤ N max ∥ W { κ ^ ∙ j ( − r ) κ 0 ∙ j − 1 } ∥ 2 , P ∥ W { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ ^ ( − r ) ϕ 0 } ∥ 2 , P . (68) \begin{aligned}
\left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iiint\left\{\widehat{\kappa}_{\bullet j}^{(-r)}-\kappa_{0\bullet j}\right\}\widehat{\kappa}_{\bullet(-j)}^{(-r)}\left\{\eta_{0\bullet j}\phi_{0}-\widehat{\eta}_{\bullet j}^{(-r)}\widehat{\phi}^{(-r)}\right\}\,dm\,dz\right]\right| \\
&\le C\max_{1\le j\le N_{\max}}\left\|\sqrt{W}\left\{\frac{\widehat{\kappa}_{\bullet j}^{(-r)}}{\kappa_{0\bullet j}}-1\right\}\right\|_{2,P}\left\|\sqrt{W}\left\{\eta_{0\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\frac{\widehat{\phi}^{(-r)}}{\phi_{0}}\right\}\right\|_{2,P}. \tag*{(68)}
\end{aligned} E [ N W j = 1 ∑ N ∭ { κ ∙ j ( − r ) − κ 0 ∙ j } κ ∙ ( − j ) ( − r ) { η 0 ∙ j ϕ 0 − η ∙ j ( − r ) ϕ ( − r ) } d m d z ] ≤ C 1 ≤ j ≤ N m a x max W { κ 0 ∙ j κ ∙ j ( − r ) − 1 } 2 , P W { η 0 ∙ j − η ∙ j ( − r ) ϕ 0 ϕ ( − r ) } 2 , P . (68)
The second product is symmetric and implies
∣ E [ W N ∑ j = 1 N ∭ { κ ^ ∙ ( − j ) ( − r ) − κ 0 ∙ ( − j ) } κ ^ ∙ j ( − r ) { η 0 ∙ j ϕ 0 − η ^ ∙ j ( − r ) ϕ ^ ( − r ) } d m d z ] ∣ ≤ C max 1 ≤ j ≤ N max ∥ W { κ ^ ∙ ( − j ) ( − r ) κ 0 ∙ ( − j ) − 1 } ∥ 2 , P ∥ W { η 0 ∙ j − η ^ ∙ j ( − r ) ϕ ^ ( − r ) ϕ 0 } ∥ 2 , P . (69) \begin{aligned}
\left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iiint\left\{\widehat{\kappa}_{\bullet(-j)}^{(-r)}-\kappa_{0\bullet(-j)}\right\}\widehat{\kappa}_{\bullet j}^{(-r)}\left\{\eta_{0\bullet j}\phi_{0}-\widehat{\eta}_{\bullet j}^{(-r)}\widehat{\phi}^{(-r)}\right\}\,dm\,dz\right]\right| \\
&\le C\max_{1\le j\le N_{\max}}\left\|\sqrt{W}\left\{\frac{\widehat{\kappa}_{\bullet(-j)}^{(-r)}}{\kappa_{0\bullet(-j)}}-1\right\}\right\|_{2,P}\left\|\sqrt{W}\left\{\eta_{0\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\frac{\widehat{\phi}^{(-r)}}{\phi_{0}}\right\}\right\|_{2,P}. \tag*{(69)}
\end{aligned} E [ N W j = 1 ∑ N ∭ { κ ∙ ( − j ) ( − r ) − κ 0 ∙ ( − j ) } κ ∙ j ( − r ) { η 0 ∙ j ϕ 0 − η ∙ j ( − r ) ϕ ( − r ) } d m d z ] ≤ C 1 ≤ j ≤ N m a x max W ⎩ ⎨ ⎧ κ 0 ∙ ( − j ) κ ∙ ( − j ) ( − r ) − 1 ⎭ ⎬ ⎫ 2 , P W { η 0 ∙ j − η ∙ j ( − r ) ϕ 0 ϕ ( − r ) } 2 , P . (69)
For the third product, insert the same true product density:
η 0 ∙ j ϕ 0 { κ ^ ∙ j ( − r ) − κ 0 ∙ j } { κ ^ ∙ ( − j ) ( − r ) − κ 0 ∙ ( − j ) } = η 0 ∙ j ϕ 0 κ 0 ∙ j κ 0 ∙ ( − j ) { κ ^ ∙ j ( − r ) κ 0 ∙ j − 1 } { κ ^ ∙ ( − j ) ( − r ) κ 0 ∙ ( − j ) − 1 } . \eta_{0\bullet j}\phi_{0}\left\{\widehat{\kappa}_{\bullet j}^{(-r)}-\kappa_{0\bullet j}\right\}\left\{\widehat{\kappa}_{\bullet(-j)}^{(-r)}-\kappa_{0\bullet(-j)}\right\}
=\eta_{0\bullet j}\phi_{0}\kappa_{0\bullet j}\kappa_{0\bullet(-j)}
\left\{\frac{\widehat{\kappa}_{\bullet j}^{(-r)}}{\kappa_{0\bullet j}}-1\right\}
\left\{\frac{\widehat{\kappa}_{\bullet(-j)}^{(-r)}}{\kappa_{0\bullet(-j)}}-1\right\}. η 0 ∙ j ϕ 0 { κ ∙ j ( − r ) − κ 0 ∙ j } { κ ∙ ( − j ) ( − r ) − κ 0 ∙ ( − j ) } = η 0 ∙ j ϕ 0 κ 0 ∙ j κ 0 ∙ ( − j ) { κ 0 ∙ j κ ∙ j ( − r ) − 1 } ⎩ ⎨ ⎧ κ 0 ∙ ( − j ) κ ∙ ( − j ) ( − r ) − 1 ⎭ ⎬ ⎫ .
The boundedness of η 0 \eta_0 η 0 and the Cauchy–Schwarz inequality yield
∣ E [ W N ∑ j = 1 N ∬ η 0 ∙ j ϕ 0 { κ ^ ∙ j ( − r ) − κ 0 ∙ j } { κ ^ ∙ ( − j ) ( − r ) − κ 0 ∙ ( − j ) } d m d z ] ∣ ≤ C max 1 ≤ j ≤ N max ∥ W { κ ^ ∙ j ( − r ) κ 0 ∙ j − 1 } ∥ 2 , P ∥ W { κ ^ ∙ ( − j ) ( − r ) κ 0 ∙ ( − j ) − 1 } ∥ 2 , P . (70) \left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iint\eta_{0\bullet j}\phi_0\left\{\widehat{\kappa}_{\bullet j}^{(-r)}-\kappa_{0\bullet j}\right\}\left\{\widehat{\kappa}_{\bullet(-j)}^{(-r)}-\kappa_{0\bullet(-j)}\right\}dm\,dz\right]\right|
\leq C\max_{1\leq j\leq N_{\max}}\left\|\sqrt{W}\left\{\frac{\widehat{\kappa}_{\bullet j}^{(-r)}}{\kappa_{0\bullet j}}-1\right\}\right\|_{2,P}\left\|\sqrt{W}\left\{\frac{\widehat{\kappa}_{\bullet(-j)}^{(-r)}}{\kappa_{0\bullet(-j)}}-1\right\}\right\|_{2,P}. \tag*{(70)} E [ N W j = 1 ∑ N ∬ η 0 ∙ j ϕ 0 { κ ∙ j ( − r ) − κ 0 ∙ j } { κ ∙ ( − j ) ( − r ) − κ 0 ∙ ( − j ) } d m d z ] ≤ C 1 ≤ j ≤ N m a x max W { κ 0 ∙ j κ ∙ j ( − r ) − 1 } 2 , P W ⎩ ⎨ ⎧ κ 0 ∙ ( − j ) κ ∙ ( − j ) ( − r ) − 1 ⎭ ⎬ ⎫ 2 , P . (70)
For the fourth product, rewrite
{ η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ∙ j ( − r ) κ ^ ∙ ( − j ) ( − r ) ϕ 0 { ξ 0 ξ ^ ( − r ) − 1 } = ϕ 0 κ 0 ∙ j κ 0 ∙ ( − j ) { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ∙ j ( − r ) κ 0 ∙ j κ ^ ∙ ( − j ) ( − r ) κ 0 ∙ ( − j ) { ξ 0 ξ ^ ( − r ) − 1 } . \begin{aligned}
\left\{\eta_{0\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\right\}\widehat{\kappa}_{\bullet j}^{(-r)}\widehat{\kappa}_{\bullet(-j)}^{(-r)}\phi_0\left\{\frac{\xi_0}{\widehat{\xi}^{(-r)}}-1\right\} \\
&=\phi_0\kappa_{0\bullet j}\kappa_{0\bullet(-j)}\left\{\eta_{0\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\right\}\frac{\widehat{\kappa}_{\bullet j}^{(-r)}}{\kappa_{0\bullet j}}\frac{\widehat{\kappa}_{\bullet(-j)}^{(-r)}}{\kappa_{0\bullet(-j)}}\left\{\frac{\xi_0}{\widehat{\xi}^{(-r)}}-1\right\}.
\end{aligned} { η 0 ∙ j − η ∙ j ( − r ) } κ ∙ j ( − r ) κ ∙ ( − j ) ( − r ) ϕ 0 { ξ ( − r ) ξ 0 − 1 } = ϕ 0 κ 0 ∙ j κ 0 ∙ ( − j ) { η 0 ∙ j − η ∙ j ( − r ) } κ 0 ∙ j κ ∙ j ( − r ) κ 0 ∙ ( − j ) κ ∙ ( − j ) ( − r ) { ξ ( − r ) ξ 0 − 1 } .
Absorbing the two bounded fitted-to-true mediator ratios into C C C gives
∣ E [ W N ∑ j = 1 N ∬ { η 0 ∙ j − η ^ ∙ j ( − r ) } κ ^ ∙ j ( − r ) κ ^ ∙ ( − j ) ( − r ) ϕ 0 { ξ 0 ξ ^ ( − r ) − 1 } d m d z ] ∣ ≤ C max 1 ≤ j ≤ N max ∥ W { η ^ ∙ j ( − r ) − η 0 ∙ j } ∥ 2 , P ∥ W { ξ 0 ξ ^ ( − r ) − 1 } ∥ 2 , P . (71) \left|\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\iint\left\{\eta_{0\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\right\}\widehat{\kappa}_{\bullet j}^{(-r)}\widehat{\kappa}_{\bullet(-j)}^{(-r)}\phi_0\left\{\frac{\xi_0}{\widehat{\xi}^{(-r)}}-1\right\}dm\,dz\right]\right|
\leq C\max_{1\leq j\leq N_{\max}}\left\|\sqrt{W}\left\{\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{0\bullet j}\right\}\right\|_{2,P}\left\|\sqrt{W}\left\{\frac{\xi_0}{\widehat{\xi}^{(-r)}}-1\right\}\right\|_{2,P}. \tag*{(71)} E [ N W j = 1 ∑ N ∬ { η 0 ∙ j − η ∙ j ( − r ) } κ ∙ j ( − r ) κ ∙ ( − j ) ( − r ) ϕ 0 { ξ ( − r ) ξ 0 − 1 } d m d z ] ≤ C 1 ≤ j ≤ N m a x max W { η ∙ j ( − r ) − η 0 ∙ j } 2 , P W { ξ ( − r ) ξ 0 − 1 } 2 , P . (71)
Adding (68) –(71) and using (9) yields
∣ E ψ ^ r − E ( W ) τ 0 ∣ = o p ( K − 1 / 2 ) \left|\mathbb{E}\widehat{\psi}_r-\mathbb{E}(W)\tau_0\right|=o_p\left(K^{-1/2}\right) E ψ r − E ( W ) τ 0 = o p ( K − 1/2 )
uniformly over folds. Since E ψ ⋆ = E ( W ) τ 0 \mathbb{E}\psi^\star=\mathbb{E}(W)\tau_0 E ψ ⋆ = E ( W ) τ 0 , we obtain
∣ E ( ψ ^ r − ψ ⋆ ) ∣ = o p ( K − 1 / 2 ) , (72) \left|\mathbb{E}(\widehat{\psi}_r-\psi^\star)\right|=o_p\left(K^{-1/2}\right), \tag*{(72)} E ( ψ r − ψ ⋆ ) = o p ( K − 1/2 ) , (72)
which verifies Lemma 2’s drift condition for τ \tau τ . Together with (67) , Lemma 2 proves (11) and variance consistency.
If h ⋆ = h 0 h^\star=h_0 h ⋆ = h 0 , Theorem 2 identifies the limiting influence functions with the canonical gradients, so their variances attain the nonparametric efficiency bounds.
Proofs for Section 5
Augmented-classification estimation of d d d
The ratio d d d in (12) can also be written as
d ( a , z , m , c , n ) = ϕ ( z ∣ a , c , n ) f Z ∣ A , M , C , N ( z ∣ a , m , c , n ) = κ ( m ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) . d(a,z,m,c,n)=\frac{\phi(z\mid a,c,n)}{f_{Z\mid A,M,C,N}(z\mid a,m,c,n)}=\frac{\kappa(m\mid a,c,n)}{\xi(m\mid a,z,c,n)}. d ( a , z , m , c , n ) = f Z ∣ A , M , C , N ( z ∣ a , m , c , n ) ϕ ( z ∣ a , c , n ) = ξ ( m ∣ a , z , c , n ) κ ( m ∣ a , c , n ) .
Thus d d d is the conditional density ratio of the distribution of Z Z Z given ( A , C , N ) (A,C,N) ( A , C , N ) to the distribution of Z Z Z given ( A , M , C , N ) (A,M,C,N) ( A , M , C , N ) . We estimate this ratio using the augmented-classification method proposed by Rudolph et al. [15 ] , adapted to the cluster-level observed data.
For each treatment arm a a a , let λ a ( ∙ ∣ c , n ) \lambda_a(\bullet\mid c,n) λ a ( ∙ ∣ c , n ) be a user-specified reference distribution on mediator vectors of dimension n n n , with support covering the support of M ∣ A = a , C = c , N = n M \mid A=a,C=c,N=n M ∣ A = a , C = c , N = n . The reference draw is required to be conditionally independent of Z Z Z given ( A , C , N ) (A,C,N) ( A , C , N ) . In a training fold, construct an augmented sample by creating, for each observed training cluster, one observed copy with label Γ = 0 \Gamma=0 Γ = 0 and one reference copy with label Γ = 1 \Gamma=1 Γ = 1 . The observed copy keeps ( Z , M , C , N , A ) (Z,M,C,N,A) ( Z , M , C , N , A ) unchanged. The reference copy keeps ( Z , C , N , A ) (Z,C,N,A) ( Z , C , N , A ) unchanged but replaces M M M by
M ~ ∼ λ A ( ∙ ∣ C , N ) , \widetilde{M} \sim\lambda_A(\bullet\mid C,N), M ∼ λ A ( ∙ ∣ C , N ) ,
drawn independently of Z Z Z conditional on ( A , C , N ) (A,C,N) ( A , C , N ) .
Let P λ P_\lambda P λ denote the distribution of the augmented data. Define the two binary classification functions
ℓ a ( z , m , c , n ) = P λ ( Γ = 1 ∣ A = a , Z = z , M = m , C = c , N = n ) , \ell_a(z,m,c,n)=P_\lambda(\Gamma=1\mid A=a,Z=z,M=m,C=c,N=n), ℓ a ( z , m , c , n ) = P λ ( Γ = 1 ∣ A = a , Z = z , M = m , C = c , N = n ) ,
and
ℓ a M ( m , c , n ) = P λ ( Γ = 1 ∣ A = a , M = m , C = c , N = n ) . \ell_a^M(m,c,n)=P_\lambda(\Gamma=1\mid A=a,M=m,C=c,N=n). ℓ a M ( m , c , n ) = P λ ( Γ = 1 ∣ A = a , M = m , C = c , N = n ) .
Then, on the common support, one has
d ( a , z , m , c , n ) = ℓ a ( z , m , c , n ) 1 − ℓ a ( z , m , c , n ) 1 − ℓ a M ( m , c , n ) ℓ a M ( m , c , n ) . d(a,z,m,c,n)=\frac{\ell_a(z,m,c,n)}{1-\ell_a(z,m,c,n)}\frac{1-\ell_a^M(m,c,n)}{\ell_a^M(m,c,n)}. d ( a , z , m , c , n ) = 1 − ℓ a ( z , m , c , n ) ℓ a ( z , m , c , n ) ℓ a M ( m , c , n ) 1 − ℓ a M ( m , c , n ) .
To verify this identity, note that under P λ P_\lambda P λ , conditional on ( A = a , C = c , N = n ) (A=a,C=c,N=n) ( A = a , C = c , N = n ) , the observed copy has joint density
ϕ ( z ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) , \phi(z\mid a,c,n)\xi(m\mid a,z,c,n), ϕ ( z ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) ,
whereas the reference copy has joint density
ϕ ( z ∣ a , c , n ) λ a ( m ∣ c , n ) . \phi(z\mid a,c,n)\lambda_a(m\mid c,n). ϕ ( z ∣ a , c , n ) λ a ( m ∣ c , n ) .
Therefore,
ℓ a ( z , m , c , n ) 1 − ℓ a ( z , m , c , n ) = λ a ( m ∣ c , n ) ξ ( m ∣ a , z , c , n ) . \frac{\ell_a(z,m,c,n)}{1-\ell_a(z,m,c,n)}=\frac{\lambda_a(m\mid c,n)}{\xi(m\mid a,z,c,n)}. 1 − ℓ a ( z , m , c , n ) ℓ a ( z , m , c , n ) = ξ ( m ∣ a , z , c , n ) λ a ( m ∣ c , n ) .
Similarly, after marginalizing over Z Z Z ,
ℓ a M ( m , c , n ) 1 − ℓ a M ( m , c , n ) = λ a ( m ∣ c , n ) κ ( m ∣ a , c , n ) . \frac{\ell_a^M(m,c,n)}{1-\ell_a^M(m,c,n)}=\frac{\lambda_a(m\mid c,n)}{\kappa(m\mid a,c,n)}. 1 − ℓ a M ( m , c , n ) ℓ a M ( m , c , n ) = κ ( m ∣ a , c , n ) λ a ( m ∣ c , n ) .
Taking the ratio of these two odds eliminates the user-specified reference distribution λ a \lambda_a λ a and yields
ℓ a ( z , m , c , n ) 1 − ℓ a ( z , m , c , n ) 1 − ℓ a M ( m , c , n ) ℓ a M ( m , c , n ) = κ ( m ∣ a , c , n ) ξ ( m ∣ a , z , c , n ) = d ( a , z , m , c , n ) . \frac{\ell_a(z,m,c,n)}{1-\ell_a(z,m,c,n)}\frac{1-\ell_a^M(m,c,n)}{\ell_a^M(m,c,n)}
=\frac{\kappa(m\mid a,c,n)}{\xi(m\mid a,z,c,n)}
=d(a,z,m,c,n). 1 − ℓ a ( z , m , c , n ) ℓ a ( z , m , c , n ) ℓ a M ( m , c , n ) 1 − ℓ a M ( m , c , n ) = ξ ( m ∣ a , z , c , n ) κ ( m ∣ a , c , n ) = d ( a , z , m , c , n ) .
Multiple reference copies may be generated for each training cluster. Provided that the two classifiers use the same class weights, the common class-prior odds cancel from their odds ratio.
In implementation, for each validation fold I r I_r I r , the two classifiers ℓ a \ell_a ℓ a and ℓ a M \ell_a^M ℓ a M are trained only on clusters outside I r I_r I r . The fold-specific estimator is
d ^ ( − r ) ( a , z , m , c , n ) = ℓ ^ a ( − r ) ( z , m , c , n ) 1 − ℓ ^ a ( − r ) ( z , m , c , n ) 1 − ℓ ^ a M , ( − r ) ( m , c , n ) ℓ ^ a M , ( − r ) ( m , c , n ) . \widehat{d}^{(-r)}(a,z,m,c,n)=\frac{\widehat{\ell}_a^{(-r)}(z,m,c,n)}{1-\widehat{\ell}_a^{(-r)}(z,m,c,n)}
\frac{1-\widehat{\ell}_a^{M,(-r)}(m,c,n)}{\widehat{\ell}_a^{M,(-r)}(m,c,n)}. d ( − r ) ( a , z , m , c , n ) = 1 − ℓ a ( − r ) ( z , m , c , n ) ℓ a ( − r ) ( z , m , c , n ) ℓ a M , ( − r ) ( m , c , n ) 1 − ℓ a M , ( − r ) ( m , c , n ) .
Proof of Proposition 1
Proof. Bayes’ rule and randomization give
π a s ( a ′ ∣ m , c , n ) π a ′ s ( a ∣ m , c , n ) = κ ( m ∣ a ′ , c , n ) κ ( m ∣ a , c , n ) . \frac{\pi_a s(a' \mid m,c,n)}{\pi_{a'}s(a \mid m,c,n)}=\frac{\kappa(m \mid a',c,n)}{\kappa(m \mid a,c,n)}. π a ′ s ( a ∣ m , c , n ) π a s ( a ′ ∣ m , c , n ) = κ ( m ∣ a , c , n ) κ ( m ∣ a ′ , c , n ) .
Multiplication by d ( a , z , m , c , n ) d(a,z,m,c,n) d ( a , z , m , c , n ) proves (15). Applying Bayes’ rule to A ∣ ( M ∙ ( − j ) , C , N ) A \mid(M_{\bullet(-j)},C,N) A ∣ ( M ∙ ( − j ) , C , N ) gives
π 1 − π s − j ( 0 ∣ m ∙ ( − j ) , c , n ) s − j ( 1 ∣ m ∙ ( − j ) , c , n ) = κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) . \frac{\pi}{1-\pi}\frac{s_{-j}(0 \mid m_{\bullet(-j)},c,n)}{s_{-j}(1 \mid m_{\bullet(-j)},c,n)}=\frac{\kappa_{\bullet(-j)}(m_{\bullet(-j)} \mid0,c,n)}{\kappa_{\bullet(-j)}(m_{\bullet(-j)} \mid1,c,n)}. 1 − π π s − j ( 1 ∣ m ∙ ( − j ) , c , n ) s − j ( 0 ∣ m ∙ ( − j ) , c , n ) = κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) .
Since
κ ( m ∣ 1 , c , n ) = κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) , \kappa(m \mid1,c,n)=\kappa_{\bullet j}^{*}(m_{\bullet j} \mid m_{\bullet(-j)},1,c,n)\kappa_{\bullet(-j)}(m_{\bullet(-j)} \mid1,c,n), κ ( m ∣ 1 , c , n ) = κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) ,
substitution into (14) proves (16).
Displays (17)–(19) follow directly from iterated expectation, with z z z or m m m held fixed when it appears as an argument of the working regression. We verify the spillover regressions. Under A = 0 A=0 A = 0 , C = c C=c C = c , N = n N=n N = n , the joint mediator density factors as
κ ( m ∣ 0 , c , n ) = κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 0 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) . \kappa(m \mid0,c,n)=\kappa_{\bullet j}^{*}(m_{\bullet j} \mid m_{\bullet(-j)},0,c,n)\kappa_{\bullet(-j)}(m_{\bullet(-j)} \mid0,c,n). κ ( m ∣ 0 , c , n ) = κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 0 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) .
Therefore, the right-hand side of (20) equals
∫ η ∙ j ( 1 , z , m , c , n ) κ ∙ j ∗ ( m ∙ j ∣ 1 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) d m , \int\eta_{\bullet j}(1,z,m,c,n)\kappa_{\bullet j}^{*}(m_{\bullet j} \mid1,c,n)\kappa_{\bullet(-j)}(m_{\bullet(-j)} \mid0,c,n)\,dm, ∫ η ∙ j ( 1 , z , m , c , n ) κ ∙ j ∗ ( m ∙ j ∣ 1 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) d m ,
which is r ∙ j ( 1 ) ( 1 , 1 , 0 , z , c , n ) r_{\bullet j}^{(1)}(1,1,0,z,c,n) r ∙ j ( 1 ) ( 1 , 1 , 0 , z , c , n ) with the working η \eta η . Conditional on A = 0 A=0 A = 0 , M ∙ j = m ∙ j M_{\bullet j}=m_{\bullet j} M ∙ j = m ∙ j , C = c C=c C = c , N = n N=n N = n , the density of M ∙ ( − j ) M_{\bullet(-j)} M ∙ ( − j ) is
κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 0 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) κ ∙ j ( m ∙ j ∣ 0 , c , n ) . \frac{\kappa_{\bullet j}^{*}(m_{\bullet j} \mid m_{\bullet(-j)},0,c,n)\kappa_{\bullet(-j)}(m_{\bullet(-j)} \mid0,c,n)}{\kappa_{\bullet j}(m_{\bullet j} \mid0,c,n)}. κ ∙ j ( m ∙ j ∣ 0 , c , n ) κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 0 , c , n ) κ ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , c , n ) .
The scalar density ratio in (21) cancels the first and last factors, proving (21). Similarly, conditional on A = 1 A=1 A = 1 , M ∙ ( − j ) = m ∙ ( − j ) M_{\bullet(-j)}=m_{\bullet(-j)} M ∙ ( − j ) = m ∙ ( − j ) , C = c C=c C = c , N = n N=n N = n , the density of M ∙ j M_{\bullet j} M ∙ j is κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) \kappa_{\bullet j}^{*}(m_{\bullet j} \mid m_{\bullet(-j)},1,c,n) κ ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) ; hence (22) follows by cancellation. Integrating (21) against κ ∙ j ( m ∙ j ∣ 1 , c , n ) \kappa_{\bullet j}(m_{\bullet j} \mid1,c,n) κ ∙ j ( m ∙ j ∣ 1 , c , n ) , or averaging (22) over M ∙ ( − j ) ∣ ( A = 0 , C = c , N = n ) M_{\bullet(-j)} \mid(A=0,C=c,N=n) M ∙ ( − j ) ∣ ( A = 0 , C = c , N = n ) , gives the two expressions in (23).
Proof of Lemma 1
Proof. At η = η 0 \eta=\eta_0 η = η 0 , the partial means in Proposition 1 are exactly the functions u ( 1 ) , u ( 2 ) , u ( 3 ) u^{(1)},u^{(2)},u^{(3)} u ( 1 ) , u ( 2 ) , u ( 3 ) and r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) r^{(1)},r^{(2)},r^{(3)},r^{(4)} r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) defined in Section 3. Substitution of (15) into (24) gives (5) term by term. Substitution of (16) into (25) gives (7) term by term. Both equalities are pointwise.
Proof of Theorem 4
Proof. We first prove (26). In this proof, denote
ρ ( M , C , N ) = π a s ( a ′ ∣ M , C , N ) π a ′ s ( a ∣ M , C , N ) , ρ 0 ( M , C , N ) = π a s 0 ( a ′ ∣ M , C , N ) π a ′ s 0 ( a ∣ M , C , N ) . \rho(M,C,N)=\frac{\pi_a\,s(a' \mid M,C,N)}{\pi_{a'}\,s(a \mid M,C,N)}, \qquad\rho_0(M,C,N)=\frac{\pi_a\,s_0(a' \mid M,C,N)}{\pi_{a'}\,s_0(a \mid M,C,N)}. ρ ( M , C , N ) = π a ′ s ( a ∣ M , C , N ) π a s ( a ′ ∣ M , C , N ) , ρ 0 ( M , C , N ) = π a ′ s 0 ( a ∣ M , C , N ) π a s 0 ( a ′ ∣ M , C , N ) .
Fix ( C , N ) (C,N) ( C , N ) and one j j j . By iterated expectation, the conditional mean of the j j j th summand in (24) is
E [ d ( a , Z , M , C , N ) ρ ( M , C , N ) { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , C , N ] + E { u 0 ∙ j ( 1 ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } + E { u 0 ∙ j ( 2 ) ( a , M , C , N ) ∣ A = a ′ , C , N } − u 0 ∙ j ( 3 ) ( a , a ′ , C , N ) . \begin{aligned}
&\mathbb{E}\left[d(a,Z,M,C,N)\rho(M,C,N)\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,Z,M,C,N)\mid A=a,C,N\right]\\
&\quad+\mathbb{E}\{u_{0\bullet j}^{(1)}(a,a',Z,C,N)\mid A=a,C,N\}\\
&\quad+\mathbb{E}\{u_{0\bullet j}^{(2)}(a,M,C,N)\mid A=a',C,N\}-u_{0\bullet j}^{(3)}(a,a',C,N).
\end{aligned} E [ d ( a , Z , M , C , N ) ρ ( M , C , N ) { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , C , N ] + E { u 0 ∙ j ( 1 ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } + E { u 0 ∙ j ( 2 ) ( a , M , C , N ) ∣ A = a ′ , C , N } − u 0 ∙ j ( 3 ) ( a , a ′ , C , N ) .
Compatibility and Proposition 1 imply that the last three terms equal
∬ η ∙ j ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) κ 0 ( m ∣ a ′ , C , N ) d z d m . \iint\eta_{\bullet j}(a,z,m,C,N)\phi_0(z\mid a,C,N)\kappa_0(m\mid a',C,N)\,dz\,dm. ∬ η ∙ j ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) κ 0 ( m ∣ a ′ , C , N ) d z d m .
The Radon–Nikodym identities for d 0 d_0 d 0 and the true classifier odds show that this display equals
E [ d 0 ( a , Z , M , C , N ) ρ 0 ( M , C , N ) η ∙ j ( a , Z , M , C , N ) ∣ A = a , C , N ] . \mathbb{E}\left[d_0(a,Z,M,C,N)\rho_0(M,C,N)\eta_{\bullet j}(a,Z,M,C,N)\mid A=a,C,N\right]. E [ d 0 ( a , Z , M , C , N ) ρ 0 ( M , C , N ) η ∙ j ( a , Z , M , C , N ) ∣ A = a , C , N ] .
After subtracting the true conditional target, the conditional drift is therefore
E [ { d ρ − d 0 ρ 0 } { η 0 ∙ j − η ∙ j } ∣ A = a , C , N ] . \mathbb{E}\left[\{d\rho-d_0\rho_0\}\{\eta_{0\bullet j}-\eta_{\bullet j}\}\mid A=a,C,N\right]. E [ { d ρ − d 0 ρ 0 } { η 0 ∙ j − η ∙ j } ∣ A = a , C , N ] .
The algebraic identity
d ρ − d 0 ρ 0 = ρ ( d − d 0 ) + d 0 ( ρ − ρ 0 ) d\rho-d_0\rho_0=\rho(d-d_0)+d_0(\rho-\rho_0) d ρ − d 0 ρ 0 = ρ ( d − d 0 ) + d 0 ( ρ − ρ 0 )
gives the first product in (26). It remains to identify the second product in (26). Since ρ ( M , C , N ) − ρ 0 ( M , C , N ) \rho(M,C,N)-\rho_0(M,C,N) ρ ( M , C , N ) − ρ 0 ( M , C , N ) is measurable with respect to ( M , C , N ) (M,C,N) ( M , C , N ) , iterated expectation gives
E [ d 0 ( a , Z , M , C , N ) { ρ ( M , C , N ) − ρ 0 ( M , C , N ) } { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , C , N ] = E [ { ρ ( M , C , N ) − ρ 0 ( M , C , N ) } E { d 0 ( a , Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , M , C , N } ∣ A = a , C , N ] . \begin{aligned}
&\mathbb{E}\left[d_0(a,Z,M,C,N)\{\rho(M,C,N)-\rho_0(M,C,N)\}\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,Z,M,C,N)\mid A=a,C,N\right]\\
&=\mathbb{E}\left[\{\rho(M,C,N)-\rho_0(M,C,N)\}\mathbb{E}\left\{d_0(a,Z,M,C,N)\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,Z,M,C,N)\mid A=a,M,C,N\right\}\mid A=a,C,N\right].
\end{aligned} E [ d 0 ( a , Z , M , C , N ) { ρ ( M , C , N ) − ρ 0 ( M , C , N )} { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , C , N ] = E [ { ρ ( M , C , N ) − ρ 0 ( M , C , N )} E { d 0 ( a , Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , M , C , N } ∣ A = a , C , N ] .
For fixed ( m , c , n ) (m,c,n) ( m , c , n ) , the inner conditional expectation is
E { d 0 ( a , Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , M = m , C = c , N = n } = ∫ κ 0 ( m ∣ a , c , n ) ξ 0 ( m ∣ a , z , c , n ) { η 0 ∙ j − η ∙ j } ( a , z , m , c , n ) ϕ 0 ( z ∣ a , c , n ) ξ 0 ( m ∣ a , z , c , n ) κ 0 ( m ∣ a , c , n ) d z = ∫ { η 0 ∙ j − η ∙ j } ( a , z , m , c , n ) ϕ 0 ( z ∣ a , c , n ) d z = { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) } ( a , m , c , n ) , \begin{aligned}
&\mathbb{E}\{d_0(a,Z,M,C,N)\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,Z,M,C,N)\mid A=a,M=m,C=c,N=n\}\\
&=\int\frac{\kappa_0(m\mid a,c,n)}{\xi_0(m\mid a,z,c,n)}\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,z,m,c,n)
\frac{\phi_0(z\mid a,c,n)\xi_0(m\mid a,z,c,n)}{\kappa_0(m\mid a,c,n)}\,dz\\
&=\int\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,z,m,c,n)\phi_0(z\mid a,c,n)\,dz\\
&=\{u_{0\bullet j}^{(2)}-u_{\bullet j}^{(2)}\}(a,m,c,n),
\end{aligned} E { d 0 ( a , Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , M = m , C = c , N = n } = ∫ ξ 0 ( m ∣ a , z , c , n ) κ 0 ( m ∣ a , c , n ) { η 0 ∙ j − η ∙ j } ( a , z , m , c , n ) κ 0 ( m ∣ a , c , n ) ϕ 0 ( z ∣ a , c , n ) ξ 0 ( m ∣ a , z , c , n ) d z = ∫ { η 0 ∙ j − η ∙ j } ( a , z , m , c , n ) ϕ 0 ( z ∣ a , c , n ) d z = { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) } ( a , m , c , n ) ,
where the last equality follows from compatibility of u ∙ j ( 2 ) u_{\bullet j}^{(2)} u ∙ j ( 2 ) with the same working regression η ∙ j \eta_{\bullet j} η ∙ j . Therefore, we obtain
E [ d 0 ( a , Z , M , C , N ) { ρ ( M , C , N ) − ρ 0 ( M , C , N ) } { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , C , N ] = E [ { ρ ( M , C , N ) − ρ 0 ( M , C , N ) } { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) } ( a , M , C , N ) ∣ A = a , C , N ] . \begin{aligned}
&\mathbb{E}\left[d_0(a,Z,M,C,N)\{\rho(M,C,N)-\rho_0(M,C,N)\}\{\eta_{0\bullet j}-\eta_{\bullet j}\}(a,Z,M,C,N)\mid A=a,C,N\right]\\
&=\mathbb{E}\left[\{\rho(M,C,N)-\rho_0(M,C,N)\}\{u_{0\bullet j}^{(2)}-u_{\bullet j}^{(2)}\}(a,M,C,N)\mid A=a,C,N\right].
\end{aligned} E [ d 0 ( a , Z , M , C , N ) { ρ ( M , C , N ) − ρ 0 ( M , C , N )} { η 0 ∙ j − η ∙ j } ( a , Z , M , C , N ) ∣ A = a , C , N ] = E [ { ρ ( M , C , N ) − ρ 0 ( M , C , N )} { u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) } ( a , M , C , N ) ∣ A = a , C , N ] .
Combining this display with the first product and then averaging over ( C , N ) (C,N) ( C , N ) , multiplying by W / N W/N W / N , and summing over j j j proves (26).
We next prove (27). Since ρ j − ρ 0 j \rho_j-\rho_{0j} ρ j − ρ 0 j is measurable with respect to ( M ∙ ( − j ) , C , N ) (M_{\bullet(-j)},C,N) ( M ∙ ( − j ) , C , N ) , the second term in
E [ { ρ j ( e ∙ j − e 0 ∙ j ) + e 0 ∙ j ( ρ j − ρ 0 j ) } { η 0 ∙ j − η ∙ j } ∣ A = 1 , C , N ] \mathbb{E}\left[\{\rho_j(e_{\bullet j}-e_{0\bullet j})+e_{0\bullet j}(\rho_j-\rho_{0j})\}\{\eta_{0\bullet j}-\eta_{\bullet j}\}\mid A=1,C,N\right] E [ { ρ j ( e ∙ j − e 0 ∙ j ) + e 0 ∙ j ( ρ j − ρ 0 j )} { η 0 ∙ j − η ∙ j } ∣ A = 1 , C , N ]
can be reduced by conditioning further on ( M ∙ ( − j ) , C , N ) (M_{\bullet(-j)},C,N) ( M ∙ ( − j ) , C , N ) . We now identify the resulting conditional mean. Fix ( m ∙ ( − j ) , c , n ) (m_{\bullet(-j)},c,n) ( m ∙ ( − j ) , c , n ) and write m = ( m ∙ j , m ∙ ( − j ) ) m=(m_{\bullet j},m_{\bullet(-j)}) m = ( m ∙ j , m ∙ ( − j ) ) . Under the true distribution, one has
p 0 ( z , m ∙ j ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ) = ϕ 0 ( z ∣ 1 , c , n ) ξ 0 ( m ∣ 1 , z , c , n ) κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) . p_0(z,m_{\bullet j}\mid A=1,M_{\bullet(-j)}=m_{\bullet(-j)},C=c,N=n)=\frac{\phi_0(z\mid1,c,n)\xi_0(m\mid1,z,c,n)}{\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid1,c,n)}. p 0 ( z , m ∙ j ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ) = κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) ϕ 0 ( z ∣ 1 , c , n ) ξ 0 ( m ∣ 1 , z , c , n ) .
Also, by the definition of e 0 ∙ j e_{0\bullet j} e 0 ∙ j , we get
e 0 ∙ j ( z , m , c , n ) = d 0 ( 1 , z , m , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) κ 0 ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) = κ 0 ( m ∣ 1 , c , n ) ξ 0 ( m ∣ 1 , z , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) κ 0 ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) . \begin{aligned}
e_{0\bullet j}(z,m,c,n)=d_0(1,z,m,c,n)\frac{\kappa_{0\bullet j}(m_{\bullet j}\mid1,c,n)}{\kappa^*_{0\bullet j}(m_{\bullet j}\mid m_{\bullet(-j)},1,c,n)} \\
&=\frac{\kappa_0(m\mid1,c,n)}{\xi_0(m\mid1,z,c,n)}\frac{\kappa_{0\bullet j}(m_{\bullet j}\mid1,c,n)}{\kappa^*_{0\bullet j}(m_{\bullet j}\mid m_{\bullet(-j)},1,c,n)}.
\end{aligned} e 0 ∙ j ( z , m , c , n ) = d 0 ( 1 , z , m , c , n ) κ 0 ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) = ξ 0 ( m ∣ 1 , z , c , n ) κ 0 ( m ∣ 1 , c , n ) κ 0 ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) .
Since
κ 0 ( m ∣ 1 , c , n ) = κ 0 ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) , \kappa_0(m\mid1,c,n)=\kappa^*_{0\bullet j}(m_{\bullet j}\mid m_{\bullet(-j)},1,c,n)\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid1,c,n), κ 0 ( m ∣ 1 , c , n ) = κ 0 ∙ j ∗ ( m ∙ j ∣ m ∙ ( − j ) , 1 , c , n ) κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , c , n ) ,
we obtain the following identity
e 0 ∙ j ( z , m , c , n ) p 0 ( z , m ∙ j ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ) = ϕ 0 ( z ∣ 1 , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) . e_{0\bullet j}(z,m,c,n)p_0(z,m_{\bullet j}\mid A=1,M_{\bullet(-j)}=m_{\bullet(-j)},C=c,N=n)
=\phi_0(z\mid1,c,n)\kappa_{0\bullet j}(m_{\bullet j}\mid1,c,n). e 0 ∙ j ( z , m , c , n ) p 0 ( z , m ∙ j ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ) = ϕ 0 ( z ∣ 1 , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) .
Therefore, we obtain
E [ e 0 ∙ j ( Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( 1 , Z , M , C , N ) ∣ A = 1 , M ∙ ( − j ) , C , N ] = ∬ { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j . \begin{aligned}
\mathbb{E}\left[e_{0\bullet j}(Z,M,C,N)\{\eta_{0\bullet j}-\eta_{\bullet j}\}(1,Z,M,C,N)\mid A=1,M_{\bullet(-j)},C,N\right] \\
&=\iint\{\eta_{0\bullet j}-\eta_{\bullet j}\}(1,z,m,C,N)\phi_0(z\mid1,C,N)\kappa_{0\bullet j}(m_{\bullet j}\mid1,C,N)\,dz\,dm_{\bullet j}.
\end{aligned} E [ e 0 ∙ j ( Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( 1 , Z , M , C , N ) ∣ A = 1 , M ∙ ( − j ) , C , N ] = ∬ { η 0 ∙ j − η ∙ j } ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j .
The part involving η 0 ∙ j \eta_{0\bullet j} η 0 ∙ j is exactly r 0 ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) r_{0\bullet j}^{(3)}(1,1,M_{\bullet(-j)},C,N) r 0 ∙ j ( 3 ) ( 1 , 1 , M ∙ ( − j ) , C , N ) . For the part involving η ∙ j \eta_{\bullet j} η ∙ j , compatibility yields
u ∙ j ( 2 ) ( 1 , m , c , n ) = ∫ η ∙ j ( 1 , z , m , c , n ) ϕ 0 ( z ∣ 1 , c , n ) d z . u_{\bullet j}^{(2)}(1,m,c,n)=\int\eta_{\bullet j}(1,z,m,c,n)\phi_0(z\mid1,c,n)\,dz. u ∙ j ( 2 ) ( 1 , m , c , n ) = ∫ η ∙ j ( 1 , z , m , c , n ) ϕ 0 ( z ∣ 1 , c , n ) d z .
Using the compatible definition of r ∙ j ( 3 ) r_{\bullet j}^{(3)} r ∙ j ( 3 ) , together with the true conditional density of M ∙ j M_{\bullet j} M ∙ j given ( A = 1 , M ∙ ( − j ) , C , N ) (A=1,M_{\bullet(-j)},C,N) ( A = 1 , M ∙ ( − j ) , C , N ) , gives
r ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , c , n ) = E [ u ∙ j ( 2 ) ( 1 , M , C , N ) κ 0 ∙ j ( M ∙ j ∣ 1 , C , N ) κ 0 ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 1 , C , N ) ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ] = ∫ u ∙ j ( 2 ) ( 1 , m , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) d m ∙ j = ∬ η ∙ j ( 1 , z , m , c , n ) ϕ 0 ( z ∣ 1 , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) d z d m ∙ j . \begin{aligned}
r_{\bullet j}^{(3)}(1,1,m_{\bullet(-j)},c,n) \\
&=\mathbb{E}\left[u_{\bullet j}^{(2)}(1,M,C,N)\frac{\kappa_{0\bullet j}(M_{\bullet j}\mid1,C,N)}{\kappa^*_{0\bullet j}(M_{\bullet j}\mid M_{\bullet(-j)},1,C,N)}\mid A=1,M_{\bullet(-j)}=m_{\bullet(-j)},C=c,N=n\right] \\
&=\int u_{\bullet j}^{(2)}(1,m,c,n)\kappa_{0\bullet j}(m_{\bullet j}\mid1,c,n)\,dm_{\bullet j} \\
&=\iint\eta_{\bullet j}(1,z,m,c,n)\phi_0(z\mid1,c,n)\kappa_{0\bullet j}(m_{\bullet j}\mid1,c,n)\,dz\,dm_{\bullet j}.
\end{aligned} r ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , c , n ) = E [ u ∙ j ( 2 ) ( 1 , M , C , N ) κ 0 ∙ j ∗ ( M ∙ j ∣ M ∙ ( − j ) , 1 , C , N ) κ 0 ∙ j ( M ∙ j ∣ 1 , C , N ) ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C = c , N = n ] = ∫ u ∙ j ( 2 ) ( 1 , m , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) d m ∙ j = ∬ η ∙ j ( 1 , z , m , c , n ) ϕ 0 ( z ∣ 1 , c , n ) κ 0 ∙ j ( m ∙ j ∣ 1 , c , n ) d z d m ∙ j .
Consequently, we get
E [ e 0 ∙ j ( Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( 1 , Z , M , C , N ) ∣ A = 1 , M ∙ ( − j ) , C , N ] = { r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) . \mathbb{E}\left[e_{0\bullet j}(Z,M,C,N)\{\eta_{0\bullet j}-\eta_{\bullet j}\}(1,Z,M,C,N)\mid A=1,M_{\bullet(-j)},C,N\right]
= \{r_{0\bullet j}^{(3)}-r_{\bullet j}^{(3)}\}(1,1,M_{\bullet(-j)},C,N). E [ e 0 ∙ j ( Z , M , C , N ) { η 0 ∙ j − η ∙ j } ( 1 , Z , M , C , N ) ∣ A = 1 , M ∙ ( − j ) , C , N ] = { r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) .
Hence the second part of the conditional drift is
E [ e 0 ∙ j ( ρ j − ρ 0 j ) { η 0 ∙ j − η ∙ j } ∣ A = 1 , C , N ] = E [ ( ρ j − ρ 0 j ) { r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 1 , C , N ] , \mathbb{E}\left[e_{0\bullet j}(\rho_j-\rho_{0j})\{\eta_{0\bullet j}-\eta_{\bullet j}\}\mid A=1,C,N\right]
= \mathbb{E}\left[(\rho_j-\rho_{0j})\{r_{0\bullet j}^{(3)}-r_{\bullet j}^{(3)}\}(1,1,M_{\bullet(-j)},C,N)\mid A=1,C,N\right], E [ e 0 ∙ j ( ρ j − ρ 0 j ) { η 0 ∙ j − η ∙ j } ∣ A = 1 , C , N ] = E [ ( ρ j − ρ 0 j ) { r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) } ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 1 , C , N ] ,
which is the second product in (27). Averaging over ( C , N ) (C,N) ( C , N ) , multiplying by W / N W/N W / N , and summing over j j j prove (27).
Proof of Theorem 5
We assume that there is a finite constant B B B such that, with probability tending to one, the true, limiting, and fold-specific values of the classifier odds, d d d , e ∙ j e_{\bullet j} e ∙ j , η \eta η , and all auxiliary regressions entering the applicable score are bounded in absolute value by B B B .
Proof. We verify the three hypotheses of Lemma 2 from the nuisance-level conditions in Assumption D. For every fold, condition on the training clusters. The fold-specific nuisance functions are then fixed and independent of a generic validation cluster.
Consider θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) . In this proof, let
ρ ^ r ( M , C , N ) = π a s ^ ( − r ) ( a ′ ∣ M , C , N ) π a ′ s ^ ( − r ) ( a ∣ M , C , N ) \widehat{\rho}_r(M,C,N)=\frac{\pi_a\widehat{s}^{(-r)}(a'\mid M,C,N)}{\pi_{a'}\widehat{s}^{(-r)}(a\mid M,C,N)} ρ r ( M , C , N ) = π a ′ s ( − r ) ( a ∣ M , C , N ) π a s ( − r ) ( a ′ ∣ M , C , N )
and define ρ ⋆ \rho^\star ρ ⋆ and ρ 0 \rho_0 ρ 0 analogously. Boundedness and (32) imply
∥ d ^ ( − r ) ρ ^ r − d ⋆ ρ ⋆ ∥ 2 , P ≤ B ∥ d ^ ( − r ) − d ⋆ ∥ 2 , P + B ∥ ρ ^ r − ρ ⋆ ∥ 2 , P = o p ( 1 ) \left\|\widehat{d}^{(-r)}\widehat{\rho}_r-d^\star\rho^\star\right\|_{2,P}
\le B\left\|\widehat{d}^{(-r)}-d^\star\right\|_{2,P}
+B\left\|\widehat{\rho}_r-\rho^\star\right\|_{2,P}
=o_p(1) d ( − r ) ρ r − d ⋆ ρ ⋆ 2 , P ≤ B d ( − r ) − d ⋆ 2 , P + B ∥ ρ r − ρ ⋆ ∥ 2 , P = o p ( 1 )
uniformly over folds. The outcome-residual part satisfies the exact algebraic identity
d ^ ( − r ) ρ ^ r { Y ∙ j − η ^ ∙ j ( − r ) } − d ⋆ ρ ⋆ { Y ∙ j − η ∙ j ⋆ } = { d ^ ( − r ) ρ ^ r − d ⋆ ρ ⋆ } { Y ∙ j − η ∙ j ⋆ } − d ^ ( − r ) ρ ^ r { η ^ ∙ j ( − r ) − η ∙ j ⋆ } . \widehat{d}^{(-r)}\widehat{\rho}_r\{Y_{\bullet j}-\widehat{\eta}_{\bullet j}^{(-r)}\}
-d^\star\rho^\star\{Y_{\bullet j}-\eta_{\bullet j}^\star\}
=
\{\widehat{d}^{(-r)}\widehat{\rho}_r-d^\star\rho^\star\}\{Y_{\bullet j}-\eta_{\bullet j}^\star\}
-\widehat{d}^{(-r)}\widehat{\rho}_r\{\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{\bullet j}^\star\}. d ( − r ) ρ r { Y ∙ j − η ∙ j ( − r ) } − d ⋆ ρ ⋆ { Y ∙ j − η ∙ j ⋆ } = { d ( − r ) ρ r − d ⋆ ρ ⋆ } { Y ∙ j − η ∙ j ⋆ } − d ( − r ) ρ r { η ∙ j ( − r ) − η ∙ j ⋆ } .
The remaining score difference is a bounded linear combination of
u ^ ∙ j ( 1 ) , ( − r ) − u ∙ j ( 1 ) , ⋆ , u ^ ∙ j ( 2 ) , ( − r ) − u ∙ j ( 2 ) , ⋆ , u ^ ∙ j ( 3 ) , ( − r ) − u ∙ j ( 3 ) , ⋆ . \widehat{u}_{\bullet j}^{(1),(-r)}-u_{\bullet j}^{(1),\star},\qquad
\widehat{u}_{\bullet j}^{(2),(-r)}-u_{\bullet j}^{(2),\star},\qquad
\widehat{u}_{\bullet j}^{(3),(-r)}-u_{\bullet j}^{(3),\star}. u ∙ j ( 1 ) , ( − r ) − u ∙ j ( 1 ) , ⋆ , u ∙ j ( 2 ) , ( − r ) − u ∙ j ( 2 ) , ⋆ , u ∙ j ( 3 ) , ( − r ) − u ∙ j ( 3 ) , ⋆ .
Using ( N − 1 ∑ j x j ) 2 ≤ N − 1 ∑ j x j 2 (N^{-1}\sum_j x_j)^2\le N^{-1}\sum_j x_j^2 ( N − 1 ∑ j x j ) 2 ≤ N − 1 ∑ j x j 2 , bounded treatment weights, and bounded nuisance functions, there is a fixed constant C C C such that
∣ ψ ^ θ , r ( O ) − ψ θ , l d ⋆ ( O ) ∣ 2 ≤ C W 2 N ∑ j = 1 N [ ( 1 + Y ∙ j 2 ) { ∣ d ^ ( − r ) ρ ^ r − d ⋆ ρ ⋆ ∣ 2 + ∣ η ^ ∙ j ( − r ) − η ∙ j ⋆ ∣ 2 } + ∑ ℓ = 1 3 ∣ u ^ ∙ j ( ℓ ) , ( − r ) − u ∙ j ( ℓ ) , ⋆ ∣ 2 ] . \left|\widehat{\psi}_{\theta,r}(O)-\psi_{\theta,\mathrm{ld}}^\star(O)\right|^2
\le C\frac{W^2}{N}\sum_{j=1}^{N}\left[(1+Y_{\bullet j}^2)\left\{|\widehat{d}^{(-r)}\widehat{\rho}_r-d^\star\rho^\star|^2+|\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{\bullet j}^\star|^2\right\}
+\sum_{\ell=1}^{3}|\widehat{u}_{\bullet j}^{(\ell),(-r)}-u_{\bullet j}^{(\ell),\star}|^2\right]. ψ θ , r ( O ) − ψ θ , ld ⋆ ( O ) 2 ≤ C N W 2 j = 1 ∑ N [ ( 1 + Y ∙ j 2 ) { ∣ d ( − r ) ρ r − d ⋆ ρ ⋆ ∣ 2 + ∣ η ∙ j ( − r ) − η ∙ j ⋆ ∣ 2 } + ℓ = 1 ∑ 3 ∣ u ∙ j ( ℓ ) , ( − r ) − u ∙ j ( ℓ ) , ⋆ ∣ 2 ] .
Again, by the same truncation argument used in the proof of Theorem 3, together with the bounded weights, the preceding score bound, and (32) provide
∥ ψ ^ θ , r − ψ θ , l d ∗ ∥ 2 , P = o p ( 1 ) (73) \left\lVert\widehat{\psi}_{\theta,r}-\psi^{\ast}_{\theta,\mathrm{ld}}\right\rVert_{2,P}=o_p(1) \tag*{(73)} ψ θ , r − ψ θ , ld ∗ 2 , P = o p ( 1 ) (73)
uniformly over folds.
Since h θ ∗ h^\ast_\theta h θ ∗ is compatible and satisfies (28), Theorem 4 gives E { ψ θ , l d ∗ ( O ) } = E ( W ) θ 0 ( a , a ′ ) \mathbb{E}\{\psi^\ast_{\theta,\mathrm{ld}}(O)\}=\mathbb{E}(W)\theta_0(a,a') E { ψ θ , ld ∗ ( O )} = E ( W ) θ 0 ( a , a ′ ) . Applying bias decomposition in Theorem 4 to the fold-specific nuisance functions and then using the triangle inequality gives
∣ E ψ ^ θ , r − E ( W ) θ 0 ( a , a ′ ) ∣ ≤ E [ W N ∑ j = 1 N E [ ∣ ρ ^ r ∣ ∣ d ^ ( − r ) − d 0 ∣ ∣ η 0 ∙ j − η ^ ∙ j ( − r ) ∣ + ∣ ρ ^ r − ρ 0 ∣ ∣ u 0 ∙ j ( 2 ) − u ^ ∙ j ( 2 ) , ( − r ) ∣ | A = a , C , N ] ] ] . \begin{aligned}
\left|\mathbb{E}\widehat{\psi}_{\theta,r}-\mathbb{E}(W)\theta_0(a,a')\right|
&\leq\mathbb{E}\left[
\frac{W}{N}\sum_{j=1}^{N}\mathbb{E}\left[
\left|\widehat{\rho}_r\right|\left|\widehat{d}^{(-r)}-d_0\right|\left|\eta_{0\mathbin{\bullet}j}-\widehat{\eta}^{(-r)}_{\mathbin{\bullet}j}\right|\right.\right.\\
&\qquad\left.\left.\left.{}+\left|\widehat{\rho}_r-\rho_0\right|\left|u^{(2)}_{0\mathbin{\bullet}j}-\widehat{u}^{(2),(-r)}_{\mathbin{\bullet}j}\right|
\,\middle|\, A=a,C,N\right]\right]\right].
\end{aligned} E ψ θ , r − E ( W ) θ 0 ( a , a ′ ) ≤ E [ N W j = 1 ∑ N E [ ∣ ρ r ∣ d ( − r ) − d 0 η 0 ∙ j − η ∙ j ( − r ) + ∣ ρ r − ρ 0 ∣ u 0 ∙ j ( 2 ) − u ∙ j ( 2 ) , ( − r ) A = a , C , N ] ] ] .
Since P ( A = a ∣ C , N ) = π a P(A=a\mid C,N)=\pi_a P ( A = a ∣ C , N ) = π a , for any nonnegative integrand in the preceding conditional expectation, one has
E { W E ( ∙ ∣ A = a , C , N ) } = 1 π a E { I ( A = a ) W ( ∙ ) } . \mathbb{E}\{W\mathbb{E}(\mathord{\bullet}\mid A=a,C,N)\}=\frac{1}{\pi_a}\mathbb{E}\{I(A=a)W(\mathord{\bullet})\}. E { W E ( ∙ ∣ A = a , C , N )} = π a 1 E { I ( A = a ) W ( ∙ )} .
Moreover, for any two square-integrable functions evaluated at the corresponding observed arguments, one has
E { I ( A = a ) W ∣ f g ∣ } ≤ { E ( W f 2 ) } 1 / 2 { E ( W g 2 ) } 1 / 2 = ∥ W f ∥ 2 , P ∥ W g ∥ 2 , P . \mathbb{E}\{I(A=a)W|fg|\}\leq\{\mathbb{E}(Wf^2)\}^{1/2}\{\mathbb{E}(Wg^2)\}^{1/2}=\|\sqrt{W}f\|_{2,P}\|\sqrt{W}g\|_{2,P}. E { I ( A = a ) W ∣ f g ∣ } ≤ { E ( W f 2 ) } 1/2 { E ( W g 2 ) } 1/2 = ∥ W f ∥ 2 , P ∥ W g ∥ 2 , P .
Since ∣ ρ ^ r ∣ |\widehat{\rho}_r| ∣ ρ r ∣ is uniformly bounded and
∣ ρ ^ r − ρ 0 ∣ = π a π a ′ ∣ s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) − s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) ∣ , |\widehat{\rho}_r-\rho_0|=\frac{\pi_a}{\pi_{a'}}\left|\frac{\widehat{s}^{(-r)}(a'\mid M,C,N)}{\widehat{s}^{(-r)}(a\mid M,C,N)}-\frac{s_0(a'\mid M,C,N)}{s_0(a\mid M,C,N)}\right|, ∣ ρ r − ρ 0 ∣ = π a ′ π a s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) − s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) ,
the preceding two displays yield
∣ E ψ ^ θ , r − E ( W ) θ 0 ( a , a ′ ) ∣ ≤ C [ ∥ W { d ^ ( − r ) − d 0 } ∥ 2 , P ∥ W { η ^ ( − r ) − η 0 } ∥ 2 , P + ∥ W { s ^ ( − r ) ( a ′ ∣ M , C , N ) s ^ ( − r ) ( a ∣ M , C , N ) − s 0 ( a ′ ∣ M , C , N ) s 0 ( a ∣ M , C , N ) } ∥ 2 , P × ∥ W { u ^ ( 2 ) , ( − r ) − u 0 ( 2 ) } ∥ 2 , P ] . \begin{aligned}
\left|\mathbb{E}\widehat{\psi}_{\theta,r}-\mathbb{E}(W)\theta_0(a,a')\right|
\leq C\Bigg[&
\left\lVert\sqrt{W}\{\widehat{d}^{(-r)}-d_0\}\right\rVert_{2,P}
\left\lVert\sqrt{W}\{\widehat{\eta}^{(-r)}-\eta_0\}\right\rVert_{2,P}\\
&+\left\lVert\sqrt{W}\left\{
\frac{\widehat{s}^{(-r)}(a'\mid M,C,N)}{\widehat{s}^{(-r)}(a\mid M,C,N)}
-\frac{s_0(a'\mid M,C,N)}{s_0(a\mid M,C,N)}
\right\}\right\rVert_{2,P}\\
&\qquad\qquad\qquad\qquad{}\times
\left\lVert\sqrt{W}\{\widehat{u}^{(2),(-r)}-u_0^{(2)}\}\right\rVert_{2,P}
\Bigg].
\end{aligned} E ψ θ , r − E ( W ) θ 0 ( a , a ′ ) ≤ C [ W { d ( − r ) − d 0 } 2 , P W { η ( − r ) − η 0 } 2 , P + W { s ( − r ) ( a ∣ M , C , N ) s ( − r ) ( a ′ ∣ M , C , N ) − s 0 ( a ∣ M , C , N ) s 0 ( a ′ ∣ M , C , N ) } 2 , P × W { u ( 2 ) , ( − r ) − u 0 ( 2 ) } 2 , P ] .
where C C C depends only on the fixed treatment probabilities, the uniform nuisance bound, and N max N_{\max} N m a x . Condition (33) makes its right-hand side o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) . Since the limiting score is centered,
∣ E { ψ ^ θ , r − ψ θ , l d ∗ } ∣ = o p ( K − 1 / 2 ) (74) \left|\mathbb{E}\{\widehat{\psi}_{\theta,r}-\psi^\ast_{\theta,\mathrm{ld}}\}\right|=o_p(K^{-1/2}) \tag*{(74)} E { ψ θ , r − ψ θ , ld ∗ } = o p ( K − 1/2 ) (74)
uniformly over folds. Equations (72) and (73) verify Lemma 2, which proves (36) and consistency of the empirical variance.
For τ \tau τ , let
ρ ^ r j ( M ∙ ( − j ) , C , N ) = π 1 − π s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) \widehat{\rho}_{rj}(M_{\mathbin{\bullet}(-j)},C,N)=\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid M_{\mathbin{\bullet}(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid M_{\mathbin{\bullet}(-j)},C,N)} ρ r j ( M ∙ ( − j ) , C , N ) = 1 − π π s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N )
and define ρ j ∗ \rho^\ast_j ρ j ∗ and ρ 0 j \rho_{0j} ρ 0 j analogously. The boundedness and (34) imply ∥ e ^ ∙ j ( − r ) ρ ^ r j − e ∙ j ∗ ρ j ∗ ∥ 2 , P = o p ( 1 ) \left\lVert\widehat{e}^{(-r)}_{\mathbin{\bullet}j}\widehat{\rho}_{rj}-e^\ast_{\mathbin{\bullet}j}\rho^\ast_j\right\rVert_{2,P}=o_p(1) e ∙ j ( − r ) ρ r j − e ∙ j ∗ ρ j ∗ 2 , P = o p ( 1 ) uniformly over folds and j ≤ N max j \le N_{\max} j ≤ N m a x . The outcome-residual difference satisfies
e ^ ∙ j ( − r ) ρ ^ r j { Y ∙ j − η ^ ∙ j ( − r ) } − e ∙ j ⋆ ρ r j ⋆ { Y ∙ j − η ∙ j ⋆ } = { e ^ ∙ j ( − r ) ρ ^ r j − e ∙ j ⋆ ρ r j ⋆ } { Y ∙ j − η ∙ j ⋆ } − e ^ ∙ j ( − r ) ρ ^ r j { η ^ ∙ j ( − r ) − η ∙ j ⋆ } . \widehat e_{\bullet j}^{(-r)}\widehat\rho_{rj}\{Y_{\bullet j}-\widehat\eta_{\bullet j}^{(-r)}\}-e_{\bullet j}^{\star}\rho_{rj}^{\star}\{Y_{\bullet j}-\eta_{\bullet j}^{\star}\}
=\{\widehat e_{\bullet j}^{(-r)}\widehat\rho_{rj}-e_{\bullet j}^{\star}\rho_{rj}^{\star}\}\{Y_{\bullet j}-\eta_{\bullet j}^{\star}\}-\widehat e_{\bullet j}^{(-r)}\widehat\rho_{rj}\{\widehat\eta_{\bullet j}^{(-r)}-\eta_{\bullet j}^{\star}\}. e ∙ j ( − r ) ρ r j { Y ∙ j − η ∙ j ( − r ) } − e ∙ j ⋆ ρ r j ⋆ { Y ∙ j − η ∙ j ⋆ } = { e ∙ j ( − r ) ρ r j − e ∙ j ⋆ ρ r j ⋆ } { Y ∙ j − η ∙ j ⋆ } − e ∙ j ( − r ) ρ r j { η ∙ j ( − r ) − η ∙ j ⋆ } .
The augmentation part of the score difference is the bounded linear combination
A π { ( r ^ ∙ j ( 1 ) , ( − r ) − r ∙ j ( 1 ) , ⋆ ) − ( r ^ ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ ) } + A π { ( r ^ ∙ j ( 2 ) , ( − r ) − r ∙ j ( 2 ) , ⋆ ) − ( r ^ ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ ) } + 1 − A 1 − π { ( r ^ ∙ j ( 3 ) , ( − r ) − r ∙ j ( 3 ) , ⋆ ) − ( r ^ ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ ) } + ( r ^ ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ ) . \frac{A}{\pi}\{(\widehat r_{\bullet j}^{(1),(-r)}-r_{\bullet j}^{(1),\star})-(\widehat r_{\bullet j}^{(4),(-r)}-r_{\bullet j}^{(4),\star})\}
+\frac{A}{\pi}\{(\widehat r_{\bullet j}^{(2),(-r)}-r_{\bullet j}^{(2),\star})-(\widehat r_{\bullet j}^{(4),(-r)}-r_{\bullet j}^{(4),\star})\}
+\frac{1-A}{1-\pi}\{(\widehat r_{\bullet j}^{(3),(-r)}-r_{\bullet j}^{(3),\star})-(\widehat r_{\bullet j}^{(4),(-r)}-r_{\bullet j}^{(4),\star})\}+(\widehat r_{\bullet j}^{(4),(-r)}-r_{\bullet j}^{(4),\star}). π A {( r ∙ j ( 1 ) , ( − r ) − r ∙ j ( 1 ) , ⋆ ) − ( r ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ )} + π A {( r ∙ j ( 2 ) , ( − r ) − r ∙ j ( 2 ) , ⋆ ) − ( r ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ )} + 1 − π 1 − A {( r ∙ j ( 3 ) , ( − r ) − r ∙ j ( 3 ) , ⋆ ) − ( r ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ )} + ( r ∙ j ( 4 ) , ( − r ) − r ∙ j ( 4 ) , ⋆ ) .
The same averaging and boundedness argument used for θ \theta θ gives
∣ ψ ^ τ , r ( O ) − ψ τ , l d ⋆ ( O ) ∣ 2 ≤ C W 2 N ∑ j = 1 N [ ( 1 + Y ∙ j 2 ) { ∣ e ^ ∙ j ( − r ) ρ ^ r j − e ∙ j ⋆ ρ r j ⋆ ∣ 2 + ∣ η ^ ∙ j ( − r ) − η ∙ j ⋆ ∣ 2 } + ∑ ℓ = 1 4 ∣ r ^ ∙ j ( ℓ ) , ( − r ) − r ∙ j ( ℓ ) , ⋆ ∣ 2 ] . \left|\widehat\psi_{\tau,r}(O)-\psi_{\tau,\mathrm{ld}}^{\star}(O)\right|^{2}
\le C\frac{W^{2}}{N}\sum_{j=1}^{N}\left[(1+Y_{\bullet j}^{2})\left\{\left|\widehat e_{\bullet j}^{(-r)}\widehat\rho_{rj}-e_{\bullet j}^{\star}\rho_{rj}^{\star}\right|^{2}+\left|\widehat\eta_{\bullet j}^{(-r)}-\eta_{\bullet j}^{\star}\right|^{2}\right\}+\sum_{\ell=1}^{4}\left|\widehat r_{\bullet j}^{(\ell),(-r)}-r_{\bullet j}^{(\ell),\star}\right|^{2}\right]. ψ τ , r ( O ) − ψ τ , ld ⋆ ( O ) 2 ≤ C N W 2 j = 1 ∑ N [ ( 1 + Y ∙ j 2 ) { e ∙ j ( − r ) ρ r j − e ∙ j ⋆ ρ r j ⋆ 2 + η ∙ j ( − r ) − η ∙ j ⋆ 2 } + ℓ = 1 ∑ 4 r ∙ j ( ℓ ) , ( − r ) − r ∙ j ( ℓ ) , ⋆ 2 ] .
The truncation argument and (34) therefore yield
∥ ψ ^ τ , r − ψ τ , l d ⋆ ∥ 2 , P = o p ( 1 ) (75) \left\|\widehat\psi_{\tau,r}-\psi_{\tau,\mathrm{ld}}^{\star}\right\|_{2,P}=o_{p}(1) \tag*{(75)} ψ τ , r − ψ τ , ld ⋆ 2 , P = o p ( 1 ) (75)
uniformly over folds. This establishes the role of all four auxiliary regressions in (34).
Compatibility and (29) give E { ψ τ , l d ⋆ ( O ) } = E ( W ) τ 0 \mathbb{E}\{\psi_{\tau,\mathrm{ld}}^{\star}(O)\}=\mathbb{E}(W)\tau_{0} E { ψ τ , ld ⋆ ( O )} = E ( W ) τ 0 . The exact drift identity, the triangle inequality, and treatment randomization now give
∣ E ψ ^ τ , r − E ( W ) τ 0 ∣ ≤ 1 π E [ A W N ∑ j = 1 N { ∣ ρ ^ r j ∣ ∣ e ^ ∙ j ( − r ) − e 0 ∙ j ∣ ∣ η 0 ∙ j − η ^ ∙ j ( − r ) ∣ + ∣ ρ ^ r j − ρ 0 j ∣ ∣ r 0 ∙ j ( 3 ) − r ^ ∙ j ( 3 ) , ( − r ) ∣ } ] . \left|\mathbb{E}\widehat\psi_{\tau,r}-\mathbb{E}(W)\tau_{0}\right|
\le\frac{1}{\pi}\mathbb{E}\left[\frac{AW}{N}\sum_{j=1}^{N}\left\{|\widehat\rho_{rj}|\left|\widehat e_{\bullet j}^{(-r)}-e_{0\bullet j}\right|\left|\eta_{0\bullet j}-\widehat\eta_{\bullet j}^{(-r)}\right|+|\widehat\rho_{rj}-\rho_{0j}|\left|r_{0\bullet j}^{(3)}-\widehat r_{\bullet j}^{(3),(-r)}\right|\right\}\right]. E ψ τ , r − E ( W ) τ 0 ≤ π 1 E [ N A W j = 1 ∑ N { ∣ ρ r j ∣ e ∙ j ( − r ) − e 0 ∙ j η 0 ∙ j − η ∙ j ( − r ) + ∣ ρ r j − ρ 0 j ∣ r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) , ( − r ) } ] .
Applying the Cauchy–Schwarz inequality displayed above and using
∣ ρ ^ r j − ρ 0 j ∣ = π 1 − π ∣ s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) ∣ |\widehat\rho_{rj}-\rho_{0j}|=\frac{\pi}{1-\pi}\left|\frac{\widehat s_{-j}^{(-r)}(0\mid\mathbf{M}_{\bullet(-j)},C,N)}{\widehat s_{-j}^{(-r)}(1\mid\mathbf{M}_{\bullet(-j)},C,N)}-\frac{s_{0,-j}(0\mid\mathbf{M}_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid\mathbf{M}_{\bullet(-j)},C,N)}\right| ∣ ρ r j − ρ 0 j ∣ = 1 − π π s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N )
yield
∣ E ψ ^ τ , r − E ( W ) τ 0 ∣ ≤ C max 1 ≤ j ≤ N max [ ∥ W { e ^ ∙ j ( − r ) − e 0 ∙ j } ∥ 2 , P ∥ W { η ^ ∙ j ( − r ) − η 0 ∙ j } ∥ 2 , P + ∥ W { s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) } ∥ 2 , P ∥ W { r ^ ∙ j ( 3 ) , ( − r ) − r 0 ∙ j ( 3 ) } ∥ 2 , P ] . \left|\mathbb{E}\widehat\psi_{\tau,r}-\mathbb{E}(W)\tau_{0}\right|
\le C\max_{1\le j\le N_{\max}}\left[\left\|\sqrt{W}\{\widehat e_{\bullet j}^{(-r)}-e_{0\bullet j}\}\right\|_{2,P}\left\|\sqrt{W}\{\widehat\eta_{\bullet j}^{(-r)}-\eta_{0\bullet j}\}\right\|_{2,P}
+\left\|\sqrt{W}\left\{\frac{\widehat s_{-j}^{(-r)}(0\mid\mathbf{M}_{\bullet(-j)},C,N)}{\widehat s_{-j}^{(-r)}(1\mid\mathbf{M}_{\bullet(-j)},C,N)}-\frac{s_{0,-j}(0\mid\mathbf{M}_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid\mathbf{M}_{\bullet(-j)},C,N)}\right\}\right\|_{2,P}\left\|\sqrt{W}\{\widehat r_{\bullet j}^{(3),(-r)}-r_{0\bullet j}^{(3)}\}\right\|_{2,P}\right]. E ψ τ , r − E ( W ) τ 0 ≤ C 1 ≤ j ≤ N m a x max W { e ∙ j ( − r ) − e 0 ∙ j } 2 , P W { η ∙ j ( − r ) − η 0 ∙ j } 2 , P + W { s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) − s 0 , − j ( 1 ∣ M ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ M ∙ ( − j ) , C , N ) } 2 , P W { r ∙ j ( 3 ) , ( − r ) − r 0 ∙ j ( 3 ) } 2 , P .
Condition (35) makes this o p ( K − 1 / 2 ) o_{p}(K^{-1/2}) o p ( K − 1/2 ) uniformly over folds. Thus Lemma 2 proves (37) and the variance result. If all nuisance limits are correct, Lemma 1 identifies the limiting influence functions with the canonical gradients.
Proof of Theorem 6
Proof. Fix r r r and condition on the training observations. The treatment randomization and iterated expectation in (24) imply
E { ψ ^ θ , r ( O ) } = E [ W N ∑ j = 1 N { E [ d ^ ( − r ) ( a , Z , M , C , N ) π a ′ s ^ ( − r ) ( a ′ ∣ M , C , N ) π a s ^ ( − r ) ( a ∣ M , C , N ) { η 0 ∙ j − η ^ ∙ j ( − r ) } ( a , Z , M , C , N ) ∣ A = a , C , N ] + E { u ^ ∙ j ( 1 ) , ( − r ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } + E { u ^ ∙ j ( 2 ) , ( − r ) ( a , M , C , N ) ∣ A = a ′ , C , N } − u ^ ∙ j ( 3 ) , ( − r ) ( a , a ′ , C , N ) } ] . (76) \mathbb{E}\{\widehat{\psi}_{\theta,r}(O)\}=\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{\mathbb{E}\left[\widehat{d}^{(-r)}(a,Z,M,C,N)\frac{\pi_{a'}\widehat{s}^{(-r)}(a' \mid M,C,N)}{\pi_{a}\widehat{s}^{(-r)}(a \mid M,C,N)}\{\eta_{0\bullet j}-\widehat{\eta}^{(-r)}_{\bullet j}\}(a,Z,M,C,N)\mid A=a,C,N\right]\right.\right.
\\
\left.\left.+\mathbb{E}\{\widehat{u}^{(1),(-r)}_{\bullet j}(a,a',Z,C,N)\mid A=a,C,N\}
+\mathbb{E}\{\widehat{u}^{(2),(-r)}_{\bullet j}(a,M,C,N)\mid A=a',C,N\}
-\widehat{u}^{(3),(-r)}_{\bullet j}(a,a',C,N)\right\}\right]. \tag*{(76)} E { ψ θ , r ( O )} = E [ N W j = 1 ∑ N { E [ d ( − r ) ( a , Z , M , C , N ) π a s ( − r ) ( a ∣ M , C , N ) π a ′ s ( − r ) ( a ′ ∣ M , C , N ) { η 0 ∙ j − η ∙ j ( − r ) } ( a , Z , M , C , N ) ∣ A = a , C , N ] + E { u ∙ j ( 1 ) , ( − r ) ( a , a ′ , Z , C , N ) ∣ A = a , C , N } + E { u ∙ j ( 2 ) , ( − r ) ( a , M , C , N ) ∣ A = a ′ , C , N } − u ∙ j ( 3 ) , ( − r ) ( a , a ′ , C , N ) } ] . (76)
Likewise, (25) gives
E { ψ ^ τ , r ( O ) } = E [ W N ∑ j = 1 N { E [ e ^ ∙ j ( − r ) ( Z , M , C , N ) π s ^ − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) 1 − π s ^ − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) { η 0 ∙ j − η ^ ∙ j ( − r ) } ( 1 , Z , M , C , N ) ∣ A = 1 , C , N ] + E { r ^ ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , Z , C , N ) ∣ A = 1 , C , N } + E { r ^ ∙ j ( 2 ) , ( − r ) ( 1 , 0 , M ∙ j , C , N ) ∣ A = 1 , C , N } + E { r ^ ∙ j ( 3 ) , ( − r ) ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 0 , C , N } − 2 r ^ ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , C , N ) } ] . (77) \mathbb{E}\{\widehat{\psi}_{\tau,r}(O)\}=\mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{\mathbb{E}\left[\widehat{e}^{(-r)}_{\bullet j}(Z,M,C,N)\frac{\pi\,\widehat{s}^{(-r)}_{-j}(0\mid M_{\bullet(-j)},C,N)}{1-\pi\,\widehat{s}^{(-r)}_{-j}(1\mid M_{\bullet(-j)},C,N)}\{\eta_{0\bullet j}-\widehat{\eta}^{(-r)}_{\bullet j}\}(1,Z,M,C,N)\mid A=1,C,N\right]\right.\right.
\\
\left.\left.+\mathbb{E}\{\widehat{r}^{(1),(-r)}_{\bullet j}(1,1,0,Z,C,N)\mid A=1,C,N\}
+\mathbb{E}\{\widehat{r}^{(2),(-r)}_{\bullet j}(1,0,M_{\bullet j},C,N)\mid A=1,C,N\}
+\mathbb{E}\{\widehat{r}^{(3),(-r)}_{\bullet j}(1,1,M_{\bullet(-j)},C,N)\mid A=0,C,N\}
-2\widehat{r}^{(4),(-r)}_{\bullet j}(1,1,0,C,N)\right\}\right]. \tag*{(77)} E { ψ τ , r ( O )} = E [ N W j = 1 ∑ N { E [ e ∙ j ( − r ) ( Z , M , C , N ) 1 − π s − j ( − r ) ( 1 ∣ M ∙ ( − j ) , C , N ) π s − j ( − r ) ( 0 ∣ M ∙ ( − j ) , C , N ) { η 0 ∙ j − η ∙ j ( − r ) } ( 1 , Z , M , C , N ) ∣ A = 1 , C , N ] + E { r ∙ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , Z , C , N ) ∣ A = 1 , C , N } + E { r ∙ j ( 2 ) , ( − r ) ( 1 , 0 , M ∙ j , C , N ) ∣ A = 1 , C , N } + E { r ∙ j ( 3 ) , ( − r ) ( 1 , 1 , M ∙ ( − j ) , C , N ) ∣ A = 0 , C , N } − 2 r ∙ j ( 4 ) , ( − r ) ( 1 , 1 , 0 , C , N ) } ] . (77)
The identities in Proposition 1 yield
κ 0 ( m ∣ a , C , N ) π a π a ′ s 0 ( a ′ ∣ m , C , N ) s 0 ( a ∣ m , C , N ) = κ 0 ( m ∣ a ′ , C , N ) , κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) π 1 − π s 0 , − j ( 0 ∣ m ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ m ∙ ( − j ) , C , N ) = κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) . (78) \kappa_{0}(m\mid a,C,N)\frac{\pi_{a}}{\pi_{a'}}\frac{s_{0}(a'\mid m,C,N)}{s_{0}(a\mid m,C,N)}
=\kappa_{0}(m\mid a',C,N),
\qquad
\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid1,C,N)\frac{\pi}{1-\pi}\frac{s_{0,-j}(0\mid m_{\bullet(-j)},C,N)}{s_{0,-j}(1\mid m_{\bullet(-j)},C,N)}
=\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N). \tag*{(78)} κ 0 ( m ∣ a , C , N ) π a ′ π a s 0 ( a ∣ m , C , N ) s 0 ( a ′ ∣ m , C , N ) = κ 0 ( m ∣ a ′ , C , N ) , κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) 1 − π π s 0 , − j ( 1 ∣ m ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ m ∙ ( − j ) , C , N ) = κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) . (78)
Conditioning on ( A , M , C , N ) (A,M,C,N) ( A , M , C , N ) , cancellation of ξ 0 \xi_{0} ξ 0 in the definition of d 0 d_{0} d 0 gives
E [ d 0 ( a , Z , M , C , N ) { η 0 ∙ j − η ^ ∙ j ( − r ) } ( a , Z , M , C , N ) ∣ A = a , M = m , C , N ] = u 0 ∙ j ( 2 ) ( a , m , C , N ) − ∫ η ^ ∙ j ( − r ) ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) d z . (79) \mathbb{E}\left[d_{0}(a,Z,M,C,N)\{\eta_{0\bullet j}-\widehat{\eta}^{(-r)}_{\bullet j}\}(a,Z,M,C,N)\mid A=a,M=m,C,N\right]
=u^{(2)}_{0\bullet j}(a,m,C,N)-\int\widehat{\eta}^{(-r)}_{\bullet j}(a,z,m,C,N)\phi_{0}(z\mid a,C,N)\,dz. \tag*{(79)} E [ d 0 ( a , Z , M , C , N ) { η 0 ∙ j − η ∙ j ( − r ) } ( a , Z , M , C , N ) ∣ A = a , M = m , C , N ] = u 0 ∙ j ( 2 ) ( a , m , C , N ) − ∫ η ∙ j ( − r ) ( a , z , m , C , N ) ϕ 0 ( z ∣ a , C , N ) d z . (79)
Conditioning on ( A , M ∙ ( − j ) , C , N ) (A,M_{\bullet(-j)},C,N) ( A , M ∙ ( − j ) , C , N ) and using (14) similarly gives
E [ e 0 ∙ j ( Z , M , C , N ) { η 0 ∙ j − η ^ ∙ j ( − r ) } ( 1 , Z , M , C , N ) ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C , N ] = r 0 ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , C , N ) − ∬ η ^ ∙ j ( − r ) ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j . (80) \mathbb{E}\left[e_{0\bullet j}(Z,M,C,N)\{\eta_{0\bullet j}-\widehat{\eta}^{(-r)}_{\bullet j}\}(1,Z,M,C,N)\mid A=1,M_{\bullet(-j)}=m_{\bullet(-j)},C,N\right]
=r^{(3)}_{0\bullet j}(1,1,m_{\bullet(-j)},C,N)-\iint\widehat{\eta}^{(-r)}_{\bullet j}(1,z,m,C,N)\phi_{0}(z\mid1,C,N)\kappa_{0\bullet j}(m_{\bullet j}\mid1,C,N)\,dz\,dm_{\bullet j}. \tag*{(80)} E [ e 0 ∙ j ( Z , M , C , N ) { η 0 ∙ j − η ∙ j ( − r ) } ( 1 , Z , M , C , N ) ∣ A = 1 , M ∙ ( − j ) = m ∙ ( − j ) , C , N ] = r 0 ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , C , N ) − ∬ η ∙ j ( − r ) ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∙ j ( m ∙ j ∣ 1 , C , N ) d z d m ∙ j . (80)
To verify (38) , expand d ^ ( − r ) = ( d ^ ( − r ) − d 0 ) + d 0 \widehat{d}^{(-r)}=(\widehat{d}^{(-r)}-d_{0})+d_{0} d ( − r ) = ( d ( − r ) − d 0 ) + d 0 in (76) , apply (79) , and express the expectation given A = a ′ A=a' A = a ′ by integration against κ 0 ( m ∣ a ′ , C , N ) \kappa_{0}(m\mid a',C,N) κ 0 ( m ∣ a ′ , C , N ) . Add and subtract u ^ ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) \widehat{u}^{(2),(-r)}_{\bullet j}(a,m,C,N) u ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) on the right-hand side of (79) , and then add and subtract the integral with the true classifier fraction in (80). (78) gives
∫ κ 0 ( m ∣ a , C , N ) π a π a ′ s 0 ( a ′ ∣ m , C , N ) s 0 ( a ∣ m , C , N ) { u ^ 0 ∙ j ( 2 ) − u ^ 0 ∙ j ( 2 ) , ( − r ) } ( a , m , C , N ) d m + ∫ κ 0 ( m ∣ a ′ , C , N ) u ^ 0 ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) d m = ∫ κ 0 ( m ∣ a ′ , C , N ) u 0 ∙ j ( 2 ) ( a , m , C , N ) d m . (81) \begin{aligned}
&\int\kappa_0(m \mid a,C,N)\frac{\pi_a}{\pi_{a'}}\frac{s_0(a' \mid m,C,N)}{s_0(a \mid m,C,N)}
\left\{\widehat{u}_{0\bullet j}^{(2)}-\widehat{u}_{0\bullet j}^{(2),(-r)}\right\}(a,m,C,N)\,dm\\
&\quad+\int\kappa_0(m \mid a',C,N)\widehat{u}_{0\bullet j}^{(2),(-r)}(a,m,C,N)\,dm\\
&=\int\kappa_0(m \mid a',C,N)u_{0\bullet j}^{(2)}(a,m,C,N)\,dm. \tag*{(81)}
\end{aligned} ∫ κ 0 ( m ∣ a , C , N ) π a ′ π a s 0 ( a ∣ m , C , N ) s 0 ( a ′ ∣ m , C , N ) { u 0 ∙ j ( 2 ) − u 0 ∙ j ( 2 ) , ( − r ) } ( a , m , C , N ) d m + ∫ κ 0 ( m ∣ a ′ , C , N ) u 0 ∙ j ( 2 ) , ( − r ) ( a , m , C , N ) d m = ∫ κ 0 ( m ∣ a ′ , C , N ) u 0 ∙ j ( 2 ) ( a , m , C , N ) d m . (81)
By Theorem 1, the expectation of the right-hand side of (81) , after multiplication by W / N W/N W / N and summation over j j j , is ν 0 θ 0 ( a , a ′ ) \nu_0\theta_0(a,a') ν 0 θ 0 ( a , a ′ ) . Substitution of (78) and (80) in (75) proves (38).
For (39), expand e ^ ∙ j ( − r ) = ( e ^ ∙ j ( − r ) − e 0 ∙ j ) + e 0 ∙ j \widehat{e}_{\bullet j}^{(-r)}=(\widehat{e}_{\bullet j}^{(-r)}-e_{0\bullet j})+e_{0\bullet j} e ∙ j ( − r ) = ( e ∙ j ( − r ) − e 0 ∙ j ) + e 0 ∙ j in (76), and apply (79) and (77). Split the right-hand side of (79) by adding and subtracting r ^ ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) \widehat{r}_{\bullet j}^{(3),(-r)}(1,1,m_{\bullet(-j)},C,N) r ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) . Split the fitted classifier ratio multiplying r 0 ∙ j ( 3 ) − r ^ ∙ j ( 3 ) , ( − r ) r_{0\bullet j}^{(3)}-\widehat{r}_{\bullet j}^{(3),(-r)} r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) , ( − r ) into its true value and its difference from the true value. Then
∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) π 1 − π s 0 , − j ( 0 ∣ m ∙ ( − j ) , C , N ) s 0 , − j ( 1 ∣ m ∙ ( − j ) , C , N ) { r 0 ∙ j ( 3 ) − r ^ ∙ j ( 3 ) , ( − r ) } ( 1 , 1 , m ∙ ( − j ) , C , N ) d m ∙ ( − j ) + ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) r ^ ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) d m ∙ ( − j ) = ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) r 0 ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , C , N ) d m ∙ ( − j ) . (82) \begin{aligned}
&\int\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid1,C,N)
\frac{\pi}{1-\pi}
\frac{s_{0,-j}(0\mid m_{\bullet(-j)},C,N)}
{s_{0,-j}(1\mid m_{\bullet(-j)},C,N)}
\left\{r_{0\bullet j}^{(3)}-\widehat{r}_{\bullet j}^{(3),(-r)}\right\}
(1,1,m_{\bullet(-j)},C,N)\,dm_{\bullet(-j)}\\
&\quad+\int\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)
\widehat{r}_{\bullet j}^{(3),(-r)}(1,1,m_{\bullet(-j)},C,N)\,dm_{\bullet(-j)}\\
&=\int\kappa_{0\bullet(-j)}(m_{\bullet(-j)}\mid0,C,N)
r_{0\bullet j}^{(3)}(1,1,m_{\bullet(-j)},C,N)\,dm_{\bullet(-j)}. \tag*{(82)}
\end{aligned} ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 1 , C , N ) 1 − π π s 0 , − j ( 1 ∣ m ∙ ( − j ) , C , N ) s 0 , − j ( 0 ∣ m ∙ ( − j ) , C , N ) { r 0 ∙ j ( 3 ) − r ∙ j ( 3 ) , ( − r ) } ( 1 , 1 , m ∙ ( − j ) , C , N ) d m ∙ ( − j ) + ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) r ∙ j ( 3 ) , ( − r ) ( 1 , 1 , m ∙ ( − j ) , C , N ) d m ∙ ( − j ) = ∫ κ 0 ∙ ( − j ) ( m ∙ ( − j ) ∣ 0 , C , N ) r 0 ∙ j ( 3 ) ( 1 , 1 , m ∙ ( − j ) , C , N ) d m ∙ ( − j ) . (82)
Theorem 1 identifies the expectation of ( W / N ) ∑ j = 1 N (W/N)\sum_{j=1}^{N} ( W / N ) ∑ j = 1 N applied to the right-hand side of (82) with ν 0 τ 0 \nu_0\tau_0 ν 0 τ 0 . Substitution in (76) proves (39).
Proof of Theorem 7
Proof. First consider θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) under (32). Recall that the argument establishing (72) in Section G.5 uses the component convergence, the bound B B B , and the second moments in Assumption (A4). In particular, the terms multiplying Y ∙ j Y_{\bullet j} Y ∙ j are uniformly bounded and converge in L 2 ( P ) L^2(P) L 2 ( P ) ; truncating Y ∙ j 2 Y_{\bullet j}^2 Y ∙ j 2 as in (61) gives convergence of their products with Y ∙ j Y_{\bullet j} Y ∙ j in L 2 ( P ) L^2(P) L 2 ( P ) . The other terms in (24) are bounded products of the components in (32). Thus (72) holds for the fitted functions considered here. The same reasoning applied to (34) gives (74). The limiting scores are square-integrable by the same bounds.
Since h θ ⋆ h_\theta^\star h θ ⋆ and h τ ⋆ h_\tau^\star h τ ⋆ retain the integral representations specified in Assumption D, Theorem 4, with (28) or (29), gives
E { ψ θ , l d ⋆ ( O ) } = ν 0 θ 0 ( a , a ′ ) , E { ψ τ , l d ⋆ ( O ) } = ν 0 τ 0 , (83) \mathbb{E}\{\psi_{\theta,\mathrm{ld}}^\star(O)\}=\nu_0\theta_0(a,a'),\qquad
\mathbb{E}\{\psi_{\tau,\mathrm{ld}}^\star(O)\}=\nu_0\tau_0, \tag*{(83)} E { ψ θ , ld ⋆ ( O )} = ν 0 θ 0 ( a , a ′ ) , E { ψ τ , ld ⋆ ( O )} = ν 0 τ 0 , (83)
under the respective conditions. The Cauchy–Schwarz inequality and (72) or (74) therefore give
max 1 ≤ r ≤ R ∣ E ψ ^ θ , r − ν 0 θ 0 ( a , a ′ ) ∣ = o p ( 1 ) , max 1 ≤ r ≤ R ∣ E ψ ^ τ , r − ν 0 τ 0 ∣ = o p ( 1 ) . (84) \max_{1\le r\le R}\left|\mathbb{E}\widehat{\psi}_{\theta,r}-\nu_0\theta_0(a,a')\right|=o_p(1),\qquad
\max_{1\le r\le R}\left|\mathbb{E}\widehat{\psi}_{\tau,r}-\nu_0\tau_0\right|=o_p(1). \tag*{(84)} 1 ≤ r ≤ R max E ψ θ , r − ν 0 θ 0 ( a , a ′ ) = o p ( 1 ) , 1 ≤ r ≤ R max E ψ τ , r − ν 0 τ 0 = o p ( 1 ) . (84)
Apply (57) separately with ψ ⋆ = ψ θ , l d ⋆ \psi^\star=\psi_{\theta,\mathrm{ld}}^\star ψ ⋆ = ψ θ , ld ⋆ and ψ ⋆ = ψ τ , l d ⋆ \psi^\star=\psi_{\tau,\mathrm{ld}}^\star ψ ⋆ = ψ τ , ld ⋆ . The conditional variance argument for (58) gives R 1 K = o p ( K − 1 / 2 ) R_{1K}=o_p(K^{-1/2}) R 1 K = o p ( K − 1/2 ) , and (83) gives R 2 K = o p ( 1 ) R_{2K}=o_p(1) R 2 K = o p ( 1 ) . The distribution of large numbers and P K ( W ) → p ν 0 > 0 P_K(W)\to_p\nu_0>0 P K ( W ) → p ν 0 > 0 then prove the stated consistency assertions.
The Cauchy–Schwarz bounds in Section G.5 apply to the products in (38) and (39). More precisely, the absolute value of the first expectation on the right-hand side of (38), maximized over r r r , is bounded by N max ( 1 + B ) / π a ′ N_{\max}(1+B)/\pi_{a'} N m a x ( 1 + B ) / π a ′ times the left-hand side of (33). For (39), the corresponding bound is N max ( 1 + B ) / ( 1 − π ) N_{\max}(1+B)/(1-\pi) N m a x ( 1 + B ) / ( 1 − π ) times the left-hand side of (35). Under (32) and (28), each product in (33) has one factor converging to zero and another bounded in probability. The analogous assertion follows from (34) and (29) for (35). Consequently, these bounds are o p ( 1 ) o_p(1) o p ( 1 ) under the respective consistency conditions. Subtracting the first expectations in (38) and (39) from (84) proves (40) and (42) .
If (33) holds and the left-hand side of (40) is o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) , (38) and the same bounds imply
max 1 ≤ r ≤ R ∣ E { ψ ^ θ , r − ψ θ , l d ∗ } ∣ = o p ( K − 1 / 2 ) . (85) \max_{1\le r\le R}\left|\mathbb{E}\{\widehat{\psi}_{\theta,r}-\psi_{\theta,ld}^{*}\}\right|=o_p(K^{-1/2}). \tag*{(85)} 1 ≤ r ≤ R max E { ψ θ , r − ψ θ , l d ∗ } = o p ( K − 1/2 ) . (85)
Likewise, (35) and the o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) version of (42) imply
max 1 ≤ r ≤ R ∣ E { ψ ^ τ , r − ψ τ , l d ∗ } ∣ = o p ( K − 1 / 2 ) . (86) \max_{1\le r\le R}\left|\mathbb{E}\{\widehat{\psi}_{\tau,r}-\psi_{\tau,ld}^{*}\}\right|=o_p(K^{-1/2}). \tag*{(86)} 1 ≤ r ≤ R max E { ψ τ , r − ψ τ , l d ∗ } = o p ( K − 1/2 ) . (86)
Equations (73) , (83) , and (85) verify (54) –(56) for θ ^ l d ( a , a ′ ) \widehat{\theta}_{ld}(a,a') θ l d ( a , a ′ ) . Equations (75) , (83) , and (86) verify them for τ ^ l d \widehat{\tau}_{ld} τ l d . Lemma 2 proves (36) , (37) , (41) , and (43) under the respective conditions. Slutsky’s theorem proves (44) . Lemma 1 identifies the variances at the true nuisance limits with those in Theorem 2.
We next verify the sufficient conditions in (46) and (49) . For fixed r r r , substitution of (45) into the expectation inside the absolute value in (40) gives
E [ W N ∑ j = 1 N { ϕ 0 − ϕ ^ ( − r ) } ( z ∣ a , C , N ) × [ u ^ ∙ j ( 1 ) , ( − r ) ( a , a ′ , z , C , N ) − ∫ κ 0 ( m ∣ a , C , N ) π a π a ′ s ^ ( − r ) ( a ′ ∣ m , C , N ) s ^ ( − r ) ( a ∣ m , C , N ) η ^ ∙ j ( − r ) ( a , z , m , C , N ) d m ] d z ] . (87) \mathbb{E}\left[\frac{W}{N}\sum_{j=1}^{N}\left\{\phi_{0}-\widehat{\phi}^{(-r)}\right\}(z\mid a,C,N)
\times\left[\widehat{u}_{\bullet j}^{(1),(-r)}(a,a',z,C,N)-\int\kappa_{0}(m\mid a,C,N)\frac{\pi_{a}}{\pi_{a'}}\frac{\widehat{s}^{(-r)}(a'\mid m,C,N)}{\widehat{s}^{(-r)}(a\mid m,C,N)}\widehat{\eta}_{\bullet j}^{(-r)}(a,z,m,C,N)\,dm\right]\,dz\right]. \tag*{(87)} E [ N W j = 1 ∑ N { ϕ 0 − ϕ ( − r ) } ( z ∣ a , C , N ) × [ u ∙ j ( 1 ) , ( − r ) ( a , a ′ , z , C , N ) − ∫ κ 0 ( m ∣ a , C , N ) π a ′ π a s ( − r ) ( a ∣ m , C , N ) s ( − r ) ( a ′ ∣ m , C , N ) η ∙ j ( − r ) ( a , z , m , C , N ) d m ] d z ] . (87)
By (78) and (17) , one has
u ^ ∙ j ( 1 ) , ( − r ) ( a , a ′ , z , C , N ) − ∫ κ 0 ( m ∣ a , C , N ) π a π a ′ s ^ ( − r ) ( a ′ ∣ m , C , N ) s ^ ( − r ) ( a ∣ m , C , N ) η ^ ∙ j ( − r ) ( a , z , m , C , N ) d m = { u ^ ∙ j ( 1 ) , ( − r ) − u 0 ∙ j ( 1 ) } ( a , a ′ , z , C , N ) − ∫ κ 0 ( m ∣ a , C , N ) π a π a ′ s ^ ( − r ) ( a ′ ∣ m , C , N ) s ^ ( − r ) ( a ∣ m , C , N ) { η ^ ∙ j ( − r ) − η 0 ∙ j } ( a , z , m , C , N ) d m − ∫ κ 0 ( m ∣ a , C , N ) π a π a ′ { s ^ ( − r ) ( a ′ ∣ m , C , N ) s ^ ( − r ) ( a ∣ m , C , N ) − s 0 ( a ′ ∣ m , C , N ) s 0 ( a ∣ m , C , N ) } η 0 ∙ j ( a , z , m , C , N ) d m . (88) \begin{aligned}
\widehat{u}_{\bullet j}^{(1),(-r)}(a,a',z,C,N)&-\int\kappa_{0}(m\mid a,C,N)\frac{\pi_{a}}{\pi_{a'}}\frac{\widehat{s}^{(-r)}(a'\mid m,C,N)}{\widehat{s}^{(-r)}(a\mid m,C,N)}\widehat{\eta}_{\bullet j}^{(-r)}(a,z,m,C,N)\,dm\\
&=\{\widehat{u}_{\bullet j}^{(1),(-r)}-u_{0\bullet j}^{(1)}\}(a,a',z,C,N)\\
&\quad-\int\kappa_{0}(m\mid a,C,N)\frac{\pi_{a}}{\pi_{a'}}\frac{\widehat{s}^{(-r)}(a'\mid m,C,N)}{\widehat{s}^{(-r)}(a\mid m,C,N)}\{\widehat{\eta}_{\bullet j}^{(-r)}-\eta_{0\bullet j}\}(a,z,m,C,N)\,dm\\
&\quad-\int\kappa_{0}(m\mid a,C,N)\frac{\pi_{a}}{\pi_{a'}}\left\{\frac{\widehat{s}^{(-r)}(a'\mid m,C,N)}{\widehat{s}^{(-r)}(a\mid m,C,N)}-\frac{s_{0}(a'\mid m,C,N)}{s_{0}(a\mid m,C,N)}\right\}\eta_{0\bullet j}(a,z,m,C,N)\,dm. \tag*{(88)}
\end{aligned} u ∙ j ( 1 ) , ( − r ) ( a , a ′ , z , C , N ) − ∫ κ 0 ( m ∣ a , C , N ) π a ′ π a s ( − r ) ( a ∣ m , C , N ) s ( − r ) ( a ′ ∣ m , C , N ) η ∙ j ( − r ) ( a , z , m , C , N ) d m = { u ∙ j ( 1 ) , ( − r ) − u 0 ∙ j ( 1 ) } ( a , a ′ , z , C , N ) − ∫ κ 0 ( m ∣ a , C , N ) π a ′ π a s ( − r ) ( a ∣ m , C , N ) s ( − r ) ( a ′ ∣ m , C , N ) { η ∙ j ( − r ) − η 0 ∙ j } ( a , z , m , C , N ) d m − ∫ κ 0 ( m ∣ a , C , N ) π a ′ π a { s ( − r ) ( a ∣ m , C , N ) s ( − r ) ( a ′ ∣ m , C , N ) − s 0 ( a ∣ m , C , N ) s 0 ( a ′ ∣ m , C , N ) } η 0 ∙ j ( a , z , m , C , N ) d m . (88)
The definition of d 0 d_{0} d 0 in (12) yields
ϕ 0 ( z ∣ a , c , n ) κ 0 ( m ∣ a , c , n ) = d 0 ( a , z , m , c , n ) ϕ 0 ( z ∣ a , c , n ) ξ 0 ( m ∣ a , z , c , n ) . (89) \phi_{0}(z\mid a,c,n)\kappa_{0}(m\mid a,c,n)=d_{0}(a,z,m,c,n)\phi_{0}(z\mid a,c,n)\xi_{0}(m\mid a,z,c,n). \tag*{(89)} ϕ 0 ( z ∣ a , c , n ) κ 0 ( m ∣ a , c , n ) = d 0 ( a , z , m , c , n ) ϕ 0 ( z ∣ a , c , n ) ξ 0 ( m ∣ a , z , c , n ) . (89)
Write ϕ 0 − ϕ ^ ( − r ) \phi_{0}-\widehat{\phi}^{(-r)} ϕ 0 − ϕ ( − r ) as ϕ 0 ( 1 − ϕ ^ ( − r ) / ϕ 0 ) \phi_{0}(1-\widehat{\phi}^{(-r)}/\phi_{0}) ϕ 0 ( 1 − ϕ ( − r ) / ϕ 0 ) in (87) . Apply the Cauchy–Schwarz inequality to the product with u ^ ∙ j ( 1 ) , ( − r ) − u 0 ∙ j ( 1 ) \widehat{u}_{\bullet j}^{(1),(-r)}-u_{0\bullet j}^{(1)} u ∙ j ( 1 ) , ( − r ) − u 0 ∙ j ( 1 ) under ϕ 0 ( z ∣ a , C , N ) \phi_{0}(z\mid a,C,N) ϕ 0 ( z ∣ a , C , N ) . For the integrals containing η ^ ∙ j ( − r ) − η 0 ∙ j \widehat{\eta}_{\bullet j}^{(-r)}-\eta_{0\bullet j} η ∙ j ( − r ) − η 0 ∙ j and the difference of the fractions in (88) , (89) changes the integration to the distribution of ( Z , M ) (Z,M) ( Z , M ) given ( A = a , C , N ) (A=a,C,N) ( A = a , C , N ) with multiplier d 0 d_{0} d 0 . Bound d 0 d_{0} d 0 , η 0 \eta_{0} η 0 , and the fitted classifier odds by B B B , and use treatment randomization to bound the conditional integrals by 1 / π a 1/\pi_{a} 1/ π a times the corresponding integrals under P P P . Since W W W is a function of ( C , N ) (C,N) ( C , N ) , it supplies a factor W \sqrt{W} W to each norm. Summation over j j j contributes at most N max N_{\max} N m a x . It follows that the left-hand side of (40) is bounded by N max ( 1 + B ) 3 / ( π a π a ′ ) N_{\max}(1+B)^{3}/(\pi_{a}\pi_{a'}) N m a x ( 1 + B ) 3 / ( π a π a ′ ) times the left-hand side of (46) .
Substitution of (48) into the expectation inside the absolute value in (42) gives
E [ W N ∑ j = 1 N { ∫ { ϕ 0 − ϕ ^ ( − r ) } ( z ∣ 1 , C , N ) [ r ^ ∘ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , z , C , N ) − ∫ η ^ ∘ j ( − r ) ( 1 , z , m , C , N ) κ 0 ∘ j ( m ∘ j ∣ 1 , C , N ) κ 0 ∘ ( − j ) ( m ∘ ( − j ) ∣ 1 , C , N ) π 1 − π s ^ − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) d m d z ] + ∫ { κ 0 ∘ j − κ ^ ∘ j ( − r ) } ( m ∘ j ∣ 1 , C , N ) [ r ^ ∘ j ( 2 ) , ( − r ) ( 1 , 0 , m ∘ j , C , N ) − ∬ η ^ ∘ j ( − r ) ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∘ ( − j ) ( m ∘ ( − j ) ∣ 1 , C , N ) π 1 − π s ^ − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) d z d m ∘ ( − j ) ] d m ∘ j + ∬ η ^ ∘ j ( − r ) ( 1 , z , m , C , N ) { ϕ ^ ( − r ) − ϕ 0 } ( z ∣ 1 , C , N ) { κ ^ ∘ j ( − r ) − κ 0 ∘ j } ( m ∘ j ∣ 1 , C , N ) × κ 0 ∘ ( − j ) ( m ∘ ( − j ) ∣ 1 , C , N ) π 1 − π s ^ − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) 1 − π s ^ − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) d z d m } ] . (90) \begin{aligned}
\mathbb{E}[\frac{W}{N}\sum_{j=1}^{N}\{\int\{\phi_0-\widehat{\phi}^{(-r)}\}(z\mid1,C,N)[\widehat{r}^{(1),(-r)}_{\circ j}(1,1,0,z,C,N)-\int\widehat{\eta}^{(-r)}_{\circ j}(1,z,m,C,N)\kappa_{0\circ j}(m_{\circ j}\mid1,C,N)\kappa_{0\circ(-j)}(m_{\circ(-j)}\mid1,C,N)\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid m_{\circ(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid m_{\circ(-j)},C,N)}\,dm\,dz] \\
&+\int\{\kappa_{0\circ j}-\widehat{\kappa}^{(-r)}_{\circ j}\}(m_{\circ j}\mid1,C,N)[\widehat{r}^{(2),(-r)}_{\circ j}(1,0,m_{\circ j},C,N)-\iint\widehat{\eta}^{(-r)}_{\circ j}(1,z,m,C,N)\phi_0(z\mid1,C,N)\kappa_{0\circ(-j)}(m_{\circ(-j)}\mid1,C,N)
\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid m_{\circ(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid m_{\circ(-j)},C,N)}\,dz\,dm_{\circ(-j)}]dm_{\circ j} \\
&+\iint\widehat{\eta}^{(-r)}_{\circ j}(1,z,m,C,N)\{\widehat{\phi}^{(-r)}-\phi_0\}(z\mid1,C,N)\{\widehat{\kappa}^{(-r)}_{\circ j}-\kappa_{0\circ j}\}(m_{\circ j}\mid1,C,N)
\times\kappa_{0\circ(-j)}(m_{\circ(-j)}\mid1,C,N)\frac{\pi}{1-\pi}\frac{\widehat{s}^{(-r)}_{-j}(0\mid m_{\circ(-j)},C,N)}{1-\pi\widehat{s}^{(-r)}_{-j}(1\mid m_{\circ(-j)},C,N)}\,dz\,dm\}]. \tag*{(90)}
\end{aligned} E [ N W j = 1 ∑ N { ∫ { ϕ 0 − ϕ ( − r ) } ( z ∣ 1 , C , N ) [ r ∘ j ( 1 ) , ( − r ) ( 1 , 1 , 0 , z , C , N ) − ∫ η ∘ j ( − r ) ( 1 , z , m , C , N ) κ 0 ∘ j ( m ∘ j ∣ 1 , C , N ) κ 0 ∘ ( − j ) ( m ∘ ( − j ) ∣ 1 , C , N ) 1 − π π s − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) d m d z ] + ∫ { κ 0 ∘ j − κ ∘ j ( − r ) } ( m ∘ j ∣ 1 , C , N ) [ r ∘ j ( 2 ) , ( − r ) ( 1 , 0 , m ∘ j , C , N ) − ∬ η ∘ j ( − r ) ( 1 , z , m , C , N ) ϕ 0 ( z ∣ 1 , C , N ) κ 0 ∘ ( − j ) ( m ∘ ( − j ) ∣ 1 , C , N ) 1 − π π s − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) d z d m ∘ ( − j ) ] d m ∘ j + ∬ η ∘ j ( − r ) ( 1 , z , m , C , N ) { ϕ ( − r ) − ϕ 0 } ( z ∣ 1 , C , N ) { κ ∘ j ( − r ) − κ 0 ∘ j } ( m ∘ j ∣ 1 , C , N ) × κ 0 ∘ ( − j ) ( m ∘ ( − j ) ∣ 1 , C , N ) 1 − π π 1 − π s − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) d z d m }] . (90)
Indeed, the contributions of r ^ ( 1 ) , ( − r ) \widehat{r}^{(1),(-r)} r ( 1 ) , ( − r ) and r ^ ( 2 ) , ( − r ) \widehat{r}^{(2),(-r)} r ( 2 ) , ( − r ) follow directly from the equality for 2 r ^ ( 4 ) , ( − r ) 2\widehat{r}^{(4),(-r)} 2 r ( 4 ) , ( − r ) in (48). The equality for r ^ ( 3 ) , ( − r ) \widehat{r}^{(3),(-r)} r ( 3 ) , ( − r ) and the identity
ϕ ^ ( − r ) ( z ∣ 1 , c , n ) κ ^ ∘ j ( − r ) ( m ∘ j ∣ 1 , c , n ) − ϕ 0 ( z ∣ 1 , c , n ) κ 0 ∘ j ( m ∘ j ∣ 1 , c , n ) = { ϕ ^ ( − r ) − ϕ 0 } ( z ∣ 1 , c , n ) κ 0 ∘ j ( m ∘ j ∣ 1 , c , n ) + ϕ 0 ( z ∣ 1 , c , n ) { κ ^ ∘ j ( − r ) − κ 0 ∘ j } ( m ∘ j ∣ 1 , c , n ) + { ϕ ^ ( − r ) − ϕ 0 } ( z ∣ 1 , c , n ) { κ ^ ∘ j ( − r ) − κ 0 ∘ j } ( m ∘ j ∣ 1 , c , n ) (91) \widehat{\phi}^{(-r)}(z\mid1,c,n)\widehat{\kappa}^{(-r)}_{\circ j}(m_{\circ j}\mid1,c,n)-\phi_0(z\mid1,c,n)\kappa_{0\circ j}(m_{\circ j}\mid1,c,n)
=\{\widehat{\phi}^{(-r)}-\phi_0\}(z\mid1,c,n)\kappa_{0\circ j}(m_{\circ j}\mid1,c,n)
+\phi_0(z\mid1,c,n)\{\widehat{\kappa}^{(-r)}_{\circ j}-\kappa_{0\circ j}\}(m_{\circ j}\mid1,c,n)
+\{\widehat{\phi}^{(-r)}-\phi_0\}(z\mid1,c,n)\{\widehat{\kappa}^{(-r)}_{\circ j}-\kappa_{0\circ j}\}(m_{\circ j}\mid1,c,n) \tag*{(91)} ϕ ( − r ) ( z ∣ 1 , c , n ) κ ∘ j ( − r ) ( m ∘ j ∣ 1 , c , n ) − ϕ 0 ( z ∣ 1 , c , n ) κ 0 ∘ j ( m ∘ j ∣ 1 , c , n ) = { ϕ ( − r ) − ϕ 0 } ( z ∣ 1 , c , n ) κ 0 ∘ j ( m ∘ j ∣ 1 , c , n ) + ϕ 0 ( z ∣ 1 , c , n ) { κ ∘ j ( − r ) − κ 0 ∘ j } ( m ∘ j ∣ 1 , c , n ) + { ϕ ( − r ) − ϕ 0 } ( z ∣ 1 , c , n ) { κ ∘ j ( − r ) − κ 0 ∘ j } ( m ∘ j ∣ 1 , c , n ) (91)
give the remaining integrals in (89).
By (77) and (20)–(21), replacing η ^ ( − r ) \widehat{\eta}^{(-r)} η ( − r ) and s ^ − j ( − r ) \widehat{s}^{(-r)}_{-j} s − j ( − r ) by η 0 \eta_0 η 0 and s 0 , − j s_{0,-j} s 0 , − j in the integrals inside the two brackets in (89) gives r 0 ∘ j ( 1 ) ( 1 , 1 , 0 , z , C , N ) r^{(1)}_{0\circ j}(1,1,0,z,C,N) r 0 ∘ j ( 1 ) ( 1 , 1 , 0 , z , C , N ) and r 0 ∘ j ( 2 ) ( 1 , 0 , m ∘ j , C , N ) r^{(2)}_{0\circ j}(1,0,m_{\circ j},C,N) r 0 ∘ j ( 2 ) ( 1 , 0 , m ∘ j , C , N ) , respectively. Adding and subtracting these integrals, and then adding and subtracting the integrals with η 0 \eta_0 η 0 and s ^ − j ( − r ) \widehat{s}^{(-r)}_{-j} s − j ( − r ) , expresses the brackets as sums involving r ^ ( 1 ) , ( − r ) − r 0 ( 1 ) \widehat{r}^{(1),(-r)}-r^{(1)}_0 r ( 1 ) , ( − r ) − r 0 ( 1 ) or r ^ ( 2 ) , ( − r ) − r 0 ( 2 ) \widehat{r}^{(2),(-r)}-r^{(2)}_0 r ( 2 ) , ( − r ) − r 0 ( 2 ) , η ^ ( − r ) − η 0 \widehat{\eta}^{(-r)}-\eta_0 η ( − r ) − η 0 , and
s ^ − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) s ^ − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) − s 0 , − j ( 0 ∣ m ∘ ( − j ) , C , N ) s 0 , − j ( 1 ∣ m ∘ ( − j ) , C , N ) . \frac{\widehat{s}^{(-r)}_{-j}(0\mid m_{\circ(-j)},C,N)}{\widehat{s}^{(-r)}_{-j}(1\mid m_{\circ(-j)},C,N)}
-
\frac{s_{0,-j}(0\mid m_{\circ(-j)},C,N)}{s_{0,-j}(1\mid m_{\circ(-j)},C,N)}. s − j ( − r ) ( 1 ∣ m ∘ ( − j ) , C , N ) s − j ( − r ) ( 0 ∣ m ∘ ( − j ) , C , N ) − s 0 , − j ( 1 ∣ m ∘ ( − j ) , C , N ) s 0 , − j ( 0 ∣ m ∘ ( − j ) , C , N ) .
The support condition in Assumption (A4) for every ( a , a ′ ) (a,a') ( a , a ′ ) gives mutual absolute continuity of κ 0 ( ∙ ∣ 0 , c , n ) \kappa_0(\bullet\mid0,c,n) κ 0 ( ∙ ∣ 0 , c , n ) and κ 0 ( ∙ ∣ 1 , c , n ) \kappa_0(\bullet\mid1,c,n) κ 0 ( ∙ ∣ 1 , c , n ) , and hence of their M ∘ ( − j ) M_{\circ(-j)} M ∘ ( − j ) marginals. Its condition for τ \tau τ therefore gives domination of ϕ 0 ( ∙ ∣ 1 , c , n ) κ 0 ∘ j ( ∙ ∣ 1 , c , n ) κ 0 ∘ ( − j ) ( ∙ ∣ 1 , c , n ) \phi_0(\bullet\mid1,c,n)\kappa_{0\circ j}(\bullet\mid1,c,n)\kappa_{0\circ(-j)}(\bullet\mid1,c,n) ϕ 0 ( ∙ ∣ 1 , c , n ) κ 0 ∘ j ( ∙ ∣ 1 , c , n ) κ 0 ∘ ( − j ) ( ∙ ∣ 1 , c , n ) by the distribution of ( Z , M ) (Z,M) ( Z , M ) given ( A = 1 , C = c , N = n ) (A=1,C=c,N=n) ( A = 1 , C = c , N = n ) . Consequently, the definition of e 0 ∘ j e_{0\circ j} e 0 ∘ j in (14) gives the almost-everywhere identity
ϕ 0 ( z ∣ 1 , c , n ) κ 0 ⋅ j ( m ⋅ j ∣ 1 , c , n ) κ 0 ⋅ ( − j ) ( m ⋅ ( − j ) ∣ 1 , c , n ) = e 0 ⋅ j ( z , m , c , n ) ϕ 0 ( z ∣ 1 , c , n ) ξ 0 ( m ∣ 1 , z , c , n ) . (92) \phi_0(z \mid1,c,n)\kappa_{0\mathbin{\cdot}j}(m_{\mathbin{\cdot}j}\mid1,c,n)\kappa_{0\mathbin{\cdot}(-j)}(m_{\mathbin{\cdot}(-j)}\mid1,c,n)
= e_{0\mathbin{\cdot}j}(z,m,c,n)\phi_0(z\mid1,c,n)\xi_0(m\mid1,z,c,n). \tag*{(92)} ϕ 0 ( z ∣ 1 , c , n ) κ 0 ⋅ j ( m ⋅ j ∣ 1 , c , n ) κ 0 ⋅ ( − j ) ( m ⋅ ( − j ) ∣ 1 , c , n ) = e 0 ⋅ j ( z , m , c , n ) ϕ 0 ( z ∣ 1 , c , n ) ξ 0 ( m ∣ 1 , z , c , n ) . (92)
Use (92) and E { W E ( ⋅ ∣ A = 1 , C , N ) } = E { A W ( ⋅ ) } / π \mathrm{E}\{\mathrm{W E}(\mathbin{\cdot}\mid A=1,C,N)\}=\mathrm{E}\{A\mathrm{W}(\mathbin{\cdot})\}/\pi E { WE ( ⋅ ∣ A = 1 , C , N )} = E { A W ( ⋅ )} / π to apply the Cauchy–Schwarz inequality under P P P . The products with r ^ ( 1 ) , ( − r ) − r 0 ( 1 ) \widehat{r}^{(1),(-r)}-r_0^{(1)} r ( 1 ) , ( − r ) − r 0 ( 1 ) and r ^ ( 2 ) , ( − r ) − r 0 ( 2 ) \widehat{r}^{(2),(-r)}-r_0^{(2)} r ( 2 ) , ( − r ) − r 0 ( 2 ) give the first two products in (49) . The products with η ^ ( − r ) − η 0 \widehat{\eta}^{(-r)}-\eta_0 η ( − r ) − η 0 and the difference of fractions give its product of the two sums of norms. The last integral in (90) gives its last product, using the bounds on η ^ ( − r ) \widehat{\eta}^{(-r)} η ( − r ) , e 0 ⋅ j e_{0\mathbin{\cdot}j} e 0 ⋅ j , and the fitted classifier odds. Thus the left-hand side of (42) is bounded by N max ( 1 + B ) 3 / { π ( 1 − π ) } N_{\max}(1+B)^3/\{\pi(1-\pi)\} N m a x ( 1 + B ) 3 / { π ( 1 − π )} times the left-hand side of (49) . These inequalities prove the stated sufficient conditions at both o p ( 1 ) o_p(1) o p ( 1 ) and o p ( K − 1 / 2 ) o_p(K^{-1/2}) o p ( K − 1/2 ) rates.
Nuisance and auxiliary function estimation
Recall that the canonical nuisance parameters are ( η , ϕ , κ , ξ ) (\eta,\phi,\kappa,\xi) ( η , ϕ , κ , ξ ) . The ratios ω ( 1 ) \omega^{(1)} ω ( 1 ) and ω ⋅ j ( 2 ) \omega_{\mathbin{\cdot}j}^{(2)} ω ⋅ j ( 2 ) and the integrated outcome regressions u ⋅ j ( 1 ) − u ⋅ j ( 3 ) u_{\mathbin{\cdot}j}^{(1)}-u_{\mathbin{\cdot}j}^{(3)} u ⋅ j ( 1 ) − u ⋅ j ( 3 ) and r ⋅ j ( 1 ) − r ⋅ j ( 4 ) r_{\mathbin{\cdot}j}^{(1)}-r_{\mathbin{\cdot}j}^{(4)} r ⋅ j ( 1 ) − r ⋅ j ( 4 ) are auxiliary functions obtained from these parameters, as defined in § 3 . The conditional mean η ⋅ j \eta_{\mathbin{\cdot}j} η ⋅ j can be estimated using a regression method appropriate for Y ⋅ j Y_{\mathbin{\cdot}j} Y ⋅ j , for example, a generalized linear model or a flexible regression method such as Super Learner, with predictors ( A , Z , M , C , N ) (A,Z,M,C,N) ( A , Z , M , C , N ) . Alternatively, within each treatment group A = a A=a A = a , regress Y ⋅ j Y_{\mathbin{\cdot}j} Y ⋅ j on ( Z , M , C , N ) (Z,M,C,N) ( Z , M , C , N ) to estimate η ⋅ j ( a , z , m , c , n ) \eta_{\mathbin{\cdot}j}(a,z,m,c,n) η ⋅ j ( a , z , m , c , n ) separately for a = 0 , 1 a=0,1 a = 0 , 1 .
To estimate κ \kappa κ , first estimate the conditional distribution of each M ⋅ j M_{\mathbin{\cdot}j} M ⋅ j given ( A , C , N ) (A,C,N) ( A , C , N ) . For example, following Cheng and Li [2 ] , specify normal conditional distributions, estimate their means by regression, and estimate their variances from the training residuals. Specify a Gaussian copula for the conditional dependence among M ⋅ 1 , … , M ⋅ N M_{\mathbin{\cdot}1},\ldots,M_{\mathbin{\cdot}N} M ⋅ 1 , … , M ⋅ N . Estimate its correlation parameters by maximizing the copula likelihood with the fitted marginal distribution functions held fixed, as described in their Appendix A.1. Evaluate κ \kappa κ by multiplying the fitted κ ⋅ j \kappa_{\mathbin{\cdot}j} κ ⋅ j over j = 1 , … , N j=1,\ldots,N j = 1 , … , N and multiplying this product by the fitted copula density evaluated at the marginal distribution functions.
To estimate ξ \xi ξ , fit the conditional distributions of M ⋅ j M_{\mathbin{\cdot}j} M ⋅ j given ( A , Z , C , N ) (A,Z,C,N) ( A , Z , C , N ) and a Gaussian copula for their conditional dependence motivated by Cheng and Li [2 ] . Similarly, to estimate ϕ \phi ϕ , fit the conditional distributions of the coordinates of Z Z Z given ( A , C , N ) (A,C,N) ( A , C , N ) and a Gaussian copula for their conditional dependence. For each of ξ \xi ξ and ϕ \phi ϕ , estimate the marginal means, residual variances, and correlation parameters for the copula separately, and calculate the joint density by the same multiplication of marginal densities and the specified copula density.
We obtain κ ⋅ j \kappa_{\mathbin{\cdot}j} κ ⋅ j and κ ⋅ ( − j ) \kappa_{\mathbin{\cdot}(-j)} κ ⋅ ( − j ) used for τ \tau τ by integrating the fitted κ \kappa κ over m ⋅ ( − j ) m_{\mathbin{\cdot}(-j)} m ⋅ ( − j ) and m ⋅ j m_{\mathbin{\cdot}j} m ⋅ j , respectively. We compute the auxiliary functions by substituting the fitted η ⋅ j \eta_{\mathbin{\cdot}j} η ⋅ j , ϕ \phi ϕ , κ \kappa κ , ξ \xi ξ , κ ⋅ j \kappa_{\mathbin{\cdot}j} κ ⋅ j , and κ ⋅ ( − j ) \kappa_{\mathbin{\cdot}(-j)} κ ⋅ ( − j ) into their defining ratios and integrals in § 3 .
For the low-dimensional estimator, we estimate s s s by regressing A A A on ( M , C , N ) (M,C,N) ( M , C , N ) and s − j s_{-j} s − j by regressing A A A on ( M ⋅ ( − j ) , C , N ) (M_{\mathbin{\cdot}(-j)},C,N) ( M ⋅ ( − j ) , C , N ) . We estimate d d d using the augmented method proposed by Rudolph et al. [15 ] , with the two classifiers and their predictors given in Section G.1. The densities κ ⋅ j \kappa_{\mathbin{\cdot}j} κ ⋅ j and κ ⋅ j ∗ \kappa_{\mathbin{\cdot}j}^{*} κ ⋅ j ∗ have the scalar response M ⋅ j M_{\mathbin{\cdot}j} M ⋅ j , conditioned on ( A , C , N ) (A,C,N) ( A , C , N ) and ( M ⋅ ( − j ) , A , C , N ) (M_{\mathbin{\cdot}(-j)},A,C,N) ( M ⋅ ( − j ) , A , C , N ) , respectively; they can be fitted through conditional mean and residual-distribution models or obtained from the same fitted κ \kappa κ . Substitution into (14) gives e ⋅ j e_{\mathbin{\cdot}j} e ⋅ j . The conditional expectations in (17)–(19) and Proposition 1 specify the responses and predictors for the auxiliary regressions. These regressions estimate u ( 1 ) , u ( 2 ) , u ( 3 ) u^{(1)},u^{(2)},u^{(3)} u ( 1 ) , u ( 2 ) , u ( 3 ) and r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) r^{(1)},r^{(2)},r^{(3)},r^{(4)} r ( 1 ) , r ( 2 ) , r ( 3 ) , r ( 4 ) from the same fitted η \eta η . Theorem 7 gives sufficient conditions on their fitted values through (33) , (40) , (35) , and (42) . For example, Table S1 records the particular fits and covariate restrictions used in our simulations.
Additional numerical details
Data-generating model
Denote q i = ( N i − 15 ) / 5 q_i=(N_i-15)/5 q i = ( N i − 15 ) /5 and X ˉ i = N i − 1 ∑ j = 1 N i X i j \bar{X}_i=N_i^{-1}\sum_{j=1}^{N_i}X_{ij} X ˉ i = N i − 1 ∑ j = 1 N i X ij . The baseline variables and treatment were generated as
N i ∼ Unif { 10 , … , 20 } , V i ∣ N i ∼ N ( 0.20 q i , 1 ) , N_i\sim\operatorname{Unif}\{10,\ldots,20\},\qquad V_i\mid N_i\sim N(0.20q_i,1), N i ∼ Unif { 10 , … , 20 } , V i ∣ N i ∼ N ( 0.20 q i , 1 ) ,
X i j = 0.55 V i + 0.15 q i + ε i j X , A i ∼ Bernoulli ( 0.5 ) , X_{ij}=0.55V_i+0.15q_i+\varepsilon_{ij}^{X},\qquad A_i\sim\operatorname{Bernoulli}(0.5), X ij = 0.55 V i + 0.15 q i + ε ij X , A i ∼ Bernoulli ( 0.5 ) ,
where the ε i j X \varepsilon_{ij}^{X} ε ij X 's were independent standard normal errors. For an N N N -vector, let Σ N ( σ , ρ ) = σ 2 { ( 1 − ρ ) I N + ρ J N } \Sigma_N(\sigma,\rho)=\sigma^2\{(1-\rho)I_N+\rho J_N\} Σ N ( σ , ρ ) = σ 2 {( 1 − ρ ) I N + ρ J N } , where J N J_N J N is the matrix of ones. The post-treatment confounder was generated from
μ Z ( a , C i ) = − 0.45 + 0.35 q i + 0.65 V i + 0.55 X ˉ i + a ( 0.15 2 + 0.07 2 q i ) , \mu_Z(a,C_i)=-0.45+0.35q_i+0.65V_i+0.55\bar{X}_i+a(0.15^2+0.07^2q_i), μ Z ( a , C i ) = − 0.45 + 0.35 q i + 0.65 V i + 0.55 X ˉ i + a ( 0.1 5 2 + 0.0 7 2 q i ) ,
Z i = μ Z ( A i , C i ) 1 N i + E i Z , E i Z ∼ N { 0 , Σ N i ( 0.85 , 0.06 ) } . Z_i=\mu_Z(A_i,C_i)\mathbf{1}_{N_i}+E_i^Z,\qquad E_i^Z\sim N\{0,\Sigma_{N_i}(0.85,0.06)\}. Z i = μ Z ( A i , C i ) 1 N i + E i Z , E i Z ∼ N { 0 , Σ N i ( 0.85 , 0.06 )} .
The marginal mediator mean was μ M ( a , C i ) = − 0.30 − 0.30 q i + 0.65 V i − 0.55 X ˉ i + 0.19 2 a \mu_M(a,C_i)=-0.30-0.30q_i+0.65V_i-0.55\bar{X}_i+0.19^2a μ M ( a , C i ) = − 0.30 − 0.30 q i + 0.65 V i − 0.55 X ˉ i + 0.1 9 2 a , and the mediator was generated through the compatible conditional representation
M i = μ M ( A i , C i ) 1 N i + 0.10 { Z i − μ Z ( A i , C i ) 1 N i } + E i M , M_i=\mu_M(A_i,C_i)\mathbf{1}_{N_i}+0.10\{Z_i-\mu_Z(A_i,C_i)\mathbf{1}_{N_i}\}+E_i^M, M i = μ M ( A i , C i ) 1 N i + 0.10 { Z i − μ Z ( A i , C i ) 1 N i } + E i M ,
E i M ∼ N { 0 , Σ N i ( 0.90 , 0.06 ) } , E_i^M\sim N\{0,\Sigma_{N_i}(0.90,0.06)\}, E i M ∼ N { 0 , Σ N i ( 0.90 , 0.06 )} ,
with E i M E_i^M E i M independent of E i Z E_i^Z E i Z . Consequently, both M i ∣ A i , C i M_i\mid A_i,C_i M i ∣ A i , C i and M i ∣ Z i , A i , C i M_i\mid Z_i,A_i,C_i M i ∣ Z i , A i , C i are exactly exchangeable Gaussian.
Let Z ˉ i , − j \bar{Z}_{i,-j} Z ˉ i , − j and M ˉ i , − j \bar{M}_{i,-j} M ˉ i , − j denote leave-one-out peer means. Conditional on all observed arguments, the outcome mean under arm a a a was
μ Y , i j ( a ) = 0.25 + 0.70 a + 0.18 q i + 0.22 V i + 0.15 X i j + 0.10 X ˉ i + 0.12 a V i − 0.10 a X ˉ i + 1.80 ( 0.50 + 0.12 V i + 0.12 a ) Z i j + ( 0.30 + 0.10 a ) Z ˉ i , − j + ( 1.10 + 0.14 V i + 0.20 a ) M i j + ( 0.40 + 0.12 a ) M ˉ i , − j + 0.06 Z i j 2 + 0.10 M i j 2 + 0.25 M i j M ˉ i , − j + ( 0.70 − 0.15 a ) Z i j M i j . \begin{aligned}
\mu_{Y,ij}(a)={}&0.25+0.70a+0.18q_i+0.22V_i+0.15X_{ij}+0.10\bar{X}_i+0.12aV_i-0.10a\bar{X}_i\\
&+1.80(0.50+0.12V_i+0.12a)Z_{ij}+(0.30+0.10a)\bar{Z}_{i,-j}\\
&+(1.10+0.14V_i+0.20a)M_{ij}+(0.40+0.12a)\bar{M}_{i,-j}\\
&+0.06Z_{ij}^{2}+0.10M_{ij}^{2}+0.25M_{ij}\bar{M}_{i,-j}+(0.70-0.15a)Z_{ij}M_{ij}.
\end{aligned} μ Y , ij ( a ) = 0.25 + 0.70 a + 0.18 q i + 0.22 V i + 0.15 X ij + 0.10 X ˉ i + 0.12 a V i − 0.10 a X ˉ i + 1.80 ( 0.50 + 0.12 V i + 0.12 a ) Z ij + ( 0.30 + 0.10 a ) Z ˉ i , − j + ( 1.10 + 0.14 V i + 0.20 a ) M ij + ( 0.40 + 0.12 a ) M ˉ i , − j + 0.06 Z ij 2 + 0.10 M ij 2 + 0.25 M ij M ˉ i , − j + ( 0.70 − 0.15 a ) Z ij M ij .
We set Y i = μ Y ( A i ) + E i Y \mathbf{Y}_i=\mu_Y(A_i)+E_i^Y Y i = μ Y ( A i ) + E i Y , where E i Y ∼ N { 0 , Σ N i ( 1.20 , 0.08 ) } E_i^Y\sim N\{0,\Sigma_{N_i}(1.20,0.08)\} E i Y ∼ N { 0 , Σ N i ( 1.20 , 0.08 )} and was independent of the preceding errors. We have the cluster-average (W i = 1 W_i=1 W i = 1 ) true values 0.468018, 0.330551, and 0.192446 for IDE, IIME, and ISME, respectively. The individual-average results and true values are reported in § 10.9 .4.2.
Stabilization
Recall that s s s and s − j s_{-j} s − j are the treatment probabilities defined in § 5 , whereas ℓ a \ell_a ℓ a and ℓ a M \ell_a^M ℓ a M are the probabilities of a reference copy in the augmented sample defined in Section G.1. Throughout the study, fitted values of s s s , s − j s_{-j} s − j , ℓ a \ell_a ℓ a , and ℓ a M \ell_a^M ℓ a M were truncated to [ 0.025 , 0.975 ] [0.025,0.975] [ 0.025 , 0.975 ] . For the canonical estimators, fitted values of ω ( 1 ) \omega^{(1)} ω ( 1 ) and ω ∙ j ( 2 ) \omega_{\bullet j}^{(2)} ω ∙ j ( 2 ) in Theorem 2 were truncated to [ 0.05 , 20 ] [0.05,20] [ 0.05 , 20 ] . For the low-dimensional estimators, fitted values of d d d , the ratio κ ∙ j / κ ⋆ j \kappa_{\bullet j}/\kappa_{\star j} κ ∙ j / κ ⋆ j in (14) , and their product e ∙ j e_{\bullet j} e ∙ j were truncated to [ 0.05 , 20 ] [0.05,20] [ 0.05 , 20 ] . The fitted ratios s ( a ′ ∣ m , c , n ) / s ( a ∣ m , c , n ) s(a'\mid m,c,n)/s(a\mid m,c,n) s ( a ′ ∣ m , c , n ) / s ( a ∣ m , c , n ) and s − j ( 0 ∣ m ∙ ( − j ) , c , n ) / s − j ( 1 ∣ m ∙ ( − j ) , c , n ) s_{-j}(0\mid m_{\bullet(-j)},c,n)/s_{-j}(1\mid m_{\bullet(-j)},c,n) s − j ( 0 ∣ m ∙ ( − j ) , c , n ) / s − j ( 1 ∣ m ∙ ( − j ) , c , n ) were also truncated to [ 0.05 , 20 ] [0.05,20] [ 0.05 , 20 ] . After combining these quantities according to (15) and (16) , the resulting estimates of ω ( 1 ) \omega^{(1)} ω ( 1 ) and ω ∙ j ( 2 ) \omega_{\bullet j}^{(2)} ω ∙ j ( 2 ) were truncated to the same interval.
We adapted the outcome-regression adjustment in Appendix A.4 of Cheng and Li [2 ] . Within each validation fold, separately for each θ ( a , a ′ ) \theta(a,a') θ ( a , a ′ ) , τ \tau τ , and choice of W W W , we added a constant to the fitted outcome regression. This constant was the weighted mean of the outcome residuals, using the coefficient of the outcome residual in Theorem 2 as its weight, including W i / N i W_i/N_i W i / N i , the treatment indicator, and the inverse randomization probability. The weighted residuals after adjustment therefore summed to zero. We added the same constant to the associated auxiliary regressions: their differences in the EIF were unchanged, while the final u ^ 0 , j ( 3 ) \hat{u}^{(3)}_{0,j} u ^ 0 , j ( 3 ) or r ^ 0 , j ( 4 ) \hat{r}^{(4)}_{0,j} r ^ 0 , j ( 4 ) increased by that constant. All other nuisance fits remained fixed at their training-fold values. In our setting, the weights also account for Z Z Z , and the adjustment is made separately for W i = 1 W_i=1 W i = 1 and W i = N i W_i=N_i W i = N i . The low-dimensional estimator uses the same adjustment with the weights represented as in Proposition 1.
Recall the estimated influence functions D ^ θ , c a n , i \widehat{D}_{\theta,\mathrm{can},i} D θ , can , i and D ^ τ , c a n , i \widehat{D}_{\tau,\mathrm{can},i} D τ , can , i defined after Theorem 3, and their low-dimensional counterparts defined after Theorem 5. For each method, we evaluated these influence functions using the adjusted fits. For IDE, we subtracted the influence function for θ ( 0 , 0 ) \theta(0,0) θ ( 0 , 0 ) from that for θ ( 1 , 0 ) \theta(1,0) θ ( 1 , 0 ) within each cluster. For IIME, we subtracted that for θ ( 1 , 0 ) \theta(1,0) θ ( 1 , 0 ) from that for τ \tau τ ; for ISME, we subtracted that for τ \tau τ from that for θ ( 1 , 1 ) \theta(1,1) θ ( 1 , 1 ) . Taking these differences before squaring retains the covariance between the estimates. The squared standard error was K − 1 K^{-1} K − 1 times P K P_K P K applied to the squared difference. Each reported 95% interval was centered at the corresponding estimated effect, with half-width equal to its standard error multiplied by K / ( K − 3 ) t K − 3 , 0.975 \sqrt{K/(K-3)}t_{K-3,0.975} K / ( K − 3 ) t K − 3 , 0.975 . The G-computation estimator used the 2.5th and 97.5th percentiles of 200 cluster-bootstrap estimates.
Nuisance fitting and theoretical guarantees
Table S1 reports the exact working models used in the final simulations.
Table S1. Nuisance estimation and misspecification scenarios in the final simulations.
For the low-dimensional estimators, we computed u ^ ( 1 ) − u ^ ( 3 ) \widehat{u}^{(1)}-\widehat{u}^{(3)} u ( 1 ) − u ( 3 ) and r ^ ( 1 ) − r ^ ( 4 ) \widehat{r}^{(1)}-\widehat{r}^{(4)} r ( 1 ) − r ( 4 ) using exact Gaussian moments of the scenario-specific parametric outcome regression under the training-fold fits of ϕ \phi ϕ and κ \kappa κ . We used the full working models for ϕ \phi ϕ and ξ \xi ξ in every low-dimensional scenario and obtained κ ^ \widehat{\kappa} κ by integrating ξ ^ \widehat{\xi} ξ against ϕ ^ \widehat{\phi} ϕ . The applicable convergence conditions for these estimated auxiliary functions are given in Theorem 7.
Table S1 (PDF p. 55).
Tables S2 and S3 distinguish two different claims: consistency and asymptotic linearity. P denotes point consistency of the effect estimator; AL denotes asymptotic linearity with the influence function used for variance estimation. Under a misspecification scenario, a multiple-robust P branch can hold while the product of one fixed nuisance error and one root-K K K estimation error remains first order. We therefore assert AL for the cross-fitted EIF estimators only under all-correct nuisance estimation. A zero entry means that the corresponding theorem does not provide statistical guarantee for the label “1.”
Additional simulation results
Additional cluster-average results
Figure S1 –Figure S6 report bias and coverage for all cluster-average scenarios and methods, including the IIME results omitted from § 6 .
Figure S6. Complete low-dimensional-study cluster-average ISME results.
The additional canonical scenarios show that simultaneous misspecification of two nuisances need not introduce persistent bias when a sufficient nuisance combination remains correct (Figure S1 –Figure S3 ). Under C-0011, both EIF implementations had small biases at the largest K K K for all three effects, consistent with correct ( κ , ξ ) (\kappa,\xi) ( κ , ξ ) . C-1010 showed a similar pattern through correct ( η , κ ) (\eta,\kappa) ( η , κ ) , although its IDE intervals were more conservative than under C-1011. The g-computation showed appreciable IDE bias in both scenarios. The scenario C-1100, like C-1101 in § 6.1 , showed small EIF bias for IDE but persistent biases for IIME and ISME, reflecting the different point-consistency requirements for the direct and mediation effects.
The IIME results in Figure S2 generally followed the point-estimation patterns seen for ISME, but their coverage patterns were not identical. Under C-1100 and C-1101, both EIF implementations underestimated IIME while overestimating ISME. Nevertheless, IIME coverage remained closer to nominal than ISME coverage at larger K K K , despite the lingering IIME bias. Under C-0101, C-0100, C-0001, and C-0000, the mediation-effect biases were more pronounced than under C-1100 and C-1101, and IIME coverage was poorer.
Figure S2. Complete canonical-study cluster-average IIME results.
Other scenarios without a point-consistency guarantee further separate small observed bias from robustness. Under C-0110 and C-0010, the EIF estimators had relatively small IDE biases at larger K K K , but their IIME and ISME biases persisted; IDE intervals were conservative in both scenarios. Under C-1001 and C-1000, appreciable biases remained for all three effects, with little separation between the two EIF implementations at the largest K K K . In these two scenarios, the g-computation estimator had smaller absolute IIME bias than either EIF implementation, illustrating that the EIF estimators did not uniformly improve point estimation outside their sufficient nuisance conditions. As in the results shown in § 6.1 , increasing K K K did not generally remove bias caused by misspecification.
Figure S4 –Figure S6 extend the low-dimensional comparisons to the LD-U scenarios. IDE results were unchanged across LD-AC and LD-U2–LD-U6 for each estimator since the nuisances entering IDE were unchanged. Under LD-U2 and LD-U3, IIME and ISME estimates remained close to their targets, consistent with their point-consistency guarantees. Their coverage nevertheless differed: LD-U2 tended to be more conservative, especially for the SL implementations, whereas LD-U3 coverage was closer to nominal at larger K K K . Under LD-U4 and LD-U6, negative IIME biases and positive ISME biases continued as K K K increased for most methods, even though coverage was often near or above nominal. LD-U5 generally had smaller mediation-effect biases than LD-U4, but, like LD-U4 and LD-U6, does not possess point-consistency guarantee for these effects.
The complete six-estimator comparison shows more variation in coverage than in bias. Under LD-T2, LD-T5, and LD-T6, the SL implementations were generally more conservative than glmnet for all three effects; RF also became conservative in several of these scenarios as K K K increased. These patterns extend the SL results described in § 6.2 . The two reference distributions gave similar results for a given classifier under LD-AC and many misspecification scenarios. However, this similarity was less uniform for the mediation effects under misspecification. At the largest K K K , the SL implementations had smaller absolute IIME and ISME biases under LD-U6 than glmnet and RF, whereas they had larger absolute biases under LD-T5. The G and RB versions of SL also separated more visibly for the mediation effects under LD-U5. Thus, no classifier or reference distribution was uniformly preferable across the misspecified scenarios.
Under all-correct nuisance estimation, the canonical and low-dimensional EIF estimators had similar IDE and IIME RMSE, whereas the canonical estimators had smaller RMSE for the ISME estimand. At the largest K K K , low-dimensional confidence intervals were wider than the parametric canonical EIF intervals, with the largest relative difference for ISME. Within the low-dimensional methods, the SL implementations generally had the widest IDE and ISME intervals. The low-dimensional ISME coverage was closer to nominal at larger K K K , while its IDE coverage was somewhat more conservative.
Figure S5. Complete low-dimensional-study cluster-average IIME results.
Individual-average results
For W i = N i W_i = N_i W i = N i , one can easily compute the true IDE, IIME, and ISME values of 0.489481, 0.342965, and 0.188840, respectively, based on the data generating process given in § 10.9 .§ 1 . Note that both the point estimator and the cluster influence function accounted for estimation of E ( N i ) E(N_i) E ( N i ) . Figure S7 –Figure S12 report the complete individual-average results. The bias and coverage patterns were broadly similar to those under cluster-average weighting; changing the target population did not materially change the comparisons among methods or scenarios.
Figure S12. Complete low-dimensional-study individual-average ISME results.
For the canonical estimators, C-0011 and C-1010 again yielded small biases at larger K K K , whereas C-1100 and C-1101 retained small IDE biases but lingering mediation-effect biases. The latter two scenarios again underestimated IIME and overestimated ISME, with IIME coverage closer to nominal than ISME coverage. The conservative IDE coverage under C-1110, C-0110, and C-0010 also persisted. Under C-0001 and C-0000, substantial mediation-effect biases and poor coverage remained as K K K increased, as in the cluster-average results.
The low-dimensional estimators likewise reproduced the cluster-average patterns. IDE remained close to its target across the LD-U scenarios, whereas LD-U4 and LD-U6 produced negative IIME biases and positive ISME biases for most methods. The relatively conservative SL intervals under LD-T2, LD-T5, and LD-T6 were also evident under individual-average weighting. Under LD-AC, differences between G and RB remained small, and coverage was close to nominal at larger K K K . The all-correct interval-width ordering was also similar across the two weighting schemes: low-dimensional intervals were wider than the parametric canonical EIF intervals for IDE and ISME, especially for the SL implementations, with little difference for IIME.
Figure S8. Complete canonical-study individual-average IIME results.
Figure S11. Complete low-dimensional-study individual-average IIME results.
PPACT trial mediation effects
In this section, we provide an additional table to § 7 for the comparison between the natural effects and the interventional effects.
Table S4. Comparison of natural mediation effects and interventional mediation effects in the PPACT trial.
Note: Natural effects were estimated using the canonical EIF estimator, also denoted the EIF1 _1 1 estimator, of Cheng and Li [2 ] . The associated nuisance parameters were estimated using the same library of machine learning algorithms employed in the interventional effect analysis.