Introduction
Clustered sampling is ubiquitous in empirical research in economics and related fields, where observations within clusters, such as states, schools, or industries, may exhibit arbitrary within-cluster dependence. Ensuring the validity of inference under clustered sampling has hence become a sine qua non . Inference procedures based on cluster-robust standard errors or cluster bootstrap methods, typically coupled with t t t - and Wald tests, are routinely employed to account for such dependence [12 ] (White 1984; Liang and Zeger 1986; Arellano 1987; Cameron, Gelbach, and Miller 2008; Cameron and Miller 2015). Additionally, bootstrap methods are often preferred when the number of clusters is small and asymptotic approximations may be unreliable, or when the estimation of cluster-robust standard errors requires tuning parameters, as in regression quantile models [7 ] . The validity of these inference procedures, however, ultimately relies on the asymptotic normality of the underlying estimator. This property can fail in the presence of extreme cluster size heterogeneity. Specifically, letting G G G denote the number of clusters, N N N the total sample size, and N g N_g N g the size of cluster g ∈ { 1 , … , G } g \in\{1,\ldots,G\} g ∈ { 1 , … , G } , asymptotic normality may fail when max 1 ≤ g ≤ G N g 2 / N ↛ 0 \max_{1\le g\le G} N_g^2/N \not\to0 max 1 ≤ g ≤ G N g 2 / N → 0 as N → ∞ N \to\infty N → ∞ [8 ] (Chiang, Sasaki, and Wang 2025). Recent theoretical advances, including Djogbenou, MacKinnon, and Nielsen (2019) and Hansen and Lee [8 ] , have relaxed restrictions on cluster growth and within-cluster dependence but continue to require the ignorability of cluster size heterogeneity.
As documented by Chiang, Sasaki, and Wang (2025), the cluster size heterogeneity condition max 1 ≤ g ≤ G N g 2 / N → 0 \max_{1\le g\le G} N_g^2/N \to0 max 1 ≤ g ≤ G N g 2 / N → 0 likely fails in a wide range of empirical applications. Substantial cluster-size heterogeneity arises in settings such as schools, firms, and municipalities. The supplementary empirical illustration revisits the municipal-school design of Akhtari, Moreira, and Trucco [1 ] , in which students are clustered within Brazilian municipalities of substantially unequal size. The Unignorably Large Cluster (ULC) problem does not require a cluster to retain a positive limiting sample share: the conditions
max g ≤ G N g N → 0 , max g ≤ G N g 2 N ↛ 0 \max_{g\le G}\frac{N_g}{N}\to0,\qquad\max_{g\le G}\frac{N_g^2}{N}\not\to0 g ≤ G max N N g → 0 , g ≤ G max N N g 2 → 0
can hold simultaneously. Under sufficiently strong within-cluster dependence, cluster scores can then remain first-order influential, so consistency may coexist with failure of conventional Gaussian approximations. To assess the empirical relevance of this issue, Chiang, Sasaki, and Wang (2025), in a pioneering effort, examines 31 articles published in the American Economic Review and Econometrica between 2020 and 2021 that implement cluster-robust inference using publicly available data. Rejection of normality in 24 of the 31 studies (77%) therein raises concerns about the validity of conventional inference in those studies.
Chiang, Sasaki, and Wang (2025) highlights several theoretical challenges arising when cluster size heterogeneity is unignorable. Estimators may be inconsistent, may fail to be asymptotically normal, their conventional cluster-robust variance estimators may fail to consistently estimate asymptotic variances, standard t t t - and Wald tests may become invalid, and the wild bootstrap offers no general remedy. The formal asymptotic analysis in Chiang, Sasaki, and Wang (2025) treats cluster sizes as random and potentially heavy-tailed, so that observation-weighted sums can have non-Gaussian stable limits. The present paper adopts a different, complementary asymptotic framework. Throughout the theoretical analysis, cluster sizes are treated as a realised triangular array, and the probabilistic arguments are finite-moment cluster-level weak law of large numbers (WLLN), central limit theorem (CLT), and empirical-process arguments in the tradition of Hansen and Lee [8 ] . Thus, this extension borrows and generalises the size-adjusted cluster-robust (SACR) reweighting device of Chiang, Sasaki, and Wang (2025) without extending the stable-law asymptotic theory therein.1 Its application to Z Z Z -estimating equations is termed Z Z Z -SACR.
Cluster reweighting, as proposed by Chiang, Sasaki, and Wang (2025), replaces observation-weighted averages in the estimating equations with averages of cluster-level scores. Conventional observation weighting gives equal weight to individuals, whereas cluster balancing gives equal total representation to each cluster. These schemes can define different estimands. Conventional observation-weighted inference can fail under ULC, whereas the Z Z Z -SACR framework supports inference under its stated conditions while accommodating arbitrary realised cluster sizes. The maintained common-zero restriction preserves the structural zero under reweighting. Indeed, in several applications, economic theory or the treatment assignment design motivates observation-level exogeneity conditions that centre each individual score at a common structural parameter (e.g., Parente and Santos Silva (2016); Hagemann [7 ] ). Averaging within clusters then gives the common cluster-level zero, so SACR reweighting preserves that structural zero. Observation-level exogeneity is a sufficient route; the generic theory permits the weaker cluster-average restriction to be imposed directly. Separation of the reweighted population moments establishes identification, while the moment and regularity conditions support consistency and asymptotic normality.
Inverse-cluster-size estimating equations have antecedents in the literature on informative cluster size. Hoffman, Sen, and Weinberg [9 ] motivates within-cluster resampling by distinguishing selection of an individual from the pooled population from selection of a cluster followed by selection of an individual within it. Cluster-weighted generalised estimating equations implement inverse-size weighting directly (Williamson, Datta, and Satten (2003); Benhin, Rao, and Scott (2005)). Cluster-weighted marginal distributions, quantiles, and regression functionals have also been studied by Nevalainen, Datta, and Oja [13 ] . The corresponding distinction between individual- and cluster-weighted treatment effects in cluster-randomised designs is discussed by Su and Ding (2021) and Bugni et al. (2025).
Building on these precedents, this paper develops pointwise and functional Z Z Z -estimation theory for independent, potentially non-identically distributed clusters when conventional cluster-size negligibility conditions fail. Primitive individual-level moment and score function class conditions in this paper support cluster-average laws of large numbers and empirical-process arguments, leading to consistency, asymptotic linearity, covariance-normalised pointwise normality, weak convergence, and functional multiplier-bootstrap inference. The common-zero restriction distinguishes preservation of a structural parameter from a change in population representation. The framework accommodates smooth and non-smooth estimating equations. Distribution Regression (DR), Quantile Regression (QR), and Expectile Regression (ER) serve as concrete examples.
The remainder of the paper is organised as follows. Section 2 formalises the ULC problem within the Z Z Z -estimation framework. Section 3 develops the general asymptotic theory, and Section 4 verifies the theory for distribution, quantile, and expectile regression as concrete examples. Section 5 presents pointwise and uniform Monte Carlo evidence, and Section 6 concludes. The Supplementary Appendix contains proofs, simulation implementation details and additional experiments, and the empirical illustration.
Notation: There are N N N observations in G G G clusters. G : = { 1 , … , G } \mathcal{G}:=\{1,\ldots,G\} G := { 1 , … , G } denotes the set of cluster indices, G g \mathcal{G}_g G g is the set of observations in cluster g g g , N g : = ∣ G g ∣ N_g:=|\mathcal{G}_g| N g := ∣ G g ∣ , and W g : = { W i : i ∈ G g } \mathcal{W}_g:=\{W_i:i\in\mathcal{G}_g\} W g := { W i : i ∈ G g } , where W i W_i W i is a vector of unit-i i i -specific observed data, with N = ∑ g = 1 G N g N=\sum_{g=1}^{G}N_g N = ∑ g = 1 G N g . g ( i ) g(i) g ( i ) denotes the unique cluster containing observation i i i , and p θ : = dim ( θ ) p_\theta:=\dim(\theta) p θ := dim ( θ ) for a vector θ \theta θ . As in empirical-process notation, P f : = ∫ f d P \mathbb{P}f:=\int f\,\mathrm{d}P P f := ∫ f d P denotes integration under the marginal target population, while P [ A ] P[A] P [ A ] denotes event probability. The operator E G \mathbb{E}_G E G denotes expectation under the stage-G G G triangular-array law; conditioning is displayed explicitly when needed. For individual-level functions f f f and cluster-level functions h h h , define P ^ N f : = 1 N ∑ i = 1 N f ( W i ) \widehat{\mathbb{P}}_Nf:=\frac{1}{N}\sum_{i=1}^{N}f(W_i) P N f := N 1 ∑ i = 1 N f ( W i ) , P ^ G h : = 1 G ∑ g = 1 G h ( W g ) \widehat{\mathbb{P}}_Gh:=\frac{1}{G}\sum_{g=1}^{G}h(\mathcal{W}_g) P G h := G 1 ∑ g = 1 G h ( W g ) , and P G h : = E G [ P ^ G h ] \mathbb{P}_Gh:=\mathbb{E}_G[\widehat{\mathbb{P}}_Gh] P G h := E G [ P G h ] . Thus P G \mathbb{P}_G P G is the cluster-averaged stage-G G G population operator. For functional parameters, T \mathcal{T} T is the index set, and ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) denotes the space of bounded functions on T \mathcal{T} T equipped with the uniform norm; for vector-valued paths, ∥ h ∥ T : = sup τ ∈ T ∥ h ( τ ) ∥ \|h\|_{\mathcal{T}}:=\sup_{\tau\in\mathcal{T}}\|h(\tau)\| ∥ h ∥ T := sup τ ∈ T ∥ h ( τ ) ∥ . ⇝ \rightsquigarrow ⇝ denotes weak convergence, and ⇝ P \rightsquigarrow_P ⇝ P denotes conditional weak convergence in probability.
Unignorably Large Clusters and Size-Adjusted Z Z Z -Estimation
This section formalises the Unignorably Large Cluster (ULC) problem within the Z Z Z -estimation framework and introduces the Z Z Z -SACR estimator as a remedy.
The ULC problem in Z Z Z -estimation
Consider a population moment restriction of the generic form
P ψ ( W ; θ o ) = 0 , (1) \mathbb{P}\psi(W;\theta_o)=0, \tag*{(1)} P ψ ( W ; θ o ) = 0 , (1)
where W W W is the vector of observable data, θ o ∈ Θ ⊂ R p θ \theta_o\in\Theta\subset\mathbb{R}^{p_\theta} θ o ∈ Θ ⊂ R p θ uniquely solves the moment restriction, and Θ \Theta Θ is a compact parameter space. Given the moment restriction above and a sample of N N N observations, a natural Z Z Z -estimator of θ o \theta_o θ o is obtained as a solution to
P ^ N ψ ( W ; θ ^ ) = o p ( 1 ) . \widehat{\mathbb{P}}_N\psi(W;\widehat{\theta})=o_p(1). P N ψ ( W ; θ ) = o p ( 1 ) .
In addition to identification, a key ingredient for establishing consistency of θ ^ \hat{\theta} θ ^ is a suitable uniform weak law of large numbers (UWLLN),
sup θ ∈ Θ ∥ ( P ^ N − P ) ψ ( W ; θ ) ∥ = o p ( 1 ) , \sup_{\theta\in\Theta} \left\|(\widehat{\mathbb{P}}_N-\mathbb{P})\psi(W;\theta)\right\|=o_p(1), θ ∈ Θ sup ( P N − P ) ψ ( W ; θ ) = o p ( 1 ) ,
for which pointwise convergence in probability of the estimating equations is necessary, namely
( P ^ N − P ) ψ ( W ; θ ) = o p ( 1 ) (\widehat{\mathbb{P}}_N-\mathbb{P})\psi(W;\theta)=o_p(1) ( P N − P ) ψ ( W ; θ ) = o p ( 1 )
for each θ ∈ Θ \theta\in\Theta θ ∈ Θ . The sampling scheme becomes crucial at this point; it determines whether consistency, asymptotic normality, and valid inference can be established. Under independently sampled clusters, unknown within-cluster dependence, and unit-level marginal laws governed by P \mathbb{P} P , suppose that sup θ ∈ Θ P ∥ ψ ( W ; θ ) ∥ 2 ≤ C < ∞ \sup_{\theta\in\Theta}\mathbb{P}\|\psi(W;\theta)\|^2 \le C < \infty sup θ ∈ Θ P ∥ ψ ( W ; θ ) ∥ 2 ≤ C < ∞ , it follows that
∥ V [ P ^ N ψ ( W ; θ o ) ] ∥ = ∥ V [ 1 N ∑ g = 1 G ( ∑ i ∈ G g ψ ( W i ; θ o ) ) ] ∥ ≤ C ∑ g = 1 G ( N g N ) 2 . \left\|\mathbb{V}\left[\widehat{\mathbb{P}}_N\psi(W;\theta_o)\right]\right\|=\left\|\mathbb{V}\left[\frac{1}{N}\sum_{g=1}^{G}\left(\sum_{i\in\mathcal{G}_g}\psi(W_i;\theta_o)\right)\right]\right\|\le C\sum_{g=1}^{G}\left(\frac{N_g}{N}\right)^2. V [ P N ψ ( W ; θ o ) ] = V N 1 g = 1 ∑ G i ∈ G g ∑ ψ ( W i ; θ o ) ≤ C g = 1 ∑ G ( N N g ) 2 .
Under unrestricted within-cluster dependence, a sufficient size-negligibility condition for pointwise convergence of the observation-weighted moment is ∑ g = 1 G ( N g N ) 2 → 0 \sum_{g=1}^{G}\left(\frac{N_g}{N}\right)^2\to0 ∑ g = 1 G ( N N g ) 2 → 0 , which is implied by max 1 ≤ g ≤ G N g N → 0 \max_{1\le g\le G}\frac{N_g}{N}\to0 max 1 ≤ g ≤ G N N g → 0 , since ∑ g = 1 G ( N g N ) 2 ≤ max 1 ≤ g ≤ G N g N \sum_{g=1}^{G}\left(\frac{N_g}{N}\right)^2\le\max_{1\le g\le G}\frac{N_g}{N} ∑ g = 1 G ( N N g ) 2 ≤ max 1 ≤ g ≤ G N N g . If this negligibility fails severely, the observation-weighted estimating equation may not even obey the pointwise WLLN needed for consistency.
Even when the law-of-large-numbers restriction is satisfied, Gaussian inference additionally requires control of cluster contributions relative to the covariance normalisation. A familiar sufficient size-negligibility benchmark is motivated by the bound ∥ V [ N P ^ N ψ ( W ; θ o ) ] ∥ ≤ C ∑ g = 1 G N g 2 N \left\|\mathbb{V}\left[\sqrt{N}\widehat{\mathbb{P}}_N\psi(W;\theta_o)\right]\right\|\le C\sum_{g=1}^{G}\frac{N_g^2}{N} V [ N P N ψ ( W ; θ o ) ] ≤ C ∑ g = 1 G N N g 2 , which may diverge as N → ∞ N\to\infty N → ∞ . The condition
N − 1 max 1 ≤ g ≤ G N g 2 → 0 N^{-1}\max_{1\le g\le G}N_g^2\to0 N − 1 1 ≤ g ≤ G max N g 2 → 0
is a sufficient size-negligibility benchmark for standard CLT arguments under suitable moment and covariance conditions. When this condition fails, the estimator may still be consistent in favourable designs, but conventional Gaussian, cluster-robust variance, and bootstrap approximations can be invalid.
Large-sample justifications for clustered estimators typically rely on cluster-size restrictions of precisely this form. [8 Assumption 2] , for example, imposes
N − 1 ( ∑ g = 1 G N g r ) 2 / r ≤ C < ∞ , together with N − 1 max 1 ≤ g ≤ G N g 2 → 0 N^{-1}\left(\sum_{g=1}^{G}N_g^r\right)^{2/r}\le C<\infty,\ \text{together with }N^{-1}\max_{1\le g\le G}N_g^2\to0 N − 1 ( g = 1 ∑ G N g r ) 2/ r ≤ C < ∞ , together with N − 1 1 ≤ g ≤ G max N g 2 → 0
as N → ∞ N \to\infty N → ∞ for some 2 ≤ r < ∞ 2 \le r < \infty 2 ≤ r < ∞ . Together with suitable moment and covariance conditions, these restrictions support the CLT of Hansen and Lee [8 ] , Theorem 2. Its WLLN uses the weaker maximum-share condition N − 1 max 1 ≤ g ≤ G N g → 0 N^{-1} \max_{1 \le g \le G} N_g \to0 N − 1 max 1 ≤ g ≤ G N g → 0 . These are sufficient conditions whose violation does not itself imply failure of asymptotic normality, which also depends on within-cluster dependence and covariance normalisation.2 The ULC problem concerns settings where the size-negligibility benchmark fails and sufficiently large cluster scores can dominate the stochastic behaviour of the estimator. In such settings, consistency can fail in severe cases, limiting distributions may be non-normal, and conventional normal, Wald, and bootstrap approximations may be unreliable. More fundamentally, Kojevnikov and Song (2023) shows that sufficiently large clusters can preclude consistent discrimination of the mean and consistent variance estimation under arbitrary within-cluster dependence. Under maintained heavy-tail assumptions, Chiang, Sasaki, and Wang (2025), Theorem 6, further shows that the wild cluster bootstrap is inconsistent.
Target definition, cluster-level exogeneity, and invariance
The primitive object remains the structural parameter θ o \theta_o θ o defined by the population moment condition (2.1). This Z Z Z -estimation framework encompasses estimating equations, first-order conditions of loss functions, and economic moment restrictions without requiring an underlying objective function; see Kosorok [10 ] , Section 2.2.5, and Chernozhukov, Fernández-Val, and Melly [4 ] , Section E.1. The target population and its weighting are logically prior to identification and inference. To make this choice explicit, for nonnegative weights { π g , G } g = 1 G \{\pi_{g,G}\}_{g=1}^{G} { π g , G } g = 1 G satisfying ∑ g = 1 G π g , G = 1 \sum_{g=1}^{G} \pi_{g,G} = 1 ∑ g = 1 G π g , G = 1 , define
M π , G ( θ ) : = ∑ g = 1 G π g , G E G [ ψ ˉ ( W g ; θ ) ] and ψ ˉ ( W g ; θ ) : = 1 N g ∑ i ∈ G g ψ ( W i ; θ ) . M_{\pi,G}(\theta) := \sum_{g=1}^{G} \pi_{g,G} \mathbb{E}_{G}\left[\bar{\psi}(W_g;\theta)\right] \quad\text{and} \quad\bar{\psi}(W_g;\theta) := \frac{1}{N_g}\sum_{i\in\mathcal{G}_g}\psi(W_i;\theta). M π , G ( θ ) := g = 1 ∑ G π g , G E G [ ψ ˉ ( W g ; θ ) ] and ψ ˉ ( W g ; θ ) := N g 1 i ∈ G g ∑ ψ ( W i ; θ ) .
The conventional observation-weighted moment uses π g , G = N g / N \pi_{g,G} = N_g/N π g , G = N g / N , whereas the cluster-balanced moment uses π g , G = 1 / G \pi_{g,G} = 1/G π g , G = 1/ G . The weighting scheme determines whose outcomes the population moment represents. Observation weighting corresponds to selecting an individual uniformly from the pooled sample, so that cluster g g g is selected with probability N g / N N_g/N N g / N . Cluster balancing instead selects a cluster uniformly and then a member uniformly within it, giving every cluster equal total representation. Neither scheme intrinsically over- or under-represents clusters, as representation is assessed relative to the research question (Nevalainen, Datta, and Oja [13 ] ; Su and Ding (2021); Bugni et al. (2025)). For example, student-weighted and municipality-balanced outcome distributions describe different policy perspectives.
Hoffman, Sen, and Weinberg [9 ] emphasises that informative cluster size can make this representational choice consequential for the estimand.3 When cluster size is related to outcome laws, changing the relative representation of clusters can change the population moment and its zero. Cluster balancing removes the additional representation of larger clusters induced by pooled observation sampling; it does not generally remove the underlying size–outcome dependence. The resulting parameter is defined by the cluster-balanced moment, rather than generally being an average of cluster-specific parameters.
The common-zero restriction maintained in this paper identifies a setting in which this change in representation is innocuous; the structural zero is preserved, where every cluster’s expected score vanishes at the same θ o \theta_o θ o . Without that restriction, weighting-specific parameters can remain meaningful scientific or causal targets under suitable identification conditions, but reweighting need not preserve the parameter. ULC poses a separate inferential question, i.e., even when observation weighting defines the desired target, sufficiently influential clusters can invalidate conventional asymptotic approximations or even parameter estimator consistency.4
The maintained economic condition in this paper is a common cluster-level zero,
E G [ ψ ˉ ( W g ; θ o ) ] = 0 , g = 1 , … , G . \mathbb{E}_{G}\left[\bar{\psi}(W_g;\theta_o)\right] = 0,\qquad g=1,\ldots,G. E G [ ψ ˉ ( W g ; θ o ) ] = 0 , g = 1 , … , G .
This condition may be imposed directly as an economic identifying restriction or derived from the assignment design. For example, random assignment at the cluster level can imply cluster-level exogeneity for an appropriately constructed score. A separate sufficient route is the individual-level restriction E G [ ψ ( W i ; θ o ) ] = 0 \mathbb{E}_{G}[\psi(W_i;\theta_o)] = 0 E G [ ψ ( W i ; θ o )] = 0 for every i = 1 , … , N i=1,\ldots,N i = 1 , … , N , e.g., Hagemann [7 ] , which yields the displayed condition by averaging within clusters. Formally, M π , G ( θ o ) = 0 M_{\pi,G}(\theta_o) = 0 M π , G ( θ o ) = 0 for every deterministic weighting scheme π \pi π ; uniqueness remains a separate property of the selected population moment. The cluster-level restriction also supplies the centring used in the inference theorems below.5
The SACR adjustment operates by giving each cluster equal weight. Algebraically related inverse-size-weighted estimating equations and functionals arise in within-cluster resampling, cluster-weighted generalised estimating equations, and marginal distributional inference ([9 , 13 ] (Williamson, Datta, and Satten 2003; Benhin, Rao, and Scott 2005)); cluster averaging is used in this paper to obtain ULC-robust Z Z Z -inference. For a candidate value θ \theta θ , define
P ^ G ψ , S A C R ( W ; θ ) : = 1 G ∑ i = 1 N N g ( i ) − 1 ψ ( W i ; θ ) = 1 G ∑ g = 1 G ψ ˉ ( W g ; θ ) . \widehat{P}_{G}^{\psi,SACR}(W;\theta) := \frac{1}{G}\sum_{i=1}^{N}N_{g(i)}^{-1}\psi(W_i;\theta) = \frac{1}{G}\sum_{g=1}^{G}\bar{\psi}(W_g;\theta). P G ψ , S A C R ( W ; θ ) := G 1 i = 1 ∑ N N g ( i ) − 1 ψ ( W i ; θ ) = G 1 g = 1 ∑ G ψ ˉ ( W g ; θ ) .
The corresponding Z Z Z -SACR estimator satisfies
P ^ G ψ S A C R ( W ; θ ^ S A C R ) = o p ( 1 ) , \widehat{\mathbb{P}}_G\psi^{SACR}(W;\widehat{\theta}^{SACR})=o_p(1), P G ψ S A C R ( W ; θ S A C R ) = o p ( 1 ) ,
or the sharper o p ( G − 1 / 2 ) o_p(G^{-1/2}) o p ( G − 1/2 ) rate when first-order asymptotic theory is invoked. The weight N g − 1 N_g^{-1} N g − 1 makes each cluster contribute one cluster-average score, so the empirical process is indexed by G G G rather than N N N .
The following proposition records the resulting target-invariance implication for deterministic reweighting.
Proposition 2.1. Let θ o ∈ Θ \theta_o \in\Theta θ o ∈ Θ be defined by the primitive population moment restriction (2.1), and suppose that the cluster-level score satisfies E G [ ψ ˉ ( W g ; θ o ) ] = 0 \mathbb{E}_G[\bar{\psi}(W_g;\theta_o)] = 0 E G [ ψ ˉ ( W g ; θ o )] = 0 for every g = 1 , … , G g=1,\ldots,G g = 1 , … , G . Let { ω g } g = 1 G \{\omega_g\}_{g=1}^G { ω g } g = 1 G be any sequence of strictly positive deterministic weights. Then, uniformly in G ≥ 1 G \ge1 G ≥ 1 , the weighted population moment vanishes at θ o \theta_o θ o , namely
E G [ 1 G ∑ i = 1 N ω g ( i ) ψ ( W i ; θ o ) ] = 1 G ∑ g = 1 G ω g N g E G [ ψ ˉ ( W g ; θ o ) ] = 0. \mathbb{E}_G\left[\frac{1}{G}\sum_{i=1}^{N}\omega_{g(i)}\psi(W_i;\theta_o)\right]=\frac{1}{G}\sum_{g=1}^{G}\omega_gN_g\mathbb{E}_G[\bar{\psi}(W_g;\theta_o)]=0. E G [ G 1 i = 1 ∑ N ω g ( i ) ψ ( W i ; θ o ) ] = G 1 g = 1 ∑ G ω g N g E G [ ψ ˉ ( W g ; θ o )] = 0. If, in addition, the weighted triangular-array population moment is separated away from zero outside every neighbourhood of θ o \theta_o θ o , then point identification occurs at the same parameter value θ o \theta_o θ o .
Proposition 2.1 separates the common-zero condition that gives target invariance from the separation condition that gives identification. Neither requires the variance restriction introduced next. Subject to finite second moments of individual scores, the SACR choice ω g = N g − 1 \omega_g=N_g^{-1} ω g = N g − 1 is a robust sufficient weighting rule under unknown within-cluster dependence. In the worst case, the variance of a cluster score ∑ i ∈ G g ψ ( W i ; θ ) \sum_{i\in\mathcal{G}_g}\psi(W_i;\theta) ∑ i ∈ G g ψ ( W i ; θ ) can scale as N g 2 N_g^2 N g 2 , so a sufficient condition for bounded variance under G \sqrt{G} G -normalisation is
1 G ∑ g = 1 G ω g 2 N g 2 = O ( 1 ) \frac{1}{G}\sum_{g=1}^{G}\omega_g^2N_g^2=O(1) G 1 g = 1 ∑ G ω g 2 N g 2 = O ( 1 )
uniformly over the relevant parameter space. This condition is satisfied by ω g = N g − 1 \omega_g=N_g^{-1} ω g = N g − 1 without imposing restrictions on the growth rate of N g N_g N g . Under stronger dependence restrictions, lighter reweighting may also be admissible, and the unweighted estimator can remain valid when cluster sizes are asymptotically negligible. Subject to the stated identification and moment conditions, the Z Z Z -SACR therefore provides a transparent cluster-balanced estimating equation that remains available under the ULC problem, where conventional unadjusted estimating equations may fail to converge.
Asymptotic Theory for the Z-SACR
Throughout this section, the asymptotic analysis uses a clustered sampling framework in which clusters are independent. The following assumption formalises this sampling structure.
Assumption 1 (Sampling). The cluster vectors W g : = { W i : i ∈ G g } W_g:=\{W_i:i\in\mathcal{G}_g\} W g := { W i : i ∈ G g } , g = 1 , … , G g=1,\ldots,G g = 1 , … , G , are independent across clusters.
Assumption 1 requires only independence across clusters and permits arbitrary, potentially non-identically distributed dependence structures within clusters. Throughout, cluster sizes { N g : g ≤ G } \{N_g:g\le G\} { N g : g ≤ G } are treated as deterministic. Equivalently, all probability statements are interpreted conditionally on the realised cluster sizes. Under this conditional interpretation, the maintained independence, common-zero, moment and covariance conditions apply under the conditional law. As G → ∞ G\to\infty G → ∞ , the total size N = N ( G ) N=N(G) N = N ( G ) and cluster sizes N g = N g ( G ) N_g=N_g(G) N g = N g ( G ) may vary with G G G ; this row dependence is suppressed in the notation.
Pointwise consistency and asymptotic normality
Define the cluster-average score
ψ ˉ ( W g ; θ ) : = 1 N g ∑ i ∈ G g ψ ( W i ; θ ) . \bar{\psi}(W_g;\theta):=\frac{1}{N_g}\sum_{i\in\mathcal{G}_g}\psi(W_i;\theta). ψ ˉ ( W g ; θ ) := N g 1 i ∈ G g ∑ ψ ( W i ; θ ) .
The following running example and subsequent continuations thereof in this section use the just-identified linear instrumental variables (IV) model to translate the generic high-level conditions into familiar restrictions. Its ordinary least squares (OLS) special case recovers the size-adjusted least-squares estimator of Chiang, Sasaki, and Wang (2025).
Running Example (Just-identified linear IV). Consider the linear model Y i = X i ′ θ o + U i Y_i=X_i^\prime\theta_o+U_i Y i = X i ′ θ o + U i , where X i , Z i ∈ R p θ X_i,Z_i\in\mathbb{R}^{p_\theta} X i , Z i ∈ R p θ , X i X_i X i may contain endogenous regressors, and Z i Z_i Z i is a vector of instruments. The just-identified IV score is ψ ( W i ; θ ) = Z i ( Y i − X i ′ θ ) \psi(W_i;\theta)=Z_i(Y_i-X_i^\prime\theta) ψ ( W i ; θ ) = Z i ( Y i − X i ′ θ ) , and the IV-SACR estimator solves
G − 1 ∑ g = 1 G N g − 1 ∑ i ∈ G g Z i ( Y i − X i ′ θ ) = 0. G^{-1}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}Z_i(Y_i-X_i^\prime\theta)=0. G − 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ Z i ( Y i − X i ′ θ ) = 0.
Setting Z i = X i Z_i=X_i Z i = X i gives exactly the OLS-SACR normal equation of Chiang, Sasaki, and Wang (2025).
The following assumptions, together with Assumption 1, are sufficient for consistency of the Z-SACR estimator. Define the unscaled centred cluster-average empirical process on Θ \Theta Θ :
U G ( θ ) : = 1 G ∑ g = 1 G { ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ ) ] } . U_G(\theta):=\frac{1}{G}\sum_{g=1}^{G}\left\{\bar{\psi}(W_g;\theta)-\mathbb{E}_G[\bar{\psi}(W_g;\theta)]\right\}. U G ( θ ) := G 1 g = 1 ∑ G { ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ )] } .
Assumption 2 (UWLLN).
(a) Θ ⊂ R p θ \Theta\subset\mathbb{R}^{p_\theta} Θ ⊂ R p θ is compact.
(b) lim M → ∞ sup G ≥ 1 sup 1 ≤ i ≤ N sup θ ∈ Θ E G [ ∥ ψ ( W i ; θ ) ∥ 1 { ∥ ψ ( W i ; θ ) ∥ > M } ] = 0. \lim_{M\to\infty}\sup_{G\ge1}\sup_{1\le i\le N}\sup_{\theta\in\Theta}\mathbb{E}_G[\|\psi(W_i;\theta)\|\mathbf{1}\{\|\psi(W_i;\theta)\|>M\}]=0. lim M → ∞ sup G ≥ 1 sup 1 ≤ i ≤ N sup θ ∈ Θ E G [ ∥ ψ ( W i ; θ ) ∥ 1 { ∥ ψ ( W i ; θ ) ∥ > M }] = 0.
(c) For every ε > 0 , \varepsilon>0, ε > 0 , \displaystyle \lim_{\delta\downarrow0}\limsup_{G\to\infty}\mathbb{P}\left[\sup_{|\theta_1-\theta_2|<\delta}|U_G(\theta_1)-U_G(\theta_2)|>\varepsilon\right]=0.$
Assumption 2 combines pointwise clustered triangular-array convergence, following the WLLN structure of Hansen and Lee [8 ] , with stochastic equicontinuity on a totally bounded parameter space. These conditions match the generic pointwise-to-uniform convergence result in Andrews [2 ] .
Define
P ^ G ψ ˉ ( W g ; θ ) : = 1 G ∑ g = 1 G ψ ˉ ( W g ; θ ) , and P G ψ ˉ ( W g ; θ ) : = E G [ P ^ G ψ ˉ ( W g ; θ ) ] = 1 G ∑ g = 1 G E G [ ψ ˉ ( W g ; θ ) ] . \widehat{\mathbb{P}}_G\bar{\psi}(W_g;\theta) := \frac{1}{G}\sum_{g=1}^{G}\bar{\psi}(W_g;\theta), \text{ and } \mathbb{P}_G\bar{\psi}(W_g;\theta) := \mathbb{E}_G\left[\widehat{\mathbb{P}}_G\bar{\psi}(W_g;\theta)\right] = \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G[\bar{\psi}(W_g;\theta)]. P G ψ ˉ ( W g ; θ ) := G 1 g = 1 ∑ G ψ ˉ ( W g ; θ ) , and P G ψ ˉ ( W g ; θ ) := E G [ P G ψ ˉ ( W g ; θ ) ] = G 1 g = 1 ∑ G E G [ ψ ˉ ( W g ; θ )] .
The distinction between P \mathbb{P} P , E G \mathbb{E}_G E G , and P G \mathbb{P}_G P G deserves some discussion. The restriction P ψ ( W ; θ o ) = 0 \mathbb{P}\psi(W;\theta_o)=0 P ψ ( W ; θ o ) = 0 defines the structural parameter in the target population. The triangular-array structure enters through cluster-specific expectations E G [ ⋅ ] \mathbb{E}_G[\cdot] E G [ ⋅ ] and their cluster average P G \mathbb{P}_G P G . Economic theory or the assignment design supplies the identifying exogeneity condition, while the clustered sampling structure determines how the resulting scores are aggregated and how their sampling variation is analysed. The maintained array-level implication is the clusterwise zero E G [ ψ ˉ ( W g ; θ o ) ] = 0 \mathbb{E}_G[\bar{\psi}(W_g;\theta_o)]=0 E G [ ψ ˉ ( W g ; θ o )] = 0 for every g = 1 , … , G g=1,\ldots,G g = 1 , … , G . An individual-level restriction E G [ ψ ( W i ; θ o ) ] = 0 \mathbb{E}_G[\psi(W_i;\theta_o)]=0 E G [ ψ ( W i ; θ o )] = 0 for every sampled observation is sufficient, but not necessary, for this condition. Away from θ o \theta_o θ o , the map θ ↦ P G ψ ˉ ( W g ; θ ) \theta\mapsto\mathbb{P}_G\bar{\psi}(W_g;\theta) θ ↦ P G ψ ˉ ( W g ; θ ) need not coincide with θ ↦ P ψ ( W ; θ ) \theta\mapsto\mathbb{P}\psi(W;\theta) θ ↦ P ψ ( W ; θ ) , so identification is imposed through array-level separation.
Assumption 3 (Identification).
(a) The primitive population moment restriction P ψ ( W ; θ o ) = 0 \mathbb{P}\psi(W;\theta_o)=0 P ψ ( W ; θ o ) = 0 holds, and the corresponding cluster-level score is centred along the triangular array:
E G [ ψ ˉ ( W g ; θ o ) ] = 0 , g = 1 , … , G . \mathbb{E}_G[\bar{\psi}(W_g;\theta_o)] = 0,\qquad g=1,\ldots,G. E G [ ψ ˉ ( W g ; θ o )] = 0 , g = 1 , … , G . (b) For every $\epsilon>0,
lim inf G → ∞ inf θ ∈ Θ : ∥ θ − θ o ∥ ≥ ϵ ∥ P G ψ ˉ ( W g ; θ ) ∥ > 0. \liminf_{G\to\infty}\inf_{\theta\in\Theta:\lVert\theta-\theta_o\rVert\ge\epsilon}\left\lVert\mathbb{P}_G\bar{\psi}(W_g;\theta)\right\rVert>0. G → ∞ lim inf θ ∈ Θ : ∥ θ − θ o ∥ ≥ ϵ inf P G ψ ˉ ( W g ; θ ) > 0.
Assumption 3 ensures identification of the target parameter. Part (a) gives the economically maintained cluster-level exogeneity condition and hence P G ψ ˉ ( W g ; θ o ) = 0 \mathbb{P}_G\bar{\psi}(W_g;\theta_o)=0 P G ψ ˉ ( W g ; θ o ) = 0 . Part (b) rules out approximate zeros of the triangular-array population moment away from θ o \theta_o θ o .
The following theorem establishes a UWLLN and consistency of the Z-SACR estimator.
Theorem 3.1. Suppose that Assumptions 1 and 2 hold.
(a) The cluster-average empirical process satisfies
sup θ ∈ Θ ∥ ( P ^ G − P G ) ψ ˉ ( W g ; θ ) ∥ → p 0. (2) \sup_{\theta\in\Theta}\left\lVert(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\psi}(W_g;\theta)\right\rVert\xrightarrow{p} 0. \tag*{(2)} θ ∈ Θ sup ( P G − P G ) ψ ˉ ( W g ; θ ) p 0. (2) (b) If, in addition, P ^ G ψ ˉ ( W g ; θ ^ S A C R ) = o p ( 1 ) \widehat{\mathbb{P}}_G\bar{\psi}(W_g;\widehat{\theta}^{SACR})=o_p(1) P G ψ ˉ ( W g ; θ S A C R ) = o p ( 1 ) and Assumption 3 holds, then θ ^ S A C R → p θ o \widehat{\theta}^{SACR}\xrightarrow{p}\theta_o θ S A C R p θ o as G → ∞ G\to\infty G → ∞ .
The IV running example verifies the two ingredients of this result in turn.
Running Example (continued: IV UWLLN). For the IV score, ψ ( W i ; θ ) = Z i U i − Z i X i ′ ( θ − θ o ) \psi(W_i;\theta)=Z_iU_i-Z_iX_i^{\prime}(\theta-\theta_o) ψ ( W i ; θ ) = Z i U i − Z i X i ′ ( θ − θ o ) . Suppose that ∥ Z i U i ∥ \lVert Z_iU_i\rVert ∥ Z i U i ∥ and ∥ Z i X i ′ ∥ \lVert Z_iX_i^{\prime}\rVert ∥ Z i X i ′ ∥ are uniformly integrable over all observations and rows G G G . Compactness of Θ \Theta Θ (Assumption 2 ( a ) 2(a) 2 ( a ) ) gives D Θ : = sup θ ∈ Θ ∥ θ − θ o ∥ < ∞ D_{\Theta}:=\sup_{\theta\in\Theta}\lVert\theta-\theta_o\rVert<\infty D Θ := sup θ ∈ Θ ∥ θ − θ o ∥ < ∞ , and
sup θ ∈ Θ ∥ ψ ( W i ; θ ) ∥ ≤ ∥ Z i U i ∥ + D Θ ∥ Z i X i ′ ∥ . \sup_{\theta\in\Theta}\lVert\psi(W_i;\theta)\rVert\leq\lVert Z_iU_i\rVert+D_{\Theta}\lVert Z_iX_i^{\prime}\rVert. θ ∈ Θ sup ∥ ψ ( W i ; θ )∥ ≤ ∥ Z i U i ∥ + D Θ ∥ Z i X i ′ ∥ .
The right-hand side is then uniformly integrable, verifying Assumption 2 ( b ) 2(b) 2 ( b ) .
Define Q ˉ g : = N g − 1 ∑ i ∈ G g Z i X i ′ \bar{Q}_g:=N_g^{-1}\sum_{i\in\mathcal{G}_g}Z_iX_i^{\prime} Q ˉ g := N g − 1 ∑ i ∈ G g Z i X i ′ . Cluster independence (Assumption 1 1 1 ) and uniform integrability (see Lemma S .2.1 ( b ) S.2.1(b) S .2.1 ( b ) ), applied to Z i U i Z_iU_i Z i U i and vec ( Z i X i ′ ) \operatorname{vec}(Z_iX_i^{\prime}) vec ( Z i X i ′ ) , give U G ( θ o ) = o p ( 1 ) U_G(\theta_o)=o_p(1) U G ( θ o ) = o p ( 1 ) and ( P ^ G − P G ) Q ˉ g = o p ( 1 ) (\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{Q}_g=o_p(1) ( P G − P G ) Q ˉ g = o p ( 1 ) . The affine identity U G ( θ ) = U G ( θ o ) − ( P ^ G − P G ) Q ˉ g ( θ − θ o ) U_G(\theta)=U_G(\theta_o)-(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{Q}_g(\theta-\theta_o) U G ( θ ) = U G ( θ o ) − ( P G − P G ) Q ˉ g ( θ − θ o ) yields
sup θ 1 , θ 2 ∈ Θ ∥ θ 1 − θ 2 ∥ < δ ∥ U G ( θ 1 ) − U G ( θ 2 ) ∥ ≤ δ ∥ ( P ^ G − P G ) Q ˉ g ∥ = o p ( 1 ) , \sup_{\substack{\theta_1,\theta_2\in\Theta\\ \lVert\theta_1-\theta_2\rVert<\delta}}\lVert U_G(\theta_1)-U_G(\theta_2)\rVert\leq\delta\lVert(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{Q}_g\rVert=o_p(1), θ 1 , θ 2 ∈ Θ ∥ θ 1 − θ 2 ∥ < δ sup ∥ U G ( θ 1 ) − U G ( θ 2 )∥ ≤ δ ∥( P G − P G ) Q ˉ g ∥ = o p ( 1 ) ,
for every fixed δ > 0 \delta>0 δ > 0 , which verifies Assumption 2 ( c ) 2(c) 2 ( c ) .
The stated primitive conditions above establish the IV UWLLN directly and verify the generic UWLLN assumptions.
The next example concerns primitives for IV identification.
Running Example (continued: IV zero preservation and separation). For the running IV model, cluster-level instrument validity (Assumption 3 ( a ) 3(a) 3 ( a ) ) is E G [ N g − 1 ∑ i ∈ G g Z i U i ] = 0 \mathbb{E}_G[N_g^{-1}\sum_{i\in\mathcal{G}_g}Z_iU_i]=0 E G [ N g − 1 ∑ i ∈ G g Z i U i ] = 0 for every g g g . Individual-level restrictions E G [ Z i U i ] = 0 \mathbb{E}_G[Z_iU_i]=0 E G [ Z i U i ] = 0 , i = 1 , … , N i=1,\ldots,N i = 1 , … , N , provide a familiar sufficient route. Define Q G : = G − 1 ∑ g = 1 G E G [ Q ˉ g ] Q_G:=G^{-1}\sum_{g=1}^{G}\mathbb{E}_G[\bar{Q}_g] Q G := G − 1 ∑ g = 1 G E G [ Q ˉ g ] , then P G ψ ˉ ( W g ; θ ) = − Q G ( θ − θ o ) \mathbb{P}_G\bar{\psi}(W_g;\theta)=-Q_G(\theta-\theta_o) P G ψ ˉ ( W g ; θ ) = − Q G ( θ − θ o ) . If lim inf G → ∞ λ min ( Q G ′ Q G ) > 0 \liminf_{G\to\infty}\lambda_{\min}(Q_G^{\prime}Q_G)>0 lim inf G → ∞ λ m i n ( Q G ′ Q G ) > 0 , then
∥ P G ψ ˉ ( W g ; θ ) ∥ 2 = ( θ − θ o ) ′ Q G ′ Q G ( θ − θ o ) ≥ λ min ( Q G ′ Q G ) ∥ θ − θ o ∥ 2 , \lVert\mathbb{P}_G\bar{\psi}(W_g;\theta)\rVert^2=(\theta-\theta_o)^{\prime}Q_G^{\prime}Q_G(\theta-\theta_o)\geq\lambda_{\min}(Q_G^{\prime}Q_G)\lVert\theta-\theta_o\rVert^2, ∥ P G ψ ˉ ( W g ; θ ) ∥ 2 = ( θ − θ o ) ′ Q G ′ Q G ( θ − θ o ) ≥ λ m i n ( Q G ′ Q G ) ∥ θ − θ o ∥ 2 ,
which verifies Assumption 3 ( b ) 3(b) 3 ( b ) . Thus cluster-level instrument validity preserves the structural zero, whereas uniform nonsingularity of the cluster-balanced first-stage cross-moment separates that zero from alternatives.
The following assumption collects the additional first-order conditions used to derive the asymptotically linear representation and the normality of the Z-SACR estimator. Define H G ( θ ) : = ∂ ∂ θ ′ P G ψ ˉ ( W g ; θ ) H_G(\theta):=\frac{\partial}{\partial\theta^{\prime}}\mathbb{P}_G\bar{\psi}(W_g;\theta) H G ( θ ) := ∂ θ ′ ∂ P G ψ ˉ ( W g ; θ ) and H G : = H G ( θ o ) H_G:=H_G(\theta_o) H G := H G ( θ o ) .
Assumption 4 (Asymptotic linearisation).
(a) The Z-SACR satisfies P ^ G ψ ˉ ( W g ; θ ^ S A C R ) = o p ( G − 1 / 2 ) \widehat{\mathbb{P}}_G\bar{\psi}(W_g;\widehat{\theta}^{SACR})=o_p(G^{-1/2}) P G ψ ˉ ( W g ; θ S A C R ) = o p ( G − 1/2 ) .
(b) θ o \theta_o θ o is in the interior of Θ \Theta Θ .
(c) The array-level expected score P G ψ ˉ ( W g ; θ ) \mathbb{P}_G\bar{\psi}(W_g;\theta) P G ψ ˉ ( W g ; θ ) is continuously differentiable in an open neighbourhood of θ o \theta_o θ o , and lim δ ↓ 0 lim sup G → ∞ sup ∥ θ − θ o ∥ ≤ δ ∥ H G ( θ ) − H G ∥ = 0 \lim_{\delta\downarrow0}\limsup_{G\to\infty}\sup_{\lVert\theta-\theta_o\rVert\leq\delta}\lVert H_G(\theta)-H_G\rVert=0 lim δ ↓ 0 lim sup G → ∞ sup ∥ θ − θ o ∥ ≤ δ ∥ H G ( θ ) − H G ∥ = 0 .
Define Ω G : = V G [ 1 G ∑ g = 1 G ψ ˉ ( W g ; θ o ) ] = G − 1 ∑ g = 1 G E G [ ψ ˉ ( W g ; θ o ) ψ ˉ ( W g ; θ o ) ′ ] \Omega_G := V_G\left[\frac{1}{\sqrt{G}}\sum_{g=1}^{G}\bar{\psi}(W_g;\theta_o)\right] = G^{-1}\sum_{g=1}^{G}\mathbb{E}_G[\bar{\psi}(W_g;\theta_o)\bar{\psi}(W_g;\theta_o)^{\prime}] Ω G := V G [ G 1 ∑ g = 1 G ψ ˉ ( W g ; θ o ) ] = G − 1 ∑ g = 1 G E G [ ψ ˉ ( W g ; θ o ) ψ ˉ ( W g ; θ o ) ′ ] . The following supplies regularity conditions on the second moment of the score and the population Jacobian at the truth.
Assumption 5 (Moments and nondegeneracy).
(a) lim M → ∞ sup G ≥ 1 sup 1 ≤ i ≤ N E G [ ∥ ψ ( W i ; θ o ) ∥ 2 1 { ∥ ψ ( W i ; θ o ) ∥ > M } ] = 0 \lim_{M\to\infty}\sup_{G\geq1}\sup_{1\leq i\leq N}\mathbb{E}_G[\|\psi(W_i;\theta_o)\|^2\mathbf{1}\{\|\psi(W_i;\theta_o)\|>M\}]=0 lim M → ∞ sup G ≥ 1 sup 1 ≤ i ≤ N E G [ ∥ ψ ( W i ; θ o ) ∥ 2 1 { ∥ ψ ( W i ; θ o ) ∥ > M }] = 0 .
(b) The sequences Ω G \Omega_G Ω G and H G H_G H G satisfy λ min ( Ω G ) ≥ λ \lambda_{\min}(\Omega_G)\geq\lambda λ m i n ( Ω G ) ≥ λ and λ min ( H G ′ H G ) ≥ λ 2 \lambda_{\min}(H_G^{\prime}H_G)\geq\lambda^2 λ m i n ( H G ′ H G ) ≥ λ 2 for some λ > 0 \lambda>0 λ > 0 .
A sufficient primitive condition for Assumption 5(a) is sup G ≥ 1 sup 1 ≤ i ≤ N E G [ ∥ ψ ( W i ; θ o ) ∥ 2 + δ ] < ∞ \sup_{G\geq1}\sup_{1\leq i\leq N}\mathbb{E}_G[\|\psi(W_i;\theta_o)\|^{2+\delta}]<\infty sup G ≥ 1 sup 1 ≤ i ≤ N E G [ ∥ ψ ( W i ; θ o ) ∥ 2 + δ ] < ∞ for some δ > 0 \delta>0 δ > 0 . Assumption 5(a) is a primitive individual-level moment condition that yields corresponding cluster-average control by Lemma S.2.1(a). Assumption 5(b) ensures Ω G \Omega_G Ω G is positive definite and keeps the population Jacobian H G H_G H G uniformly invertible.
Define the first-order centred cluster-average empirical process by
G G ( θ ) : = G U G ( θ ) = G ( P b G − P G ) ψ ˉ ( W g ; θ ) . \mathbb{G}_G(\theta):=\sqrt{G}\,U_G(\theta)=\sqrt{G}(\mathbb{P}_{bG}-P_G)\bar{\psi}(W_g;\theta). G G ( θ ) := G U G ( θ ) = G ( P b G − P G ) ψ ˉ ( W g ; θ ) .
The following assumption strengthens the preceding level-scale control to the G \sqrt{G} G scale locally around θ o \theta_o θ o . Unlike the global condition in Assumption 2(c), it is tailored to the empirical-process remainder in the asymptotic linearisation.
Assumption 6 (Local first-order stochastic equicontinuity). For every ε > 0 \varepsilon>0 ε > 0 ,
lim δ ↓ 0 lim sup G → ∞ P [ sup ∥ θ − θ o ∥ ≤ δ ∥ G G ( θ ) − G G ( θ o ) ∥ > ε ] = 0. \lim_{\delta\downarrow0}\limsup_{G\to\infty}\mathbb{P}\left[\sup_{\|\theta-\theta_o\|\leq\delta}\|\mathbb{G}_G(\theta)-\mathbb{G}_G(\theta_o)\|>\varepsilon\right]=0. δ ↓ 0 lim G → ∞ lim sup P [ ∥ θ − θ o ∥ ≤ δ sup ∥ G G ( θ ) − G G ( θ o ) ∥ > ε ] = 0.
The sample score in a Z Z Z -estimation problem need not be continuously differentiable; regression quantiles are the canonical example (Koenker and Bassett Jr, 1978). The differentiability requirement in Assumption 4(c) is therefore imposed only on the array-level expected score. To control the non-smooth sample component, Assumption 6 imposes local stochastic equicontinuity on the first-order process G G \mathbb{G}_G G G . This condition ensures that the empirical-process remainder is negligible at the G − 1 / 2 G^{-1/2} G − 1/2 scale in a neighbourhood of θ o \theta_o θ o , even when the sample score is non-smooth.
Define V G : = H G − 1 Ω G ( H G − 1 ) ′ V_G:=H_G^{-1}\Omega_G(H_G^{-1})^{\prime} V G := H G − 1 Ω G ( H G − 1 ) ′ . With consistency and the preceding regularity conditions in hand, the following establishes the asymptotic linearity and asymptotic normality of the Z Z Z -SACR.
Theorem 3.2. Under Assumptions 1 to 6, as G → ∞ G\to\infty G → ∞ ,
(a) the Z-SACR admits the asymptotically linear representation
G ( θ ^ S A C R − θ o ) = − H G − 1 G P b G ψ ˉ ( W g ; θ o ) + o p ( 1 ) ; (3) \sqrt{G}(\widehat{\theta}^{SACR}-\theta_o)=-H_G^{-1}\sqrt{G}\mathbb{P}_{bG}\bar{\psi}(W_g;\theta_o)+o_p(1); \tag*{(3)} G ( θ S A C R − θ o ) = − H G − 1 G P b G ψ ˉ ( W g ; θ o ) + o p ( 1 ) ; (3)
(b) and
V G − 1 / 2 G ( θ ^ S A C R − θ o ) ⟶ d N ( 0 , I ) . (4) V_G^{-1/2}\sqrt{G}(\widehat{\theta}^{SACR}-\theta_o)\mathrel{\stackrel{d}{\longrightarrow}}\mathcal{N}(0,I). \tag*{(4)} V G − 1/2 G ( θ S A C R − θ o ) ⟶ d N ( 0 , I ) . (4)
Theorem 3.2(a) expresses G ( θ ^ S A C R − θ o ) \sqrt{G}(\widehat{\theta}^{SACR}-\theta_o) G ( θ S A C R − θ o ) , to first order, as a scaled average of the mean-zero cluster-average scores ψ ˉ ( W g ; θ o ) \bar{\psi}(\mathbf{W}_g;\theta_o) ψ ˉ ( W g ; θ o ) . Since these summands are independent across clusters by Assumption 1, Theorem 3.2(b) supplies their pointwise CLT. Importantly, V G V_G V G can be a drifting covariance, for which neither Ω G \Omega_G Ω G nor H G H_G H G need converge to a fixed matrix; the covariance sequence need only remain well behaved under Assumption 5(b).
For the IV running example, the remaining first-order checks take a particularly simple form.
Running Example (continued: IV linearisation and covariance). For the just-identified IV score, H G = − Q G H_G=-Q_G H G = − Q G , ψ ˉ ( W g ; θ o ) = N g − 1 ∑ i ∈ G g Z i U i \bar{\psi}(\mathbf{W}_g;\theta_o)=N_g^{-1}\sum_{i\in\mathcal{G}_g}Z_iU_i ψ ˉ ( W g ; θ o ) = N g − 1 ∑ i ∈ G g Z i U i , and Ω G = G − 1 ∑ g = 1 G E G [ ψ ˉ ( W g ; θ o ) ψ ˉ ( W g ; θ o ) ′ ] \Omega_G=G^{-1}\sum_{g=1}^{G}\mathbb{E}_G[\bar{\psi}(\mathbf{W}_g;\theta_o)\bar{\psi}(\mathbf{W}_g;\theta_o)'] Ω G = G − 1 ∑ g = 1 G E G [ ψ ˉ ( W g ; θ o ) ψ ˉ ( W g ; θ o ) ′ ] . Because the score is affine in θ \theta θ , its population linearisation P G ψ ˉ ( W g ; θ ) = P G ψ ˉ ( W g ; θ o ) − Q G ( θ − θ o ) P_G\bar{\psi}(\mathbf{W}_g;\theta)=P_G\bar{\psi}(\mathbf{W}_g;\theta_o)-Q_G(\theta-\theta_o) P G ψ ˉ ( W g ; θ ) = P G ψ ˉ ( W g ; θ o ) − Q G ( θ − θ o ) is exact. The generic covariance matrix therefore reduces to the familiar just-identified IV sandwich V G = Q G − 1 Ω G ( Q G − 1 ) ′ V_G=Q_G^{-1}\Omega_G(Q_G^{-1})' V G = Q G − 1 Ω G ( Q G − 1 ) ′ . The relevance condition established in the running example’s zero-preservation and separation discussion verifies the Jacobian lower bound in Assumption 5(b) for all sufficiently large G G G , which suffices for the asymptotic arguments. The separate lower bound on λ min ( Ω G ) \lambda_{\min}(\Omega_G) λ m i n ( Ω G ) prevents degeneracy of the cluster-average score variance.
Running Example (continued: IV stochastic equicontinuity). Suppose additionally that
sup G ≥ 1 sup i ≤ N ( G ) E G ∥ Z i X i ′ ∥ F 2 ≤ C < ∞ , \sup_{G\geq1}\sup_{i\leq N(G)}\mathbb{E}_G\lVert Z_iX_i'\rVert_F^2\leq C<\infty, G ≥ 1 sup i ≤ N ( G ) sup E G ∥ Z i X i ′ ∥ F 2 ≤ C < ∞ ,
where ∥ ⋅ ∥ F \lVert\cdot\rVert_F ∥ ⋅ ∥ F *is the Frobenius norm. The affine IV score satisfies ψ ˉ ( W g ; θ ) − ψ ˉ ( W g ; θ o ) = − Q ˉ g ( θ − θ o ) \bar{\psi}(\mathbf{W}_g;\theta)-\bar{\psi}(\mathbf{W}_g;\theta_o)=-\bar{Q}_g(\theta-\theta_o) ψ ˉ ( W g ; θ ) − ψ ˉ ( W g ; θ o ) = − Q ˉ g ( θ − θ o ) . Consequently,
sup ∥ θ − θ o ∥ ≤ δ ∥ G G ( θ ) − G G ( θ o ) ∥ = sup ∥ θ − θ o ∥ ≤ δ ∥ G ( P ^ G − P G ) Q ˉ g ( θ − θ o ) ∥ ≤ δ ∥ G ( P ^ G − P G ) Q ˉ g ∥ F = O p ( δ ) , \sup_{\lVert\theta-\theta_o\rVert\leq\delta}\lVert\mathbb{G}_G(\theta)-\mathbb{G}_G(\theta_o)\rVert
=
\sup_{\lVert\theta-\theta_o\rVert\leq\delta}
\left\lVert\sqrt{G}(\widehat{P}_G-P_G)\bar{Q}_g(\theta-\theta_o)\right\rVert
\leq
\delta\left\lVert\sqrt{G}(\widehat{P}_G-P_G)\bar{Q}_g\right\rVert_F
=O_p(\delta), ∥ θ − θ o ∥ ≤ δ sup ∥ G G ( θ ) − G G ( θ o )∥ = ∥ θ − θ o ∥ ≤ δ sup G ( P G − P G ) Q ˉ g ( θ − θ o ) ≤ δ G ( P G − P G ) Q ˉ g F = O p ( δ ) ,
Indeed, cluster independence and Jensen’s inequality give
E G ∥ G ( P ^ G − P G ) Q ˉ g ∥ F 2 ≤ 1 G ∑ g = 1 G 1 N g ∑ i ∈ G g E G ∥ Z i X i ′ ∥ F 2 ≤ C . \mathbb{E}_G\left\lVert\sqrt{G}(\widehat{P}_G-P_G)\bar{Q}_g\right\rVert_F^2
\leq
\frac{1}{G}\sum_{g=1}^{G}\frac{1}{N_g}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\lVert Z_iX_i'\rVert_F^2
\leq C. E G G ( P G − P G ) Q ˉ g F 2 ≤ G 1 g = 1 ∑ G N g 1 i ∈ G g ∑ E G ∥ Z i X i ′ ∥ F 2 ≤ C .
Markov’s inequality therefore bounds the probability in Assumption 6 by C δ 2 / ε 2 C\delta^2/\varepsilon^2 C δ 2 / ε 2 , which vanishes as δ ↓ 0 \delta\downarrow0 δ ↓ 0 .
Remark 3.1 (Structural role of SACR reweighting). *The conventional observation-weighted moment can be written as
1 N ∑ g = 1 G ∑ i ∈ G g ψ ( W i ; θ ) = ∑ g = 1 G N g N ψ ˉ ( W g ; θ ) , \frac{1}{N}\sum_{g=1}^{G}\sum_{i\in G_g}\psi(W_i;\theta)=\sum_{g=1}^{G}\frac{N_g}{N}\bar{\psi}(W_g;\theta), N 1 g = 1 ∑ G i ∈ G g ∑ ψ ( W i ; θ ) = g = 1 ∑ G N N g ψ ˉ ( W g ; θ ) ,
so a large cluster receives weight N g / N N_g/N N g / N and may remain first-order influential. By contrast, the SACR moment G − 1 ∑ g = 1 G ψ ˉ ( W g ; θ ) G^{-1}\sum_{g=1}^{G}\bar{\psi}(W_g;\theta) G − 1 ∑ g = 1 G ψ ˉ ( W g ; θ ) assigns every cluster the same deterministic weight 1 / G 1/G 1/ G . The effective triangular array therefore contains one independent cluster-average summand per cluster, with maximal deterministic weight 1 / G 1/G 1/ G . Treating each cluster average as one effective observation then makes the Hansen–Lee cluster-size restrictions [8 Assumption 2] automatic. This is the structural reason SACR removes cluster-size dominance and yields the G \sqrt{G} G -linear representation in Theorem 3.2.
The G \sqrt{G} G -normalisation does not by itself establish an efficiency cost relative to conventional estimation because under clustered dependence, the latter need not attain a N \sqrt{N} N rate. If all clusters have the same size, the SACR and conventional estimating equations differ only by a constant multiplier and have the same roots. With unequal sizes, cluster balancing can sacrifice efficiency when additional observations within a cluster provide substantial information, but can improve efficiency when strong within-cluster dependence makes that information largely redundant. There is therefore no general efficiency ranking; the comparison depends on cluster sizes, dependence, and the score and Jacobian structure.
The above representation motivates the natural plug-in estimator V ^ G S A C R = H ^ G − 1 Ω ^ G ( H ^ G − 1 ) ′ \widehat{V}^{SACR}_{G}=\widehat{H}_{G}^{-1}\widehat{\Omega}_{G}(\widehat{H}_{G}^{-1})^{\prime} V G S A C R = H G − 1 Ω G ( H G − 1 ) ′ , where H ^ G : = 1 G ∑ g = 1 G H ^ g \widehat{H}_{G}:=\frac{1}{G}\sum_{g=1}^{G}\widehat{H}_{g} H G := G 1 ∑ g = 1 G H g and H ^ g : = N g − 1 ∑ i ∈ G g ∂ ∂ θ ′ ψ ( W i ; θ ^ S A C R ) \widehat{H}_{g}:=N_{g}^{-1}\sum_{i\in G_{g}}\frac{\partial}{\partial\theta^{\prime}}\psi(W_i;\widehat{\theta}^{SACR}) H g := N g − 1 ∑ i ∈ G g ∂ θ ′ ∂ ψ ( W i ; θ S A C R ) , for smooth scores, with the usual model-specific population-Jacobian analogue used for non-smooth scores, and
Ω ^ G : = 1 G ∑ g = 1 G ( ψ ˉ ( W g ; θ ^ S A C R ) − P ^ G ψ ˉ ( W g ; θ ^ S A C R ) ) ( ψ ˉ ( W g ; θ ^ S A C R ) − P ^ G ψ ˉ ( W g ; θ ^ S A C R ) ) ′ . \widehat{\Omega}_{G}:=\frac{1}{G}\sum_{g=1}^{G}\left(\bar{\psi}(W_g;\widehat{\theta}^{SACR})-\widehat{P}_{G}\bar{\psi}(W_g;\widehat{\theta}^{SACR})\right)\left(\bar{\psi}(W_g;\widehat{\theta}^{SACR})-\widehat{P}_{G}\bar{\psi}(W_g;\widehat{\theta}^{SACR})\right)^{\prime}. Ω G := G 1 g = 1 ∑ G ( ψ ˉ ( W g ; θ S A C R ) − P G ψ ˉ ( W g ; θ S A C R ) ) ( ψ ˉ ( W g ; θ S A C R ) − P G ψ ˉ ( W g ; θ S A C R ) ) ′ .
Under model-specific consistency conditions for the Jacobian and score-covariance estimators, this estimates the covariance of G ( θ ^ S A C R − θ o ) \sqrt{G}(\widehat{\theta}^{SACR}-\theta_o) G ( θ S A C R − θ o ) ; the covariance of θ ^ S A C R \widehat{\theta}^{SACR} θ S A C R itself is G − 1 V ^ G S A C R G^{-1}\widehat{V}^{SACR}_{G} G − 1 V G S A C R . The Z-SACR modification enters through the cluster-size-adjusted score covariance, since ψ ˉ ( W g ; θ ^ S A C R ) \bar{\psi}(W_g;\widehat{\theta}^{SACR}) ψ ˉ ( W g ; θ S A C R ) is already a cluster-average score.
Functional consistency and weak convergence
The ULC problem is increasingly important in clustered sampling settings involving economic parameters of interest such as distributional effects and tests of stochastic dominance, which are inherently functional in nature. It therefore becomes useful to extend the asymptotic theory from a finite-dimensional point parameter θ o ∈ R p θ \theta_o\in\mathbb{R}^{p_\theta} θ o ∈ R p θ to a functional parameter θ o ( τ ) \theta_o(\tau) θ o ( τ ) , τ ∈ T \tau\in\mathcal{T} τ ∈ T . Write the primitive estimating equation process as
Ψ ( W ; θ ) : = { ψ τ ( W ; θ ) : τ ∈ T } , \Psi(W;\theta):=\{\psi_\tau(W;\theta):\tau\in\mathcal{T}\}, Ψ ( W ; θ ) := { ψ τ ( W ; θ ) : τ ∈ T } ,
where T T T denotes a compact index set. The population moment restriction is P Ψ ( W ; θ o ) = 0 \mathbb{P}\Psi(W;\theta_o)=0 P Ψ ( W ; θ o ) = 0 in ℓ ∞ ( T ) \ell^\infty(T) ℓ ∞ ( T ) . At the cluster level, define the SACR process
Ψ ˉ ( W g ; θ ) : = { ψ ˉ τ ( W g ; θ ( τ ) ) : τ ∈ T } where ψ ˉ τ ( W g ; θ ( τ ) ) : = N g − 1 ∑ i ∈ G g ψ τ ( W i ; θ ( τ ) ) . \bar{\Psi}(W_g;\theta):=\{\bar{\psi}_{\tau}(W_g;\theta(\tau)):\tau\in T\}\quad\text{where}\quad\bar{\psi}_{\tau}(W_g;\theta(\tau)):=N_g^{-1}\sum_{i\in\mathcal{G}_g}\psi_{\tau}(W_i;\theta(\tau)). Ψ ˉ ( W g ; θ ) := { ψ ˉ τ ( W g ; θ ( τ )) : τ ∈ T } where ψ ˉ τ ( W g ; θ ( τ )) := N g − 1 i ∈ G g ∑ ψ τ ( W i ; θ ( τ )) .
Thus, P ^ G Ψ ˉ ( W g ; θ ) \widehat{\mathbb{P}}_G\bar{\Psi}(W_g;\theta) P G Ψ ˉ ( W g ; θ ) and P G Ψ ˉ ( W g ; θ ) \mathbb{P}_G\bar{\Psi}(W_g;\theta) P G Ψ ˉ ( W g ; θ ) are the functional analogues of the cluster-average moments used in Theorem 3.2. § 4 supplies the concrete DR, QR, and ER applications with the model-specific index notation.
The functional theory uses three related (score) function classes. The global class indexed jointly by ( τ , t ) (\tau,t) ( τ , t ) delivers the functional UWLLN, the class evaluated at the true path delivers the functional CLT, and a localised increment class delivers first-order stochastic equicontinuity. Let ψ τ , j \psi_{\tau,j} ψ τ , j denote the j j j th coordinate of ψ τ \psi_\tau ψ τ , and define
F j : = { ψ τ , j ( ⋅ ; t ) : τ ∈ T , t ∈ Θ } , F o , j : = { ψ τ , j ( ⋅ ; θ o ( τ ) ) : τ ∈ T } , \mathcal{F}_j:=\{\psi_{\tau,j}(\cdot;t):\tau\in T,\ t\in\Theta\},\qquad\mathcal{F}_{o,j}:=\{\psi_{\tau,j}(\cdot;\theta_o(\tau)):\tau\in T\}, F j := { ψ τ , j ( ⋅ ; t ) : τ ∈ T , t ∈ Θ } , F o , j := { ψ τ , j ( ⋅ ; θ o ( τ )) : τ ∈ T } ,
and, for δ > 0 \delta>0 δ > 0 ,
F j Δ ( δ ) : = { ψ τ , j ( ⋅ ; t ) − ψ τ , j ( ⋅ ; θ o ( τ ) ) : τ ∈ T , ∥ t − θ o ( τ ) ∥ ≤ δ } . \mathcal{F}^{\Delta}_j(\delta):=\{\psi_{\tau,j}(\cdot;t)-\psi_{\tau,j}(\cdot;\theta_o(\tau)):\tau\in T,\ \lVert t-\theta_o(\tau)\rVert\le\delta\}. F j Δ ( δ ) := { ψ τ , j ( ⋅ ; t ) − ψ τ , j ( ⋅ ; θ o ( τ )) : τ ∈ T , ∥ t − θ o ( τ )∥ ≤ δ } .
For every coordinate j j j , let F j F_j F j be a measurable envelope of F j \mathcal{F}_j F j , so that
sup τ ∈ T , t ∈ Θ ∣ ψ τ , j ( w ; t ) ∣ ≤ F j ( w ) , F ˉ G g , j : = { N g − 1 ∑ i ∈ G g F j ( W i ) 2 } 1 / 2 , \sup_{\tau\in T,\ t\in\Theta}|\psi_{\tau,j}(w;t)|\le F_j(w),\qquad\bar{F}_{Gg,j}:=\left\{N_g^{-1}\sum_{i\in\mathcal{G}_g}F_j(W_i)^2\right\}^{1/2}, τ ∈ T , t ∈ Θ sup ∣ ψ τ , j ( w ; t ) ∣ ≤ F j ( w ) , F ˉ G g , j := ⎩ ⎨ ⎧ N g − 1 i ∈ G g ∑ F j ( W i ) 2 ⎭ ⎬ ⎫ 1/2 ,
and define the normalised cluster array and its envelope by
Z G g , j ( τ , t ) : = G − 1 / 2 ψ ˉ τ , j ( W g ; t ) , F G g , j : = G − 1 / 2 F ˉ G g , j . Z_{Gg,j}(\tau,t):=G^{-1/2}\bar{\psi}_{\tau,j}(W_g;t),\qquad F_{Gg,j}:=G^{-1/2}\bar{F}_{Gg,j}. Z G g , j ( τ , t ) := G − 1/2 ψ ˉ τ , j ( W g ; t ) , F G g , j := G − 1/2 F ˉ G g , j .
Assumption 2’ (Functional manageability and moments).
(a) For every coordinate j j j , the array { Z G g , j ( τ , t ) : ( τ , t ) ∈ T × Θ , g ≤ G } \{Z_{Gg,j}(\tau,t):(\tau,t)\in T\times\Theta,\ g\le G\} { Z G g , j ( τ , t ) : ( τ , t ) ∈ T × Θ , g ≤ G } is manageable with respect to { F G g , j : g ≤ G } \{F_{Gg,j}:g\le G\} { F G g , j : g ≤ G } , uniformly in G G G . Each coordinate is measurable at every fixed index. The full array, its truth-indexed restriction, and, for every δ > 0 \delta>0 δ > 0 , the local difference array Z G g , j ( τ , t ) − Z G g , j ( τ , θ o ( τ ) ) Z_{Gg,j}(\tau,t)-Z_{Gg,j}(\tau,\theta_o(\tau)) Z G g , j ( τ , t ) − Z G g , j ( τ , θ o ( τ )) indexed by { ( τ , t ) ∈ T × Θ : ∥ t − θ o ( τ ) ∥ ≤ δ } \{(\tau,t)\in T\times\Theta:\lVert t-\theta_o(\tau)\rVert\le\delta\} {( τ , t ) ∈ T × Θ : ∥ t − θ o ( τ )∥ ≤ δ } are almost measurable Suslin (AMS), with common Suslin approximating index sets and jointly measurable parametrisations across coordinates.
(b) For every coordinate j j j , the squared individual envelopes are uniformly integrable:
lim M → ∞ sup G ≥ 1 sup g ≤ G sup i ∈ G g E G [ F j ( W i ) 2 1 { F j ( W i ) 2 > M } ] = 0. \lim_{M\to\infty}\sup_{G\ge1}\sup_{g\le G}\sup_{i\in\mathcal{G}_g}\mathbb{E}_G[F_j(W_i)^2\mathbf{1}\{F_j(W_i)^2>M\}]=0. M → ∞ lim G ≥ 1 sup g ≤ G sup i ∈ G g sup E G [ F j ( W i ) 2 1 { F j ( W i ) 2 > M }] = 0.
(c) For every coordinate j j j ,
lim δ ↓ 0 lim sup G → ∞ sup τ ∈ T , t ∈ Θ : ∥ t − θ o ( τ ) ∥ ≤ δ 1 G ∑ g = 1 G E G [ ( ψ ˉ τ , j ( W g ; t ) − ψ ˉ τ , j ( W g ; θ o ( τ ) ) ) 2 ] = 0. \lim_{\delta\downarrow0}\limsup_{G \to\infty}\sup_{\tau\in\mathcal{T},\ t \in\Theta:\ \lVert t-\theta_o(\tau)\rVert\le\delta}\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[\left(\bar{\psi}_{\tau,j}(W_g;t)-\bar{\psi}_{\tau,j}(W_g;\theta_o(\tau))\right)^2\right]=0. δ ↓ 0 lim G → ∞ lim sup τ ∈ T , t ∈ Θ : ∥ t − θ o ( τ )∥ ≤ δ sup G 1 g = 1 ∑ G E G [ ( ψ ˉ τ , j ( W g ; t ) − ψ ˉ τ , j ( W g ; θ o ( τ )) ) 2 ] = 0.
Manageability in Assumption 2’(a) is understood in the sense of Pollard [15 ] . The almost-measurable-Suslin condition is that of Kosorok [11 ] , applied jointly to the finitely many coordinates with the approximation error measured by the sum of squared differences over both clusters and coordinates. A sufficient verification is joint separability; see Kosorok [10 ] .
Let θ o ( ⋅ ) \theta_o(\cdot) θ o ( ⋅ ) denote the true functional parameter. The object of interest is the process
G ( θ ^ − θ o ) ( ⋅ ) , \sqrt{G}(\hat{\theta}-\theta_o)(\cdot), G ( θ ^ − θ o ) ( ⋅ ) ,
viewed as a random element of ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) . For a parameter path θ ( ⋅ ) \theta(\cdot) θ ( ⋅ ) , extend the first-order process notation pointwise in τ \tau τ by
G G ( θ ) ( τ ) : = G ( P ^ G − P G ) ψ ˉ τ ( W g ; θ ( τ ) ) . \mathbb{G}_G(\theta)(\tau):=\sqrt{G}(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\psi}_{\tau}(W_g;\theta(\tau)). G G ( θ ) ( τ ) := G ( P G − P G ) ψ ˉ τ ( W g ; θ ( τ )) .
Throughout this subsection, Θ T \Theta_{\mathcal{T}} Θ T denotes the admissible collection of parameter paths equipped with the norm ∥ θ ∥ T : = sup τ ∈ T ∥ θ ( τ ) ∥ \lVert\theta\rVert_{\mathcal{T}}:=\sup_{\tau\in\mathcal{T}}\lVert\theta(\tau)\rVert ∥ θ ∥ T := sup τ ∈ T ∥ θ ( τ )∥ , with each admissible path satisfying θ ( τ ) ∈ Θ \theta(\tau)\in\Theta θ ( τ ) ∈ Θ uniformly over τ ∈ T \tau\in\mathcal{T} τ ∈ T .
Lemma 3.1 (Functional UWLLN and local first-order stochastic equicontinuity). Under Assumption 1 and assumption 2’,
(a) sup θ ∈ Θ T ∥ ( P ^ G − P G ) Ψ ˉ ( W g ; θ ) ∥ T = o p ( 1 ) \sup_{\theta\in\Theta_{\mathcal{T}}}\lVert(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\Psi}(W_g;\theta)\rVert_{\mathcal{T}}=o_p(1) sup θ ∈ Θ T ∥( P G − P G ) Ψ ˉ ( W g ; θ ) ∥ T = o p ( 1 ) ; and
(b) for every ε > 0 \varepsilon>0 ε > 0 ,
lim δ ↓ 0 lim sup G → ∞ P [ sup ∥ θ − θ o ∥ T < δ ∥ G G ( θ ) − G G ( θ o ) ∥ T > ε ] = 0. (5) \lim_{\delta\downarrow0}\limsup_{G\to\infty}\mathbb{P}\left[\sup_{\lVert\theta-\theta_o\rVert_{\mathcal{T}}<\delta}\lVert\mathbb{G}_G(\theta)-\mathbb{G}_G(\theta_o)\rVert_{\mathcal{T}}>\varepsilon\right]=0. \tag*{(5)} δ ↓ 0 lim G → ∞ lim sup P [ ∥ θ − θ o ∥ T < δ sup ∥ G G ( θ ) − G G ( θ o ) ∥ T > ε ] = 0. (5)
Part (a) is the functional counterpart of Theorem 3.1(a). Part (b) strengthens the pointwise local condition in Assumption 6 by making its first-order control uniform over τ \tau τ .
The remaining functional identification requirements concern the population map itself.
Assumption 3’ (Functional identification and separation).
(a) The primitive functional moment restriction is preserved through cluster-level centring along the triangular array: E G [ ψ ˉ τ ( W g ; θ o ( τ ) ) ] = 0 \mathbb{E}_G[\bar{\psi}_{\tau}(W_g;\theta_o(\tau))]=0 E G [ ψ ˉ τ ( W g ; θ o ( τ ))] = 0 , g = 1 , … , G g=1,\ldots,G g = 1 , … , G , τ ∈ T \tau\in\mathcal{T} τ ∈ T .
(b) For every $\epsilon>0,
lim inf G → ∞ inf θ ∈ Θ T : ∥ θ − θ o ∥ T ≥ ϵ ∥ P G Ψ ˉ ( W g ; θ ) ∥ T > 0. \liminf_{G\to\infty}\inf_{\theta\in\Theta_{\mathcal{T}}:\lVert\theta-\theta_o\rVert_{\mathcal{T}}\ge\epsilon}\lVert\mathbb{P}_G\bar{\Psi}(W_g;\theta)\rVert_{\mathcal{T}}>0. G → ∞ lim inf θ ∈ Θ T : ∥ θ − θ o ∥ T ≥ ϵ inf ∥ P G Ψ ˉ ( W g ; θ ) ∥ T > 0.
The following functional linearisation conditions enable the first-order transfer from the score process to the estimator.
Assumption 4′ (Functional asymptotic linearisation).
(a) The process estimator θ ^ ( ⋅ ) \hat{\theta}(\cdot) θ ^ ( ⋅ ) satisfies
P ^ G Ψ ‾ ( W g ; θ ^ ) = o p ( G − 1 / 2 ) in ℓ ∞ ( T ) . (6) \widehat{\mathbb{P}}_{G}\overline{\Psi}(\mathbf{W}_{g};\hat{\theta}) = o_{p}(G^{-1/2}) \quad\text{in }\ell^{\infty}(\mathcal{T}). \tag*{(6)} P G Ψ ( W g ; θ ^ ) = o p ( G − 1/2 ) in ℓ ∞ ( T ) . (6) (b) The true path θ o ( ⋅ ) \theta_{o}(\cdot) θ o ( ⋅ ) is uniformly interior to Θ \Theta Θ : inf τ ∈ T dist ( θ o ( τ ) , ∂ Θ ) > 0 \inf_{\tau\in\mathcal{T}}\operatorname{dist}(\theta_{o}(\tau),\partial\Theta)>0 inf τ ∈ T dist ( θ o ( τ ) , ∂ Θ ) > 0 , where dist ( ⋅ , ∂ Θ ) \operatorname{dist}(\cdot,\partial\Theta) dist ( ⋅ , ∂ Θ ) denotes Euclidean distance to the boundary of Θ \Theta Θ .
(c) The map t ↦ P G ψ ‾ τ ( W g ; t ) t\mapsto\mathbb{P}_{G}\overline{\psi}_{\tau}(\mathbf{W}_{g};t) t ↦ P G ψ τ ( W g ; t ) is continuously differentiable in a neighbourhood of θ o ( τ ) \theta_{o}(\tau) θ o ( τ ) , uniformly over τ ∈ T \tau\in\mathcal{T} τ ∈ T , with H G ( τ ) : = ∂ ∂ t ′ P G ψ ‾ τ ( W g ; t ) ∣ t = θ o ( τ ) H_{G}(\tau):=\left.\frac{\partial}{\partial t^{\prime}}\mathbb{P}_{G}\overline{\psi}_{\tau}(\mathbf{W}_{g};t)\right|_{t=\theta_{o}(\tau)} H G ( τ ) := ∂ t ′ ∂ P G ψ τ ( W g ; t ) t = θ o ( τ ) , and
lim δ ↓ 0 lim sup G → ∞ sup τ ∈ T sup ∥ t − θ o ( τ ) ∥ ≤ δ ∥ ∂ ∂ t ′ P G ψ ‾ τ ( W g ; t ) − H G ( τ ) ∥ = 0. \lim_{\delta\downarrow0}\limsup_{G\to\infty}\sup_{\tau\in\mathcal{T}}\sup_{\lVert t-\theta_{o}(\tau)\rVert\leq\delta}\left\lVert\frac{\partial}{\partial t^{\prime}}\mathbb{P}_{G}\overline{\psi}_{\tau}(\mathbf{W}_{g};t)-H_{G}(\tau)\right\rVert=0. δ ↓ 0 lim G → ∞ lim sup τ ∈ T sup ∥ t − θ o ( τ )∥ ≤ δ sup ∂ t ′ ∂ P G ψ τ ( W g ; t ) − H G ( τ ) = 0. Moreover, H G ( τ ) → H ( τ ) H_{G}(\tau)\to H(\tau) H G ( τ ) → H ( τ ) uniformly on T \mathcal{T} T , and, for some λ > 0 \lambda>0 λ > 0 , inf τ ∈ T λ min ( H ( τ ) ′ H ( τ ) ) ≥ λ 2 \inf_{\tau\in\mathcal{T}}\lambda_{\min}(H(\tau)^{\prime}H(\tau))\geq\lambda^{2} inf τ ∈ T λ m i n ( H ( τ ) ′ H ( τ )) ≥ λ 2 .
For the truth-indexed score process, write
Ω G ( τ , τ ′ ) : = 1 G ∑ g = 1 G E G [ ψ ‾ τ ( W g ; θ o ( τ ) ) ψ ‾ τ ′ ( W g ; θ o ( τ ′ ) ) ′ ] . \Omega_{G}(\tau,\tau^{\prime}):=\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_{G}\left[\overline{\psi}_{\tau}(\mathbf{W}_{g};\theta_{o}(\tau))\overline{\psi}_{\tau^{\prime}}(\mathbf{W}_{g};\theta_{o}(\tau^{\prime}))^{\prime}\right]. Ω G ( τ , τ ′ ) := G 1 g = 1 ∑ G E G [ ψ τ ( W g ; θ o ( τ )) ψ τ ′ ( W g ; θ o ( τ ′ ) ) ′ ] .
For a ∈ R p θ a\in\mathbb{R}^{p_{\theta}} a ∈ R p θ , define the semimetric
ρ G , a ( τ , τ ′ ) : = [ 1 G ∑ g = 1 G E G [ ( a ′ { ψ ‾ τ ( W g ; θ o ( τ ) ) − ψ ‾ τ ′ ( W g ; θ o ( τ ′ ) ) } ) 2 ] ] 1 / 2 . \rho_{G,a}(\tau,\tau^{\prime}):=\left[\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_{G}\left[\left(a^{\prime}\left\{\overline{\psi}_{\tau}(\mathbf{W}_{g};\theta_{o}(\tau))-\overline{\psi}_{\tau^{\prime}}(\mathbf{W}_{g};\theta_{o}(\tau^{\prime}))\right\}\right)^{2}\right]\right]^{1/2}. ρ G , a ( τ , τ ′ ) := [ G 1 g = 1 ∑ G E G [ ( a ′ { ψ τ ( W g ; θ o ( τ )) − ψ τ ′ ( W g ; θ o ( τ ′ )) } ) 2 ] ] 1/2 .
For any positive semidefinite matrix-valued kernel K K K , let G K \mathbb{G}_{K} G K denote a centred Gaussian process with covariance kernel K K K , and define its semimetric by
ρ K , a ( τ , τ ′ ) : = ∥ a ′ ( G K ( τ ) − G K ( τ ′ ) ) ∥ L 2 = [ a ′ { K ( τ , τ ) + K ( τ ′ , τ ′ ) − K ( τ , τ ′ ) − K ( τ ′ , τ ) } a ] 1 / 2 . \rho_{K,a}(\tau,\tau^{\prime}):=\lVert a^{\prime}(\mathbb{G}_{K}(\tau)-\mathbb{G}_{K}(\tau^{\prime}))\rVert_{L^{2}}=\left[a^{\prime}\{K(\tau,\tau)+K(\tau^{\prime},\tau^{\prime})-K(\tau,\tau^{\prime})-K(\tau^{\prime},\tau)\}a\right]^{1/2}. ρ K , a ( τ , τ ′ ) := ∥ a ′ ( G K ( τ ) − G K ( τ ′ )) ∥ L 2 = [ a ′ { K ( τ , τ ) + K ( τ ′ , τ ′ ) − K ( τ , τ ′ ) − K ( τ ′ , τ )} a ] 1/2 .
Assumption 5′ (Functional covariance kernel).
(a) The covariance kernels Ω G \Omega_{G} Ω G converge pointwise on T × T \mathcal{T}\times\mathcal{T} T × T to a positive semidefinite kernel Ω \Omega Ω .
(b) For any deterministic sequences τ G , τ G ′ ∈ T \tau_{G},\tau_{G}^{\prime}\in\mathcal{T} τ G , τ G ′ ∈ T and a G ∈ R p θ a_{G}\in\mathbb{R}^{p_{\theta}} a G ∈ R p θ with ∥ a G ∥ ≤ 1 \lVert a_{G}\rVert\leq1 ∥ a G ∥ ≤ 1 , ρ Ω , a G ( τ G , τ G ′ ) → 0 \rho_{\Omega,a_{G}}(\tau_{G},\tau_{G}^{\prime})\to0 ρ Ω , a G ( τ G , τ G ′ ) → 0 implies ρ G , a G ( τ G , τ G ′ ) → 0 \rho_{G,a_{G}}(\tau_{G},\tau_{G}^{\prime})\to0 ρ G , a G ( τ G , τ G ′ ) → 0 .
Assumption 5′ supplies the covariance and metric conditions in Kosorok (2003, Theorem 1), equivalently Kosorok (2007, Theorem 11.16). It is intentionally stronger than the condition required for
Theorem 3.2. The pointwise result permits a drifting covariance sequence and achieves asymptotic pivotality through studentisation. By contrast, weak convergence in ℓ ∞ ( T ) \ell^\infty(T) ℓ ∞ ( T ) to a fixed Gaussian process requires convergence of the covariance kernel itself.
Assumptions 2’ and 5’, together with Assumption 1, provide the functional CLT for the cluster empirical process at θ o ( ⋅ ) \theta_o(\cdot) θ o ( ⋅ ) . For a scalar-valued target path, such as a quantile function, distribution function, quantile or distribution treatment effect, let η o ( τ ) : = φ τ ( θ o ( τ ) ) \eta_o(\tau) := \varphi_\tau(\theta_o(\tau)) η o ( τ ) := φ τ ( θ o ( τ )) , η ^ ( τ ) : = φ τ ( θ ^ ( τ ) ) \hat{\eta}(\tau) := \varphi_\tau(\hat{\theta}(\tau)) η ^ ( τ ) := φ τ ( θ ^ ( τ )) , and define the induced map
Φ φ : ℓ ∞ ( T ) p θ → ℓ ∞ ( T ) , Φ φ ( θ ) ( τ ) : = φ τ ( θ ( τ ) ) . \Phi_\varphi: \ell^\infty(T)^{p_\theta} \to\ell^\infty(T), \qquad\Phi_\varphi(\theta)(\tau) := \varphi_\tau(\theta(\tau)). Φ φ : ℓ ∞ ( T ) p θ → ℓ ∞ ( T ) , Φ φ ( θ ) ( τ ) := φ τ ( θ ( τ )) .
Theorem 3.3 (Weak Convergence). Suppose that Assumptions 1, 2’, 3’, 4’, and 5’ hold. Then
(a) G ( θ ^ − θ o ) ( ⋅ ) ⇝ Z θ ( ⋅ ) : = − H ( ⋅ ) − 1 G 0 ( ⋅ ) \sqrt{G}(\hat{\theta}-\theta_o)(\cdot) \rightsquigarrow Z_\theta(\cdot) := -H(\cdot)^{-1}\mathbb{G}_0(\cdot) G ( θ ^ − θ o ) ( ⋅ ) ⇝ Z θ ( ⋅ ) := − H ( ⋅ ) − 1 G 0 ( ⋅ ) in ℓ ∞ ( T ) \ell^\infty(T) ℓ ∞ ( T ) , where G 0 \mathbb{G}_0 G 0 is the tight mean-zero Gaussian process with covariance kernel Ω ( τ , τ ′ ) \Omega(\tau,\tau') Ω ( τ , τ ′ ) from Assumption 5’.
(b) Additionally, suppose Φ φ \Phi_\varphi Φ φ is Hadamard differentiable at θ o \theta_o θ o , tangentially to a subspace containing the support of Z θ Z_\theta Z θ , with derivative Φ φ , θ o ′ \Phi'_{\varphi,\theta_o} Φ φ , θ o ′ . Then G ( η ^ − η o ) ( ⋅ ) ⇝ Φ φ , θ o ′ [ Z θ ] ( ⋅ ) \sqrt{G}(\hat{\eta}-\eta_o)(\cdot) \rightsquigarrow\Phi'_{\varphi,\theta_o}[Z_\theta](\cdot) G ( η ^ − η o ) ( ⋅ ) ⇝ Φ φ , θ o ′ [ Z θ ] ( ⋅ ) in ℓ ∞ ( T ) \ell^\infty(T) ℓ ∞ ( T ) .
Multiplier bootstrap and uniform inference
The weak-convergence result in Theorem 3.3 characterises the asymptotic distribution of the full parameter path τ ↦ θ ^ ( τ ) \tau\mapsto\hat{\theta}(\tau) τ ↦ θ ^ ( τ ) . Inference for such path-valued parameters is conveniently conducted via the bootstrap, e.g., [4 ] . It is therefore imperative to establish bootstrap validity in the same functional setting.
Let { ξ g : g ≥ 1 } \{\xi_g : g \ge1\} { ξ g : g ≥ 1 } be an i.i.d. sequence of centred bootstrap multipliers with a fixed law, independent of the data, and normalised so that E [ ξ g ] = 0 \mathbb{E}[\xi_g] = 0 E [ ξ g ] = 0 and E [ ξ g 2 ] = 1 \mathbb{E}[\xi_g^2] = 1 E [ ξ g 2 ] = 1 . Conditional on the sample, these equalities are also written E ∗ [ ξ g ] = 0 \mathbb{E}^*[\xi_g] = 0 E ∗ [ ξ g ] = 0 and E ∗ [ ξ g 2 ] = 1 \mathbb{E}^*[\xi_g^2] = 1 E ∗ [ ξ g 2 ] = 1 , where E ∗ \mathbb{E}^* E ∗ denotes conditional expectation given the sample; see [11 condition F] . When the estimator is computed from a weighted objective function or weighted estimating equation, it is often convenient to use non-negative cluster weights ξ ~ g \tilde{\xi}_g ξ ~ g with E ∗ [ ξ ~ g ] = 1 \mathbb{E}^*[\tilde{\xi}_g] = 1 E ∗ [ ξ ~ g ] = 1 and V a r ∗ ( ξ ~ g ) = 1 \mathop{\mathrm{Var}}^*(\tilde{\xi}_g) = 1 Var ∗ ( ξ ~ g ) = 1 ; the centred multiplier entering the first-order representation is then ξ g = ξ ~ g − 1 \xi_g = \tilde{\xi}_g - 1 ξ g = ξ ~ g − 1 .
A cluster-level multiplier perturbation of the empirical estimating equation is obtained by setting ξ ~ g = 1 + ξ g \widetilde{\xi}_g = 1 + \xi_g ξ g = 1 + ξ g and defining
Ψ ^ G ∗ ( θ ) ( τ ) : = P ^ G ψ ‾ τ ( W g ; θ ( τ ) ) + 1 G ∑ g = 1 G ξ g ψ ‾ τ ( W g ; θ ( τ ) ) = 1 G ∑ g = 1 G ξ ~ g ψ ‾ τ ( W g ; θ ( τ ) ) , τ ∈ T . \widehat{\Psi}_G^*(\theta)(\tau) := \widehat{P}_G\overline{\psi}_{\tau}(W_g;\theta(\tau)) + \frac{1}{G}\sum_{g=1}^{G}\xi_g\overline{\psi}_{\tau}(W_g;\theta(\tau)) = \frac{1}{G}\sum_{g=1}^{G}\widetilde{\xi}_g\overline{\psi}_{\tau}(W_g;\theta(\tau)), \quad\tau\in\mathcal{T}. Ψ G ∗ ( θ ) ( τ ) := P G ψ τ ( W g ; θ ( τ )) + G 1 g = 1 ∑ G ξ g ψ τ ( W g ; θ ( τ )) = G 1 g = 1 ∑ G ξ g ψ τ ( W g ; θ ( τ )) , τ ∈ T .
For nonnegative ξ ~ g \widetilde{\xi}_g ξ g , this is the estimating equation induced by weighting each cluster’s contribution to the original criterion or estimating equation. For general centred ξ g \xi_g ξ g , it is understood as a direct multiplier perturbation of the Z Z Z -equation. Let θ ^ ∗ ( ⋅ ) \widehat{\theta}^*(\cdot) θ ∗ ( ⋅ ) denote a corresponding bootstrap Z Z Z -SACR estimator satisfying
Ψ ^ G ∗ ( θ ^ ∗ ) = o P ∗ ( G − 1 / 2 ) in ℓ ∞ ( T ) \widehat{\Psi}_G^*(\widehat{\theta}^*) = o_{P^*}(G^{-1/2}) \quad\text{in } \ell^\infty(\mathcal{T}) Ψ G ∗ ( θ ∗ ) = o P ∗ ( G − 1/2 ) in ℓ ∞ ( T )
in outer probability. Define the corresponding first-order multiplier moment process by
M G ∗ ( θ ) ( τ ) : = G − 1 / 2 ∑ g = 1 G ξ g ψ ‾ τ ( W g ; θ ( τ ) ) , τ ∈ T . \mathbb{M}_G^*(\theta)(\tau) := G^{-1/2}\sum_{g=1}^{G}\xi_g\overline{\psi}_{\tau}(W_g;\theta(\tau)), \quad\tau\in\mathcal{T}. M G ∗ ( θ ) ( τ ) := G − 1/2 g = 1 ∑ G ξ g ψ τ ( W g ; θ ( τ )) , τ ∈ T .
The following theorem establishes bootstrap validity for the functional Z Z Z -SACR estimator and its Hadamard differentiable map.
Theorem 3.4 (Functional Bootstrap Validity). Suppose the conditions of Theorem 3.3(a) hold, and let the chosen multiplier weights satisfy the conditional independence, centring, and unit-variance conditions described above. Let η ^ ∗ ( τ ) : = φ τ ( θ ^ ∗ ( τ ) ) \widehat{\eta}^*(\tau) := \varphi_\tau(\widehat{\theta}^*(\tau)) η ∗ ( τ ) := φ τ ( θ ∗ ( τ )) . Then the following hold.
(a) Conditionally on the sample, G ( θ ^ ∗ − θ ^ ) ( ⋅ ) ⇝ P Z θ ( ⋅ ) \sqrt{G}(\widehat{\theta}^*-\widehat{\theta})(\cdot) \rightsquigarrow_P Z_\theta(\cdot) G ( θ ∗ − θ ) ( ⋅ ) ⇝ P Z θ ( ⋅ ) in ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) .
(b) Additionally, suppose Φ φ \Phi_\varphi Φ φ is Hadamard differentiable at θ o \theta_o θ o , tangentially to a subspace containing the support of Z θ Z_\theta Z θ , with derivative Φ φ , θ o ′ \Phi'_{\varphi,\theta_o} Φ φ , θ o ′ . Then, conditionally on the sample, G ( η ^ ∗ − η ^ ) ( ⋅ ) ⇝ P Φ φ , θ o ′ [ Z θ ] ( ⋅ ) \sqrt{G}(\widehat{\eta}^*-\widehat{\eta})(\cdot) \rightsquigarrow_P \Phi'_{\varphi,\theta_o}[Z_\theta](\cdot) G ( η ∗ − η ) ( ⋅ ) ⇝ P Φ φ , θ o ′ [ Z θ ] ( ⋅ ) in ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) .
Consequently, the conditional distributions of G ( θ ^ ∗ − θ ^ ) ( ⋅ ) \sqrt{G}(\hat{\theta}^{*}-\hat{\theta})(\cdot) G ( θ ^ ∗ − θ ^ ) ( ⋅ ) and G ( η ^ ∗ − η ^ ) ( ⋅ ) \sqrt{G}(\hat{\eta}^{*}-\hat{\eta})(\cdot) G ( η ^ ∗ − η ^ ) ( ⋅ ) consistently approximate the Gaussian limits of G ( θ ^ − θ o ) ( ⋅ ) \sqrt{G}(\hat{\theta}-\theta_o)(\cdot) G ( θ ^ − θ o ) ( ⋅ ) and G ( η ^ − η o ) ( ⋅ ) \sqrt{G}(\hat{\eta}-\eta_o)(\cdot) G ( η ^ − η o ) ( ⋅ ) .
Theorem 3.4(b) supplies the theoretical framework for inference on Hadamard differentiable maps of SACR estimators. The following remark specialises to distributional functionals and treatment effects as maps of distribution functions, which are the functional parameters of interest in Section S.6.
If φ τ \varphi_{\tau} φ τ is continuously differentiable in θ \theta θ , uniformly over τ ∈ T \tau\in\mathcal{T} τ ∈ T , then the derivative takes the pointwise form
Φ φ , θ o ′ [ h ] ( τ ) = φ ˙ τ ( θ o ( τ ) ) ′ h ( τ ) . \Phi_{\varphi,\theta_o}^{\prime}[h](\tau)=\dot{\varphi}_{\tau}(\theta_o(\tau))^{\prime}h(\tau). Φ φ , θ o ′ [ h ] ( τ ) = φ ˙ τ ( θ o ( τ ) ) ′ h ( τ ) .
Theorem 3.4 establishes bootstrap validity for the functional Z Z Z -SACR estimator under ULC. The next corollary specialises the functional-regime result to fixed values of τ \tau τ . Hence its limit uses the fixed functional limit H ( τ ) H(\tau) H ( τ ) from Theorem 3.3; the more general pointwise theory in Theorem 3.2 allows drifting H G H_G H G and Ω G \Omega_G Ω G .
Corollary 3.1 (Pointwise Bootstrap Validity). Under the conditions of Theorem 3.4, for each fixed τ ∈ T \tau\in\mathcal{T} τ ∈ T ,
G ( θ ^ ∗ ( τ ) − θ ^ ( τ ) ) ⇝ P − H ( τ ) − 1 G 0 ( τ ) . \sqrt{G}\left(\hat{\theta}^{*}(\tau)-\hat{\theta}(\tau)\right)\rightsquigarrow_{P}-H(\tau)^{-1}G_0(\tau). G ( θ ^ ∗ ( τ ) − θ ^ ( τ ) ) ⇝ P − H ( τ ) − 1 G 0 ( τ ) .
For uniform inference on the scalar path η o ( ⋅ ) \eta_o(\cdot) η o ( ⋅ ) , Theorem 3.4(b) provides a bootstrap approximation to the Gaussian limit of G ( η ^ − η o ) ( ⋅ ) \sqrt{G}(\hat{\eta}-\eta_o)(\cdot) G ( η ^ − η o ) ( ⋅ ) . Fix 0 < q 2 < q 1 < 1 0<q_2<q_1<1 0 < q 2 < q 1 < 1 , let z q z_q z q denote the q q q -quantile of a standard normal variable, and let q ^ q ∗ ( τ ) \hat{q}^{*}_q(\tau) q ^ q ∗ ( τ ) denote the conditional q q q -quantile of G { η ^ ∗ ( τ ) − η ^ ( τ ) } \sqrt{G}\{\hat{\eta}^{*}(\tau)-\hat{\eta}(\tau)\} G { η ^ ∗ ( τ ) − η ^ ( τ )} . Following Chernozhukov, Fernández-Val, and Melly [4 ] , define the quantile-spread studentiser
s ^ ( τ ) : = q ^ q 1 ∗ ( τ ) − q ^ q 2 ∗ ( τ ) z q 1 − z q 2 . \hat{s}(\tau):=\frac{\hat{q}^{*}_{q_1}(\tau)-\hat{q}^{*}_{q_2}(\tau)}{z_{q_1}-z_{q_2}}. s ^ ( τ ) := z q 1 − z q 2 q ^ q 1 ∗ ( τ ) − q ^ q 2 ∗ ( τ ) .
The implementation uses q 1 = 0.75 q_1=0.75 q 1 = 0.75 and q 2 = 0.25 q_2=0.25 q 2 = 0.25 , giving the normalised bootstrap interquartile range (IQR). Uniform confidence bands can then be constructed using the conditional quantiles of
T G ∗ : = sup τ ∈ T ∣ s ^ ( τ ) − 1 G { η ^ ∗ ( τ ) − η ^ ( τ ) } ∣ . T_G^{*}:=\sup_{\tau\in\mathcal{T}}\left|\hat{s}(\tau)^{-1}\sqrt{G}\{\hat{\eta}^{*}(\tau)-\hat{\eta}(\tau)\}\right|. T G ∗ := τ ∈ T sup s ^ ( τ ) − 1 G { η ^ ∗ ( τ ) − η ^ ( τ )} .
Let c 1 − α ∗ c^*_{1-\alpha} c 1 − α ∗ denote the conditional ( 1 − α ) (1-\alpha) ( 1 − α ) -quantile of T G ∗ T_G^* T G ∗ . Then
[ η ^ ( τ ) − G − 1 / 2 c 1 − α ∗ s ^ ( τ ) , η ^ ( τ ) + G − 1 / 2 c 1 − α ∗ s ^ ( τ ) ] , τ ∈ T , \left[\widehat{\eta}(\tau)-G^{-1/2}c^*_{1-\alpha}\widehat{s}(\tau),\widehat{\eta}(\tau)+G^{-1/2}c^*_{1-\alpha}\widehat{s}(\tau)\right],\qquad\tau\in\mathcal{T}, [ η ( τ ) − G − 1/2 c 1 − α ∗ s ( τ ) , η ( τ ) + G − 1/2 c 1 − α ∗ s ( τ ) ] , τ ∈ T ,
forms an asymptotically valid ( 1 − α ) (1-\alpha) ( 1 − α ) uniform confidence band for the path η o ( ⋅ ) \eta_o(\cdot) η o ( ⋅ ) . Under the conditions of Theorem 3.4(b), continuity of the limiting sample paths, and pointwise standard deviations bounded and bounded away from zero, Chernozhukov, Fernández-Val, and Melly [4 ] gives uniform consistency of this studentiser and validity of the resulting band.
Model-Specific Verification
This section verifies the theoretical framework established in the preceding section for distribution regression, quantile regression, and expectile regression estimating equations. The model-specific results establish identification and global score-class manageability with suitable envelopes. Together with the remaining generic sampling and regularity conditions, these verifications support the pointwise and functional inference results developed above. The supplementary empirical illustration in Section S.6 also considers quantile and expectile effects obtained from estimated distributions, a construction requiring additional distribution-to-functional map conditions; see Theorem 3.4(b) and Remark 3.4.
DR involves a smooth finite-dimensional score indexed by a threshold, QR introduces a non-smooth threshold score, and ER combines a threshold indicator with an unbounded residual envelope. These examples encompass binary response models, Least Absolute Deviations (LAD)/median regression, and OLS as special cases. The examples below are understood under the generic compactness and interiority requirements in Assumptions 2 and 4 and assumption 4′. The following assumption collects the remaining common linear-index restrictions used across the DR, QR, and ER examples. These conditions provide model-specific primitive counterparts to the rank and X X X -moment requirements entering Assumption 5.
Assumption LI 1 (Common linear-index conditions).
(a)
1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ ] is uniformly positive definite. \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'\right]\text{ is uniformly positive definite.} G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ is uniformly positive definite. (b) For some δ > 0 \delta>0 δ > 0 ,
sup G ≥ 1 sup 1 ≤ i ≤ N E G [ ∥ X i ∥ 2 + δ ] < ∞ . \sup_{G\geq1}\sup_{1\leq i\leq N}\mathbb{E}_G[\lVert X_i\rVert^{2+\delta}]<\infty. G ≥ 1 sup 1 ≤ i ≤ N sup E G [∥ X i ∥ 2 + δ ] < ∞.
Distribution regression
For distribution regression, the target is the conditional distribution function
F Y ∣ X ( y ∣ x ) = Λ ( x ′ θ o ( y ) ) , F_{Y\mid X}(y\mid x)=\Lambda(x'\theta_o(y)), F Y ∣ X ( y ∣ x ) = Λ ( x ′ θ o ( y )) ,
where Λ \Lambda Λ is a known link function, such as the logistic link, and y y y ranges over a compact subset Y \mathcal{Y} Y of the support of Y Y Y .6 The model implies the conditional moment restriction
P ψ y ( W ; θ o ( y ) ) : = E [ X ( 1 { Y ≤ y } − Λ ( X ′ θ o ( y ) ) ) ] = 0 , y ∈ Y . \mathbb{P}\psi_y(W;\theta_o(y)) := \mathbb{E}\left[X\left(\mathbb{1}\{Y \le y\}-\Lambda(X'\theta_o(y))\right)\right]=0,\qquad y\in\mathcal{Y}. P ψ y ( W ; θ o ( y )) := E [ X ( 1 { Y ≤ y } − Λ ( X ′ θ o ( y )) ) ] = 0 , y ∈ Y .
Accordingly, the primitive score process is ψ y ( W ; θ ( y ) ) = X ( 1 { Y ≤ y } − Λ ( X ′ θ ( y ) ) ) \psi_y(W;\theta(y))=X\left(\mathbb{1}\{Y\le y\}-\Lambda(X'\theta(y))\right) ψ y ( W ; θ ( y )) = X ( 1 { Y ≤ y } − Λ ( X ′ θ ( y )) ) , y ∈ Y y\in\mathcal{Y} y ∈ Y . At the cluster level, the corresponding SACR score is
ψ ˉ y ( W g ; θ ( y ) ) = N g − 1 ∑ i ∈ G g X i ( 1 { Y i ≤ y } − Λ ( X i ′ θ ( y ) ) ) . \bar{\psi}_y(W_g;\theta(y))=N_g^{-1}\sum_{i\in\mathcal{G}_g}X_i\left(\mathbb{1}\{Y_i\le y\}-\Lambda(X_i'\theta(y))\right). ψ ˉ y ( W g ; θ ( y )) = N g − 1 i ∈ G g ∑ X i ( 1 { Y i ≤ y } − Λ ( X i ′ θ ( y )) ) .
For each coordinate j j j , define the full jointly indexed DR score class
F j D R : = { w ↦ x j ( 1 { u ≤ y } − Λ ( x ′ t ) ) : ( y , t ) ∈ Y × Θ } , w = ( u , x ) , \mathcal{F}^{DR}_j:=\left\{w\mapsto x_j\left(\mathbb{1}\{u\le y\}-\Lambda(x't)\right):(y,t)\in\mathcal{Y}\times\Theta\right\},\qquad w=(u,x), F j D R := { w ↦ x j ( 1 { u ≤ y } − Λ ( x ′ t ) ) : ( y , t ) ∈ Y × Θ } , w = ( u , x ) ,
and its normalised cluster-average array, unnormalised envelope, and corresponding normalised envelope by
Z G g , j D R ( y , t ) : = G − 1 / 2 ψ ˉ y , j ( W g ; t ) , F ˉ G g , j D R : = { N g − 1 ∑ i ∈ G g X i j 2 } 1 / 2 , F G g , j D R : = G − 1 / 2 F ˉ G g , j D R . Z^{DR}_{Gg,j}(y,t):=G^{-1/2}\bar{\psi}_{y,j}(W_g;t),\qquad\bar{F}^{DR}_{Gg,j}:=\left\{N_g^{-1}\sum_{i\in\mathcal{G}_g}X_{ij}^{2}\right\}^{1/2},\qquad F^{DR}_{Gg,j}:=G^{-1/2}\bar{F}^{DR}_{Gg,j}. Z G g , j D R ( y , t ) := G − 1/2 ψ ˉ y , j ( W g ; t ) , F ˉ G g , j D R := ⎩ ⎨ ⎧ N g − 1 i ∈ G g ∑ X ij 2 ⎭ ⎬ ⎫ 1/2 , F G g , j D R := G − 1/2 F ˉ G g , j D R .
The corresponding Jacobian and covariance kernel are
H G D R ( y ) = − 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ Λ ′ ( X i ′ θ o ( y ) ) ] , H^{DR}_G(y)=-\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'\Lambda'(X_i'\theta_o(y))\right], H G D R ( y ) = − G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ Λ ′ ( X i ′ θ o ( y )) ,
and
Ω G D R ( y , y ′ ) = 1 G ∑ g = 1 G E G [ ψ ˉ y ( W g ; θ o ( y ) ) ψ ˉ y ′ ( W g ; θ o ( y ′ ) ) ′ ] . \Omega^{DR}_G(y,y')=\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[\bar{\psi}_y(W_g;\theta_o(y))\bar{\psi}_{y'}(W_g;\theta_o(y'))'\right]. Ω G D R ( y , y ′ ) = G 1 g = 1 ∑ G E G [ ψ ˉ y ( W g ; θ o ( y )) ψ ˉ y ′ ( W g ; θ o ( y ′ ) ) ′ ] .
The next assumptions establish identification of the DR-SACR population moment and functional separation.
Assumption DR 1 (Identification and array preservation). The link Λ : R → ( 0 , 1 ) \Lambda:\mathbb{R}\to(0,1) Λ : R → ( 0 , 1 ) is known and strictly increasing.
F Y ∣ X ( y ∣ X ) = Λ ( X ′ θ o ( y ) ) , y ∈ Y F_{Y\mid X}(y\mid X)=\Lambda(X'\theta_o(y)),\qquad y\in\mathcal{Y} F Y ∣ X ( y ∣ X ) = Λ ( X ′ θ o ( y )) , y ∈ Y in the target population, and this restriction is preserved along the sampled triangular array:
F Y i ∣ X i ( y ∣ X i ) : = E G [ 1 { Y i ≤ y } ∣ X i ] = Λ ( X i ′ θ o ( y ) ) almost surely F_{Y_i\mid X_i}(y\mid X_i):=\mathbb{E}_G[\mathbb{1}\{Y_i\le y\}\mid X_i]=\Lambda(X_i'\theta_o(y))\qquad\text{almost surely} F Y i ∣ X i ( y ∣ X i ) := E G [ 1 { Y i ≤ y } ∣ X i ] = Λ ( X i ′ θ o ( y )) almost surely for every G G G , every sampled observation i ≤ N i\le N i ≤ N , and every y ∈ Y y\in\mathcal{Y} y ∈ Y . Moreover, the coefficient path
y ↦ θ o ( y ) y \mapsto\theta_o(y) y ↦ θ o ( y ) is Borel measurable.
The following condition strengthens strict link monotonicity in the cluster-balanced population and supports uniform functional separation.
Assumption DR 2 (Regularity).
(a) The derivative Λ ′ \Lambda' Λ ′ is bounded and uniformly continuous on R \mathbb{R} R .
(b) For
Q G D R ( y , t , s ) : = 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ Λ ′ ( X i ′ { θ o ( y ) + s ( t − θ o ( y ) ) } ) ] , Q_G^{DR}(y,t,s) := \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'\Lambda'\left(X_i'\{\theta_o(y)+s(t-\theta_o(y))\}\right)\right], Q G D R ( y , t , s ) := G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ Λ ′ ( X i ′ { θ o ( y ) + s ( t − θ o ( y ))} ) , there is a constant c H > 0 c_H>0 c H > 0 such that
inf G inf y ∈ Y inf t ∈ Θ inf s ∈ [ 0 , 1 ] λ min ( Q G D R ( y , t , s ) ) ≥ c H . \inf_{G}\inf_{y\in\mathcal{Y}}\inf_{t\in\Theta}\inf_{s\in[0,1]}\lambda_{\min}\left(Q_G^{DR}(y,t,s)\right)\ge c_H. G inf y ∈ Y inf t ∈ Θ inf s ∈ [ 0 , 1 ] inf λ m i n ( Q G D R ( y , t , s ) ) ≥ c H .
The following supplies the DR functional identification result.
Proposition 4.1 (DR-SACR identification). Under Assumptions LI 1 and DR 1, θ o ( y ) \theta_o(y) θ o ( y ) is the unique solution to the DR-SACR population moment
1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ X i ( 1 { Y i ≤ y } − Λ ( X i ′ θ ( y ) ) ) ] = 0 for each y ∈ Y . \frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[X_i\left(\mathbb{1}\{Y_i\le y\}-\Lambda(X_i'\theta(y))\right)\right]=0 \quad\text{for each } y\in\mathcal{Y}. G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ X i ( 1 { Y i ≤ y } − Λ ( X i ′ θ ( y )) ) ] = 0 for each y ∈ Y . If, in addition, Assumption DR 2 holds, the DR-SACR population moments satisfy the functional separation condition in Assumption 3’(b).
The next result verifies the global complexity and envelope conditions for the DR score process.
Proposition 4.2 (DR global-class verification). Suppose that Assumptions 2(a), DR 1, LI 1, and DR 2(a) hold. Then Assumption 2’(a) and (b) hold for the DR score process.
Together with the generic approximate-root, interiority, Jacobian, covariance, and remaining functional conditions, the preceding DR conditions yield the conclusions of Theorems 3.1 to 3.4.
Linear quantile regression
For a growing number of clusters, the closest process-bootstrap benchmark is Hagemann [7 ] , which establishes cluster-robust inference for the quantile-regression process under bounded-cluster-size restrictions. Inverse-cluster-size-weighted quantile estimating equations also appear in the censored-data setting of Kim et al. (2024), where cluster sizes are relatively small. The results below use cluster-average scores within the general functional Z-SACR framework with a growing number of arbitrarily sized clusters.
For linear quantile regression,
Q Y ∣ X ( τ ∣ X ) = X ′ θ o ( τ ) , τ ∈ T ⊂ ( 0 , 1 ) , Q_{Y\mid X}(\tau\mid X)=X'\theta_o(\tau),\qquad\tau\in\mathcal{T}\subset(0,1), Q Y ∣ X ( τ ∣ X ) = X ′ θ o ( τ ) , τ ∈ T ⊂ ( 0 , 1 ) ,
and the model implies the conditional moment restriction
P ψ τ ( W ; θ o ( τ ) ) = E [ X ( τ − 1 { Y ≤ X ′ θ o ( τ ) } ) ] = 0 , τ ∈ T . \mathbb{P}\psi_{\tau}(W;\theta_o(\tau))=\mathbb{E}\left[X\left(\tau-\mathbb{1}\{Y\le X'\theta_o(\tau)\}\right)\right]=0,\qquad\tau\in\mathcal{T}. P ψ τ ( W ; θ o ( τ )) = E [ X ( τ − 1 { Y ≤ X ′ θ o ( τ )} ) ] = 0 , τ ∈ T .
Accordingly, the primitive score process is ψ τ ( W ; θ ( τ ) ) = X ( τ − 1 { Y ≤ X ′ θ ( τ ) } ) \psi_{\tau}(W;\theta(\tau))=X\left(\tau-\mathbb{1}\{Y\le X'\theta(\tau)\}\right) ψ τ ( W ; θ ( τ )) = X ( τ − 1 { Y ≤ X ′ θ ( τ )} ) , τ ∈ T \tau\in\mathcal{T} τ ∈ T . At the cluster level, the corresponding SACR score is
ψ ˉ τ ( W g ; θ ( τ ) ) = N g − 1 ∑ i ∈ G g X i ( τ − 1 { Y i ≤ X i ′ θ ( τ ) } ) . \bar{\psi}_{\tau}(W_g;\theta(\tau))=N_g^{-1}\sum_{i\in\mathcal{G}_g}X_i\left(\tau-\mathbb{1}\{Y_i\le X_i'\theta(\tau)\}\right). ψ ˉ τ ( W g ; θ ( τ )) = N g − 1 i ∈ G g ∑ X i ( τ − 1 { Y i ≤ X i ′ θ ( τ )} ) .
For each coordinate j j j , define the full jointly indexed QR score class
F j Q R : = { w ↦ x j ( τ − 1 { u ≤ x ′ t } ) : ( τ , t ) ∈ T × Θ } , w = ( u , x ) , \mathcal{F}^{QR}_j:=\left\{w\mapsto x_j\left(\tau-\mathbb{1}\{u\le x't\}\right):(\tau,t)\in\mathcal{T}\times\Theta\right\},\qquad w=(u,x), F j QR := { w ↦ x j ( τ − 1 { u ≤ x ′ t } ) : ( τ , t ) ∈ T × Θ } , w = ( u , x ) ,
and its normalised cluster-average array, unnormalised envelope, and corresponding normalised envelope by
Z G g , j Q R ( τ , t ) : = G − 1 / 2 ψ ˉ τ , j ( W g ; t ) , F ˉ G g , j Q R : = { N g − 1 ∑ i ∈ G g X i j 2 } 1 / 2 , F G g , j Q R : = G − 1 / 2 F ˉ G g , j Q R . Z^{QR}_{Gg,j}(\tau,t):=G^{-1/2}\bar{\psi}_{\tau,j}(W_g;t),\qquad\bar{F}^{QR}_{Gg,j}:=\left\{N_g^{-1}\sum_{i\in\mathcal{G}_g}X_{ij}^{2}\right\}^{1/2},\qquad F^{QR}_{Gg,j}:=G^{-1/2}\bar{F}^{QR}_{Gg,j}. Z G g , j QR ( τ , t ) := G − 1/2 ψ ˉ τ , j ( W g ; t ) , F ˉ G g , j QR := ⎩ ⎨ ⎧ N g − 1 i ∈ G g ∑ X ij 2 ⎭ ⎬ ⎫ 1/2 , F G g , j QR := G − 1/2 F ˉ G g , j QR .
The corresponding Jacobian and covariance kernel are
H G Q R ( τ ) = − 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ f Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) ] , H^{QR}_{G}(\tau)=-\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_{G}\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'f_{Y_i\mid X_i}\left(X_i'\theta_o(\tau)\mid X_i\right)\right], H G QR ( τ ) = − G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ f Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) ,
and
Ω G Q R ( τ , τ ′ ) = 1 G ∑ g = 1 G E G [ ψ ˉ τ ( W g ; θ o ( τ ) ) ψ ˉ τ ′ ( W g ; θ o ( τ ′ ) ) ′ ] . \Omega^{QR}_{G}(\tau,\tau')=\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_{G}\left[\bar{\psi}_{\tau}(W_g;\theta_o(\tau))\bar{\psi}_{\tau'}(W_g;\theta_o(\tau'))'\right]. Ω G QR ( τ , τ ′ ) = G 1 g = 1 ∑ G E G [ ψ ˉ τ ( W g ; θ o ( τ )) ψ ˉ τ ′ ( W g ; θ o ( τ ′ ) ) ′ ] .
The next assumptions establish identification of the QR-SACR population moment and provide primitive conditions for verifying the remaining high-level Z Z Z -SACR requirements.
Assumption QR 1 (Identification and regularity).
(a) Q Y ∣ X ( τ ∣ X ) = X ′ θ o ( τ ) Q_{Y\mid X}(\tau\mid X)=X'\theta_o(\tau) Q Y ∣ X ( τ ∣ X ) = X ′ θ o ( τ ) almost surely for every τ ∈ T \tau\in\mathcal{T} τ ∈ T in the target population, and this conditional-quantile restriction is preserved along the sampled triangular array:
F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) = τ almost surely under the stage- G law F_{Y_i\mid X_i}\left(X_i'\theta_o(\tau)\mid X_i\right)=\tau\qquad\text{almost surely under the stage-}G\text{ law} F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) = τ almost surely under the stage- G law for every G G G , every sampled observation i ≤ N i\le N i ≤ N , and every τ ∈ T \tau\in\mathcal{T} τ ∈ T .
(b) The conditional distribution of Y i Y_i Y i given X i X_i X i is absolutely continuous, with a density f Y i ∣ X i ( ⋅ ∣ X i ) f_{Y_i\mid X_i}(\cdot\mid X_i) f Y i ∣ X i ( ⋅ ∣ X i ) bounded above by a common constant C f < ∞ C_f<\infty C f < ∞ , almost surely in X i X_i X i , uniformly over G G G and i ≤ N i\le N i ≤ N . For some r > 0 r>0 r > 0 and a common nondecreasing modulus ω ( u ) → 0 \omega(u)\to0 ω ( u ) → 0 as $u\downarrow 0,
∣ f Y i ∣ X i ( z ∣ X i ) − f Y i ∣ X i ( z ′ ∣ X i ) ∣ ≤ ω ( ∣ z − z ′ ∣ ) \lvert f_{Y_i\mid X_i}(z\mid X_i)-f_{Y_i\mid X_i}(z'\mid X_i)\rvert\le\omega(\lvert z-z'\rvert) ∣ f Y i ∣ X i ( z ∣ X i ) − f Y i ∣ X i ( z ′ ∣ X i )∣ ≤ ω (∣ z − z ′ ∣) whenever both z z z and z ′ z' z ′ lie within r ∥ X i ∥ r\lVert X_i\rVert r ∥ X i ∥ of X i ′ θ o ( τ ) X_i'\theta_o(\tau) X i ′ θ o ( τ ) , uniformly over G G G , i i i , τ \tau τ , almost surely in X i X_i X i .
(c) For
Q G Q R ( τ , t , s ) : = 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ f Y i ∣ X i ( X i ′ { θ o ( τ ) + s ( t − θ o ( τ ) ) } ∣ X i ) ] , Q_G^{QR}(\tau,t,s):=\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'f_{Y_i\mid X_i}\left(X_i'\{\theta_o(\tau)+s(t-\theta_o(\tau))\}\mid X_i\right)\right], Q G QR ( τ , t , s ) := G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ f Y i ∣ X i ( X i ′ { θ o ( τ ) + s ( t − θ o ( τ ))} ∣ X i ) ,
there is a constant c f > 0 c_f>0 c f > 0 such that
inf G inf τ ∈ T inf t ∈ Θ inf s ∈ [ 0 , 1 ] λ min ( Q G Q R ( τ , t , s ) ) ≥ c f . \inf_G\inf_{\tau\in\mathcal{T}}\inf_{t\in\Theta}\inf_{s\in[0,1]}\lambda_{\min}\left(Q_G^{QR}(\tau,t,s)\right)\ge c_f. G inf τ ∈ T inf t ∈ Θ inf s ∈ [ 0 , 1 ] inf λ m i n ( Q G QR ( τ , t , s ) ) ≥ c f .
The following supplies the QR functional identification result.
Proposition 4.3 (QR-SACR identification). Under Assumptions LI 1 and QR 1, θ o ( τ ) \theta_o(\tau) θ o ( τ ) is the unique solution to the QR-SACR population moment
1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i ( τ − 1 { Y i ≤ X i ′ θ ( τ ) } ) ] = 0 , τ ∈ T . \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_i\left(\tau-1\{Y_i\le X_i'\theta(\tau)\}\right)\right]=0,\qquad\tau\in\mathcal{T}. G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i ( τ − 1 { Y i ≤ X i ′ θ ( τ )} ) = 0 , τ ∈ T .
The population moments also satisfy the functional separation condition in Assumption 3’(b).
The next result verifies the global complexity and envelope conditions for the QR score process.
Proposition 4.4 (QR global-class verification). Suppose that Assumptions 2(a), QR 1, and LI 1 hold. Then Assumption 2’(a) and (b) hold for the QR score process.
Together with the generic approximate-root, interiority, Jacobian, covariance, and remaining functional conditions, the preceding QR conditions yield the conclusions of Theorems 3.1 to 3.4.
Linear expectile regression
For linear expectile regression,
E τ ( Y ∣ X ) = X ′ θ o ( τ ) , τ ∈ T ⊂ [ τ ‾ , τ ‾ ] ⊂ ( 0 , 1 ) , \mathbb{E}_{\tau}(Y\mid X)=X'\theta_o(\tau),\qquad\tau\in\mathcal{T}\subset[\underline{\tau},\overline{\tau}]\subset(0,1), E τ ( Y ∣ X ) = X ′ θ o ( τ ) , τ ∈ T ⊂ [ τ , τ ] ⊂ ( 0 , 1 ) ,
where E τ ( Y ∣ X ) \mathbb{E}_{\tau}(Y\mid X) E τ ( Y ∣ X ) denotes the τ \tau τ th conditional expectile of Y Y Y given X X X , the model implies the conditional moment restriction
P ψ τ ( W ; θ o ( τ ) ) = E [ X ( Y − X ′ θ o ( τ ) ) { τ − 1 { Y < X ′ θ o ( τ ) } } ] = 0 , τ ∈ T . \mathbb{P}\psi_{\tau}(W;\theta_o(\tau))=\mathbb{E}\left[X\left(Y-X'\theta_o(\tau)\right)\left\{\tau-1\{Y<X'\theta_o(\tau)\}\right\}\right]=0,\qquad\tau\in\mathcal{T}. P ψ τ ( W ; θ o ( τ )) = E [ X ( Y − X ′ θ o ( τ ) ) { τ − 1 { Y < X ′ θ o ( τ )} } ] = 0 , τ ∈ T .
Accordingly, the primitive score process is
ψ τ ( W ; θ ( τ ) ) = X ( Y − X ′ θ ( τ ) ) { τ − 1 { Y < X ′ θ ( τ ) } } , τ ∈ T . \psi_{\tau}(W;\theta(\tau))=X\left(Y-X'\theta(\tau)\right)\left\{\tau-1\{Y<X'\theta(\tau)\}\right\},\qquad\tau\in\mathcal{T}. ψ τ ( W ; θ ( τ )) = X ( Y − X ′ θ ( τ ) ) { τ − 1 { Y < X ′ θ ( τ )} } , τ ∈ T .
When 0.5 ∈ T 0.5\in\mathcal{T} 0.5 ∈ T , the score at that index satisfies ψ 0.5 ( W ; θ ) = 1 2 X ( Y − X ′ θ ) \psi_{0.5}(W;\theta)=\frac{1}{2}X(Y-X'\theta) ψ 0.5 ( W ; θ ) = 2 1 X ( Y − X ′ θ ) , which is one half of the OLS score obtained by setting Z = X Z=X Z = X in the running IV example.
At the cluster level, the corresponding SACR score is
ψ ˉ τ ( W g ; θ ( τ ) ) = N g − 1 ∑ i ∈ G g X i ( Y i − X i ′ θ ( τ ) ) ∣ τ − 1 { Y i < X i ′ θ ( τ ) } ∣ . \bar{\psi}_{\tau}(\mathbf{W}_g;\theta(\tau)) = N_g^{-1}\sum_{i\in\mathcal{G}_g} X_i(Y_i-X_i'\theta(\tau))\left|\tau-\mathbb{1}\{Y_i<X_i'\theta(\tau)\}\right|. ψ ˉ τ ( W g ; θ ( τ )) = N g − 1 i ∈ G g ∑ X i ( Y i − X i ′ θ ( τ )) ∣ τ − 1 { Y i < X i ′ θ ( τ )} ∣ .
For each coordinate j j j , let B Θ : = sup t ∈ Θ ∥ t ∥ B_{\Theta}:=\sup_{t\in\Theta}\lVert t\rVert B Θ := sup t ∈ Θ ∥ t ∥ and define the full jointly indexed ER score class and its measurable envelope by
F j E R : = { w ↦ x j ( u − x ′ t ) ∣ τ − 1 { u < x ′ t } ∣ : ( τ , t ) ∈ T × Θ } , w = ( u , x ) , \mathcal{F}^{ER}_j:=\left\{w\mapsto x_j(u-x't)\left|\tau-\mathbb{1}\{u<x't\}\right|:(\tau,t)\in\mathcal{T}\times\Theta\right\},\qquad w=(u,x), F j E R := { w ↦ x j ( u − x ′ t ) ∣ τ − 1 { u < x ′ t } ∣ : ( τ , t ) ∈ T × Θ } , w = ( u , x ) ,
F j E R ( W ) : = ∣ X j ∣ ( ∣ Y ∣ + B Θ ∥ X ∥ ) . F^{ER}_j(\mathbf{W}):=\lvert X_j\rvert(\lvert Y\rvert+B_{\Theta}\lVert X\rVert). F j E R ( W ) := ∣ X j ∣ (∣ Y ∣ + B Θ ∥ X ∥) .
Define the corresponding normalised cluster-average array, unnormalised envelope, and normalised envelope by
Z G g , j E R ( τ , t ) : = G − 1 / 2 ψ ˉ τ , j ( W g ; t ) , F ˉ G g , j E R : = { N g − 1 ∑ i ∈ G g { F j E R ( W i ) } 2 } 1 / 2 , F G g , j E R : = G − 1 / 2 F G g , j E R . Z^{ER}_{Gg,j}(\tau,t):=G^{-1/2}\bar{\psi}_{\tau,j}(\mathbf{W}_g;t),\qquad\bar{F}^{ER}_{Gg,j}:=\left\{N_g^{-1}\sum_{i\in\mathcal{G}_g}\left\{F^{ER}_j(W_i)\right\}^{2}\right\}^{1/2},\qquad F^{ER}_{Gg,j}:=G^{-1/2}F^{ER}_{Gg,j}. Z G g , j E R ( τ , t ) := G − 1/2 ψ ˉ τ , j ( W g ; t ) , F ˉ G g , j E R := ⎩ ⎨ ⎧ N g − 1 i ∈ G g ∑ { F j E R ( W i ) } 2 ⎭ ⎬ ⎫ 1/2 , F G g , j E R := G − 1/2 F G g , j E R .
The corresponding Jacobian and covariance kernel are
H G E R ( τ ) = − 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ E G [ ∣ τ − 1 { Y i < X i ′ θ o ( τ ) } ∣ | X i ] ] , H^{ER}_G(\tau)=-\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'\mathbb{E}_G\left[\left|\tau-\mathbb{1}\{Y_i<X_i'\theta_o(\tau)\}\right|\middle|X_i\right]\right], H G E R ( τ ) = − G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ E G [ ∣ τ − 1 { Y i < X i ′ θ o ( τ )} ∣ ∣ X i ] ,
and
Ω G E R ( τ , τ ′ ) = 1 G ∑ g = 1 G E G [ ψ ˉ τ ( W g ; θ o ( τ ) ) ψ ˉ τ ′ ( W g ; θ o ( τ ′ ) ) ′ ] . \Omega^{ER}_G(\tau,\tau')=\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[\bar{\psi}_{\tau}(\mathbf{W}_g;\theta_o(\tau))\bar{\psi}_{\tau'}(\mathbf{W}_g;\theta_o(\tau'))'\right]. Ω G E R ( τ , τ ′ ) = G 1 g = 1 ∑ G E G [ ψ ˉ τ ( W g ; θ o ( τ )) ψ ˉ τ ′ ( W g ; θ o ( τ ′ ) ) ′ ] .
The next assumptions establish identification of the ER-SACR population moment and provide primitive conditions for verifying the remaining high-level Z Z Z -SACR requirements.
Assumption ER 1 (Identification and array preservation). E τ ( Y ∣ X ) = X ′ θ o ( τ ) \mathbb{E}_{\tau}(Y\mid X)=X'\theta_o(\tau) E τ ( Y ∣ X ) = X ′ θ o ( τ ) almost surely for every τ ∈ T ⊂ [ τ ‾ , τ ˉ ] ⊂ ( 0 , 1 ) \tau\in\mathcal{T}\subset[\underline{\tau},\bar{\tau}]\subset(0,1) τ ∈ T ⊂ [ τ , τ ˉ ] ⊂ ( 0 , 1 ) , in the target population, and the corresponding first-order restriction is preserved along the sampled triangular array:
E G [ ( Y i − X i ′ θ o ( τ ) ) ∣ τ − 1 { Y i < X i ′ θ o ( τ ) } ∣ ∣ X i ] = 0 almost surely under the stage- G law \mathbb{E}_G\left[\left.(Y_i-X_i'\theta_o(\tau))\left|\tau-\mathbb{1}\{Y_i<X_i'\theta_o(\tau)\}\right|\right|X_i\right]=0\qquad\text{almost surely under the stage-}G\text{ law} E G [ ( Y i − X i ′ θ o ( τ )) ∣ τ − 1 { Y i < X i ′ θ o ( τ )} ∣ ∣ X i ] = 0 almost surely under the stage- G law for every G G G , every sampled observation i ≤ N i\le N i ≤ N , and every τ ∈ T \tau\in\mathcal{T} τ ∈ T .
Assumption ER 2 (Regularity). For some δ > 0 \delta>0 δ > 0 , sup G ≥ 1 sup i ≤ N E G [ ( ∥ X i ∥ ∣ Y i ∣ ) 2 + δ + ∥ X i ∥ 4 + 2 δ ] < ∞ \sup_{G\ge1}\sup_{i\le N}\mathbb{E}_G\left[(\lVert X_i\rVert\lvert Y_i\rvert)^{2+\delta}+\lVert X_i\rVert^{4+2\delta}\right]<\infty sup G ≥ 1 sup i ≤ N E G [ (∥ X i ∥ ∣ Y i ∣ ) 2 + δ + ∥ X i ∥ 4 + 2 δ ] < ∞ .
The following supplies the ER functional identification result.
Proposition 4.5 (ER-SACR identification). Under Assumption ER 1 , Assumption ER 2 , and
Assumption LI 1, θ o ( τ ) \theta_o(\tau) θ o ( τ ) is the unique solution to the ER-SACR population moment
1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i ( Y i − X i ′ θ ( τ ) ) ∣ τ − 1 { Y i < X i ′ θ ( τ ) } ] = 0 , τ ∈ T . \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_{\mathcal{G}}\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_i(Y_i-X_i'\theta(\tau))\left\lvert\tau-1\{Y_i<X_i'\theta(\tau)\}\right.\right]=0,\qquad\tau\in\mathcal{T}. G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i ( Y i − X i ′ θ ( τ )) ∣ τ − 1 { Y i < X i ′ θ ( τ )} = 0 , τ ∈ T .
The population moments also satisfy the functional separation condition in Assumption 3 ′ ( b ) 3'(b) 3 ′ ( b ) .
The next result verifies the global complexity and envelope conditions for the ER score process.
Proposition 4.6 (ER global-class verification). Suppose that Assumptions 2(a), ER 1, LI 1, and ER 2 hold. Then Assumption 2 ′ ( a ) 2'(a) 2 ′ ( a ) and (b) hold for the ER score process.
Together with the generic approximate-root, interiority, Jacobian, covariance, and remaining functional conditions, the preceding ER conditions yield the conclusions of Theorems 3.1 to 3.4.
Monte Carlo Evidence
This section studies the finite-sample performance of conventional cluster-robust and Z Z Z -SACR inference for both pointwise coefficients and distributional treatment-effect functions. The two experiments use the same no-covariate design, cluster-level treatment assignment, deterministic cluster-size arrays, cluster grid, and Monte Carlo settings. They differ only in whether the target is evaluated at a central index or over a functional domain. The data-generating processes (DGPs) impose common coefficient paths and centre the relevant scores within every cluster, so observation weighting and cluster weighting share the same structural target. The experiments therefore isolate the inferential consequences of cluster-size dominance rather than differences between estimands. Coverage and rejection comparisons concern complete inference procedures, which differ in estimator weighting, resampling scheme, and, where applicable, scale estimation; their differences cannot be attributed solely to SACR weighting. Point-estimate bias and root mean squared error (RMSE) comparisons assess the estimators independently of the bootstrap procedure.
Simulation design and inference procedures
Let X i = ( 1 , T g ) ′ X_i=(1,T_g)' X i = ( 1 , T g ) ′ for i ∈ G g i\in\mathcal{G}_g i ∈ G g , with T g = 1 { g ≡ 0 ( m o d 5 ) } T_g=1\{g\equiv0\pmod{5}\} T g = 1 { g ≡ 0 ( mod 5 )} . The DR design is Y i = X i ′ θ o + logit { Φ ( U g i ) } Y_i=X_i'\theta_o+\operatorname{logit}\{\Phi(U_{gi})\} Y i = X i ′ θ o + logit { Φ ( U g i )} , while QR and ER use Y i = X i ′ θ o + X i ′ γ U g i Y_i=X_i'\theta_o+X_i'\gamma U_{gi} Y i = X i ′ θ o + X i ′ γ U g i , where θ o = ( − 0.5 , 0.5 ) ′ \theta_o=(-0.5,0.5)' θ o = ( − 0.5 , 0.5 ) ′ , γ = ( 0.2 , 1 ) ′ \gamma=(0.2,1)' γ = ( 0.2 , 1 ) ′ , and U g i = 0.96 Z g + 0.04 ε g i U_{gi}=\sqrt{0.96}Z_g+\sqrt{0.04}\varepsilon_{gi} U g i = 0.96 Z g + 0.04 ε g i . The common and individual shocks are mutually independent standard normals, independent across clusters. The errors therefore have standard logistic margins for DR and standard normal margins for QR and ER.
Cluster sizes are deterministic Pareto quantiles, N g = ⌈ ( 1 − ( g − 1 / 2 ) / G ) − 1 / β ⌉ N_g=\left\lceil\left(1-(g-1/2)/G\right)^{-1/\beta}\right\rceil N g = ⌈ ( 1 − ( g − 1/2 ) / G ) − 1/ β ⌉ . For β = 2.5 \beta=2.5 β = 2.5 , max g N g 2 / N → 0 \max_g N_g^2/N\to0 max g N g 2 / N → 0 ; for β = 1 \beta=1 β = 1 , this condition fails although max g N g / N → 0 \max_g N_g/N\to0 max g N g / N → 0 . The latter regime isolates failure of conventional Gaussian inference while retaining cluster-size negligibility sufficient for a consistency argument. Both regimes use G ∈ { 150 , 300 , 600 } G \in\{150, 300, 600\} G ∈ { 150 , 300 , 600 } , 1,000 Monte Carlo replications, and 499 bootstrap draws.
Analytical procedures use the relevant cluster-score sandwich, with a kernel density estimate for conventional QR. Conventional bootstrap procedures resample whole clusters in pairs; SACR bootstrap procedures apply independent Exp(1) cluster weights to the size-adjusted estimating equations. Section S.5.1 gives the coefficient paths and implementation details. Supplementary experiments in Sections S.5.3 and S.5.4 examine weaker dependence and a common exponential-weight bootstrap. Section S.5.5 considers a single dominant cluster.
Pointwise coefficient inference
The pointwise targets are the DR treatment coefficient at the cluster-balanced population median threshold y = Q Y ( 0.5 ) y = Q_Y(0.5) y = Q Y ( 0.5 ) , the QR coefficient at τ = 0.5 \tau= 0.5 τ = 0.5 , and the ER coefficient at τ = 0.5 \tau= 0.5 τ = 0.5 . The tables report signed bias, RMSE, average standard error (Avg. SE), and rejection frequencies for nominal 5% two-sided Wald tests using the critical value 1.96. SACR-A and CR-A use analytical standard errors; SACR-B and CR-B use bootstrap standard errors. For both procedures, each coefficient’s bootstrap standard error is the sample standard deviation of its bootstrap coefficient draws.
The paired tables isolate the effect of the cluster-size regime. Bias and RMSE are shared across each estimator’s analytical and bootstrap rows because the inference procedure does not change the point estimates. In the asymptotically ignorable β = 2.5 \beta= 2.5 β = 2.5 regime, both estimators’ RMSEs decline with G G G , with lower RMSE for SACR. Both bootstrap procedures have rejection frequencies close to the nominal 5% level throughout the ignorable regime; the larger QR departures occur with analytical standard errors at G = 150 G = 150 G = 150 . Under β = 1 \beta= 1 β = 1 , SACR RMSE continues to fall, whereas conventional RMSE remains substantial. At G = 600 G = 600 G = 600 , SACR-B rejects 5.5%, 7.5%, and 6.0% for DR, QR, and ER, respectively, all reasonably close to nominal. Its QR rejection rates remain reasonably close to nominal at all three cluster counts. Conventional analytical inference rejects 36.4%, 91.0%, and 62.7% at G = 600 G = 600 G = 600 . The pairs bootstrap reduces these frequencies to 16.5%, 33.9%, and 24.7%, but substantial size distortions remain.
Table 2. Pointwise Monte Carlo results: ULC Pareto regime (β = 1 \beta= 1 β = 1 ). The table has the same column blocks, targets, cluster grid, Monte Carlo replications, bootstrap draws, and inference procedures as Table 1; only the Pareto tail index differs.
Uniform inference on distributional treatment effects
The functional experiment studies Δ F ( y ) = F Y ∣ T = 1 ( y ) − F Y ∣ T = 0 ( y ) \Delta_F(y) = F_{Y\mid T=1}(y) - F_{Y\mid T=0}(y) Δ F ( y ) = F Y ∣ T = 1 ( y ) − F Y ∣ T = 0 ( y ) , Δ Q ( τ ) = Q Y ∣ T = 1 ( τ ) − Q Y ∣ T = 0 ( τ ) \Delta_Q(\tau) = Q_{Y\mid T=1}(\tau) - Q_{Y\mid T=0}(\tau) Δ Q ( τ ) = Q Y ∣ T = 1 ( τ ) − Q Y ∣ T = 0 ( τ ) , and Δ E ( τ ) = E τ ( Y ∣ T = 1 ) − E τ ( Y ∣ T = 0 ) \Delta_E(\tau) = E_\tau(Y \mid T = 1) - E_\tau(Y \mid T = 0) Δ E ( τ ) = E τ ( Y ∣ T = 1 ) − E τ ( Y ∣ T = 0 ) . The Distribution Function (DF) is evaluated on the central unconditional population support Y 0.1 , 0.9 = [ Q Y ( 0.1 ) , Q Y ( 0.9 ) ] \mathcal{Y}_{0.1,0.9} = [Q_Y(0.1), Q_Y(0.9)] Y 0.1 , 0.9 = [ Q Y ( 0.1 ) , Q Y ( 0.9 )] , while the Quantile Function (QF) and Expectile Function (EF) are evaluated over τ ∈ [ 0.1 , 0.9 ] \tau\in[0.1, 0.9] τ ∈ [ 0.1 , 0.9 ] . In this saturated binary-treatment design, the group DF, quantile, and expectile differences are equivalent to the corresponding fitted DR, QR, and ER treatment-effect functions, including under inverse-cluster-size weighting. The evaluation grids grow at rate G 0.3 G^{0.3} G 0.3 . Uniform inference is bootstrap-based following Chernozhukov, Fernández-Val, and Melly [4 ] . Both conventional and SACR uniform bands use the normalised bootstrap IQR studentiser described in § 3.3 .
Integrated bias (IBias) is the root mean squared displacement of the mean estimated path from the true path across grid points; integrated RMSE (IRMSE) also includes variation across Monte Carlo samples. Avg. Band is the mean uniform-band half-width, and U-Rej. is the frequency with which the band fails to cover the complete true path.8 The tables report SACR-B followed by CR-B; each bootstrap draw is shared across the complete functional grid.
The three column blocks report DF, QF, and EF treatment effects over the domains defined above. The grids use the curvature-based rule described in the text. For G = 150 , 300 , 600 G = 150, 300, 600 G = 150 , 300 , 600 , respectively, the grid sizes are L G ( F ) = 12 , 15 , 18 L_G^{(F)} = 12, 15, 18 L G ( F ) = 12 , 15 , 18 , L G ( Q ) = 43 , 52 , 64 L_G^{(Q)} = 43, 52, 64 L G ( Q ) = 43 , 52 , 64 , and L G ( E ) = 41 , 50 , 61 L_G^{(E)} = 41, 50, 61 L G ( E ) = 41 , 50 , 61 . Both procedures use normalised bootstrap IQR studentisation.
Table 4. Uniform Monte Carlo results for distributional treatment effects: ULC Pareto regime (β = 1 \beta= 1 β = 1 ).
The table has the same functional targets, supports, grids, and bootstrap procedures as Table 3 ; only the Pareto tail index differs.
Table 3. Uniform Monte Carlo results for distributional treatment effects: asymptotically ignorable Pareto regime (β = 2.5 \beta= 2.5 β = 2.5 ).
The functional results reinforce the pointwise comparison. In the β = 2.5 \beta= 2.5 β = 2.5 regime, SACR-B QF rejection frequencies remain close to 5%; DF and EF distortions at smaller G G G diminish as the number of clusters increases, reaching 7.8% and 8.2%, respectively, at G = 600 G = 600 G = 600 . Under β = 1 \beta= 1 β = 1 , SACR-B integrated RMSE declines with G G G , and its uniform rejection frequencies at G = 600 G = 600 G = 600 are 4.1%, 4.1%, and 7.0% for DF, QF, and EF, all reasonably close to nominal. The corresponding CR-B frequencies are 15.9%, 15.2%, and 18.6%, alongside substantially larger integrated RMSE and wider bands. Thus size adjustment improves both estimation and uniform inference under severe cluster-size heterogeneity, with the remaining SACR distortions concentrated at smaller cluster counts.
A supplementary empirical illustration in Section S.6 revisits the municipal-school setting of Akhtari, Moreira, and Trucco [1 ] , where students are clustered within municipalities of substantially unequal size. The empirical illustration reports pointwise coefficient comparisons and distributional effects, together with the identifying assumptions and implementation details. The application illustrates the consequences of cluster weighting in practice.
Conclusion
Unignorably large clusters can invalidate conventional inference for Z Z Z -estimators. This paper develops pointwise and functional inference for cluster-balanced estimating equations under independent clusters and unrestricted within-cluster dependence, subject to the stated conditions. Building on the SACR reweighting of Chiang, Sasaki, and Wang (2025), the theory covers smooth and non-smooth scores, with distribution, quantile, and expectile regression as applications. Pointwise normality permits a drifting covariance sequence, whereas functional weak convergence and bootstrap inference require a stable covariance kernel. A common cluster-level zero preserves the structural target under size adjustment, while separation of the reweighted moments ensures identification. Without that common zero, cluster balancing can change the estimand; the appropriate population representation then depends on the research question. Monte Carlo experiments illustrate pointwise and uniform inference when the competing moments share a common structural zero.
Acknowledgments: This paper grew out of the first author’s master’s thesis.
Declaration of AI use: During the preparation of this manuscript, the authors used OpenAI’s ChatGPT and Codex for research assistance. The authors reviewed and edited all AI-assisted output and take full responsibility for the content of the manuscript.
Supplement to “Z-Estimation under Unignorably Large Clusters”
Agossou Michel Yomakou; Abderrahim Taamouti; Emmanuel S. Tsyawo
The Supplementary Appendix contains the target-invariance proof (Section S.1), proofs of the general Z Z Z -SACR theory (Section S.2), and proofs and supporting lemmas for the model-specific results (Section S.3). These are followed by the distribution-to-expectile map and its proof (Section S.4). Section S.5 collects Monte Carlo implementation details and supplementary experiments, while Section S.6 presents the empirical illustration. Section S.7 concludes with a classification distinguishing cluster-size informativeness from asymptotic size influence.
Proof of prop:invid
By the maintained cluster-level exogeneity condition, for every cluster g g g ,
E G [ ψ ˉ ( W g ; θ o ) ] = 0 , \mathbb{E}_{G}\left[\bar{\psi}(\mathbf{W}_{g};\theta_{o})\right]=0, E G [ ψ ˉ ( W g ; θ o ) ] = 0 ,
and therefore E G [ ∑ i ∈ G g ψ ( W i ; θ o ) ] = N g E G [ ψ ˉ ( W g ; θ o ) ] = 0 \mathbb{E}_{G}\left[\sum_{i\in\mathcal{G}_{g}}\psi(W_{i};\theta_{o})\right]=N_{g}\mathbb{E}_{G}\left[\bar{\psi}(\mathbf{W}_{g};\theta_{o})\right]=0 E G [ ∑ i ∈ G g ψ ( W i ; θ o ) ] = N g E G [ ψ ˉ ( W g ; θ o ) ] = 0 . Since ω g \omega_g ω g is deterministic, summing the cluster-level equalities with these weights gives
E G [ 1 G ∑ i = 1 N ω g ( i ) ψ ( W i ; θ o ) ] = 1 G ∑ g = 1 G ω g N g E G [ ψ ˉ ( W g ; θ o ) ] = 0. \mathbb{E}_{G}\left[\frac{1}{G}\sum_{i=1}^{N}\omega_{g(i)}\psi(W_i;\theta_o)\right]=\frac{1}{G}\sum_{g=1}^{G}\omega_gN_g\mathbb{E}_{G}\left[\bar{\psi}(\mathbf{W}_g;\theta_o)\right]=0. E G [ G 1 i = 1 ∑ N ω g ( i ) ψ ( W i ; θ o ) ] = G 1 g = 1 ∑ G ω g N g E G [ ψ ˉ ( W g ; θ o ) ] = 0.
This establishes the population zero for every deterministic weighting scheme. Since no common zero is imposed for θ ≠ θ o \theta\ne\theta_o θ = θ o , uniqueness away from the truth is supplied by the weighted triangular-array separation condition.
Proofs for the General Z-SACR Theory
The following supplies three reusable Z-SACR-compatible results based on theoretical results established in Hansen and Lee [8 ] .
Lemma S.2.1 (Cluster-average moment transfer and pointwise limits). For each G G G , let X g i , G ∈ R p X_{gi,G} \in\mathbb{R}^{p} X g i , G ∈ R p , i = 1 , … , N g i=1,\ldots,N_g i = 1 , … , N g , and define X ˉ g , G : = 1 N g ∑ i = 1 N g X g i , G \bar{X}_{g,G} := \frac{1}{N_g}\sum_{i=1}^{N_g}X_{gi,G} X ˉ g , G := N g 1 ∑ i = 1 N g X g i , G . For r ≥ 1 r \ge1 r ≥ 1 , suppose that
lim M → ∞ sup G ≥ 1 sup g ≤ G sup i ≤ N g E G [ ∥ X g i , G ∥ r 1 { ∥ X g i , G ∥ > M } ] = 0. (7) \lim_{M\to\infty}\sup_{G\ge1}\sup_{g\le G}\sup_{i\le N_g}\mathbb{E}_G[\lVert X_{gi,G}\rVert^r\mathbf{1}\{\lVert X_{gi,G}\rVert>M\}]=0. \tag*{(7)} M → ∞ lim G ≥ 1 sup g ≤ G sup i ≤ N g sup E G [∥ X g i , G ∥ r 1 {∥ X g i , G ∥ > M }] = 0. (7) Then
(a) lim M → ∞ sup G ≥ 1 sup g ≤ G E G [ ∥ X ˉ g , G ∥ r 1 { ∥ X ˉ g , G ∥ > M } ] = 0 \lim_{M\to\infty}\sup_{G\ge1}\sup_{g\le G}\mathbb{E}_G[\lVert\bar{X}_{g,G}\rVert^r\mathbf{1}\{\lVert\bar{X}_{g,G}\rVert>M\}]=0 lim M → ∞ sup G ≥ 1 sup g ≤ G E G [∥ X ˉ g , G ∥ r 1 {∥ X ˉ g , G ∥ > M }] = 0 ;
(b) 1 G ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ] } → p 0 \frac{1}{G}\sum_{g=1}^{G}\{\bar{X}_{g,G}-\mathbb{E}_G[\bar{X}_{g,G}]\}\xrightarrow{p}0 G 1 ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ]} p 0 ; if the cluster vectors are independent with r = 1 r=1 r = 1 ; and
(c) for Ω X , G : = 1 G ∑ g = 1 G V G ( X ˉ g , G ) \Omega_{X,G}:=\frac{1}{G}\sum_{g=1}^{G}V_G(\bar{X}_{g,G}) Ω X , G := G 1 ∑ g = 1 G V G ( X ˉ g , G ) , one has sup G ∥ Ω X , G ∥ < ∞ \sup_G\lVert\Omega_{X,G}\rVert<\infty sup G ∥ Ω X , G ∥ < ∞ and 1 G ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ] } = O p ( 1 ) \frac{1}{\sqrt{G}}\sum_{g=1}^{G}\{\bar{X}_{g,G}-\mathbb{E}_G[\bar{X}_{g,G}]\}=O_p(1) G 1 ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ]} = O p ( 1 ) , if the cluster vectors are independent with r = 2 r=2 r = 2 . If, in addition, λ min ( Ω X , G ) ≥ λ > 0 \lambda_{\min}(\Omega_{X,G})\ge\lambda>0 λ m i n ( Ω X , G ) ≥ λ > 0 , then
Ω X , G − 1 / 2 1 G ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ] } → d N ( 0 , I p ) . \Omega_{X,G}^{-1/2}\frac{1}{\sqrt{G}}\sum_{g=1}^{G}\{\bar{X}_{g,G}-\mathbb{E}_G[\bar{X}_{g,G}]\}\xrightarrow{d}N(0,I_p). Ω X , G − 1/2 G 1 g = 1 ∑ G { X ˉ g , G − E G [ X ˉ g , G ]} d N ( 0 , I p ) .
Proof. Part (a): This part follows from the argument of Hansen and Lee [8 ] , applied uniformly over the triangular array. For the remaining parts, treat each X ˉ g , G \bar{X}_{g,G} X ˉ g , G as one observation in a singleton effective cluster.
Part (b): The effective total sample size is G G G , and every effective cluster size equals one. Thus the cluster-size condition of Hansen and Lee [8 ] is G − 1 → 0 G^{-1}\to0 G − 1 → 0 , which, together with part (a) for r = 1 r=1 r = 1 , proves part (b).
Part (c): For r = 2 r=2 r = 2 , the two conditions in Hansen and Lee [8 ] reduce to
( ∑ g = 1 G 1 2 ) 2 / 2 G = 1 and max g ≤ G 1 2 G = 1 G → 0. \frac{\left(\sum_{g=1}^{G}1^2\right)^{2/2}}{G}=1
\quad\text{and}\quad
\max_{g\le G}\frac{1^2}{G}=\frac{1}{G}\to0. G ( ∑ g = 1 G 1 2 ) 2/2 = 1 and g ≤ G max G 1 2 = G 1 → 0.
Part (a) supplies the squared uniform-integrability condition in that paper. It also implies uniformly bounded second moments of the cluster averages, so sup G ∥ Ω X , G ∥ < ∞ \sup_G\lVert\Omega_{X,G}\rVert<\infty sup G ∥ Ω X , G ∥ < ∞ and 1 G ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ] } = O p ( 1 ) \frac{1}{\sqrt{G}}\sum_{g=1}^{G}\{\bar{X}_{g,G}-\mathbb{E}_G[\bar{X}_{g,G}]\}=O_p(1) G 1 ∑ g = 1 G { X ˉ g , G − E G [ X ˉ g , G ]} = O p ( 1 ) by Chebyshev’s inequality. Under the stated lower eigenvalue bound, Hansen and Lee [8 ] gives the CLT. ∎
Proof of Thm:uwlln
Proof of Theorem 3.1(a). Let
ψ ˉ ( W g ; θ ) : = 1 N g ∑ i ∈ G g ψ ( W i ; θ ) . \bar{\psi}(\mathbf{W}_g;\theta) := \frac{1}{N_g}\sum_{i\in\mathcal{G}_g}\psi(\mathbf{W}_i;\theta). ψ ˉ ( W g ; θ ) := N g 1 i ∈ G g ∑ ψ ( W i ; θ ) .
Then P ^ G ψ ˉ ( W g ; θ ) = G − 1 ∑ g = 1 G ψ ˉ ( W g ; θ ) \widehat{\mathbb{P}}_G\bar{\psi}(\mathbf{W}_g;\theta)=G^{-1}\sum_{g=1}^{G}\bar{\psi}(\mathbf{W}_g;\theta) P G ψ ˉ ( W g ; θ ) = G − 1 ∑ g = 1 G ψ ˉ ( W g ; θ ) and P G ψ ˉ ( W g ; θ ) = G − 1 ∑ g = 1 G E G [ ψ ˉ ( W g ; θ ) ] \mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\theta)=G^{-1}\sum_{g=1}^{G}\mathbb{E}_G[\bar{\psi}(\mathbf{W}_g;\theta)] P G ψ ˉ ( W g ; θ ) = G − 1 ∑ g = 1 G E G [ ψ ˉ ( W g ; θ )] .
For each fixed θ ∈ Θ \theta\in\Theta θ ∈ Θ , apply Lemma S.2.1(b) to X g i , G = ψ ( W i ; θ ) X_{gi,G}=\psi(\mathbf{W}_i;\theta) X g i , G = ψ ( W i ; θ ) . Independence follows from Assumption 1, and the primitive uniform-integrability condition is Assumption 2(b). Thus, pointwise convergence holds:
1 G ∑ g = 1 G ( ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ ) ] ) ⟶ p 0. \frac{1}{G}\sum_{g=1}^{G}\left(\bar{\psi}(\mathbf{W}_g;\theta)-\mathbb{E}_G[\bar{\psi}(\mathbf{W}_g;\theta)]\right)\overset{p}{\longrightarrow}0. G 1 g = 1 ∑ G ( ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ )] ) ⟶ p 0.
Because Θ \Theta Θ is compact, it is totally bounded under the Euclidean metric. This pointwise convergence and the stochastic equicontinuity imposed in Assumption 2(c) therefore verify, respectively, P-WCON and SE in Andrews [2 ] . Applying that result componentwise to the finite-dimensional process U G \mathbf{U}_G U G gives
sup θ ∈ Θ ∥ 1 G ∑ g = 1 G ( ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ ) ] ) ∥ ⟶ p 0. \sup_{\theta\in\Theta}\left\|\frac{1}{G}\sum_{g=1}^{G}\left(\bar{\psi}(\mathbf{W}_g;\theta)-\mathbb{E}_G[\bar{\psi}(\mathbf{W}_g;\theta)]\right)\right\|\overset{p}{\longrightarrow}0. θ ∈ Θ sup G 1 g = 1 ∑ G ( ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ )] ) ⟶ p 0.
Since ( P ^ G − P G ) ψ ˉ ( W g ; θ ) = G − 1 ∑ g = 1 G { ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ ) ] } (\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\psi}(\mathbf{W}_g;\theta)=G^{-1}\sum_{g=1}^{G}\{\bar{\psi}(\mathbf{W}_g;\theta)-\mathbb{E}_G[\bar{\psi}(\mathbf{W}_g;\theta)]\} ( P G − P G ) ψ ˉ ( W g ; θ ) = G − 1 ∑ g = 1 G { ψ ˉ ( W g ; θ ) − E G [ ψ ˉ ( W g ; θ )]} , this establishes Equation (2) .
Proof of Theorem 3.1(b). Define the population cluster-average moment map
Ψ G ( θ ) : = P G ψ ˉ ( W g ; θ ) . \Psi_G(\theta):=\mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\theta). Ψ G ( θ ) := P G ψ ˉ ( W g ; θ ) .
By the triangle inequality,
∥ Ψ G ( θ ^ S A C R ) ∥ = ∥ P G ψ ˉ ( W g ; θ ^ S A C R ) ∥ ≤ ∥ ( P ^ G − P G ) ψ ˉ ( W g ; θ ^ S A C R ) ∥ + ∥ P ^ G ψ ˉ ( W g ; θ ^ S A C R ) ∥ ≤ sup θ ∈ Θ ∥ ( P ^ G − P G ) ψ ˉ ( W g ; θ ) ∥ + ∥ P ^ G ψ ˉ ( W g ; θ ^ S A C R ) ∥ . \begin{aligned}
\|\Psi_G(\widehat{\theta}^{SACR})\|
&=\|\mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\widehat{\theta}^{SACR})\|\\
&\leq\|(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\psi}(\mathbf{W}_g;\widehat{\theta}^{SACR})\|+\|\widehat{\mathbb{P}}_G\bar{\psi}(\mathbf{W}_g;\widehat{\theta}^{SACR})\|\\
&\leq\sup_{\theta\in\Theta}\|(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\psi}(\mathbf{W}_g;\theta)\|+\|\widehat{\mathbb{P}}_G\bar{\psi}(\mathbf{W}_g;\widehat{\theta}^{SACR})\|.
\end{aligned} ∥ Ψ G ( θ S A C R ) ∥ = ∥ P G ψ ˉ ( W g ; θ S A C R ) ∥ ≤ ∥ ( P G − P G ) ψ ˉ ( W g ; θ S A C R ) ∥ + ∥ P G ψ ˉ ( W g ; θ S A C R ) ∥ ≤ θ ∈ Θ sup ∥ ( P G − P G ) ψ ˉ ( W g ; θ ) ∥ + ∥ P G ψ ˉ ( W g ; θ S A C R ) ∥.
The first term is o p ( 1 ) o_p(1) o p ( 1 ) by part (a) above, and the second term is o p ( 1 ) o_p(1) o p ( 1 ) by the approximate estimating equation condition in part (b) of the statement of the theorem. Hence
∥ Ψ G ( θ ^ S A C R ) ∥ = o p ( 1 ) . \|\Psi_G(\widehat{\theta}^{SACR})\|=o_p(1). ∥ Ψ G ( θ S A C R ) ∥ = o p ( 1 ) .
It remains to translate this approximate population zero into parameter consistency. Fix any ϵ > 0 \epsilon> 0 ϵ > 0 and set
c ϵ : = lim inf G → ∞ inf θ ∈ Θ : ∥ θ − θ o ∥ ≥ ϵ ∥ Ψ G ( θ ) ∥ . c_{\epsilon} := \liminf_{G \to\infty}\inf_{\theta\in\Theta:\lVert\theta-\theta_o\rVert\ge\epsilon}\lVert\Psi_G(\theta)\rVert. c ϵ := G → ∞ lim inf θ ∈ Θ : ∥ θ − θ o ∥ ≥ ϵ inf ∥ Ψ G ( θ )∥ .
By the separation condition in Assumption 3(b), c ϵ > 0 c_{\epsilon} > 0 c ϵ > 0 . Therefore, for all sufficiently large G G G ,
inf θ ∈ Θ : ∥ θ − θ o ∥ ≥ ϵ ∥ Ψ G ( θ ) ∥ ≥ c ϵ / 2. \inf_{\theta\in\Theta:\lVert\theta-\theta_o\rVert\ge\epsilon}\lVert\Psi_G(\theta)\rVert\ge c_{\epsilon}/2. θ ∈ Θ : ∥ θ − θ o ∥ ≥ ϵ inf ∥ Ψ G ( θ )∥ ≥ c ϵ /2.
On the event { ∥ θ ^ S A C R − θ o ∥ ≥ ϵ } \{\lVert\widehat{\theta}^{SACR}-\theta_o\rVert\ge\epsilon\} {∥ θ S A C R − θ o ∥ ≥ ϵ } , this implies ∥ Ψ G ( θ ^ S A C R ) ∥ ≥ c ϵ / 2 \lVert\Psi_G(\widehat{\theta}^{SACR})\rVert\ge c_{\epsilon}/2 ∥ Ψ G ( θ S A C R )∥ ≥ c ϵ /2 . Consequently,
P [ ∥ θ ^ S A C R − θ o ∥ ≥ ϵ ] ≤ P [ ∥ Ψ G ( θ ^ S A C R ) ∥ ≥ c ϵ / 2 ] → 0. \mathbb{P}\left[\lVert\widehat{\theta}^{SACR}-\theta_o\rVert\ge\epsilon\right]
\le\mathbb{P}\left[\lVert\Psi_G(\widehat{\theta}^{SACR})\rVert\ge c_{\epsilon}/2\right] \to0. P [ ∥ θ S A C R − θ o ∥ ≥ ϵ ] ≤ P [ ∥ Ψ G ( θ S A C R )∥ ≥ c ϵ /2 ] → 0.
Since ϵ > 0 \epsilon> 0 ϵ > 0 is arbitrary, θ ^ S A C R → p θ o \widehat{\theta}^{SACR} \mathrel{\mathop{\to}\limits^{p}} \theta_o θ S A C R → p θ o .
Proof of Theorem 3.2
Proof of Theorem 3.2(a). By Theorem 3.1(b), θ ^ S A C R → p θ o \widehat{\theta}^{SACR} \mathrel{\mathop{\to}\limits^{p}} \theta_o θ S A C R → p θ o . Define
E G : = G P ^ G ψ ‾ ( W g ; θ ^ S A C R ) = o p ( 1 ) \mathcal{E}_G := \sqrt{G}\widehat{\mathbb{P}}_G\overline{\psi}(\mathbf{W}_g;\widehat{\theta}^{SACR}) = o_p(1) E G := G P G ψ ( W g ; θ S A C R ) = o p ( 1 )
and, on the event that the line segment joining θ ^ S A C R \widehat{\theta}^{SACR} θ S A C R and θ o \theta_o θ o lies in the differentiability neighbourhood, define
H ~ G : = ∫ 0 1 H G ( θ o + u ( θ ^ S A C R − θ o ) ) d u . \widetilde{H}_G := \int_0^1 H_G\left(\theta_o+u(\widehat{\theta}^{SACR}-\theta_o)\right)\,du. H G := ∫ 0 1 H G ( θ o + u ( θ S A C R − θ o ) ) d u .
The population zero in Assumption 3(a) and the fundamental theorem of calculus give, on the event that H ~ G \widetilde{H}_G H G is invertible, the exact identity
G ( θ ^ S A C R − θ o ) = − H ~ G − 1 [ G G ( θ o ) + { G G ( θ ^ S A C R ) − G G ( θ o ) } ⏟ R G − E G ] . (8) \sqrt{G}(\widehat{\theta}^{SACR}-\theta_o)
=
-\widetilde{H}_G^{-1}
\left[
\mathbb{G}_G(\theta_o)
+
\underbrace{\{\mathbb{G}_G(\widehat{\theta}^{SACR})-\mathbb{G}_G(\theta_o)\}}_{\mathcal{R}_G}
-\mathcal{E}_G
\right]. \tag*{(8)} G ( θ S A C R − θ o ) = − H G − 1 G G ( θ o ) + R G { G G ( θ S A C R ) − G G ( θ o )} − E G . (8)
which holds thanks to consistency and interiority in Assumption 4(b) with probability approaching one. In addition, local stochastic equicontinuity in Assumption 6 gives R G = o p ( 1 ) \mathcal{R}_G=o_p(1) R G = o p ( 1 ) , while Assumption 4(a) gives E G = o p ( 1 ) \mathcal{E}_G=o_p(1) E G = o p ( 1 ) .
Consistency and Assumption 4(c) also give H ~ G − H G = o p ( 1 ) \widetilde{H}_G-H_G=o_p(1) H G − H G = o p ( 1 ) . The uniform lower singular-value bound in Assumption 5(b) therefore makes H ~ G \widetilde{H}_G H G and H G H_G H G uniformly invertible with probability approaching one, and the inverse identity gives H ~ G − 1 − H G − 1 = o p ( 1 ) \widetilde{H}_G^{-1}-H_G^{-1}=o_p(1) H G − 1 − H G − 1 = o p ( 1 ) . Moreover, Lemma S.2.1(c), applied to X g i , G = ψ ( W i ; θ o ) X_{gi,G}=\psi(W_i;\theta_o) X g i , G = ψ ( W i ; θ o ) , gives G G ( θ o ) = O p ( 1 ) \mathbb{G}_G(\theta_o)=O_p(1) G G ( θ o ) = O p ( 1 ) . Consequently,
( H ~ G − 1 − H G − 1 ) G G ( θ o ) = o p ( 1 ) . (\widetilde{H}_G^{-1}-H_G^{-1})\mathbb{G}_G(\theta_o)=o_p(1). ( H G − 1 − H G − 1 ) G G ( θ o ) = o p ( 1 ) .
Equation (S.2.2) therefore gives the asymptotically linear representation
G ( θ ^ S A C R − θ o ) = − H G − 1 G G ( θ o ) + o p ( 1 ) (9) \sqrt{G}(\widehat{\theta}^{SACR}-\theta_o)=-H_G^{-1}\mathbb{G}_G(\theta_o)+o_p(1) \tag*{(9)} G ( θ S A C R − θ o ) = − H G − 1 G G ( θ o ) + o p ( 1 ) (9)
as claimed.
Proof of Theorem 3.2(b). Let Z g , G : = ψ ˉ ( W g ; θ o ) Z_{g,G}:=\bar{\psi}(\mathbf{W}_g;\theta_o) Z g , G := ψ ˉ ( W g ; θ o ) . After SACR reweighting, this array has one summand per cluster:
G P ^ G ψ ˉ ( W g ; θ o ) = 1 G ∑ g = 1 G Z g , G . \sqrt{G}\widehat{\mathbb{P}}_G\bar{\psi}(\mathbf{W}_g;\theta_o)=\frac{1}{\sqrt{G}}\sum_{g=1}^{G}Z_{g,G}. G P G ψ ˉ ( W g ; θ o ) = G 1 g = 1 ∑ G Z g , G .
Independence holds by Assumption 1, and the zero-mean condition follows because Assumption 3(a) implies E G [ Z g , G ] = E G [ ψ ˉ ( W g ; θ o ) ] = 0 \mathbb{E}_G[Z_{g,G}]=\mathbb{E}_G[\bar{\psi}(\mathbf{W}_g;\theta_o)]=0 E G [ Z g , G ] = E G [ ψ ˉ ( W g ; θ o )] = 0 for every g g g . The individual squared uniform-integrability condition in Assumption 5(a), together with the lower covariance bound in Assumption 5(b), verifies Lemma S.2.1(c) for X g i , G = ψ ( W i ; θ o ) X_{gi,G}=\psi(\mathbf{W}_i;\theta_o) X g i , G = ψ ( W i ; θ o ) . Therefore
Ω G − 1 / 2 G P ^ G ψ ˉ ( W g ; θ o ) ⟶ d N ( 0 , I ) . \Omega_G^{-1/2}\sqrt{G}\widehat{\mathbb{P}}_G\bar{\psi}(\mathbf{W}_g;\theta_o)\mathrel{\stackrel{d}{\longrightarrow}}\mathcal{N}(0,I). Ω G − 1/2 G P G ψ ˉ ( W g ; θ o ) ⟶ d N ( 0 , I ) .
To combine this standardised CLT with the asymptotically linear representation (S.2.3), define
V G : = H G − 1 Ω G ( H G − 1 ) ′ . V_G:=H_G^{-1}\Omega_G(H_G^{-1})'. V G := H G − 1 Ω G ( H G − 1 ) ′ .
It remains only to justify standardisation by the drifting matrix sequence. The uniform-integrability condition in Assumption 2(b) implies sup G , θ ∥ P G ψ ˉ ( W g ; θ ) ∥ < ∞ \sup_{G,\theta}\|\mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\theta)\|<\infty sup G , θ ∥ P G ψ ˉ ( W g ; θ ) ∥ < ∞ . Interiority permits a fixed r > 0 r>0 r > 0 such that θ o + r v ∈ Θ \theta_o+rv\in\Theta θ o + r v ∈ Θ for every unit vector v v v , and Assumption 4(c) permits this r r r to be chosen so that sup ∥ θ − θ o ∥ ≤ r ∥ H G ( θ ) − H G ∥ ≤ 1 \sup_{\|\theta-\theta_o\|\le r}\|H_G(\theta)-H_G\|\le1 sup ∥ θ − θ o ∥ ≤ r ∥ H G ( θ ) − H G ∥ ≤ 1 for all sufficiently large G G G . Define M G ( θ ) : = P G ψ ˉ ( W g ; θ ) M_G(\theta):=\mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\theta) M G ( θ ) := P G ψ ˉ ( W g ; θ ) . Applying the fundamental theorem of calculus along t ↦ θ o + t v t\mapsto\theta_o+tv t ↦ θ o + t v gives
M G ( θ o + r v ) − M G ( θ o ) = ∫ 0 r H G ( θ o + t v ) v d t . M_G(\theta_o+rv)-M_G(\theta_o)=\int_0^r H_G(\theta_o+tv)v\,\mathrm{d}t. M G ( θ o + r v ) − M G ( θ o ) = ∫ 0 r H G ( θ o + t v ) v d t .
Consequently,
∥ M G ( θ o + r v ) − M G ( θ o ) − r H G v ∥ = ∥ ∫ 0 r H G ( θ o + t v ) v d t − ∫ 0 r H G v d t ∥ ≤ ∫ 0 r ∥ H G ( θ o + t v ) − H G ∥ d t ≤ r . \begin{aligned}
\|M_G(\theta_o+rv)-M_G(\theta_o)-rH_Gv\|
&=\left\|\int_0^r H_G(\theta_o+tv)v\,\mathrm{d}t-\int_0^r H_Gv\,\mathrm{d}t\right\|\\
&\le\int_0^r\|H_G(\theta_o+tv)-H_G\|\,\mathrm{d}t\le r.
\end{aligned} ∥ M G ( θ o + r v ) − M G ( θ o ) − r H G v ∥ = ∫ 0 r H G ( θ o + t v ) v d t − ∫ 0 r H G v d t ≤ ∫ 0 r ∥ H G ( θ o + t v ) − H G ∥ d t ≤ r .
Taking K < ∞ K<\infty K < ∞ to be a uniform bound on ∥ M G ( θ ) ∥ \|M_G(\theta)\| ∥ M G ( θ ) ∥ , the triangle inequality yields
r ∥ H G v ∥ ≤ ∥ P G ψ ˉ ( W g ; θ o + r v ) − P G ψ ˉ ( W g ; θ o ) ∥ + r ≤ 2 K + r . r\|H_Gv\|\le\|\mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\theta_o+rv)-\mathbb{P}_G\bar{\psi}(\mathbf{W}_g;\theta_o)\|+r\le2K+r. r ∥ H G v ∥ ≤ ∥ P G ψ ˉ ( W g ; θ o + r v ) − P G ψ ˉ ( W g ; θ o ) ∥ + r ≤ 2 K + r .
Thus sup G ∥ H G ∥ < ∞ \sup_G\|H_G\|<\infty sup G ∥ H G ∥ < ∞ . Moreover, λ min ( V G ) ≥ λ min ( Ω G ) ∥ H G ∥ 2 \lambda_{\min}(V_G)\ge\frac{\lambda_{\min}(\Omega_G)}{\|H_G\|^2} λ m i n ( V G ) ≥ ∥ H G ∥ 2 λ m i n ( Ω G ) , so inf G λ min ( V G ) > 0 \inf_G\lambda_{\min}(V_G)>0 inf G λ m i n ( V G ) > 0 and sup G ∥ V G − 1 / 2 ∥ < ∞ \sup_G\|V_G^{-1/2}\|<\infty sup G ∥ V G − 1/2 ∥ < ∞ . Let U G : = Ω G − 1 / 2 G P ^ G ψ ˉ ( W g ; θ o ) U_G:=\Omega_G^{-1/2}\sqrt{G}\widehat{\mathbb{P}}_G\bar{\psi}(\mathbf{W}_g;\theta_o) U G := Ω G − 1/2 G P G ψ ˉ ( W g ; θ o ) , and Q G : = V G − 1 / 2 H G − 1 Ω G 1 / 2 Q_G:=V_G^{-1/2}H_G^{-1}\Omega_G^{1/2} Q G := V G − 1/2 H G − 1 Ω G 1/2 . Then U G ⟶ d N ( 0 , I ) U_G\mathrel{\stackrel{d}{\longrightarrow}}\mathcal{N}(0,I) U G ⟶ d N ( 0 , I ) and Q G Q G ′ = I Q_GQ_G^{\prime}=I Q G Q G ′ = I . The linear representation therefore yields
V G − 1 / 2 G ( θ ^ S A C R − θ o ) = − V G − 1 / 2 H G 1 / 2 Ω G 1 / 2 Ω G − 1 / 2 G P ‾ G ψ ‾ ( W g ; θ o ) + o p ( 1 ) = − Q G U G + o p ( 1 ) . V_G^{-1/2}\sqrt{G}(\widehat{\theta}^{SACR}-\theta_o)=-V_G^{-1/2}H_G^{1/2}\Omega_G^{1/2}\Omega_G^{-1/2}\sqrt{G}\,\overline{\mathbb{P}}_G\overline{\psi}(W_g;\theta_o)+o_p(1)
=-Q_GU_G+o_p(1). V G − 1/2 G ( θ S A C R − θ o ) = − V G − 1/2 H G 1/2 Ω G 1/2 Ω G − 1/2 G P G ψ ( W g ; θ o ) + o p ( 1 ) = − Q G U G + o p ( 1 ) .
Every Q G Q_G Q G is orthogonal, so rotational invariance gives the studentised form
V G − 1 / 2 G ( θ ^ S A C R − θ o ) → d N ( 0 , I ) . V_G^{-1/2}\sqrt{G}(\widehat{\theta}^{SACR}-\theta_o)\xrightarrow{d}\mathcal{N}(0,I). V G − 1/2 G ( θ S A C R − θ o ) d N ( 0 , I ) .
Proof of lem:functionalep
Part (a): Fix a coordinate j j j . By Assumption 2 ′ 2' 2 ′ (a), the normalised full cluster array is manageable with envelope F G g , j F_{Gg,j} F G g , j . Its joint almost-measurable-Suslin parametrisation justifies the symmetrisation and interchange of expectations, as in Kosorok [11 ] . Pollard [15 ] ’s maximal inequality (Pollard [15 ] ), with p = 2 p=2 p = 2 , and the cluster-envelope moment bound supplied by Lemma S.2.1(a), applied to the squared individual envelopes under Assumption 2 ′ 2' 2 ′ (b), give
E G ∗ [ sup τ ∈ T , t ∈ Θ ∣ ∑ g = 1 G ( Z G g , j ( τ , t ) − E G Z G g , j ( τ , t ) ) ∣ 2 ] ≤ C ∑ g = 1 G E G F G g , j 2 = C G ∑ g = 1 G E G F ‾ G g , j 2 = O ( 1 ) . \mathbb{E}_G^*\left[\sup_{\tau\in\mathcal{T},\,t\in\Theta}\left|\sum_{g=1}^{G}\left(Z_{Gg,j}(\tau,t)-\mathbb{E}_G Z_{Gg,j}(\tau,t)\right)\right|^2\right]\leq C\sum_{g=1}^{G}\mathbb{E}_G F_{Gg,j}^2=\frac{C}{G}\sum_{g=1}^{G}\mathbb{E}_G\overline{F}_{Gg,j}^2=O(1). E G ∗ τ ∈ T , t ∈ Θ sup g = 1 ∑ G ( Z G g , j ( τ , t ) − E G Z G g , j ( τ , t ) ) 2 ≤ C g = 1 ∑ G E G F G g , j 2 = G C g = 1 ∑ G E G F G g , j 2 = O ( 1 ) .
Consequently,
sup τ ∈ T , t ∈ Θ ∣ 1 G ∑ g = 1 G { ψ ‾ τ , j ( W g ; t ) − E G ψ ‾ τ , j ( W g ; t ) } ∣ = O p ( G − 1 / 2 ) = o p ( 1 ) . \sup_{\tau\in\mathcal{T},\,t\in\Theta}\left|\frac{1}{G}\sum_{g=1}^{G}\left\{\overline{\psi}_{\tau,j}(W_g;t)-\mathbb{E}_G\overline{\psi}_{\tau,j}(W_g;t)\right\}\right|=O_p(G^{-1/2})=o_p(1). τ ∈ T , t ∈ Θ sup G 1 g = 1 ∑ G { ψ τ , j ( W g ; t ) − E G ψ τ , j ( W g ; t ) } = O p ( G − 1/2 ) = o p ( 1 ) .
Taking the maximum over the finitely many coordinates and observing that every admissible path is contained in the joint index set proves part (a).
Part (b): Cluster-array manageability is imposed directly in Assumption 2 ′ 2' 2 ′ (a). Fix a coordinate j j j , and write
d G g , j ( τ , t ) : = ψ ‾ τ , j ( W g ; t ) − ψ ‾ τ , j ( W g ; θ o ( τ ) ) . d_{Gg,j}(\tau,t):=\overline{\psi}_{\tau,j}(W_g;t)-\overline{\psi}_{\tau,j}(W_g;\theta_o(\tau)). d G g , j ( τ , t ) := ψ τ , j ( W g ; t ) − ψ τ , j ( W g ; θ o ( τ )) .
Define
I δ : = { ( τ , t ) ∈ T × Θ : ∥ t − θ o ( τ ) ∥ ≤ δ } , R G , j 2 ( δ ) : = sup ( τ , t ) ∈ I δ 1 G ∑ g = 1 G d G g , j ( τ , t ) 2 . \mathcal{I}_{\delta}:=\{(\tau,t)\in\mathcal{T}\times\Theta:\lVert t-\theta_o(\tau)\rVert\leq\delta\},\qquad R^2_{G,j}(\delta):=\sup_{(\tau,t)\in\mathcal{I}_{\delta}}\frac{1}{G}\sum_{g=1}^{G}d_{Gg,j}(\tau,t)^2. I δ := {( τ , t ) ∈ T × Θ : ∥ t − θ o ( τ )∥ ≤ δ } , R G , j 2 ( δ ) := ( τ , t ) ∈ I δ sup G 1 g = 1 ∑ G d G g , j ( τ , t ) 2 .
Manageability and the joint AMS property follow from Assumption 2 ′ 2' 2 ′ (a), using stability of manageability under differences. By Lemma S.2.1(a), applied to the squared individual envelopes, Assumption 2 ′ 2' 2 ′ (b) implies uniform integrability of the squared cluster envelopes, supplying the required envelope moment and Lindeberg bounds. Applying the truncation and squared-increment argument in the proof of Pollard [15 ] gives, for each fixed δ > 0 \delta> 0 δ > 0 ,
sup ( τ , t ) ∈ I δ ∣ 1 G ∑ g = 1 G ( d G g , j ( τ , t ) 2 − E G [ d G g , j ( τ , t ) 2 ] ) ∣ = o p ( 1 ) . (10) \sup_{(\tau,t)\in\mathcal{I}_{\delta}}\left|\frac{1}{G}\sum_{g=1}^{G}\left(d_{Gg,j}(\tau,t)^2-\mathbb{E}_{G}[d_{Gg,j}(\tau,t)^2]\right)\right|=o_p(1). \tag*{(10)} ( τ , t ) ∈ I δ sup G 1 g = 1 ∑ G ( d G g , j ( τ , t ) 2 − E G [ d G g , j ( τ , t ) 2 ] ) = o p ( 1 ) . (10)
Here Lindeberg truncation permits application of the squared-process maximal inequality without imposing fourth moments on the original envelope. Consequently, Assumption 2’(c) implies
lim δ ↓ 0 lim sup G → ∞ P [ R G , j ( δ ) > ε ] = 0 , ε > 0. (11) \lim_{\delta\downarrow0}\limsup_{G\to\infty}\mathbb{P}\left[R_{G,j}(\delta)>\varepsilon\right]=0,\qquad\varepsilon>0. \tag*{(11)} δ ↓ 0 lim G → ∞ lim sup P [ R G , j ( δ ) > ε ] = 0 , ε > 0. (11)
The local maximal-inequality argument in the same proof of Pollard [15 ] therefore yields
lim δ ↓ 0 lim sup G → ∞ P [ sup ( τ , t ) ∈ I δ ∣ 1 G ∑ g = 1 G ( d G g , j ( τ , t ) − E G d G g , j ( τ , t ) ) ∣ > ε ] = 0. \lim_{\delta\downarrow0}\limsup_{G\to\infty}\mathbb{P}\left[\sup_{(\tau,t)\in\mathcal{I}_{\delta}}\left|\frac{1}{\sqrt{G}}\sum_{g=1}^{G}\left(d_{Gg,j}(\tau,t)-\mathbb{E}_{G}d_{Gg,j}(\tau,t)\right)\right|>\varepsilon\right]=0. δ ↓ 0 lim G → ∞ lim sup P [ ( τ , t ) ∈ I δ sup G 1 g = 1 ∑ G ( d G g , j ( τ , t ) − E G d G g , j ( τ , t ) ) > ε ] = 0.
Taking the maximum over the finitely many coordinates and restricting the supremum to admissible parameter paths proves part (b). In particular, R G , j ( δ G ) = o p ( 1 ) R_{G,j}(\delta_G)=o_p(1) R G , j ( δ G ) = o p ( 1 ) for every deterministic sequence δ G ↓ 0 \delta_G\downarrow0 δ G ↓ 0 .
Proof of Thm:weakconvergence
Part (a):
The proof is organised around the exact functional counterpart of the pointwise decomposition in (S.2.2). Define the scaled approximate-zero process
E G : = G P G Ψ ( W g ; θ ^ ) = o p ( 1 ) in ℓ ∞ ( T ) \mathcal{E}_G:=\sqrt{G}\mathbb{P}_G\Psi(W_g;\hat{\theta})=o_p(1)\qquad\text{in }\ell^\infty(\mathcal{T}) E G := G P G Ψ ( W g ; θ ^ ) = o p ( 1 ) in ℓ ∞ ( T )
and, whenever the line segments joining θ o ( τ ) \theta_o(\tau) θ o ( τ ) and θ ^ ( τ ) \hat{\theta}(\tau) θ ^ ( τ ) lie in the differentiability neighbourhood, let
H ~ G ( τ ) : = ∫ 0 1 ∂ ∂ t ′ P G ψ τ ( W g ; t ′ ) ∣ t ′ = θ o ( τ ) + u { θ ^ ( τ ) − θ o ( τ ) } d u . \widetilde{H}_G(\tau):=\int_0^1\left.\frac{\partial}{\partial t'}\mathbb{P}_G\psi_\tau(W_g;t')\right|_{t'=\theta_o(\tau)+u\{\hat{\theta}(\tau)-\theta_o(\tau)\}}\,du. H G ( τ ) := ∫ 0 1 ∂ t ′ ∂ P G ψ τ ( W g ; t ′ ) t ′ = θ o ( τ ) + u { θ ^ ( τ ) − θ o ( τ )} d u .
The population zero in Assumption 3’(a) and the fundamental theorem of calculus then give the exact identity
E G = G G ( θ o ) + { G G ( θ ^ ) − G G ( θ o ) } + H ~ G G ( θ ^ − θ o ) . \mathcal{E}_G=\mathbb{G}_G(\theta_o)+\{\mathbb{G}_G(\hat{\theta})-\mathbb{G}_G(\theta_o)\}+\widetilde{H}_G\sqrt{G}(\hat{\theta}-\theta_o). E G = G G ( θ o ) + { G G ( θ ^ ) − G G ( θ o )} + H G G ( θ ^ − θ o ) .
Consequently, on the event that H ~ G \widetilde{H}_G H G is uniformly invertible,
G ( θ ^ − θ o ) = − H ~ G − 1 [ G G ( θ o ) + { G G ( θ ^ ) − G G ( θ o ) } ⏟ R G − E G ] . (12) \sqrt{G}(\hat{\theta}-\theta_o)=-\widetilde{H}_G^{-1}\left[\mathbb{G}_G(\theta_o)+\underbrace{\{\mathbb{G}_G(\hat{\theta})-\mathbb{G}_G(\theta_o)\}}_{\mathcal{R}_G}-\mathcal{E}_G\right]. \tag*{(12)} G ( θ ^ − θ o ) = − H G − 1 G G ( θ o ) + R G { G G ( θ ^ ) − G G ( θ o )} − E G . (12)
The first two steps below verify the event and dispatch the two remainder terms. The weak limit is established only afterwards for the surviving first-order term.
Step 1: Establish uniform consistency of the estimator. The argument is the sup-norm analogue of Theorem 3.1(b). The functional UWLLN in Lemma 3.1(a), together with the approximate empirical zero (6) , implies
∥ P G Ψ ˉ ( W g ; θ ^ ) ∥ T ≤ sup θ ∈ Θ T ∥ ( P ^ G − P G ) Ψ ˉ ( W g ; θ ) ∥ T + ∥ P ^ G Ψ ˉ ( W g ; θ ^ ) ∥ T = o p ( 1 ) . \left\|\mathbb{P}_G\bar{\Psi}(W_g;\widehat{\theta})\right\|_{\mathcal{T}} \leq\sup_{\theta\in\Theta_{\mathcal{T}}}\left\|(\widehat{\mathbb{P}}_G-\mathbb{P}_G)\bar{\Psi}(W_g;\theta)\right\|_{\mathcal{T}}+\left\|\widehat{\mathbb{P}}_G\bar{\Psi}(W_g;\widehat{\theta})\right\|_{\mathcal{T}}=o_p(1). P G Ψ ˉ ( W g ; θ ) T ≤ θ ∈ Θ T sup ( P G − P G ) Ψ ˉ ( W g ; θ ) T + P G Ψ ˉ ( W g ; θ ) T = o p ( 1 ) .
By arguments analogous to the concluding part of the proof of Theorem 3.1(b), the functional separation condition in Assumption 3′(b) therefore yields
∥ θ ^ − θ o ∥ T = o p ( 1 ) . \left\|\widehat{\theta}-\theta_o\right\|_{\mathcal{T}}=o_p(1). θ − θ o T = o p ( 1 ) .
Indeed, failure of this conclusion along a subsequence would place θ ^ \widehat{\theta} θ , with non-vanishing probability, in a set on which functional separation bounds the population moment uniformly away from zero, contradicting the preceding display.
Step 2: Validate the decomposition (S.2.6) and dispatch its degenerate terms. Uniform consistency, uniform interiority in Assumption 4′(b), and uniform differentiability in Assumption 4′(c) imply that the line segments used to define H ~ G \widetilde{H}_G H G lie in the differentiability neighbourhood with probability approaching one. Thus (S.2.6) is an exact identity on such an event. The same conditions give
∥ H ~ G − H ∥ T = o p ( 1 ) . \left\|\widetilde{H}_G-H\right\|_{\mathcal{T}}=o_p(1). H G − H T = o p ( 1 ) .
The local stochastic equicontinuity result in Lemma 3.1(b), evaluated at the uniformly consistent path, dispatches the remainder in (S.2.6):
∥ R G ∥ T = o p ( 1 ) \left\|R_G\right\|_{\mathcal{T}}=o_p(1) ∥ R G ∥ T = o p ( 1 )
The approximate-zero error satisfies ∥ E G ∥ T = o p ( 1 ) \left\|E_G\right\|_{\mathcal{T}}=o_p(1) ∥ E G ∥ T = o p ( 1 ) directly by (6) . Finally, the uniform lower singular-value bound on H H H , together with ∥ H ~ G − H ∥ T = o p ( 1 ) \left\|\widetilde{H}_G-H\right\|_{\mathcal{T}}=o_p(1) H G − H T = o p ( 1 ) (Assumption 4′(c)), ensures that H ~ G \widetilde{H}_G H G is uniformly invertible and ∥ H ~ G − 1 ∥ T = O p ( 1 ) \left\|\widetilde{H}_G^{-1}\right\|_{\mathcal{T}}=O_p(1) H G − 1 T = O p ( 1 ) . Hence the exact decomposition reduces to
G ( θ ^ − θ o ) ( ⋅ ) = − H ~ G ( ⋅ ) − 1 G G ( θ o ) ( ⋅ ) + o p ( 1 ) in ℓ ∞ ( T ) . (13) \sqrt{G}(\widehat{\theta}-\theta_o)(\cdot)=-\widetilde{H}_G(\cdot)^{-1}G_G(\theta_o)(\cdot)+o_p(1)\quad\text{in }\ell^\infty(\mathcal{T}). \tag*{(13)} G ( θ − θ o ) ( ⋅ ) = − H G ( ⋅ ) − 1 G G ( θ o ) ( ⋅ ) + o p ( 1 ) in ℓ ∞ ( T ) . (13)
Step 3: Establish weak convergence of the surviving score process. For each a ∈ R p θ a\in\mathbb{R}^{p_\theta} a ∈ R p θ , apply Kosorok [11 ] , equivalently Kosorok [10 ] , to the scalar array
X G a ( τ ) : = G − 1 / 2 a ′ ψ ~ τ ( W g ; θ o ( τ ) ) . X_G^a(\tau):=G^{-1/2}a^{\prime}\widetilde{\psi}_{\tau}(W_g;\theta_o(\tau)). X G a ( τ ) := G − 1/2 a ′ ψ τ ( W g ; θ o ( τ )) .
Finite sums and nonnegative rescalings of manageable arrays are manageable by Pollard [15 ] , with envelope ∑ j ∣ a j ∣ F G g , j \sum_j |a_j|F_{Gg,j} ∑ j ∣ a j ∣ F G g , j . The truth-indexed vector array is almost measurable Suslin by Assumption 2′(a). Its common Suslin approximating sets also serve each scalar projection, since
∑ g ∣ a ′ ( z g − z ~ g ) ∣ 2 ≤ ∥ a ∥ 2 ∑ g ∥ z g − z ~ g ∥ 2 , \sum_g |a^{\prime}(z_g-\widetilde{z}_g)|^2\leq\|a\|^2\sum_g\|z_g-\widetilde{z}_g\|^2, g ∑ ∣ a ′ ( z g − z g ) ∣ 2 ≤ ∥ a ∥ 2 g ∑ ∥ z g − z g ∥ 2 ,
and scalar projection preserves joint measurability on those sets. Condition B follows from Assumption 5’(a), condition E is Assumption 5’(b), and the envelope bound and Lindeberg conditions C and D follow from Assumption 2’(b) via Lemma S.2.1(a), applied to the squared individual envelopes, since the truth-indexed class is a subclass of the full class. Thus every scalar projection is asymptotically tight.
For joint finite-dimensional convergence, fix τ 1 , … , τ k ∈ T \tau_1,\ldots,\tau_k \in\mathcal{T} τ 1 , … , τ k ∈ T and stack the individual scores as
X g i , G : = ( ψ τ 1 ( W i ; θ o ( τ 1 ) ) ′ , … , ψ τ k ( W i ; θ o ( τ k ) ) ′ ) ′ . X_{gi,G} := \left(\psi_{\tau_1}(W_i;\theta_o(\tau_1))^\prime,\ldots,\psi_{\tau_k}(W_i;\theta_o(\tau_k))^\prime\right)^\prime. X g i , G := ( ψ τ 1 ( W i ; θ o ( τ 1 ) ) ′ , … , ψ τ k ( W i ; θ o ( τ k ) ) ′ ) ′ .
The corresponding cluster average X ˉ g , G \bar{X}_{g,G} X ˉ g , G stacks the cluster scores at these indices. Since ∥ X g i , G ∥ 2 ≤ k ∑ j = 1 p θ F j ( W i ) 2 \lVert X_{gi,G}\rVert^2 \le k\sum_{j=1}^{p_\theta}F_j(W_i)^2 ∥ X g i , G ∥ 2 ≤ k ∑ j = 1 p θ F j ( W i ) 2 , Assumption 2’(b) implies the primitive squared uniform-integrability condition (S.2.1) with r = 2 r=2 r = 2 . Moreover, Assumption 3’(a) gives E G [ X ˉ g , G ] = 0 \mathbb{E}_G[\bar{X}_{g,G}]=0 E G [ X ˉ g , G ] = 0 , and Assumption 1 gives independence across g g g . Let Ω k , G \Omega_{k,G} Ω k , G denote the covariance matrix of G − 1 / 2 ∑ g X ˉ g , G G^{-1/2}\sum_g\bar{X}_{g,G} G − 1/2 ∑ g X ˉ g , G . Its blocks are Ω G ( τ ℓ , τ m ) \Omega_G(\tau_\ell,\tau_m) Ω G ( τ ℓ , τ m ) , so Assumption 5’(a) gives Ω k , G → Ω k \Omega_{k,G}\to\Omega_k Ω k , G → Ω k .
For any fixed b ∈ R k p θ b\in\mathbb{R}^{kp_\theta} b ∈ R k p θ , if b ′ Ω k b > 0 b^\prime\Omega_kb>0 b ′ Ω k b > 0 , then b ′ Ω k , G b b^\prime\Omega_{k,G}b b ′ Ω k , G b is eventually bounded away from zero and Lemma S.2.1(c), applied to the scalar projection b ′ X g i , G b^\prime X_{gi,G} b ′ X g i , G , gives the standardised CLT. If b ′ Ω k b = 0 b^\prime\Omega_kb=0 b ′ Ω k b = 0 , Chebyshev’s inequality instead gives G − 1 / 2 ∑ g b ′ X ˉ g , G ⟶ p 0 G^{-1/2}\sum_g b^\prime\bar{X}_{g,G}\overset{p}{\longrightarrow}0 G − 1/2 ∑ g b ′ X ˉ g , G ⟶ p 0 . The Cramér–Wold device therefore yields convergence of the stacked cluster-score vector to N ( 0 , Ω k ) \mathcal{N}(0,\Omega_k) N ( 0 , Ω k ) . Because p θ p_\theta p θ is finite, coordinatewise tightness and joint finite-dimensional convergence give
G G ( θ o ) ⇝ G 0 in ℓ ∞ ( T ) , (14) \mathbb{G}_G(\theta_o)\rightsquigarrow\mathbb{G}_0 \quad\text{in }\ell^\infty(\mathcal{T}), \tag*{(14)} G G ( θ o ) ⇝ G 0 in ℓ ∞ ( T ) , (14)
with covariance kernel Ω \Omega Ω . For every scalar projection, the same theorem also gives total boundedness in its limiting semimetric and a uniformly continuous version of the projected Gaussian limit. This proves weak convergence of the sole surviving first-order term in (S.2.6).
Step 4: Conclude from the decomposition. The score functional CLT in (S.2.8) implies ∥ G G ( θ o ) ∥ T = O p ( 1 ) \lVert\mathbb{G}_G(\theta_o)\rVert_{\mathcal{T}}=O_p(1) ∥ G G ( θ o ) ∥ T = O p ( 1 ) . The inverse identity and the uniform convergence established in Step 2 therefore give
∥ ( H ~ G − 1 − H − 1 ) G G ( θ o ) ∥ T = o p ( 1 ) . \left\lVert\left(\widetilde{H}_G^{-1}-H^{-1}\right)\mathbb{G}_G(\theta_o)\right\rVert_{\mathcal{T}}=o_p(1). ( H G − 1 − H − 1 ) G G ( θ o ) T = o p ( 1 ) .
It follows from (S.2.7) that
G ( θ ^ − θ o ) ( ⋅ ) = − H ( ⋅ ) − 1 G G ( θ o ) ( ⋅ ) + o p ( 1 ) . \sqrt{G}(\hat{\theta}-\theta_o)(\cdot)=-H(\cdot)^{-1}\mathbb{G}_G(\theta_o)(\cdot)+o_p(1). G ( θ ^ − θ o ) ( ⋅ ) = − H ( ⋅ ) − 1 G G ( θ o ) ( ⋅ ) + o p ( 1 ) .
The continuous mapping theorem applied to (S.2.8) now yields
G ( θ ^ − θ o ) ( ⋅ ) ⇝ − H ( ⋅ ) − 1 G 0 ( ⋅ ) in ℓ ∞ ( T ) . \sqrt{G}(\hat{\theta}-\theta_o)(\cdot)\rightsquigarrow-H(\cdot)^{-1}\mathbb{G}_0(\cdot)\quad\text{in }\ell^\infty(\mathcal{T}). G ( θ ^ − θ o ) ( ⋅ ) ⇝ − H ( ⋅ ) − 1 G 0 ( ⋅ ) in ℓ ∞ ( T ) .
Part (b): Since η ^ = Φ ϕ ( θ ^ ) \hat{\eta}=\Phi_\phi(\hat{\theta}) η ^ = Φ ϕ ( θ ^ ) , the ordinary functional delta method of Chernozhukov, Fernández-Val, and Melly [4 ] , applied to Theorem 3.3, gives
G ( η ^ − η o ) ( ⋅ ) ⇝ Φ ϕ , θ o ′ [ Z θ ] ( ⋅ ) in ℓ ∞ ( T ) . \sqrt{G}(\hat{\eta}-\eta_o)(\cdot)\rightsquigarrow\Phi_{\phi,\theta_o}^{\prime}[Z_\theta](\cdot)\quad\text{in }\ell^\infty(\mathcal{T}). G ( η ^ − η o ) ( ⋅ ) ⇝ Φ ϕ , θ o ′ [ Z θ ] ( ⋅ ) in ℓ ∞ ( T ) .
Proof of Thm:bootconst
Part (a):
Define
Ψ ^ G ( θ , τ ) : = P ^ G ψ ˉ τ ( W g ; θ ( τ ) ) , Ψ G ( θ , τ ) : = P G ψ ˉ τ ( W g ; θ ( τ ) ) . \widehat{\Psi}_G(\theta,\tau):=\widehat{\mathbb{P}}_G\bar{\psi}_{\tau}(W_g;\theta(\tau)),\qquad\Psi_G(\theta,\tau):=\mathbb{P}_G\bar{\psi}_{\tau}(W_g;\theta(\tau)). Ψ G ( θ , τ ) := P G ψ ˉ τ ( W g ; θ ( τ )) , Ψ G ( θ , τ ) := P G ψ ˉ τ ( W g ; θ ( τ )) .
The difference between the bootstrap and original approximate-zero errors is
E G ∗ : = G ( Ψ ^ G ∗ ( θ ^ ∗ ) − Ψ ^ G ( θ ^ ) ) = o P ∗ ( 1 ) in ℓ ∞ ( T ) \mathcal{E}_G^*:=\sqrt{G}\left(\widehat{\Psi}_G^*(\widehat{\theta}^*)-\widehat{\Psi}_G(\widehat{\theta})\right)=o_{P^*}(1)\quad\text{in }\ell^\infty(\mathcal{T}) E G ∗ := G ( Ψ G ∗ ( θ ∗ ) − Ψ G ( θ ) ) = o P ∗ ( 1 ) in ℓ ∞ ( T )
in outer probability. On the event that the line segments joining θ ^ \widehat{\theta} θ and θ ^ ∗ \widehat{\theta}^* θ ∗ lie in the differentiability neighbourhood, define
H ~ G ∗ ( τ ) : = ∫ 0 1 ∂ ∂ t Ψ G ( t , τ ) ∣ t = θ ^ ( τ ) + u { θ ^ ∗ ( τ ) − θ ^ ( τ ) } d u . \widetilde{H}_G^*(\tau):=\int_0^1\left.\frac{\partial}{\partial t}\Psi_G(t,\tau)\right|_{t=\widehat{\theta}(\tau)+u\{\widehat{\theta}^*(\tau)-\widehat{\theta}(\tau)\}}du. H G ∗ ( τ ) := ∫ 0 1 ∂ t ∂ Ψ G ( t , τ ) t = θ ( τ ) + u { θ ∗ ( τ ) − θ ( τ )} d u .
and
R G ∗ : = G [ ( Ψ ^ G − Ψ G ) ( θ ^ ∗ ) − ( Ψ ^ G − Ψ G ) ( θ ^ ) ] + M G ∗ ( θ ^ ∗ ) − M G ∗ ( θ o ) . \mathcal{R}_G^*:=\sqrt{G}\left[\left(\widehat{\Psi}_G-\Psi_G\right)(\widehat{\theta}^*)-\left(\widehat{\Psi}_G-\Psi_G\right)(\widehat{\theta})\right]
+\mathbb{M}_G^*(\widehat{\theta}^*)-\mathbb{M}_G^*(\theta_o). R G ∗ := G [ ( Ψ G − Ψ G ) ( θ ∗ ) − ( Ψ G − Ψ G ) ( θ ) ] + M G ∗ ( θ ∗ ) − M G ∗ ( θ o ) .
Subtracting the original approximate estimating equation from its bootstrap analogue and applying the fundamental theorem of calculus gives, on the event that H ~ G ∗ \widetilde{H}_G^* H G ∗ is uniformly invertible, the exact decomposition
G ( θ ^ ∗ − θ ^ ) = − ( H ~ G ∗ ) − 1 [ M G ∗ ( θ o ) + R G ∗ − E G ∗ ] . (15) \sqrt{G}(\widehat{\theta}^*-\widehat{\theta})=-(\widetilde{H}_G^*)^{-1}\left[\mathbb{M}_G^*(\theta_o)+\mathcal{R}_G^*-\mathcal{E}_G^*\right]. \tag*{(15)} G ( θ ∗ − θ ) = − ( H G ∗ ) − 1 [ M G ∗ ( θ o ) + R G ∗ − E G ∗ ] . (15)
The first three steps below validate this identity and dispatch its remainder terms before the conditional limit of the surviving multiplier score is established.
Step 1: Control the multiplier perturbation globally. The same conditional maximal inequality, now applied to the full jointly indexed class, yields
E ∗ [ sup τ , t ∣ G − 1 / 2 ∑ g = 1 G ξ g ψ ˉ τ , j ( W g ; t ) ∣ 2 ] ≤ C G − 1 ∑ g = 1 G F ‾ G g , j 2 = O p ( 1 ) . \mathbb{E}^*\left[\sup_{\tau,t}\left|G^{-1/2}\sum_{g=1}^{G}\xi_g\bar{\psi}_{\tau,j}(W_g;t)\right|^2\right]\leq C G^{-1}\sum_{g=1}^{G}\overline{F}_{Gg,j}^{\,2}=O_p(1). E ∗ τ , t sup G − 1/2 g = 1 ∑ G ξ g ψ ˉ τ , j ( W g ; t ) 2 ≤ C G − 1 g = 1 ∑ G F G g , j 2 = O p ( 1 ) .
The inequality uses E ∗ ξ g 2 = 1 \mathbb{E}^*\xi_g^2=1 E ∗ ξ g 2 = 1 , and the last equality follows by Markov’s inequality from the cluster-envelope moment bound supplied by Lemma S.2.1(a) under Assumption 2′(b). Dividing by G \sqrt{G} G , applying conditional Markov’s inequality, maximising over the finitely many coordinates, and restricting to admissible paths gives
sup θ ∈ Θ T ∥ G − 1 ∑ g = 1 G ξ g Ψ ‾ ( W g ; θ ) ∥ T = o P ∗ ( 1 ) in outer probability . (16) \sup_{\theta\in\Theta_{\mathcal{T}}}\left\|G^{-1}\sum_{g=1}^{G}\xi_g\overline{\Psi}(W_g;\theta)\right\|_{\mathcal{T}}=o_{P^*}(1)\quad\text{in outer probability}. \tag*{(16)} θ ∈ Θ T sup G − 1 g = 1 ∑ G ξ g Ψ ( W g ; θ ) T = o P ∗ ( 1 ) in outer probability . (16)
Step 2: Localise the bootstrap estimator. The functional UWLLN in Lemma 3.1(a) gives
sup θ ∈ Θ T ∥ Ψ ^ G ( θ ) − Ψ G ( θ ) ∥ T = o p ( 1 ) , \sup_{\theta\in\Theta_T}\left\|\widehat{\Psi}_G(\theta)-\Psi_G(\theta)\right\|_T=o_p(1), θ ∈ Θ T sup Ψ G ( θ ) − Ψ G ( θ ) T = o p ( 1 ) ,
whereas (S.2.10) gives
sup θ ∈ Θ T ∥ Ψ ^ G ∗ ( θ ) − Ψ ^ G ( θ ) ∥ T = o P ∗ ( 1 ) \sup_{\theta\in\Theta_T}\left\|\widehat{\Psi}_G^*(\theta)-\widehat{\Psi}_G(\theta)\right\|_T=o_{P^*}(1) θ ∈ Θ T sup Ψ G ∗ ( θ ) − Ψ G ( θ ) T = o P ∗ ( 1 )
in outer probability. The triangle inequality therefore yields
sup θ ∈ Θ T ∥ Ψ ^ G ∗ ( θ ) − Ψ G ( θ ) ∥ T = o p ( 1 ) + o P ∗ ( 1 ) \sup_{\theta\in\Theta_T}\left\|\widehat{\Psi}_G^*(\theta)-\Psi_G(\theta)\right\|_T=o_p(1)+o_{P^*}(1) θ ∈ Θ T sup Ψ G ∗ ( θ ) − Ψ G ( θ ) T = o p ( 1 ) + o P ∗ ( 1 )
in outer probability. Since ∥ Ψ ^ G ∗ ( θ ^ ∗ ) ∥ T = o P ∗ ( G − 1 / 2 ) \left\|\widehat{\Psi}_G^*(\widehat{\theta}^*)\right\|_T=o_{P^*}(G^{-1/2}) Ψ G ∗ ( θ ∗ ) T = o P ∗ ( G − 1/2 ) , the bootstrap approximate-zero condition implies
∥ Ψ G ( θ ^ ∗ ) ∥ T ≤ sup θ ∈ Θ T ∥ Ψ ^ G ∗ ( θ ) − Ψ G ( θ ) ∥ T + ∥ Ψ ^ G ∗ ( θ ^ ∗ ) ∥ T = o P ∗ ( 1 ) \left\|\Psi_G(\widehat{\theta}^*)\right\|_T\leq\sup_{\theta\in\Theta_T}\left\|\widehat{\Psi}_G^*(\theta)-\Psi_G(\theta)\right\|_T+\left\|\widehat{\Psi}_G^*(\widehat{\theta}^*)\right\|_T=o_{P^*}(1) Ψ G ( θ ∗ ) T ≤ θ ∈ Θ T sup Ψ G ∗ ( θ ) − Ψ G ( θ ) T + Ψ G ∗ ( θ ∗ ) T = o P ∗ ( 1 )
in outer probability. Functional separation in Assumption 3’(b) then gives
∥ θ ^ ∗ − θ o ∥ T = o P ∗ ( 1 ) \left\|\widehat{\theta}^*-\theta_o\right\|_T=o_{P^*}(1) θ ∗ − θ o T = o P ∗ ( 1 )
in outer probability. The corresponding consistency ∥ θ ^ − θ o ∥ T = o p ( 1 ) \left\|\widehat{\theta}-\theta_o\right\|_T=o_p(1) θ − θ o T = o p ( 1 ) was established in the proof of Theorem 3.3.
Step 3: Validate the decomposition and dispatch remainders. Conditionally on the data, multiplication by ∣ ξ g ∣ \lvert\xi_g\rvert ∣ ξ g ∣ is a nonnegative rescaling and therefore preserves manageability in Pollard’s definition. The full and local vector arrays have the common almost-measurable-Suslin parametrisations required by Assumption 2’(a). Multiplication by the finitely many row multipliers preserves these parametrisations on the product probability space, i.e., the approximation error is bounded by max g ≤ G ∣ ξ g ∣ 2 \max_{g\leq G}\lvert\xi_g\rvert^2 max g ≤ G ∣ ξ g ∣ 2 times its original value. The empirical-radius conclusion (S.2.5) established in Lemma 3.1 also controls the multiplier-weighted local radius. With d G g , j d_{Gg,j} d G g , j and I δ \mathcal{I}_\delta I δ defined there, for any K < ∞ K<\infty K < ∞ ,
sup ( τ , t ) ∈ I δ 1 G ∑ g = 1 G ξ g 2 d G g , j ( τ , t ) 2 ≤ K 2 R G , j 2 ( δ ) + 4 G ∑ g = 1 G ξ g 2 1 { ∣ ξ g ∣ > K } F ‾ G g , j 2 . \sup_{(\tau,t)\in\mathcal{I}_\delta}\frac{1}{G}\sum_{g=1}^{G}\xi_g^2d_{Gg,j}(\tau,t)^2\leq K^2R_{G,j}^2(\delta)+\frac{4}{G}\sum_{g=1}^{G}\xi_g^2\mathbf{1}\{\lvert\xi_g\rvert>K\}\overline{F}_{Gg,j}^2. ( τ , t ) ∈ I δ sup G 1 g = 1 ∑ G ξ g 2 d G g , j ( τ , t ) 2 ≤ K 2 R G , j 2 ( δ ) + G 4 g = 1 ∑ G ξ g 2 1 {∣ ξ g ∣ > K } F G g , j 2 .
Independence of the multipliers and data, together with the envelope moment bound, bounds the expectation of the second term by
C E [ ξ 2 1 { ∣ ξ ∣ > K } ] ⟶ 0 as K ⟶ ∞ . C\mathbb{E}\left[\xi^2\mathbf{1}\{\lvert\xi\rvert>K\}\right]\longrightarrow0\qquad\text{as }K\longrightarrow\infty. C E [ ξ 2 1 {∣ ξ ∣ > K } ] ⟶ 0 as K ⟶ ∞.
Thus, first localising for fixed K K K and then letting $K\longrightarrow\infty makes this weighted radius negligible. Applying the symmetrisation and maximal-inequality argument used in the proof of Kosorok [11 ] Theorem 2), followed by conditional Markov’s inequality, gives
lim δ ↓ 0 lim sup G → ∞ P [ P ∗ { sup ∥ θ − θ o ∥ T < δ ∥ M G ∗ ( θ ) − M G ∗ ( θ o ) ∥ T > ε } > η ] = 0 ( ε , η > 0 ) . (17) \lim_{\delta\downarrow0}\limsup_{G\to\infty}\mathbb{P}\left[\mathbb{P}^{*}\left\{\sup_{\lVert\theta-\theta_o\rVert_T<\delta}\lVert\mathbb{M}^{*}_{G}(\theta)-\mathbb{M}^{*}_{G}(\theta_o)\rVert_T>\varepsilon\right\}>\eta\right]=0\qquad(\varepsilon,\eta>0). \tag*{(17)} δ ↓ 0 lim G → ∞ lim sup P [ P ∗ { ∥ θ − θ o ∥ T < δ sup ∥ M G ∗ ( θ ) − M G ∗ ( θ o ) ∥ T > ε } > η ] = 0 ( ε , η > 0 ) . (17)
The two consistency results from Step 2 and the uniform differentiability condition in Assumption 4’(c) validate (S.2.9) with probability approaching one and give
∥ H ~ G ∗ − H ∥ T = o P ∗ ( 1 ) \lVert\widetilde{H}^{*}_{G}-H\rVert_T=o_{\mathbb{P}^{*}}(1) ∥ H G ∗ − H ∥ T = o P ∗ ( 1 )
in outer probability. The empirical local stochastic equicontinuity in Lemma 3.1(b) gives
G ∥ ( Ψ ^ G − Ψ G ) ( θ ^ ∗ ) − ( Ψ ^ G − Ψ G ) ( θ ^ ) ∥ T = o P ∗ ( 1 ) , \sqrt{G}\left\lVert(\widehat{\Psi}_{G}-\Psi_{G})(\widehat{\theta}^{*})-(\widehat{\Psi}_{G}-\Psi_{G})(\widehat{\theta})\right\rVert_T=o_{\mathbb{P}^{*}}(1), G ( Ψ G − Ψ G ) ( θ ∗ ) − ( Ψ G − Ψ G ) ( θ ) T = o P ∗ ( 1 ) ,
and (S.2.11) gives
∥ M G ∗ ( θ ^ ∗ ) − M G ∗ ( θ o ) ∥ T = o P ∗ ( 1 ) . \left\lVert\mathbb{M}^{*}_{G}(\widehat{\theta}^{*})-\mathbb{M}^{*}_{G}(\theta_o)\right\rVert_T=o_{\mathbb{P}^{*}}(1). M G ∗ ( θ ∗ ) − M G ∗ ( θ o ) T = o P ∗ ( 1 ) .
Hence ∥ R G ∗ ∥ T = o P ∗ ( 1 ) \lVert\mathcal{R}^{*}_{G}\rVert_T=o_{\mathbb{P}^{*}}(1) ∥ R G ∗ ∥ T = o P ∗ ( 1 ) , while the two approximate-zero conditions give ∥ E G ∗ ∥ T = o P ∗ ( 1 ) \lVert\mathcal{E}^{*}_{G}\rVert_T=o_{\mathbb{P}^{*}}(1) ∥ E G ∗ ∥ T = o P ∗ ( 1 ) , all in outer probability. The uniform lower singular-value bound on H H H now makes H ~ G ∗ \widetilde{H}^{*}_{G} H G ∗ uniformly invertible with probability approaching one. Thus (S.2.9) reduces to
G ( θ ^ ∗ − θ ^ ) = − ( H ~ G ∗ ) − 1 M G ∗ ( θ o ) + o P ∗ ( 1 ) in ℓ ∞ ( T ) , (18) \sqrt{G}(\widehat{\theta}^{*}-\widehat{\theta})=-(\widetilde{H}^{*}_{G})^{-1}\mathbb{M}^{*}_{G}(\theta_o)+o_{\mathbb{P}^{*}}(1)\quad\text{in }\ell^{\infty}(\mathcal{T}), \tag*{(18)} G ( θ ∗ − θ ) = − ( H G ∗ ) − 1 M G ∗ ( θ o ) + o P ∗ ( 1 ) in ℓ ∞ ( T ) , (18)
in outer probability.
Step 4: Establish weak convergence of the surviving multiplier score. Apply Kosorok [11 ] , corresponding to Kosorok [10 ] , to every scalar projection of the truth-indexed array used in the proof of Theorem 3.3. Its conditions are the same manageability, envelope, Lindeberg, covariance, metric, and measurability conditions already verified there, together with the mean-zero, unit-variance multiplier condition. Moreover, Assumption 3’(a) gives E G ψ ˉ τ ( W g ; θ o ( τ ) ) = 0 \mathbb{E}_{G}\bar{\psi}_{\tau}(W_g;\theta_o(\tau))=0 E G ψ ˉ τ ( W g ; θ o ( τ )) = 0 for every g g g and τ \tau τ , so the theorem applies with the true component means equal to zero. Equivalently, Kosorok [10 ] applies with μ ^ G g = μ G g = 0 \widehat{\mu}_{Gg}=\mu_{Gg}=0 μ G g = μ G g = 0 , making its conditions G–I immediate. Conditional joint finite-dimensional convergence follows by the Crámer–Wold device, and finite coordinatewise tightness then yields
M G ∗ ( θ o ) ⇝ P G 0 in ℓ ∞ ( T ) . (19) \mathbb{M}^{*}_{G}(\theta_o)\rightsquigarrow_{\mathbb{P}}G_0\quad\text{in }\ell^{\infty}(\mathcal{T}). \tag*{(19)} M G ∗ ( θ o ) ⇝ P G 0 in ℓ ∞ ( T ) . (19)
This is the sole surviving first-order term in (S.2.9).
Step 5: Conclude from the decomposition. The conditional multiplier limit implies ∥ M G ∗ ( θ o ) ∥ T = O P ∗ ( 1 ) \lVert\mathbb{M}^{*}_{G}(\theta_o)\rVert_T=O_{\mathbb{P}^{*}}(1) ∥ M G ∗ ( θ o ) ∥ T = O P ∗ ( 1 ) in outer probability. The inverse identity and the uniform convergence established in Step 3 therefore give
∥ ( ( H ~ G ∗ ) − 1 − H − 1 ) M G ∗ ( θ o ) ∥ T = o P ∗ ( 1 ) . \left\lVert\left((\widetilde{H}^{*}_{G})^{-1}-H^{-1}\right)\mathbb{M}^{*}_{G}(\theta_o)\right\rVert_T=o_{\mathbb{P}^{*}}(1). ( ( H G ∗ ) − 1 − H − 1 ) M G ∗ ( θ o ) T = o P ∗ ( 1 ) .
Combining (S.2.12) and (S.2.13), and applying conditional Slutsky, gives
G ( θ ^ ∗ − θ ^ ) ( ⋅ ) ⇝ P − H ( ⋅ ) − 1 G 0 ( ⋅ ) in ℓ ∞ ( T ) , \sqrt{G}(\widehat{\theta}^{*}-\widehat{\theta})(\cdot)\rightsquigarrow_{P}-H(\cdot)^{-1}G_{0}(\cdot)\quad\text{in }\ell^{\infty}(\mathcal{T}), G ( θ ∗ − θ ) ( ⋅ ) ⇝ P − H ( ⋅ ) − 1 G 0 ( ⋅ ) in ℓ ∞ ( T ) ,
which proves part (a).
Part (b): For the bootstrap statement, the ordinary weak convergence in Theorem 3.3 and the conditional weak convergence in Theorem 3.4 meet the hypotheses of the bootstrap functional delta method in Chernozhukov, Fernández-Val, and Melly [4 ] . Hence
G ( η ^ ∗ − η ^ ) ( ⋅ ) ⇝ P Φ ϕ , θ o ′ [ Z θ ] ( ⋅ ) in ℓ ∞ ( T ) . \sqrt{G}(\widehat{\eta}^{*}-\widehat{\eta})(\cdot)\rightsquigarrow_{P}\Phi_{\phi,\theta_{o}}^{\prime}[Z_{\theta}](\cdot)\quad\text{in }\ell^{\infty}(\mathcal{T}). G ( η ∗ − η ) ( ⋅ ) ⇝ P Φ ϕ , θ o ′ [ Z θ ] ( ⋅ ) in ℓ ∞ ( T ) .
Proof of Corollary 3.1
For each fixed τ ∈ T \tau\in\mathcal{T} τ ∈ T , the evaluation map z ↦ z ( τ ) z \mapsto z(\tau) z ↦ z ( τ ) is a bounded linear map on ℓ ∞ ( T ) \ell^{\infty}(\mathcal{T}) ℓ ∞ ( T ) . The result follows from Theorem 3.4 by the continuous mapping theorem.
Proofs for the Model-Specific Results
For ease of reference, the coordinate-score classes studied in this section are
F j D R = { ( u , x ) ↦ x j ( 1 { u ≤ y } − Λ ( x ′ t ) ) : ( y , t ) ∈ Y × Θ } , \mathcal{F}_{j}^{DR}=\left\{(u,x)\mapsto x_{j}\left(\mathbb{1}\{u\leq y\}-\Lambda(x^{\prime}t)\right):(y,t)\in\mathcal{Y}\times\Theta\right\}, F j D R = { ( u , x ) ↦ x j ( 1 { u ≤ y } − Λ ( x ′ t ) ) : ( y , t ) ∈ Y × Θ } ,
F j Q R = { ( u , x ) ↦ x j ( τ − 1 { u ≤ x ′ t } ) : ( τ , t ) ∈ T × Θ } , \mathcal{F}_{j}^{QR}=\left\{(u,x)\mapsto x_{j}\left(\tau-\mathbb{1}\{u\leq x^{\prime}t\}\right):(\tau,t)\in\mathcal{T}\times\Theta\right\}, F j QR = { ( u , x ) ↦ x j ( τ − 1 { u ≤ x ′ t } ) : ( τ , t ) ∈ T × Θ } ,
F j E R = { ( u , x ) ↦ x j ( u − x ′ t ) ∣ τ − 1 { u < x ′ t } ∣ : ( τ , t ) ∈ T × Θ } . \mathcal{F}_{j}^{ER}=\left\{(u,x)\mapsto x_{j}(u-x^{\prime}t)\left|\tau-\mathbb{1}\{u<x^{\prime}t\}\right|:(\tau,t)\in\mathcal{T}\times\Theta\right\}. F j E R = { ( u , x ) ↦ x j ( u − x ′ t ) ∣ τ − 1 { u < x ′ t } ∣ : ( τ , t ) ∈ T × Θ } .
We first establish their manageability and its preservation under cluster averaging, then prove the model-specific identification and verification results.
Lemma S.3.1 (Manageability of DR, QR, and ER score classes). Let w = ( u , x ) w=(u,x) w = ( u , x ) , let Θ ⊂ R p θ \Theta\subset\mathbb{R}^{p_{\theta}} Θ ⊂ R p θ and T ⊂ ( 0 , 1 ) \mathcal{T}\subset(0,1) T ⊂ ( 0 , 1 ) be compact, and set B Θ : = sup t ∈ Θ ∥ t ∥ B_{\Theta}:=\sup_{t\in\Theta}\|t\| B Θ := sup t ∈ Θ ∥ t ∥ . For a monotone link Λ \Lambda Λ taking values in [ 0 , 1 ] [0,1] [ 0 , 1 ] , the DR, QR, and ER coordinate-score classes defined above are manageable in the sense of Pollard [15 ] , with respective envelopes ∣ x j ∣ |x_{j}| ∣ x j ∣ , ∣ x j ∣ |x_{j}| ∣ x j ∣ , and ∣ x j ∣ { ∣ u ∣ + B Θ ∥ x ∥ } |x_{j}|\{|u|+B_{\Theta}\|x\|\} ∣ x j ∣ { ∣ u ∣ + B Θ ∥ x ∥ } .
Proof. The linear-index, residual, and constant classes, respectively,
{ ( u , x ) ↦ x ′ t : t ∈ Θ } , { ( u , x ) ↦ u − x ′ t : t ∈ Θ } , { ( u , x ) ↦ τ : τ ∈ T } , \{(u,x)\mapsto x^{\prime}t:t\in\Theta\},\qquad\{(u,x)\mapsto u-x^{\prime}t:t\in\Theta\},\qquad\{(u,x)\mapsto\tau:\tau\in\mathcal{T}\}, {( u , x ) ↦ x ′ t : t ∈ Θ } , {( u , x ) ↦ u − x ′ t : t ∈ Θ } , {( u , x ) ↦ τ : τ ∈ T } ,
are contained in finite-dimensional vector spaces of measurable functions and are therefore Vapnik–Chervonenkis (VC) subgraph classes by Kosorok [10 ] . The threshold and halfspace indicator classes are, respectively, { ( u , x ) ↦ 1 { u ≤ y } : y ∈ R } \{(u,x)\mapsto\mathbb{1}\{u\leq y\}:y\in\mathbb{R}\} {( u , x ) ↦ 1 { u ≤ y } : y ∈ R } , { ( u , x ) ↦ 1 { u ≤ x ′ t } : t ∈ Θ } \{(u,x)\mapsto\mathbb{1}\{u\leq x^{\prime}t\}:t\in\Theta\} {( u , x ) ↦ 1 { u ≤ x ′ t } : t ∈ Θ } , and { ( u , x ) ↦ 1 { u < x ′ t } : t ∈ Θ } \{(u,x)\mapsto\mathbb{1}\{u<x^{\prime}t\}:t\in\Theta\} {( u , x ) ↦ 1 { u < x ′ t } : t ∈ Θ } . Their defining sets are VC by Kosorok (2007, Lemma 9.12(i)), using complements for strict boundaries. By Kosorok [10 ] , the indicator classes are therefore VC-subgraph. The monotone-link class
{ ( u , x ) ↦ Λ ( x ′ t ) : t ∈ Θ } \{(u,x)\mapsto\Lambda(x^{\prime}t):t\in\Theta\} {( u , x ) ↦ Λ ( x ′ t ) : t ∈ Θ }
is also VC-subgraph by the monotone-composition result in Kosorok [10 ] . All these VC indices are bounded independently of the observations and sample size. Hence Kosorok [10 ] gives uniform polynomial covering bounds for each component class F \mathcal{F} F , with its indicated envelope F F F :
sup Q N ( ε ∥ F ∥ Q , 2 , F , L 2 ( Q ) ) ≤ A ε − v , 0 < ε < 1 , \sup_{Q}N(\varepsilon\lVert F\rVert_{Q,2},\mathcal{F},L_{2}(Q))\le A\varepsilon^{-v},\qquad0<\varepsilon<1, Q sup N ( ε ∥ F ∥ Q , 2 , F , L 2 ( Q )) ≤ A ε − v , 0 < ε < 1 ,
where the supremum is over finitely supported probability measures with ∥ F ∥ Q , 2 > 0 \lVert F\rVert_{Q,2}>0 ∥ F ∥ Q , 2 > 0 , and A , v A,v A , v are finite constants.
The singleton class { x j } \{x_j\} { x j } has envelope ∣ x j ∣ |x_j| ∣ x j ∣ and covering number one. The covering inequalities in the proofs of Kosorok [10 ] preserve polynomial bounds under finite sums and products, with sum and product envelopes. Restricting a product index set to the common parameters used in a score does not increase its entropy bound, apart from a constant adjustment of the covering radius. For ER, use
∣ τ − 1 { u < x ′ t } ∣ = τ + ( 1 − 2 τ ) 1 { u < x ′ t } . |\tau-1\{u<x^{\prime}t\}|=\tau+(1-2\tau)1\{u<x^{\prime}t\}. ∣ τ − 1 { u < x ′ t } ∣ = τ + ( 1 − 2 τ ) 1 { u < x ′ t } .
These operations give polynomial bounds for the DR, QR, and ER coordinate-score classes with envelopes bounded by constant multiples of ∣ x j ∣ |x_j| ∣ x j ∣ , ∣ x j ∣ |x_j| ∣ x j ∣ , and ∣ x j ∣ { ∣ u ∣ + B Θ ∥ x ∥ } |x_j|\{|u|+B_{\Theta}\lVert x\rVert\} ∣ x j ∣ { ∣ u ∣ + B Θ ∥ x ∥} , respectively. The scores themselves are bounded by the stated envelopes; absorbing the fixed envelope multiples into A A A gives polynomial bounds relative to those envelopes.
To connect these bounds to Pollard [15 ] , fix w 1 , … , w n w_1,\ldots,w_n w 1 , … , w n and any nonnegative rescaling vector a = ( a 1 , … , a n ) a=(a_1,\ldots,a_n) a = ( a 1 , … , a n ) . Define the rescaled evaluation set and its envelope norm by
V a : = { ( a 1 f ( w 1 ) , … , a n f ( w n ) ) : f ∈ F } , E a : = { ∑ i a i 2 F ( w i ) 2 } 1 / 2 . V_a:=\{(a_1f(w_1),\ldots,a_nf(w_n)):f\in\mathcal{F}\},\qquad E_a:=\left\{\sum_i a_i^2F(w_i)^2\right\}^{1/2}. V a := {( a 1 f ( w 1 ) , … , a n f ( w n )) : f ∈ F } , E a := { i ∑ a i 2 F ( w i ) 2 } 1/2 .
If E a = 0 E_a=0 E a = 0 , all rescaled evaluation vectors coincide and the conclusion is immediate. Otherwise, put Q a : = ∑ i a i 2 δ w i / ∑ i a i 2 Q_a:=\sum_i a_i^2\delta_{w_i}/\sum_i a_i^2 Q a := ∑ i a i 2 δ w i / ∑ i a i 2 . For any f , g ∈ F f,g\in\mathcal{F} f , g ∈ F ,
∥ f − g ∥ Q a , 2 ∥ F ∥ Q a , 2 = { ∑ i a i 2 [ f ( w i ) − g ( w i ) ] 2 } 1 / 2 E a . \frac{\lVert f-g\rVert_{Q_a,2}}{\lVert F\rVert_{Q_a,2}}=\frac{\{\sum_i a_i^2[f(w_i)-g(w_i)]^2\}^{1/2}}{E_a}. ∥ F ∥ Q a , 2 ∥ f − g ∥ Q a , 2 = E a { ∑ i a i 2 [ f ( w i ) − g ( w i ) ] 2 } 1/2 .
Thus separation at radius ε E a \varepsilon E_a ε E a in V a V_a V a is equivalent to separation at radius ε ∥ F ∥ Q a , 2 \varepsilon\lVert F\rVert_{Q_a,2} ε ∥ F ∥ Q a , 2 in L 2 ( Q a ) L_2(Q_a) L 2 ( Q a ) . Writing D D D for the packing number, the covering–packing inequality (Pollard, [15 ] , following Definition 3.3) gives
D ( ε E a , V a ) ≤ N ( ε 2 ∥ F ∥ Q a , 2 , F , L 2 ( Q a ) ) ≤ A ( ε / 2 ) − v = 2 v A ε − v . D(\varepsilon E_a,V_a)\leq N\left(\frac{\varepsilon}{2}\|F\|_{Q_{a,2},\mathcal{F},L_2(Q_a)}\right)
\leq A(\varepsilon/2)^{-v}=2^vA\varepsilon^{-v}. D ( ε E a , V a ) ≤ N ( 2 ε ∥ F ∥ Q a , 2 , F , L 2 ( Q a ) ) ≤ A ( ε /2 ) − v = 2 v A ε − v .
Absorbing 2 v 2^v 2 v into C C C gives the bound C ε − v C\varepsilon^{-v} C ε − v , uniformly in the observations, a a a , and n n n . Taking C ≥ 1 C\geq1 C ≥ 1 , the deterministic capacity bound λ ( ε ) = C ε − v \lambda(\varepsilon)=C\varepsilon^{-v} λ ( ε ) = C ε − v satisfies ∫ 0 1 log λ ( ε ) d ε < ∞ \int_0^1\sqrt{\log\lambda(\varepsilon)}\,d\varepsilon<\infty ∫ 0 1 log λ ( ε ) d ε < ∞ . Thus all the asserted classes are Euclidean, and hence manageable, with the indicated envelopes. □
The following lemma transfers manageability of individual-level evaluation arrays to cluster-average arrays. Let F \mathcal{F} F be a class of measurable functions with nonnegative measurable envelope F F F . For any deterministic cluster partition with N g ≥ 1 N_g\geq1 N g ≥ 1 , define
f ˉ G g : = N g − 1 ∑ i ∈ G g f ( W i ) , F ˉ G g : = { N g − 1 ∑ i ∈ G g F ( W i ) 2 } 1 / 2 . \bar{f}_{G_g}:=N_g^{-1}\sum_{i\in\mathcal{G}_g}f(W_i),\qquad
\bar{F}_{G_g}:=\left\{N_g^{-1}\sum_{i\in\mathcal{G}_g}F(W_i)^2\right\}^{1/2}. f ˉ G g := N g − 1 i ∈ G g ∑ f ( W i ) , F ˉ G g := ⎩ ⎨ ⎧ N g − 1 i ∈ G g ∑ F ( W i ) 2 ⎭ ⎬ ⎫ 1/2 .
Lemma S.3.2 (Manageability under cluster averaging). Suppose the individual evaluation arrays of F \mathcal{F} F are manageable in the sense of Pollard (1990, Definition 7.9), uniformly over sample sizes, evaluation points, and nonnegative rescaling vectors, with a deterministic capacity bound λ \lambda λ . Then the cluster-average array { f ˉ G g : f ∈ F , g ≤ G } \{\bar{f}_{G_g}:f\in\mathcal{F},\ g\leq G\} { f ˉ G g : f ∈ F , g ≤ G } is manageable with envelope { F ˉ G g : g ≤ G } \{\bar{F}_{G_g}:g\leq G\} { F ˉ G g : g ≤ G } and capacity bound λ \lambda λ , uniformly over cluster sizes. Consequently, { G − 1 / 2 f ˉ G g : f ∈ F , g ≤ G } \{G^{-1/2}\bar{f}_{G_g}:f\in\mathcal{F},\ g\leq G\} { G − 1/2 f ˉ G g : f ∈ F , g ≤ G } is manageable with envelope { G − 1 / 2 F ˉ G g : g ≤ G } \{G^{-1/2}\bar{F}_{G_g}:g\leq G\} { G − 1/2 F ˉ G g : g ≤ G } .
Proof. Jensen’s inequality gives ∣ f ˉ G g ∣ ≤ F ˉ G g |\bar{f}_{G_g}|\leq\bar{F}_{G_g} ∣ f ˉ G g ∣ ≤ F ˉ G g . Fix the evaluation points and any nonnegative rescaling vector α = ( α 1 , … , α G ) \alpha=(\alpha_1,\ldots,\alpha_G) α = ( α 1 , … , α G ) . For f , h ∈ F f,h\in\mathcal{F} f , h ∈ F , clusterwise Jensen’s inequality gives
∑ g = 1 G α g 2 ( f ˉ G g − h ˉ G g ) 2 ≤ ∑ g = 1 G α g 2 N g ∑ i ∈ G g ( f ( W i ) − h ( W i ) ) 2 . (20) \sum_{g=1}^{G}\alpha_g^2(\bar{f}_{G_g}-\bar{h}_{G_g})^2\leq\sum_{g=1}^{G}\frac{\alpha_g^2}{N_g}\sum_{i\in\mathcal{G}_g}(f(W_i)-h(W_i))^2. \tag*{(20)} g = 1 ∑ G α g 2 ( f ˉ G g − h ˉ G g ) 2 ≤ g = 1 ∑ G N g α g 2 i ∈ G g ∑ ( f ( W i ) − h ( W i ) ) 2 . (20)
Set β i : = α g ( i ) / N g ( i ) \beta_i:=\alpha_{g(i)}/\sqrt{N_{g(i)}} β i := α g ( i ) / N g ( i ) . The right-hand side is the squared distance between the individual evaluation vectors rescaled by β \beta β , and the envelope scales agree:
∑ g = 1 G α g 2 F ˉ G g 2 = ∑ i = 1 N β i 2 F ( W i ) 2 . \sum_{g=1}^{G}\alpha_g^2\bar{F}_{G_g}^2=\sum_{i=1}^{N}\beta_i^2F(W_i)^2. g = 1 ∑ G α g 2 F ˉ G g 2 = i = 1 ∑ N β i 2 F ( W i ) 2 .
Choose one function f ∈ F f\in\mathcal{F} f ∈ F for each vector in a packing of the rescaled cluster-average vectors. By (S.3.1), the corresponding rescaled individual evaluation vectors are separated by at least the same distance, so they form a packing at the same radius. The assumed individual capacity bound therefore also bounds the cluster packing numbers at radius ε { ∑ g α g 2 F ˉ G g 2 } 1 / 2 \varepsilon\{\sum_g\alpha_g^2\bar{F}_{G_g}^2\}^{1/2} ε { ∑ g α g 2 F ˉ G g 2 } 1/2 . If the envelope norm is zero, all rescaled vectors coincide and the conclusion is immediate. Otherwise, the cluster-average array satisfies condition (ii) of Pollard (1990, Definition 7.9, p. 38) with the same capacity bound λ \lambda λ . Since this function is unchanged, its integrability condition (i) also carries over. Absorbing G − 1 / 2 G^{-1/2} G − 1/2 into the arbitrary weights α g \alpha_g α g proves the normalised-array conclusion.
Proof of prop:DRidentification
Fix y ∈ Y y \in\mathcal{Y} y ∈ Y . By the law of iterated expectations and Assumption DR 1,
E G [ X i ( 1 { Y i ≤ y } − Λ ( X i ′ θ o ( y ) ) ) ] = E G [ X i ( F Y i ∣ X i ( y ∣ X i ) − Λ ( X i ′ θ o ( y ) ) ) ] = 0 \mathbb{E}_G\left[X_i\left(\mathbb{1}\{Y_i \le y\}-\Lambda(X_i'\theta_o(y))\right)\right] = \mathbb{E}_G\left[X_i\left(F_{Y_i\mid X_i}(y \mid X_i)-\Lambda(X_i'\theta_o(y))\right)\right] = 0 E G [ X i ( 1 { Y i ≤ y } − Λ ( X i ′ θ o ( y )) ) ] = E G [ X i ( F Y i ∣ X i ( y ∣ X i ) − Λ ( X i ′ θ o ( y )) ) ] = 0
for every sampled observation. Averaging these equalities within each cluster gives clusterwise exogeneity; averaging again with the SACR weights N g − 1 N_g^{-1} N g − 1 gives zero for the DR-SACR population moment at θ o ( y ) \theta_o(y) θ o ( y ) , so the cluster-size adjustment does not move the target zero.
It remains to show uniqueness. Suppose, for contradiction, that the DR-SACR population moment is also zero at some θ ( y ) ≠ θ o ( y ) \theta(y) \ne\theta_o(y) θ ( y ) = θ o ( y ) , and set δ = θ o ( y ) − θ ( y ) \delta= \theta_o(y)-\theta(y) δ = θ o ( y ) − θ ( y ) , so δ ≠ 0 \delta\ne0 δ = 0 . Subtracting the zero moment at θ o ( y ) \theta_o(y) θ o ( y ) gives
1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ X i ( Λ ( X i ′ θ o ( y ) ) − Λ ( X i ′ θ ( y ) ) ) ] = 0. \frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[X_i\left(\Lambda(X_i'\theta_o(y))-\Lambda(X_i'\theta(y))\right)\right]=0. G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ X i ( Λ ( X i ′ θ o ( y )) − Λ ( X i ′ θ ( y )) ) ] = 0.
Premultiplying by δ ′ \delta' δ ′ yields
1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ ( X i ′ δ ) ( Λ ( X i ′ θ o ( y ) ) − Λ ( X i ′ θ ( y ) ) ) ] = 0. \frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[(X_i'\delta)\left(\Lambda(X_i'\theta_o(y))-\Lambda(X_i'\theta(y))\right)\right]=0. G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ ( X i ′ δ ) ( Λ ( X i ′ θ o ( y )) − Λ ( X i ′ θ ( y )) ) ] = 0.
Indeed, for real numbers a a a and b b b , strict monotonicity of Λ \Lambda Λ implies ( a − b ) { Λ ( a ) − Λ ( b ) } ≥ 0 (a-b)\{\Lambda(a)-\Lambda(b)\}\ge0 ( a − b ) { Λ ( a ) − Λ ( b )} ≥ 0 , with equality only if a = b a=b a = b . Taking a = X i ′ θ o ( y ) a=X_i'\theta_o(y) a = X i ′ θ o ( y ) and b = X i ′ θ ( y ) b=X_i'\theta(y) b = X i ′ θ ( y ) , each summand is nonnegative and vanishes only when X i ′ δ = 0 X_i'\delta=0 X i ′ δ = 0 . Hence the sum can be zero only if X i ′ δ = 0 X_i'\delta=0 X i ′ δ = 0 almost surely under the cluster-balanced array. Therefore
1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g ( X i ′ δ ) 2 ] = 0 , \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}(X_i'\delta)^2\right]=0, G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ ( X i ′ δ ) 2 = 0 ,
or equivalently,
δ ′ { 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ ] } δ = 0. \delta'\left\{\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'\right]\right\}\delta=0. δ ′ ⎩ ⎨ ⎧ G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ ⎭ ⎬ ⎫ δ = 0.
The uniform positive definiteness condition in Assumption LI 1(a) forces δ = 0 \delta=0 δ = 0 , a contradiction. Thus, the DR-SACR population moment has the same unique zero as the primitive DR moment for each fixed y y y .
The functional strengthening follows directly from Assumption-DR 2(b). For any y ∈ Y y \in\mathcal{Y} y ∈ Y , t ∈ Θ t \in\Theta t ∈ Θ , and δ = θ o ( y ) − t \delta= \theta_o(y) - t δ = θ o ( y ) − t , the fundamental theorem of calculus gives
δ ′ P G Ψ ˉ y ( W g ; t ) = ∫ 0 1 δ ′ Q G D R ( y , t , 1 − s ) δ d s ≥ c H ∥ δ ∥ 2 . \delta' \mathbb{P}_G \bar{\Psi}_y(\boldsymbol{W}_g;t) = \int_0^1 \delta' Q_G^{DR}(y,t,1-s)\delta\,ds \ge c_H\|\delta\|^2. δ ′ P G Ψ ˉ y ( W g ; t ) = ∫ 0 1 δ ′ Q G D R ( y , t , 1 − s ) δ d s ≥ c H ∥ δ ∥ 2 .
By Cauchy–Schwarz,
c H ∥ δ ∥ 2 ≤ δ ′ P G Ψ ˉ y ( W g ; t ) ≤ ∥ δ ∥ ∥ P G Ψ ˉ y ( W g ; t ) ∥ . c_H\|\delta\|^2 \le\delta' \mathbb{P}_G\bar{\Psi}_y(\boldsymbol{W}_g;t) \le\|\delta\|\|\mathbb{P}_G\bar{\Psi}_y(\boldsymbol{W}_g;t)\|. c H ∥ δ ∥ 2 ≤ δ ′ P G Ψ ˉ y ( W g ; t ) ≤ ∥ δ ∥∥ P G Ψ ˉ y ( W g ; t ) ∥.
Dividing by ∥ δ ∥ \|\delta\| ∥ δ ∥ when δ ≠ 0 \delta\ne0 δ = 0 gives ∥ P G Ψ ˉ y ( W g ; t ) ∥ ≥ c H ∥ t − θ o ( y ) ∥ \|\mathbb{P}_G\bar{\Psi}_y(\boldsymbol{W}_g;t)\| \ge c_H\|t-\theta_o(y)\| ∥ P G Ψ ˉ y ( W g ; t ) ∥ ≥ c H ∥ t − θ o ( y ) ∥ ; the same bound is immediate when δ = 0 \delta= 0 δ = 0 . If an admissible path θ \theta θ satisfies ∥ θ − θ o ∥ Y ≥ ϵ \|\theta-\theta_o\|_Y \ge\epsilon ∥ θ − θ o ∥ Y ≥ ϵ , choose y y y such that ∥ θ ( y ) − θ o ( y ) ∥ ≥ ϵ / 2 \|\theta(y)-\theta_o(y)\| \ge\epsilon/2 ∥ θ ( y ) − θ o ( y ) ∥ ≥ ϵ /2 . The preceding bound then gives
∥ P G Ψ ˉ ( W g ; θ ) ∥ Y ≥ c H ϵ / 2 \|\mathbb{P}_G\bar{\Psi}(\boldsymbol{W}_g;\theta)\|_Y \ge c_H\epsilon/2 ∥ P G Ψ ˉ ( W g ; θ ) ∥ Y ≥ c H ϵ /2
uniformly in G G G , which is the functional separation condition in Assumption 3′.
Proof of prop:DRverify
The individual coordinate-score class is manageable with envelope ∣ X j ∣ |X_j| ∣ X j ∣ by Lemma S.3.1. Lemma S.3.2 therefore gives manageability of the normalised cluster-average array Z G g , j Z_{Gg,j} Z G g , j with the corresponding normalised root mean square (RMS) envelope F G g , j F_{Gg,j} F G g , j . Joint measurability of the DR score and Borel measurability of θ o ( ⋅ ) \theta_o(\cdot) θ o ( ⋅ ) under Assumption DR 1 give jointly measurable full, truth-indexed, and local-difference arrays on Borel subsets of Euclidean spaces; hence these arrays are jointly AMS, taking the entire respective index spaces as the Suslin approximating sets. Finally, Assumption LI 1(b) establishes uniform integrability of the squared envelope.
Proof of prop:QRidentification
Fix τ ∈ T \tau\in\mathcal{T} τ ∈ T . By the law of iterated expectations,
E G [ X i ( τ − 1 { Y i ≤ X i ′ θ o ( τ ) } ) ] = E G [ X i ( τ − F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) ) ] = 0 , \mathbb{E}_G\left[X_i\left(\tau-\mathbf{1}\{Y_i\le X_i'\theta_o(\tau)\}\right)\right]
= \mathbb{E}_G\left[X_i\left(\tau-F_{Y_i|X_i}(X_i'\theta_o(\tau)\mid X_i)\right)\right] = 0, E G [ X i ( τ − 1 { Y i ≤ X i ′ θ o ( τ )} ) ] = E G [ X i ( τ − F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) ) ] = 0 ,
where the last equality follows from the array-preservation restriction in Assumption QR 1(a). Averaging these zero moments with the SACR weights N g − 1 N_g^{-1} N g − 1 gives the displayed QR-SACR population moment at θ o ( τ ) \theta_o(\tau) θ o ( τ ) , and hence gives clusterwise exogeneity.
It remains to show uniqueness. Suppose, for contradiction, that some θ ( τ ) ∈ Θ \theta(\tau)\in\Theta θ ( τ ) ∈ Θ with θ ( τ ) ≠ θ o ( τ ) \theta(\tau)\ne\theta_o(\tau) θ ( τ ) = θ o ( τ ) also solves the QR-SACR population moment, and set δ = θ o ( τ ) − θ ( τ ) \delta=\theta_o(\tau)-\theta(\tau) δ = θ o ( τ ) − θ ( τ ) , so δ ≠ 0 \delta\ne0 δ = 0 . Subtracting the zero moment at θ o ( τ ) \theta_o(\tau) θ o ( τ ) gives
1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ X i ( F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) − F Y i ∣ X i ( X i ′ θ ( τ ) ∣ X i ) ) ] = 0. \frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[X_i\left(F_{Y_i|X_i}(X_i'\theta_o(\tau)\mid X_i)-F_{Y_i|X_i}(X_i'\theta(\tau)\mid X_i)\right)\right]=0. G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ X i ( F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) − F Y i ∣ X i ( X i ′ θ ( τ ) ∣ X i ) ) ] = 0.
Premultiplying by δ ′ \delta' δ ′ yields
1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ ( X i ′ δ ) ( F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) − F Y i ∣ X i ( X i ′ θ ( τ ) ∣ X i ) ) ] = 0. \frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[(X_i'\delta)\left(F_{Y_i\mid X_i}(X_i'\theta_o(\tau)\mid X_i)-F_{Y_i\mid X_i}(X_i'\theta(\tau)\mid X_i)\right)\right]=0. G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ ( X i ′ δ ) ( F Y i ∣ X i ( X i ′ θ o ( τ ) ∣ X i ) − F Y i ∣ X i ( X i ′ θ ( τ ) ∣ X i ) ) ] = 0.
Using the integral representation of the conditional cumulative distribution function (CDF) inside the preceding display gives
δ ′ P G ψ ˉ τ Q R ( W g ; θ ( τ ) ) = ∫ 0 1 δ ′ Q G Q R ( τ , θ ( τ ) , 1 − s ) δ d s ≥ c f ∥ δ ∥ 2 . \delta'P_G\bar{\psi}^{QR}_{\tau}(W_g;\theta(\tau))=\int_0^1\delta'Q_G^{QR}(\tau,\theta(\tau),1-s)\delta\,ds\geq c_f\|\delta\|^2. δ ′ P G ψ ˉ τ QR ( W g ; θ ( τ )) = ∫ 0 1 δ ′ Q G QR ( τ , θ ( τ ) , 1 − s ) δ d s ≥ c f ∥ δ ∥ 2 .
by Assumption QR 1(c). A second zero would make the left-hand side zero, forcing δ = 0 \delta=0 δ = 0 , a contradiction. Hence θ o ( τ ) \theta_o(\tau) θ o ( τ ) is the unique QR-SACR population zero for each fixed τ \tau τ . The same inequality and Cauchy–Schwarz imply
∥ P G ψ ˉ τ Q R ( W g ; t ) ∥ ≥ c f ∥ t − θ o ( τ ) ∥ (21) \left\|P_G\bar{\psi}^{QR}_{\tau}(W_g;t)\right\|\geq c_f\|t-\theta_o(\tau)\| \tag*{(21)} P G ψ ˉ τ QR ( W g ; t ) ≥ c f ∥ t − θ o ( τ ) ∥ (21)
uniformly over G G G , τ ∈ T \tau\in\mathcal{T} τ ∈ T , and t ∈ Θ t\in\Theta t ∈ Θ . If an admissible path θ \theta θ satisfies ∥ θ − θ o ∥ T ≥ ϵ \|\theta-\theta_o\|_{\mathcal{T}}\geq\epsilon ∥ θ − θ o ∥ T ≥ ϵ , choose τ \tau τ such that ∥ θ ( τ ) − θ o ( τ ) ∥ ≥ ϵ / 2 \|\theta(\tau)-\theta_o(\tau)\|\geq\epsilon/2 ∥ θ ( τ ) − θ o ( τ ) ∥ ≥ ϵ /2 . It follows that
∥ P G Ψ ˉ ( W g ; θ ) ∥ T ≥ c f ϵ / 2 \left\|P_G\bar{\Psi}(W_g;\theta)\right\|_{\mathcal{T}}\geq c_f\epsilon/2 P G Ψ ˉ ( W g ; θ ) T ≥ c f ϵ /2
uniformly in G G G , proving the functional separation condition.
Proof of Proposition 4.4
The individual coordinate-score class is manageable with envelope ∣ X j ∣ |X_j| ∣ X j ∣ by Lemma S.3.1. Lemma S.3.2 therefore gives manageability of the normalised cluster-average array Z G g , j Q R Z^{QR}_{Gg,j} Z G g , j QR with the corresponding normalised RMS envelope F G g , j Q R F^{QR}_{Gg,j} F G g , j QR . Joint measurability of the QR score and continuity of θ o ( ⋅ ) \theta_o(\cdot) θ o ( ⋅ ) from Lemma S.3.3 give jointly measurable full, truth-indexed, and local-difference arrays on Borel subsets of Euclidean spaces; hence these arrays are jointly AMS, taking the entire respective index spaces as the Suslin approximating sets. Finally, Assumption LI 1(b) establishes uniform integrability of the squared envelope.
Proof of Proposition 4.5
Fix τ ∈ T \tau\in\mathcal{T} τ ∈ T and let B Θ : = sup t ∈ Θ ∥ t ∥ < ∞ B_{\Theta}:=\sup_{t\in\Theta}\|t\|<\infty B Θ := sup t ∈ Θ ∥ t ∥ < ∞ . Assumption ER 2 ensures integrability of ∥ X i ∥ ∣ Y i ∣ + B Θ ∥ X i ∥ 2 \|X_i\||Y_i|+B_{\Theta}\|X_i\|^2 ∥ X i ∥∣ Y i ∣ + B Θ ∥ X i ∥ 2 , so the score expectations below are well defined. The array-preservation restriction in Assumption ER 1 gives
E G [ ( Y i − X i ′ θ o ( τ ) ) ∣ τ − 1 { Y i < X i ′ θ o ( τ ) } ∣ ∣ X i ] = 0 \mathbb{E}_G\left[(Y_i-X_i'\theta_o(\tau))\left|\tau-1\{Y_i<X_i'\theta_o(\tau)\}\right|\mid X_i\right]=0 E G [ ( Y i − X i ′ θ o ( τ )) ∣ τ − 1 { Y i < X i ′ θ o ( τ )} ∣ ∣ X i ] = 0
for every i i i . This is the first-order characterisation of the conditional expectile; see Newey and Powell (1987, Theorem 1). Multiplying by X i X_i X i , applying the law of iterated expectations, and averaging with the SACR weights N g − 1 N_g^{-1} N g − 1 gives the displayed ER-SACR population moment at θ o ( τ ) \theta_o(\tau) θ o ( τ ) , and hence gives clusterwise exogeneity.
For uniqueness, define the conditional expectile score
q τ ( m ∣ X i ) : = E G [ ( Y i − m ) ∣ τ − 1 { Y i < m } ∣ ∣ X i ] . q_\tau(m \mid X_i) := \mathbb{E}_G\left[(Y_i-m)\left|\tau-\mathbf{1}\{Y_i<m\}\right| \mid X_i\right]. q τ ( m ∣ X i ) := E G [ ( Y i − m ) ∣ τ − 1 { Y i < m } ∣ ∣ X i ] .
For fixed y y y , the score s τ ( y , m ) : = ( y − m ) ∣ τ − 1 { y < m } ∣ s_\tau(y,m):=(y-m)\left|\tau-\mathbf{1}\{y<m\}\right| s τ ( y , m ) := ( y − m ) ∣ τ − 1 { y < m } ∣ is continuous and piecewise linear, with slopes − τ -\tau − τ for m < y m<y m < y and − ( 1 − τ ) -(1-\tau) − ( 1 − τ ) for m > y m>y m > y . Thus, for any m 1 , m 2 m_1,m_2 m 1 , m 2 ,
s τ ( y , m 1 ) − s τ ( y , m 2 ) = − ∫ m 2 m 1 [ τ 1 { y ≥ u } + ( 1 − τ ) 1 { y < u } ] d u . s_\tau(y,m_1)-s_\tau(y,m_2)=-\int_{m_2}^{m_1}\left[\tau\mathbf{1}\{y\ge u\}+(1-\tau)\mathbf{1}\{y<u\}\right]\,du. s τ ( y , m 1 ) − s τ ( y , m 2 ) = − ∫ m 2 m 1 [ τ 1 { y ≥ u } + ( 1 − τ ) 1 { y < u } ] d u .
Taking conditional expectations given X i X_i X i and applying conditional Fubini, justified by the bounded integrand on the finite interval, gives
q τ ( m 1 ∣ X i ) − q τ ( m 2 ∣ X i ) = − ∫ m 2 m 1 [ τ P [ Y i ≥ u ∣ X i ] + ( 1 − τ ) P [ Y i < u ∣ X i ] ] d u . q_\tau(m_1\mid X_i)-q_\tau(m_2\mid X_i)=-\int_{m_2}^{m_1}\left[\tau\mathbb{P}[Y_i\ge u\mid X_i]+(1-\tau)\mathbb{P}[Y_i<u\mid X_i]\right]\,du. q τ ( m 1 ∣ X i ) − q τ ( m 2 ∣ X i ) = − ∫ m 2 m 1 [ τ P [ Y i ≥ u ∣ X i ] + ( 1 − τ ) P [ Y i < u ∣ X i ] ] d u .
The bracket lies between min { τ , 1 − τ } \min\{\tau,1-\tau\} min { τ , 1 − τ } and max { τ , 1 − τ } \max\{\tau,1-\tau\} max { τ , 1 − τ } , and hence
( m 1 − m 2 ) { q τ ( m 1 ∣ X i ) − q τ ( m 2 ∣ X i ) } ≤ − min { τ , 1 − τ } ( m 1 − m 2 ) 2 . (22) (m_1-m_2)\{q_\tau(m_1\mid X_i)-q_\tau(m_2\mid X_i)\}\le-\min\{\tau,1-\tau\}(m_1-m_2)^2. \tag*{(22)} ( m 1 − m 2 ) { q τ ( m 1 ∣ X i ) − q τ ( m 2 ∣ X i )} ≤ − min { τ , 1 − τ } ( m 1 − m 2 ) 2 . (22)
Suppose, for contradiction, that another θ ( τ ) ≠ θ o ( τ ) \theta(\tau)\ne\theta_o(\tau) θ ( τ ) = θ o ( τ ) solves the ER-SACR population moment, and set δ = θ o ( τ ) − θ ( τ ) \delta=\theta_o(\tau)-\theta(\tau) δ = θ o ( τ ) − θ ( τ ) , so δ ≠ 0 \delta\ne0 δ = 0 . Write c τ : = min { τ , 1 − τ } > 0 c_\tau:=\min\{\tau,1-\tau\}>0 c τ := min { τ , 1 − τ } > 0 , and let c X > 0 c_X>0 c X > 0 denote the uniform lower eigenvalue bound in Assumption LI 1(a). Subtracting the zero moment at θ o ( τ ) \theta_o(\tau) θ o ( τ ) gives
1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ X i ( q τ ( X i ′ θ o ( τ ) ∣ X i ) − q τ ( X i ′ θ ( τ ) ∣ X i ) ) ] = 0. \frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[X_i\left(q_\tau(X_i'\theta_o(\tau)\mid X_i)-q_\tau(X_i'\theta(\tau)\mid X_i)\right)\right]=0. G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ X i ( q τ ( X i ′ θ o ( τ ) ∣ X i ) − q τ ( X i ′ θ ( τ ) ∣ X i ) ) ] = 0.
Premultiplication by δ ′ \delta' δ ′ , followed by (S.3.3), gives the contradiction
0 = 1 G ∑ g = 1 G N g − 1 ∑ i ∈ G g E G [ ( X i ′ δ ) ( q τ ( X i ′ θ o ( τ ) ∣ X i ) − q τ ( X i ′ θ ( τ ) ∣ X i ) ) ] ≤ − c τ 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g ( X i ′ δ ) 2 ] = − c τ δ ′ { 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ ] } δ ≤ − c τ c X ∥ δ ∥ 2 < 0. \begin{aligned}
0=\frac{1}{G}\sum_{g=1}^{G}N_g^{-1}\sum_{i\in\mathcal{G}_g}\mathbb{E}_G\left[(X_i'\delta)\left(q_\tau(X_i'\theta_o(\tau)\mid X_i)-q_\tau(X_i'\theta(\tau)\mid X_i)\right)\right] \\
&\le-c_\tau\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}(X_i'\delta)^2\right] \\
&=-c_\tau\delta'\left\{\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i'\right]\right\}\delta\le-c_\tau c_X\|\delta\|^2<0.
\end{aligned} 0 = G 1 g = 1 ∑ G N g − 1 i ∈ G g ∑ E G [ ( X i ′ δ ) ( q τ ( X i ′ θ o ( τ ) ∣ X i ) − q τ ( X i ′ θ ( τ ) ∣ X i ) ) ] ≤ − c τ G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ ( X i ′ δ ) 2 = − c τ δ ′ ⎩ ⎨ ⎧ G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ ⎭ ⎬ ⎫ δ ≤ − c τ c X ∥ δ ∥ 2 < 0.
Hence θ o ( τ ) \theta_o(\tau) θ o ( τ ) is the unique ER-SACR population zero for each fixed τ \tau τ .
Finally, suppose that functional separation failed. Then for some ϵ > 0 \epsilon>0 ϵ > 0 one could find admissible paths θ G ( ⋅ ) ∈ Θ T \theta_G(\cdot) \in\Theta_{\mathcal{T}} θ G ( ⋅ ) ∈ Θ T with ∥ θ G − θ o ∥ T ≥ ϵ \|\theta_G-\theta_o\|_{\mathcal{T}} \ge\epsilon ∥ θ G − θ o ∥ T ≥ ϵ and
∥ P G Ψ ‾ ( W g ; θ G ) ∥ T → 0 \|\mathbb{P}_G\overline{\Psi}(W_g;\theta_G)\|_{\mathcal{T}} \to0 ∥ P G Ψ ( W g ; θ G ) ∥ T → 0
along a subsequence. Choose τ G ∈ T \tau_G \in\mathcal{T} τ G ∈ T with ∥ θ G ( τ G ) − θ o ( τ G ) ∥ ≥ ϵ / 2 \|\theta_G(\tau_G)-\theta_o(\tau_G)\| \ge\epsilon/2 ∥ θ G ( τ G ) − θ o ( τ G ) ∥ ≥ ϵ /2 , and set c T : = min { τ , 1 − τ ‾ } > 0 c_{\mathcal{T}} := \min\{\tau,\ 1-\overline{\tau}\} > 0 c T := min { τ , 1 − τ } > 0 . For δ = θ o ( τ ) − t \delta= \theta_o(\tau)-t δ = θ o ( τ ) − t , (S.3.3) and Assumption LI 1(a) imply
δ ′ P G ψ ‾ τ ( W g ; t ) ≥ c T δ ′ { 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i X i ′ ] } δ ≥ c T c X ∥ δ ∥ 2 \delta^{\prime}\mathbb{P}_G\overline{\psi}_{\tau}(W_g;t) \ge c_{\mathcal{T}}\delta^{\prime}\left\{\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_iX_i^{\prime}\right]\right\}\delta\ge c_{\mathcal{T}}c_X\|\delta\|^2 δ ′ P G ψ τ ( W g ; t ) ≥ c T δ ′ ⎩ ⎨ ⎧ G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i X i ′ ⎭ ⎬ ⎫ δ ≥ c T c X ∥ δ ∥ 2
where c X > 0 c_X > 0 c X > 0 is the uniform eigenvalue bound. Cauchy–Schwarz gives
∥ P G ψ ‾ τ ( W g ; t ) ∥ ≥ c T c X ∥ t − θ o ( τ ) ∥ . (23) \|\mathbb{P}_G\overline{\psi}_{\tau}(W_g;t)\| \ge c_{\mathcal{T}}c_X\|t-\theta_o(\tau)\|. \tag*{(23)} ∥ P G ψ τ ( W g ; t ) ∥ ≥ c T c X ∥ t − θ o ( τ ) ∥. (23)
At τ G \tau_G τ G , the right-hand side is at least c T c X ϵ / 2 c_{\mathcal{T}}c_X\epsilon/2 c T c X ϵ /2 , contradicting convergence to zero. Thus the functional separation condition holds.
The following lemma establishes Lipschitz continuity of the QR and ER coefficient paths, supplying the path regularity used in the subsequent score-process verification proofs.
Lemma S.3.3 (Lipschitz continuity of QR and ER coefficient paths). (a) Under Assumptions LI 1 and QR 1, the QR coefficient path satisfies
∥ θ o ( τ ′ ) − θ o ( τ ) ∥ ≤ C Q R ∣ τ ′ − τ ∣ , τ , τ ′ ∈ T , \|\theta_o(\tau^{\prime})-\theta_o(\tau)\| \le C_{QR}|\tau^{\prime}-\tau|,\qquad\tau,\tau^{\prime}\in\mathcal{T}, ∥ θ o ( τ ′ ) − θ o ( τ ) ∥ ≤ C QR ∣ τ ′ − τ ∣ , τ , τ ′ ∈ T , for some finite constant C Q R C_{QR} C QR .
(b) Under Assumptions LI 1 and ER 1 and Assumption ER 2, the ER coefficient path satisfies
∥ θ o ( τ ′ ) − θ o ( τ ) ∥ ≤ C E R ∣ τ ′ − τ ∣ , τ , τ ′ ∈ T , \|\theta_o(\tau^{\prime})-\theta_o(\tau)\| \le C_{ER}|\tau^{\prime}-\tau|,\qquad\tau,\tau^{\prime}\in\mathcal{T}, ∥ θ o ( τ ′ ) − θ o ( τ ) ∥ ≤ C E R ∣ τ ′ − τ ∣ , τ , τ ′ ∈ T , for some finite constant C E R C_{ER} C E R .
Proof. Part (a): To establish regularity of the coefficient path, define
μ G : = 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i ] . \mu_G := \frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_i\right]. μ G := G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i .
The QR score is affine in τ \tau τ , and the population moment vanishes at θ o ( τ ′ ) \theta_o(\tau^{\prime}) θ o ( τ ′ ) for index τ ′ \tau^{\prime} τ ′ . Consequently,
P G ψ ‾ τ ( W g ; θ o ( τ ′ ) ) = P G ψ ‾ τ ′ ( W g ; θ o ( τ ′ ) ) + ( τ − τ ′ ) μ G = ( τ − τ ′ ) μ G . \mathbb{P}_G\overline{\psi}_{\tau}(W_g;\theta_o(\tau^{\prime}))=\mathbb{P}_G\overline{\psi}_{\tau^{\prime}}(W_g;\theta_o(\tau^{\prime}))+(\tau-\tau^{\prime})\mu_G=(\tau-\tau^{\prime})\mu_G. P G ψ τ ( W g ; θ o ( τ ′ )) = P G ψ τ ′ ( W g ; θ o ( τ ′ )) + ( τ − τ ′ ) μ G = ( τ − τ ′ ) μ G .
Applying the identification bound (S.3.2) at t = θ o ( τ ′ ) t=\theta_o(\tau') t = θ o ( τ ′ ) gives
c f ∥ θ o ( τ ′ ) − θ o ( τ ) ∥ ≤ ∣ τ − τ ′ ∣ ∥ μ G ∥ . c_f\lVert\theta_o(\tau')-\theta_o(\tau)\rVert\leq\lvert\tau-\tau'\rvert\lVert\mu_G\rVert. c f ∥ θ o ( τ ′ ) − θ o ( τ )∥ ≤ ∣ τ − τ ′ ∣ ∥ μ G ∥ .
By Assumption LI 1(b), sup G ∥ μ G ∥ ≤ sup G , i E G ∥ X i ∥ < ∞ \sup_G\lVert\mu_G\rVert\leq\sup_{G,i}\mathbb{E}_G\lVert X_i\rVert<\infty sup G ∥ μ G ∥ ≤ sup G , i E G ∥ X i ∥ < ∞ . Thus C Q R : = c f − 1 sup G ∥ μ G ∥ C_{QR}:=c_f^{-1}\sup_G\lVert\mu_G\rVert C QR := c f − 1 sup G ∥ μ G ∥ is a finite Lipschitz constant, uniform over τ , τ ′ ∈ T \tau,\tau'\in\mathcal{T} τ , τ ′ ∈ T . This is the cluster-average analogue of Hagemann [7 ] , obtained directly from identification without differentiating the path.
Part (b): Let B Θ : = sup t ∈ Θ ∥ t ∥ B_\Theta:=\sup_{t\in\Theta}\lVert t\rVert B Θ := sup t ∈ Θ ∥ t ∥ , and let c X > 0 c_X>0 c X > 0 denote the uniform eigenvalue bound in Assumption LI 1(a). Set c τ : = inf τ ∈ T min { τ , 1 − τ } > 0 c_\tau:=\inf_{\tau\in\mathcal{T}}\min\{\tau,1-\tau\}>0 c τ := inf τ ∈ T min { τ , 1 − τ } > 0 . For Lipschitz continuity, the ER score satisfies the exact identity
ψ τ ( W ; t ) − ψ τ ′ ( W ; t ) = ( τ − τ ′ ) X ∣ Y − X ′ t ∣ . \psi_\tau(W;t)-\psi_{\tau'}(W;t)=(\tau-\tau')X\lvert Y-X't\rvert. ψ τ ( W ; t ) − ψ τ ′ ( W ; t ) = ( τ − τ ′ ) X ∣ Y − X ′ t ∣ .
The zero moment at index τ ′ \tau' τ ′ therefore gives
P G ψ ‾ τ ( W g ; θ o ( τ ′ ) ) = ( τ − τ ′ ) 1 G ∑ g = 1 G E G [ N g − 1 ∑ i ∈ G g X i ∣ Y i − X i ′ θ o ( τ ′ ) ∣ ] . \mathbb{P}_G\overline{\psi}_{\tau}(W_g;\theta_o(\tau'))=(\tau-\tau')\frac{1}{G}\sum_{g=1}^{G}\mathbb{E}_G\left[N_g^{-1}\sum_{i\in\mathcal{G}_g}X_i\lvert Y_i-X_i'\theta_o(\tau')\rvert\right]. P G ψ τ ( W g ; θ o ( τ ′ )) = ( τ − τ ′ ) G 1 g = 1 ∑ G E G N g − 1 i ∈ G g ∑ X i ∣ Y i − X i ′ θ o ( τ ′ )∣ .
Define
M E R : = sup G , i E G [ ∥ X i ∥ ∣ Y i ∣ + B Θ ∥ X i ∥ 2 ] < ∞ , M_{ER}:=\sup_{G,i}\mathbb{E}_G\left[\lVert X_i\rVert\lvert Y_i\rvert+B_\Theta\lVert X_i\rVert^2\right]<\infty, M E R := G , i sup E G [ ∥ X i ∥ ∣ Y i ∣ + B Θ ∥ X i ∥ 2 ] < ∞ ,
where finiteness follows from Assumption ER 2. Applying the identification bound (S.3.4) at t = θ o ( τ ′ ) t=\theta_o(\tau') t = θ o ( τ ′ ) , followed by the triangle inequality, yields
c τ c X ∥ θ o ( τ ′ ) − θ o ( τ ) ∥ ≤ ∥ P G ψ ‾ τ ( W g ; θ o ( τ ′ ) ) ∥ ≤ M E R ∣ τ − τ ′ ∣ . c_\tau c_X\lVert\theta_o(\tau')-\theta_o(\tau)\rVert\leq\lVert\mathbb{P}_G\overline{\psi}_{\tau}(W_g;\theta_o(\tau'))\rVert\leq M_{ER}\lvert\tau-\tau'\rvert. c τ c X ∥ θ o ( τ ′ ) − θ o ( τ )∥ ≤ ∥ P G ψ τ ( W g ; θ o ( τ ′ ))∥ ≤ M E R ∣ τ − τ ′ ∣ .
Thus C E R : = M E R / ( c τ c X ) C_{ER}:=M_{ER}/(c_\tau c_X) C E R := M E R / ( c τ c X ) is a finite uniform Lipschitz constant.
Proof of prop:ERverify
The individual coordinate-score class is manageable with envelope F j E R F_j^{ER} F j E R by Lemma S.3.1. Lemma S.3.2 therefore gives manageability of the normalised cluster-average array Z G g , j E R Z_{Gg,j}^{ER} Z G g , j E R with the corresponding normalised RMS envelope F G g , j E R F_{Gg,j}^{ER} F G g , j E R . Continuity of the ER score in ( τ , t ) (\tau,t) ( τ , t ) , together with continuity of θ o ( ⋅ ) \theta_o(\cdot) θ o ( ⋅ ) from Lemma S.3.3, makes the full, truth-indexed, and local-difference arrays separable; hence they are AMS by Kosorok [10 ] . Finally, Assumption ER 2 gives uniformly bounded ( 2 + δ ) (2+\delta) ( 2 + δ ) th moments of F j E R F_j^{ER} F j E R , establishing the required squared-envelope uniform integrability.
Expectile paths from counterfactual distributions
The preceding verification concerns the direct ER estimating equation. For expectile treatment effects computed instead from estimated counterfactual distributions, the required scalar functional is the distribution-to-expectile map whose limiting behaviour is handled in this section using the Z-map framework of Chernozhukov, Fernández-Val, and Melly [4 ] . Let T = [ τ ‾ , τ ‾ ] ⊂ ( 0 , 1 ) \mathcal{T} = [\underline{\tau},\overline{\tau}] \subset(0,1) T = [ τ , τ ] ⊂ ( 0 , 1 ) , and let F F F be a distribution with a finite first moment. Set w τ ( y , e ) = τ 1 { y ≥ e } + ( 1 − τ ) 1 { y < e } w_{\tau}(y,e) = \tau1\{y \ge e\} + (1-\tau)1\{y < e\} w τ ( y , e ) = τ 1 { y ≥ e } + ( 1 − τ ) 1 { y < e } , equivalently, w τ ( y , e ) = τ + ( 1 − 2 τ ) 1 { y < e } w_{\tau}(y,e) = \tau+ (1-2\tau)1\{y < e\} w τ ( y , e ) = τ + ( 1 − 2 τ ) 1 { y < e } , and write e τ ( F ) e_{\tau}(F) e τ ( F ) for the unique zero in e ∈ R e \in\mathbb{R} e ∈ R of Ψ F ( e , τ ) : = ∫ w τ ( y , e ) ( e − y ) d F ( y ) \Psi_F(e,\tau) := \int w_{\tau}(y,e)(e-y)\,dF(y) Ψ F ( e , τ ) := ∫ w τ ( y , e ) ( e − y ) d F ( y ) . The following imposes a regularity condition on the distribution-to-expectile map.
Condition (Expectile-map regularity). F F F is continuous at e τ ( F ) e_{\tau}(F) e τ ( F ) for every τ ∈ T \tau\in\mathcal{T} τ ∈ T .
Write E ( F ) : = ( e τ ( F ) ) τ ∈ T \mathcal{E}(F) := (e_{\tau}(F))_{\tau\in\mathcal{T}} E ( F ) := ( e τ ( F ) ) τ ∈ T , and define the candidate derivative map E F ′ : L 1 ( R ) → ℓ ∞ ( T ) \mathcal{E}'_F : L^1(\mathbb{R}) \to\ell^\infty(\mathcal{T}) E F ′ : L 1 ( R ) → ℓ ∞ ( T ) by
E F ′ [ h ] ( τ ) = − a τ ( F ) − 1 ∫ w τ ( y , e τ ( F ) ) h ( y ) d y , a τ ( F ) : = ∫ w τ ( y , e τ ( F ) ) d F ( y ) \mathcal{E}'_F[h](\tau) = -a_{\tau}(F)^{-1}\int w_{\tau}(y,e_{\tau}(F))h(y)\,dy,\qquad a_{\tau}(F) := \int w_{\tau}(y,e_{\tau}(F))\,dF(y) E F ′ [ h ] ( τ ) = − a τ ( F ) − 1 ∫ w τ ( y , e τ ( F )) h ( y ) d y , a τ ( F ) := ∫ w τ ( y , e τ ( F )) d F ( y )
and the tangent set
D F : = { h ∈ L 1 ( R ) : there exist t n ↓ 0 and h n → h in L 1 , with F + t n h n a finite-first-moment distribution function } . \mathbb{D}_F := \left\{h \in L^1(\mathbb{R}) : \begin{array}{l}\text{there exist }t_n \downarrow0\text{ and }h_n \to h\text{ in }L^1,\\
\text{with }F+t_nh_n\text{ a finite-first-moment distribution function}\end{array}\right\}. D F := { h ∈ L 1 ( R ) : there exist t n ↓ 0 and h n → h in L 1 , with F + t n h n a finite-first-moment distribution function } .
Proposition S.4.1 (Expectile map for distribution paths). Under the preceding expectile-map regularity condition:
(a) With differences of distribution functions equipped with the L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) norm, E \mathcal{E} E is Hadamard differentiable at F F F , tangentially to D F \mathbb{D}_F D F , with derivative E F ′ \mathcal{E}'_F E F ′ .
(b) If G ( F ^ − F ) ⇝ H \sqrt{G}(\widehat{F}-F) \rightsquigarrow\mathbb{H} G ( F − F ) ⇝ H in L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) , then G { E ( F ^ ) − E ( F ) } ⇝ E F ′ [ H ] \sqrt{G}\{\mathcal{E}(\widehat{F})-\mathcal{E}(F)\} \rightsquigarrow\mathcal{E}'_F[\mathbb{H}] G { E ( F ) − E ( F )} ⇝ E F ′ [ H ] in ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) .
(c) Under the premise of part (b), if G ( F ^ ∗ − F ^ ) ⇝ P H \sqrt{G}(\widehat{F}^{*}-\widehat{F}) \rightsquigarrow_{\mathbb{P}} \mathbb{H} G ( F ∗ − F ) ⇝ P H in L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) , then G { E ( F ^ ∗ ) − E ( F ^ ) } ⇝ P E F ′ [ H ] \sqrt{G}\{\mathcal{E}(\widehat{F}^{*})-\mathcal{E}(\widehat{F})\} \rightsquigarrow_{\mathbb{P}} \mathcal{E}'_F[\mathbb{H}] G { E ( F ∗ ) − E ( F )} ⇝ P E F ′ [ H ] in ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) . The same conclusion holds for approximate bootstrap roots e ~ τ ∗ \widetilde{e}^{*}_{\tau} e τ ∗ satisfying sup τ ∈ T ∣ Ψ F ^ ∗ ( e ~ τ ∗ , τ ) ∣ = o P ∗ ( G − 1 / 2 ) \sup_{\tau\in\mathcal{T}}|\Psi_{\widehat{F}^{*}}(\widetilde{e}^{*}_{\tau},\tau)| = o_{\mathbb{P}^{*}}(G^{-1/2}) sup τ ∈ T ∣ Ψ F ∗ ( e τ ∗ , τ ) ∣ = o P ∗ ( G − 1/2 ) , in outer probability.
Unlike quantiles, expectiles depend on integrated probabilities on both sides of the candidate expectile. The L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) premise in Proposition S.4.1 states the population condition for full-distribution inference and is not implied by compact-threshold DR convergence. The empirical illustration in Section S.6 therefore implements an approximation explicitly on the observed support of the outcome.
Proof of Proposition S.4.1
Part (a). The finite first moment makes Ψ F ( e , τ ) \Psi_F(e,\tau) Ψ F ( e , τ ) finite and continuous in ( e , τ ) (e,\tau) ( e , τ ) on R × U \mathbb{R} \times\mathcal{U} R × U for any open interval U \mathcal{U} U containing T \mathcal{T} T and contained in ( 0 , 1 ) (0,1) ( 0 , 1 ) . For each τ \tau τ , e ↦ Ψ F ( e , τ ) e \mapsto\Psi_F(e,\tau) e ↦ Ψ F ( e , τ ) is strictly increasing, crosses zero, and is the gradient of a convex antiderivative. Existence and uniqueness, together with strict monotonicity of the expectile path, follow from Newey and Powell [14 ] .
Monotonicity places the entire root path in the compact interval [ e τ ( F ) , e τ ( F ) ] [e_{\tau}(F),e_{\tau}(F)] [ e τ ( F ) , e τ ( F )] . To verify continuity, suppose to the contrary that τ n → τ \tau_n \to\tau τ n → τ while, for some ε > 0 \varepsilon> 0 ε > 0 ,
∣ e τ n ( F ) − e τ ( F ) ∣ ≥ ε . \lvert e_{\tau_n}(F)-e_{\tau}(F)\rvert\ge\varepsilon. ∣ e τ n ( F ) − e τ ( F )∣ ≥ ε .
The Bolzano–Weierstrass Theorem gives a subsequence satisfying e τ n k ( F ) → e ∗ e_{\tau_{n_k}}(F) \to e^* e τ n k ( F ) → e ∗ . Joint continuity and the root restrictions give 0 = lim k Ψ F ( e τ n k ( F ) , τ n k ) = Ψ F ( e ∗ , τ ) 0=\lim_k \Psi_F(e_{\tau_{n_k}}(F),\tau_{n_k})=\Psi_F(e^*,\tau) 0 = lim k Ψ F ( e τ n k ( F ) , τ n k ) = Ψ F ( e ∗ , τ ) , so uniqueness requires e ∗ = e τ ( F ) e^*=e_{\tau}(F) e ∗ = e τ ( F ) , a contradiction. Thus τ ↦ e τ ( F ) \tau\mapsto e_{\tau}(F) τ ↦ e τ ( F ) is continuous. Because F F F is continuous at every point in the image of this path, it is continuous on that compact root interval. At each root,
∂ e Ψ F ( e τ ( F ) , τ ) = a τ ( F ) ≥ min { τ , 1 − τ } > 0 , ∂ τ Ψ F ( e τ ( F ) , τ ) = − ∫ ∣ e τ ( F ) − y ∣ d F ( y ) . \partial_e\Psi_F(e_{\tau}(F),\tau)=a_{\tau}(F)\ge\min\{\tau,1-\tau\}>0,\qquad
\partial_{\tau}\Psi_F(e_{\tau}(F),\tau)=-\int\lvert e_{\tau}(F)-y\rvert\,dF(y). ∂ e Ψ F ( e τ ( F ) , τ ) = a τ ( F ) ≥ min { τ , 1 − τ } > 0 , ∂ τ Ψ F ( e τ ( F ) , τ ) = − ∫ ∣ e τ ( F ) − y ∣ d F ( y ) .
Dominated convergence, using the compact root interval and the finite first moment, makes both derivatives continuous along the root path. These properties connect the construction to the indexed Z Z Z -map framework of Chernozhukov, Fernández-Val, and Melly [4 ] . The next step is to establish differentiability in the stated L 1 L^1 L 1 topology.
To make the derivative explicit, consider an admissible perturbation F t = F + t h t F_t=F+th_t F t = F + t h t , where h t → h h_t\to h h t → h in L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) , and write e t = e τ ( F t ) e_t=e_{\tau}(F_t) e t = e τ ( F t ) and e = e τ ( F ) e=e_{\tau}(F) e = e τ ( F ) . Since both are exact roots,
0 = Ψ F t ( e t , τ ) − Ψ F ( e , τ ) = { Ψ F t ( e t , τ ) − Ψ F t ( e , τ ) } ⏟ Δ e , t ( τ ) + { Ψ F t ( e , τ ) − Ψ F ( e , τ ) } ⏟ Δ F , t ( τ ) . (24) \begin{aligned}
0&=\Psi_{F_t}(e_t,\tau)-\Psi_F(e,\tau)\\
&=\underbrace{\{\Psi_{F_t}(e_t,\tau)-\Psi_{F_t}(e,\tau)\}}_{\Delta_{e,t}(\tau)}
+\underbrace{\{\Psi_{F_t}(e,\tau)-\Psi_F(e,\tau)\}}_{\Delta_{F,t}(\tau)}. \tag*{(24)}
\end{aligned} 0 = Ψ F t ( e t , τ ) − Ψ F ( e , τ ) = Δ e , t ( τ ) { Ψ F t ( e t , τ ) − Ψ F t ( e , τ )} + Δ F , t ( τ ) { Ψ F t ( e , τ ) − Ψ F ( e , τ )} . (24)
Define a τ ( F t , u ) : = ∫ w τ ( y , u ) d F t ( y ) a_{\tau}(F_t,u):=\int w_{\tau}(y,u)\,dF_t(y) a τ ( F t , u ) := ∫ w τ ( y , u ) d F t ( y ) . For fixed y y y , the function u ↦ w τ ( y , u ) ( u − y ) u\mapsto w_{\tau}(y,u)(u-y) u ↦ w τ ( y , u ) ( u − y ) is continuous and piecewise linear, with derivative w τ ( y , u ) w_{\tau}(y,u) w τ ( y , u ) almost everywhere. Hence
w τ ( y , e t ) ( e t − y ) − w τ ( y , e ) ( e − y ) = ∫ e e t w τ ( y , u ) d u . w_{\tau}(y,e_t)(e_t-y)-w_{\tau}(y,e)(e-y)=\int_e^{e_t}w_{\tau}(y,u)\,du. w τ ( y , e t ) ( e t − y ) − w τ ( y , e ) ( e − y ) = ∫ e e t w τ ( y , u ) d u .
Integrating against d F t ( y ) dF_t(y) d F t ( y ) and applying Fubini’s theorem, justified by 0 ≤ w τ ≤ 1 0\le w_{\tau}\le1 0 ≤ w τ ≤ 1 , gives
Δ e , t ( τ ) = ∫ e e t a τ ( F t , u ) d u = a ˉ t , τ ( e t − e ) , \Delta_{e,t}(\tau)=\int_e^{e_t}a_{\tau}(F_t,u)\,du=\bar a_{t,\tau}(e_t-e), Δ e , t ( τ ) = ∫ e e t a τ ( F t , u ) d u = a ˉ t , τ ( e t − e ) ,
where a ˉ t , τ \bar a_{t,\tau} a ˉ t , τ denotes the average of a τ ( F t , ⋅ ) a_{\tau}(F_t,\cdot) a τ ( F t , ⋅ ) between e e e and e t e_t e t , with the natural value a τ ( F t , e ) a_{\tau}(F_t,e) a τ ( F t , e ) when e t = e e_t=e e t = e . Set H t : = F t − F = t h t H_t:=F_t-F=th_t H t := F t − F = t h t . Splitting at e e e and applying Lebesgue–Stieltjes integration by parts, the distribution-perturbation component Δ F , t ( τ ) \Delta_{F,t}(\tau) Δ F , t ( τ ) in (S.4.1) satisfies
Δ F , t ( τ ) = ( 1 − τ ) ∫ ( − ∞ , e ) ( e − y ) d H t ( y ) + τ ∫ [ e , ∞ ) ( e − y ) d H t ( y ) = ( 1 − τ ) ∫ − ∞ e H t ( y ) d y + τ ∫ e ∞ H t ( y ) d y = ∫ R w τ ( y , e ) H t ( y ) d y = t ∫ w τ ( y , e ) h t ( y ) d y . \begin{aligned}
\Delta_{F,t}(\tau) &= (1-\tau)\int_{(-\infty,e)} (e-y)\,dH_t(y)+\tau\int_{[e,\infty)}(e-y)\,dH_t(y)\\
&=(1-\tau)\int_{-\infty}^{e}H_t(y)\,dy+\tau\int_{e}^{\infty}H_t(y)\,dy\\
&=\int_{\mathbb{R}}w_\tau(y,e)H_t(y)\,dy\\
&=t\int w_\tau(y,e)h_t(y)\,dy.
\end{aligned} Δ F , t ( τ ) = ( 1 − τ ) ∫ ( − ∞ , e ) ( e − y ) d H t ( y ) + τ ∫ [ e , ∞ ) ( e − y ) d H t ( y ) = ( 1 − τ ) ∫ − ∞ e H t ( y ) d y + τ ∫ e ∞ H t ( y ) d y = ∫ R w τ ( y , e ) H t ( y ) d y = t ∫ w τ ( y , e ) h t ( y ) d y .
The terms at e e e vanish because e − y = 0 e-y=0 e − y = 0 there. The tail terms vanish because the finite first moments of F t F_t F t and F F F imply the corresponding tail limits y ( 1 − F t ( y ) ) , y ( 1 − F ( y ) ) → 0 y(1-F_t(y)),y(1-F(y))\to0 y ( 1 − F t ( y )) , y ( 1 − F ( y )) → 0 as y → ∞ y\to\infty y → ∞ and ∣ y ∣ F t ( y ) , ∣ y ∣ F ( y ) → 0 |y|F_t(y),|y|F(y)\to0 ∣ y ∣ F t ( y ) , ∣ y ∣ F ( y ) → 0 as y → − ∞ y\to-\infty y → − ∞ . Substitution of these two expressions into Equation (S.4.1) gives
e t − e t = − a ˉ t , τ − 1 ∫ w τ ( y , e ) h t ( y ) d y . \frac{e_t-e}{t}=-\bar{a}_{t,\tau}^{-1}\int w_\tau(y,e)h_t(y)\,dy. t e t − e = − a ˉ t , τ − 1 ∫ w τ ( y , e ) h t ( y ) d y .
Set c : = min { τ , 1 − τ } > 0 c:=\min\{\tau,1-\tau\}>0 c := min { τ , 1 − τ } > 0 . Since a ˉ t , τ ≥ c \bar{a}_{t,\tau}\ge c a ˉ t , τ ≥ c and 0 ≤ w τ ≤ 1 0\le w_\tau\le1 0 ≤ w τ ≤ 1 , the same identity yields the bound
sup τ ∈ T ∣ e τ ( F t ) − e τ ( F ) ∣ ≤ c − 1 ∥ F t − F ∥ L 1 = O ( ∣ t ∣ ) . (25) \sup_{\tau\in\mathcal{T}}|e_\tau(F_t)-e_\tau(F)|\le c^{-1}\|F_t-F\|_{L^1}=O(|t|). \tag*{(25)} τ ∈ T sup ∣ e τ ( F t ) − e τ ( F ) ∣ ≤ c − 1 ∥ F t − F ∥ L 1 = O ( ∣ t ∣ ) . (25)
Let I : = [ e τ ( F ) , e τ ‾ ( F ) ] I:=[e_{\tau}(F),e_{\overline{\tau}}(F)] I := [ e τ ( F ) , e τ ( F )] , and put
ω F ( r ) : = sup x ∈ I { F ( x + 2 r ) − F ( x − 2 r ) } . \omega_F(r):=\sup_{x\in I}\{F(x+2r)-F(x-2r)\}. ω F ( r ) := x ∈ I sup { F ( x + 2 r ) − F ( x − 2 r )} .
Continuity of F F F at every point of the compact interval I I I implies ω F ( r ) → 0 \omega_F(r)\to0 ω F ( r ) → 0 as r ↓ 0 r\downarrow0 r ↓ 0 . For fixed r > 0 r>0 r > 0 , (S.4.2) places every u u u between e τ ( F t ) e_\tau(F_t) e τ ( F t ) and e τ ( F ) e_\tau(F) e τ ( F ) within distance r r r of e τ ( F ) e_\tau(F) e τ ( F ) for all sufficiently small ∣ t ∣ |t| ∣ t ∣ . Monotonicity of the CDFs, averaging over [ u − r , u ] [u-r,u] [ u − r , u ] and [ u , u + r ] [u,u+r] [ u , u + r ] , then gives
∣ F t ( u − ) − F ( e τ ( F ) ) ∣ ≤ r − 1 ∥ F t − F ∥ L 1 + ω F ( r ) . |F_t(u-)-F(e_\tau(F))|\le r^{-1}\|F_t-F\|_{L^1}+\omega_F(r). ∣ F t ( u − ) − F ( e τ ( F )) ∣ ≤ r − 1 ∥ F t − F ∥ L 1 + ω F ( r ) .
Because a τ ( F t , u ) = τ + ( 1 − 2 τ ) F t ( u − ) a_\tau(F_t,u)=\tau+(1-2\tau)F_t(u-) a τ ( F t , u ) = τ + ( 1 − 2 τ ) F t ( u − ) and a τ ( F ) = τ + ( 1 − 2 τ ) F ( e τ ( F ) ) a_\tau(F)=\tau+(1-2\tau)F(e_\tau(F)) a τ ( F ) = τ + ( 1 − 2 τ ) F ( e τ ( F )) , averaging this inequality along the root segment, first letting t → 0 t\to0 t → 0 and then r ↓ 0 r\downarrow0 r ↓ 0 , proves
sup τ ∈ T ∣ a ˉ t , τ − a τ ( F ) ∣ ⟶ 0. \sup_{\tau\in\mathcal{T}}|\bar{a}_{t,\tau}-a_\tau(F)|\longrightarrow0. τ ∈ T sup ∣ a ˉ t , τ − a τ ( F ) ∣ ⟶ 0.
Finally, the integral operator is linear and bounded on L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) , uniformly in ( e , τ ) (e,\tau) ( e , τ ) , so h t → h h_t\to h h t → h in L 1 L^1 L 1 gives the asserted derivative uniformly over T \mathcal{T} T . Hence
e t ( τ ) − e τ ( F ) t ⟶ − a τ ( F ) − 1 ∫ w τ ( y , e τ ( F ) ) h ( y ) d y \frac{e_t(\tau)-e_\tau(F)}{t}\longrightarrow-a_\tau(F)^{-1}\int w_\tau(y,e_\tau(F))h(y)\,dy t e t ( τ ) − e τ ( F ) ⟶ − a τ ( F ) − 1 ∫ w τ ( y , e τ ( F )) h ( y ) d y
in ℓ ∞ ( T ) \ell^\infty(\mathcal{T}) ℓ ∞ ( T ) . The limit is E F ′ [ h ] \mathcal{E}'_F[h] E F ′ [ h ] . Since the admissible perturbation sequence was arbitrary, and ∥ E F ′ [ h ] ∥ T ≤ c − 1 ∥ h ∥ L 1 \left\|\mathcal{E}'_{F}[h]\right\|_{T} \le c^{-1}\|h\|_{L^{1}} ∥ E F ′ [ h ] ∥ T ≤ c − 1 ∥ h ∥ L 1 , this proves the asserted Hadamard differentiability with a continuous linear derivative.
Part (b). Applying the ordinary functional delta method to part (a) and the stated L 1 L^{1} L 1 weak convergence gives the conclusion; see Chernozhukov, Fernández-Val, and Melly [4 ] .
Part (c). Applying the bootstrap functional delta method to part (a), under the premises of parts (b) and (c), gives the conclusion; see Chernozhukov, Fernández-Val, and Melly [4 ] . The uniform slope lower bound also bounds the displacement of an approximate bootstrap root from the exact root by its absolute score residual divided by c c c . Thus the stipulated uniformly negligible residuals do not affect the bootstrap limit.
Supplementary Monte Carlo Experiments and Implementation
Data-generating processes and inference implementation
The baseline DGPs, treatment assignment, cluster-size arrays, and simulation settings are given in § 5.1 . Their true coefficient paths are
θ o D R ( y ) = ( y + 0.5 , − 0.5 ) ′ , \theta_{o}^{DR}(y) = (y + 0.5, -0.5)^{\prime}, θ o D R ( y ) = ( y + 0.5 , − 0.5 ) ′ ,
θ o Q R ( τ ) = θ o + γ Φ − 1 ( τ ) , \theta_{o}^{QR}(\tau) = \theta_{o} + \gamma\Phi^{-1}(\tau), θ o QR ( τ ) = θ o + γ Φ − 1 ( τ ) ,
θ o E R ( τ ) = θ o + γ E τ ( U ) , \theta_{o}^{ER}(\tau) = \theta_{o} + \gamma\mathbb{E}_{\tau}(U), θ o E R ( τ ) = θ o + γ E τ ( U ) ,
where U U U is standard normal and E τ ( U ) \mathbb{E}_{\tau}(U) E τ ( U ) is its τ \tau τ -expectile. In particular, the DR conditional CDF is F Y ∣ T ( y ∣ T g ) = Λ ( X i ′ θ o D R ( y ) ) F_{Y\mid T}(y\mid T_g)=\Lambda(X_i^{\prime}\theta_{o}^{DR}(y)) F Y ∣ T ( y ∣ T g ) = Λ ( X i ′ θ o D R ( y )) , with logistic link Λ \Lambda Λ .
For the baseline Pareto arrays, β = 2.5 \beta=2.5 β = 2.5 gives N ≍ G N\asymp G N ≍ G and max g N g 2 / N ≍ G − 0.2 → 0 \max_g N_g^2/N\asymp G^{-0.2}\to0 max g N g 2 / N ≍ G − 0.2 → 0 . When β = 1 \beta=1 β = 1 , N ≍ G log G N\asymp G\log G N ≍ G log G , max g N g 2 / N ≍ G / log G \max_g N_g^2/N\asymp G/\log G max g N g 2 / N ≍ G / log G , and max g N g / N ≍ 1 / log G → 0 \max_g N_g/N\asymp1/\log G\to0 max g N g / N ≍ 1/ log G → 0 .
For conventional QR inference, CR-A uses an observation-weighted Jacobian estimated by a normal-kernel residual-density plug-in with bandwidth h = 1.06 s r N − 1 / 5 h=1.06s_rN^{-1/5} h = 1.06 s r N − 1/5 , where s r s_r s r is the QR residual standard deviation. The remaining inference procedures are described in § 5.2 and § 5.3 .
Evaluation grids and performance measures
Let R M C R_{\mathrm{MC}} R MC denote the number of Monte Carlo replications, θ ^ s \hat{\theta}_s θ ^ s the estimate in replication s s s , and θ o \theta_o θ o its true value. The pointwise measures are
Bias = R M C − 1 ∑ s = 1 R M C ( θ ^ s − θ o ) , \operatorname{Bias}=R_{\mathrm{MC}}^{-1}\sum_{s=1}^{R_{\mathrm{MC}}}(\hat{\theta}_s-\theta_o), Bias = R MC − 1 s = 1 ∑ R MC ( θ ^ s − θ o ) ,
RMSE = { R M C − 1 ∑ s = 1 R M C ( θ ^ s − θ o ) 2 } 1 / 2 . \operatorname{RMSE}=\left\{R_{\mathrm{MC}}^{-1}\sum_{s=1}^{R_{\mathrm{MC}}}(\hat{\theta}_s-\theta_o)^2\right\}^{1/2}. RMSE = { R MC − 1 s = 1 ∑ R MC ( θ ^ s − θ o ) 2 } 1/2 .
For each target f ∈ { F , Q , E } f\in\{F,Q,E\} f ∈ { F , Q , E } , the evaluation grid contains L G ( f ) = 1 + ⌈ 5 M f 1 / 2 S f G 0.3 2 2 ⌉ L_G^{(f)}=1+\left\lceil\frac{5M_f^{1/2}S_fG^{0.3}}{2\sqrt{2}}\right\rceil L G ( f ) = 1 + ⌈ 2 2 5 M f 1/2 S f G 0.3 ⌉ equally spaced points, where S f S_f S f is the length of the evaluation set and M f M_f M f is its target-specific population. curvature envelope.9 For a given target f f f and cluster count G G G , write L = L G ( f ) L = L_{G}^{(f)} L = L G ( f ) . For grid points z 1 , … , z L z_{1}, \ldots, z_{L} z 1 , … , z L , the reported integrated bias and integrated RMSE are
IBias = ( L − 1 ∑ ℓ = 1 L ( Δ ^ ( z ℓ ) − Δ ( z ℓ ) ) 2 ) 1 / 2 , IRMSE = ( ( R M C L ) − 1 ∑ s = 1 R M C ∑ ℓ = 1 L ( Δ ^ s ( z ℓ ) − Δ ( z ℓ ) ) 2 ) 1 / 2 . \operatorname{IBias} = \left(L^{-1}\sum_{\ell=1}^{L}\left(\widehat{\Delta}(z_{\ell})-\Delta(z_{\ell})\right)^{2}\right)^{1/2}, \qquad\operatorname{IRMSE} = \left((R_{\mathrm{MC}}L)^{-1}\sum_{s=1}^{R_{\mathrm{MC}}}\sum_{\ell=1}^{L}\left(\widehat{\Delta}_{s}(z_{\ell})-\Delta(z_{\ell})\right)^{2}\right)^{1/2}. IBias = ( L − 1 ℓ = 1 ∑ L ( Δ ( z ℓ ) − Δ ( z ℓ ) ) 2 ) 1/2 , IRMSE = ( ( R MC L ) − 1 s = 1 ∑ R MC ℓ = 1 ∑ L ( Δ s ( z ℓ ) − Δ ( z ℓ ) ) 2 ) 1/2 .
Here Δ ^ ( z ℓ ) = R M C − 1 ∑ s = 1 R M C Δ ^ s ( z ℓ ) \widehat{\Delta}(z_{\ell}) = R_{\mathrm{MC}}^{-1}\sum_{s=1}^{R_{\mathrm{MC}}}\widehat{\Delta}_{s}(z_{\ell}) Δ ( z ℓ ) = R MC − 1 ∑ s = 1 R MC Δ s ( z ℓ ) . IBias measures systematic displacement of the estimated path, whereas IRMSE also incorporates its Monte Carlo variation. Avg. Band and U-Rej. are defined in § 5.3 .
Sensitivity to within-cluster dependence
Table S.1 repeats the pointwise comparison at latent correlations 0.25 and 0.50. In the ignorable β = 2.5 \beta= 2.5 β = 2.5 regime, SACR/conventional RMSE ratios range from 0.94 to 1.01 at ρ = 0.25 \rho= 0.25 ρ = 0.25 , and from 0.89 to 0.96 at ρ = 0.50 \rho= 0.50 ρ = 0.50 , indicating little efficiency cost in these designs. Under β = 1 \beta= 1 β = 1 , SACR retains substantial RMSE and rejection-rate advantages. At G = 600 G = 600 G = 600 , its bootstrap rejection rates across both correlations and all models lie between 5.1% and 5.7%, compared with 18.8% to 26.4% conventionally. The principal smaller-sample distortion is ER at ρ = 0.25 \rho= 0.25 ρ = 0.25 and G = 150 G = 150 G = 150 in the ignorable regime, where SACR-B rejects 9.3% of the time.
Table S.1. Pointwise sensitivity to within-cluster dependence
Relative RMSE is SACR RMSE divided by conventional RMSE; values above one indicate an efficiency cost. Rejection frequencies concern nominal 5% pointwise tests. Both SACR-B and CR-B use bootstrap sample standard deviations. ρ \rho ρ is the Gaussian latent error correlation; DR errors use the existing logistic transformation. Each cell uses 1,000 Monte Carlo replications and 499 bootstrap draws.
Matched exponential-weight bootstrap
Table S.2 holds the bootstrap scheme and scale construction common across the two estimators: each replication uses the same Exp(1) cluster draws, and both procedures use bootstrap sample standard deviations. Under β = 1 \beta= 1 β = 1 at G = 600 G = 600 G = 600 , conventional rejection rates are 43.5%, 75.1%, and 60.6% for DR, QR, and ER, compared with 5.0%, 6.5%, and 5.8% for SACR. SACR also has shorter average intervals in these cells. Thus its ULC advantage persists when the bootstrap draws and scale construction are aligned. In the ignorable regime, SACR rejection frequencies are close to nominal across models and cluster counts. This diagnostic supplements the pairs-bootstrap benchmark and does not assert conventional exponential-bootstrap validity under ULC.
Table S.2. Pointwise inference with matched exponential cluster weights
SACR-EXP-SD and CR-EXP-SD use identical exponential cluster draws within each Monte Carlo replication and bootstrap sample standard deviations. Relative length is SACR average interval length divided by conventional average interval length; values below one indicate shorter SACR intervals. The procedures use inverse-cluster-size and individual weighting, respectively. The latent correlation is 0.96, with 1,000 replications, 499 bootstrap draws, and nominal 95% normal intervals. All cluster counts and both Pareto regimes are reported; full diagnostics remain in the replication outputs.
Single-dominant-cluster experiments
The single-dominant-cluster experiments retain the DGPs and inference settings of § 5.2 and 5.3, replacing only the cluster-size array. The Pareto β = 1 \beta= 1 β = 1 regime violates the negligibility condition for conventional N \sqrt{N} N -inference, but its largest cluster’s sample share still vanishes, leaving a route to consistency. The share in this case, on the contrary, stays bounded away from zero, removing this negligibility route for the conventional estimator as well.
Let s d o m = 0.3 s_{\mathrm{dom}} = 0.3 s dom = 0.3 denote the target sample share of the dominant cluster. Define
N = round ( G − 1 1 − s d o m ) , N G = N − ( G − 1 ) , N g = 1 , g = 1 , … , G − 1. N = \operatorname{round}\left(\frac{G - 1}{1 - s_{\mathrm{dom}}}\right), \qquad N_G = N - (G - 1), \qquad N_g = 1,\quad g = 1,\ldots,G - 1. N = round ( 1 − s dom G − 1 ) , N G = N − ( G − 1 ) , N g = 1 , g = 1 , … , G − 1.
The dominant cluster is placed at g = G g = G g = G . It is therefore treated under the common assignment rule T g = 1 { g ≡ 0 ( m o d 5 ) } T_g = \mathbb{1}\{g \equiv0 \pmod{5}\} T g = 1 { g ≡ 0 ( mod 5 )} , matching the treatment status of the largest cluster in the ordered Pareto designs. Integer rounding makes the realised dominant share approximately, rather than exactly, 0.3. Because N G / N → s d o m N_G/N \to s_{\mathrm{dom}} N G / N → s dom , the dominant cluster remains asymptotically non-negligible and max g N g 2 / N \max_g N_g^2/N max g N g 2 / N diverges at rate G G G . Under SACR, however, its cluster-average score still receives weight 1 / G 1/G 1/ G , and its contribution to the cluster-balanced covariance kernel is therefore O ( G − 1 ) O(G^{-1}) O ( G − 1 ) . The stable proportions of treated and control clusters among the remaining clusters preserve the rank and covariance limits required by the SACR theory.
Pointwise coefficient inference
The pointwise single-dominant-cluster experiment described in Section S.5.5 follows the targets, procedures, and performance measures in § 5.2 .
Table S.3. Pointwise Monte Carlo results under the single-dominant-cluster design (s d o m = 0.3 s_{\mathrm{dom}} = 0.3 s dom = 0.3 )
Targets, cluster counts, and inference procedures are as in Table 1 ; only the cluster-size array differs. Bias and RMSE are shared within each estimator’s A/B pair.
Table 1. Pointwise Monte Carlo results: asymptotically ignorable Pareto regime (β = 2.5 \beta= 2.5 β = 2.5 ). The three column blocks report DR at y = Q Y ( 0.5 ) y = Q_Y(0.5) y = Q Y ( 0.5 ) , QR at τ = 0.5 \tau= 0.5 τ = 0.5 (LAD), and ER at τ = 0.5 \tau= 0.5 τ = 0.5 (OLS), for G = 150 , 300 , 600 G = 150, 300, 600 G = 150 , 300 , 600 . SACR-A and CR-A use analytical standard errors; SACR-B and CR-B use bootstrap standard deviations. Bias and RMSE are shared within each estimator’s A/B pair because the two rows differ only in their inference procedure.
The single-dominant-cluster results sharpen the contrast with the Pareto designs. Pointwise SACR RMSE declines as G G G rises. At G = 600 G = 600 G = 600 , SACR-B rejection rates are 4.1%, 5.5%, and 5.1% for DR, QR, and ER, all close to nominal. Conventional RMSE remains large: CR-A rejects 63.0%, 96.5%, and 85.5%. The pairs bootstrap reduces these rates but still rejects 9.9%, 47.6%, and 28.1%, respectively.
Uniform inference on distributional treatment effects
The functional single-dominant-cluster experiment described in Section S.5.5 follows the targets, grids, procedures, and performance measures in § 5.3 .
For the treatment-effect paths, SACR-B integrated RMSE also declines with G G G . At G = 600 G = 600 G = 600 , its uniform rejection frequencies are 7.1%, 3.4%, and 7.7% for DF, QF, and EF, compared with 20.6%, 7.8%, and 12.9% for CR-B. The conventional QF rejection rate is reasonably close to nominal, but this alone does not indicate comparable performance, because its average band half-width is 5.2497, against 0.9085 for SACR-B, while its integrated RMSE is 1.0453, against 0.1519. Thus conventional bootstrap bands can be extremely wide without resolving the substantial estimation error induced by the dominant treated cluster.
Table S.4. Uniform Monte Carlo results for distributional treatment effects under the single-dominant-cluster design (s d o m = 0.3 s_{\mathrm{dom}} = 0.3 s dom = 0.3 ).
Targets, supports, grids, and bootstrap procedures are as in Table 3 ; only the cluster-size array differs.
Empirical Illustration
The illustration revisits Akhtari, Moreira, and Trucco [1 ] ’s study of political turnover and student achievement in Brazilian municipal schools, replicating the fourth- and eighth-grade specifications in Table 3 , Panels A and B, Columns 3–6. Observations are clustered by municipality, with substantially unequal cluster sizes. Pointwise OLS and LAD comparisons maintain common-coefficient RD specifications; distributional effects use municipality- and student-weighted cut-off distributions.10
Empirical design, identification, and cluster-size diagnostics
The replicated specification is
score g i , t + 1 = κ + θ 1 1 { IVM g t < 0 } + θ 2 IVM g t + θ 3 1 { IVM g t < 0 } × IVM g t + θ 4 score g i , t b a s e + C g i , t ′ Γ + U g i , t , \operatorname{score}_{gi,t+1} = \kappa+ \theta_1 \mathbf{1}\{\operatorname{IVM}_{gt} < 0\} + \theta_2 \operatorname{IVM}_{gt} + \theta_3 \mathbf{1}\{\operatorname{IVM}_{gt} < 0\} \times\operatorname{IVM}_{gt}
+ \theta_4 \operatorname{score}^{\mathrm{base}}_{gi,t} + C'_{gi,t}\Gamma+ U_{gi,t}, score g i , t + 1 = κ + θ 1 1 { IVM g t < 0 } + θ 2 IVM g t + θ 3 1 { IVM g t < 0 } × IVM g t + θ 4 score g i , t base + C g i , t ′ Γ + U g i , t ,
where g g g , t t t , i i i index municipalities, election cycles, and students. The outcome is the post-election test score; score g i , t b a s e \operatorname{score}^{\mathrm{base}}_{gi,t} score g i , t base is the baseline school mean, and C g i , t C_{gi,t} C g i , t contains the remaining controls. The incumbent party’s vote margin IVM g t \operatorname{IVM}_{gt} IVM g t is the running variable, with negative margins indicating turnover. Each municipality cluster contains all included student–cycle observations, permitting dependence across students and election cycles. Columns 3–4 use the fixed window 0.07, and Columns 5–6 use 0.11; the second specification in each pair adds full school and student controls and an election-cycle indicator. These windows provide a limited bandwidth-sensitivity comparison. The coefficient θ 1 \theta_1 θ 1 represents the common local effect of turnover.
The identifying logic is the sharp regression-discontinuity (RD) comparison at the zero vote-margin cut-off. Let R R R denote the incumbent party’s vote margin, let D = 1 { R < 0 } D = 1\{R < 0\} D = 1 { R < 0 } denote political turnover, and let X = ( X b a s e , C ′ ) ′ X = (X_{\mathrm{base}}, C')' X = ( X base , C ′ ) ′ collect the predetermined baseline score, the remaining school and student characteristics, and the election-cycle indicator. Continuity at R = 0 R = 0 R = 0 of the potential-outcome distributions and the distribution of X X X , together with consistency and the absence of precise manipulation of the cut-off, identifies the two side-specific potential-outcome distributions at the cut-off. In the maintained additive specification, the conditional mean discontinuity is θ 1 \theta_1 θ 1 .
For bandwidth h h h , let V g t i ( h ) V_{gti}(h) V g t i ( h ) denote the regressor vector in the specification above. The maintained implication required by the Z-SACR theory is the municipality-level score restriction
E G [ 1 N g ∑ ( t , i ) ∈ G g V g t i ( h ) U g i , t ] = 0 ( g = 1 , … , G ) . \mathbb{E}_{G}\left[\frac{1}{N_g}\sum_{(t,i)\in\mathcal{G}_g} V_{gti}(h)U_{gi,t}\right] = 0 \qquad(g = 1,\ldots,G). E G N g 1 ( t , i ) ∈ G g ∑ V g t i ( h ) U g i , t = 0 ( g = 1 , … , G ) .
The RD design identifies the cut-off comparison, while the exact score restriction is the modelling condition that represents it through a common coefficient. Provided the corresponding cluster-balanced second-moment matrix is nonsingular, deterministic cluster-size adjustment preserves this coefficient.
At τ = 0.5 \tau= 0.5 τ = 0.5 , ER coincides with OLS and QR with LAD. For LAD, the maintained additive local-linear conditional-median specification and continuity of potential-outcome conditional medians at the cut-off identify θ 1 \theta_1 θ 1 as a conditional-median RD contrast common across covariates, distinct from the marginal Quantile Treatment Effect (QTE) at the median obtained by inverting cut-off distributions.
The conventional OLS coefficients and observation and municipality counts reproduce the original specifications to the reported precision. Pointwise inference follows the pairs and exponential-weight cluster-bootstrap procedures in § 5.2 , using 499 draws and bootstrap SD standard errors rather than the original study’s variance calculation.
Across specifications, max g N g 2 / N \max_g N_g^2/N max g N g 2 / N ranges from 10.4 to 39.7, indicating substantial cluster-size concentration, while the largest municipality shares range from 0.80% to 1.16%. Tables S.5 and S.6 report the squared-size diagnostics by specification.
Pointwise coefficient evidence
Table S.5. OLS estimates across fixed-bandwidth specifications
Notes: The table reports ER estimates at τ = 0.5 \tau= 0.5 τ = 0.5 , which coincide with OLS estimates under the linear conditional mean specification.
Table S.6. LAD estimates across fixed-bandwidth specifications
Notes: The table reports QR estimates at τ = 0.5 \tau= 0.5 τ = 0.5 , that is, LAD estimates under the maintained local-linear conditional-median specification.
The pointwise results in Tables S.5 and S.6 show that cluster-size adjustment can change both magnitudes and inferential conclusions under a potentially ULC setting where the conventional observation-weighted inference may be misleading. For fourth-grade scores, Z-SACR produces more negative LAD and OLS estimates than the corresponding standard estimates across all fixed-bandwidth specifications. For eighth-grade scores, the adjustment moves the estimates in the opposite direction: the negative effects are attenuated, and none of the Z-SACR estimates is significant at 5%; conventional estimates remain significant in the wider-bandwidth specifications.
Following the distributional RD approach of Frandsen, Frölich, and Melly [6 ] , effects are constructed from side-specific conditional cut-off CDFs standardised over a common covariate distribution.
The distributional analysis uses the fully specified fourth-grade Column 6 sample with the original study’s uniform bandwidth h = 0.11 h = 0.11 h = 0.11 . Write Y ( d ) Y(d) Y ( d ) for the potential score under political-turnover state d d d , with covariates X X X defined above. The causal standardisation below requires these covariates to be determined before turnover or unaffected by it. For each threshold y y y and covariate value x x x , define the two conditional cut-off distributions separately under weighting scheme m ∈ { S , C } m \in\{S,C\} m ∈ { S , C } . Since political turnover is defined by D = 1 { R < 0 } D = 1\{R < 0\} D = 1 { R < 0 } , the treated distribution uses the left-hand limit r ↑ 0 r \uparrow0 r ↑ 0 , and the control distribution uses the right-hand limit r ↓ 0 r \downarrow0 r ↓ 0 .
F 1 0 , m ( y ∣ x ) = lim r ↑ 0 P m [ Y ≤ y ∣ R = r , X = x ] , F 0 0 , m ( y ∣ x ) = lim r ↓ 0 P m [ Y ≤ y ∣ R = r , X = x ] . F_1^{0,m}(y \mid x) = \lim_{r \uparrow0} \mathbb{P}_m[Y \le y \mid R = r, X = x], \qquad F_0^{0,m}(y \mid x) = \lim_{r \downarrow0} \mathbb{P}_m[Y \le y \mid R = r, X = x]. F 1 0 , m ( y ∣ x ) = r ↑ 0 lim P m [ Y ≤ y ∣ R = r , X = x ] , F 0 0 , m ( y ∣ x ) = r ↓ 0 lim P m [ Y ≤ y ∣ R = r , X = x ] .
Here P S \mathbb{P}_S P S and P C \mathbb{P}_C P C denote municipality- and student-weighted laws. Under the RD continuity conditions, the marginal potential-outcome cut-off CDFs are obtained using the common limiting covariate distribution H 0 m H_0^m H 0 m :
F d 0 , m ( y ) = ∫ F d 0 , m ( y ∣ x ) d H 0 m ( x ) , d ∈ { 0 , 1 } . F_d^{0,m}(y) = \int F_d^{0,m}(y \mid x)\,dH_0^m(x), \qquad d \in\{0,1\}. F d 0 , m ( y ) = ∫ F d 0 , m ( y ∣ x ) d H 0 m ( x ) , d ∈ { 0 , 1 } .
All four local RD effects are functionals of this pair of marginal cut-off distributions: Δ F 0 , m ( y ) = F 1 0 , m ( y ) − F 0 0 , m ( y ) \Delta_F^{0,m}(y) = F_1^{0,m}(y)-F_0^{0,m}(y) Δ F 0 , m ( y ) = F 1 0 , m ( y ) − F 0 0 , m ( y ) , Δ μ 0 , m = E m [ Y ( 1 ) − Y ( 0 ) ∣ R = 0 ] = − ∫ R Δ F 0 , m ( y ) d y \Delta_\mu^{0,m} = \mathbb{E}_m[Y(1)-Y(0) \mid R = 0] = -\int_{\mathbb{R}} \Delta_F^{0,m}(y)\,dy Δ μ 0 , m = E m [ Y ( 1 ) − Y ( 0 ) ∣ R = 0 ] = − ∫ R Δ F 0 , m ( y ) d y , Δ Q 0 , m ( τ ) = ( F 1 0 , m ) − 1 ( τ ) − ( F 0 0 , m ) − 1 ( τ ) \Delta_Q^{0,m}(\tau) = (F_1^{0,m})^{-1}(\tau)-(F_0^{0,m})^{-1}(\tau) Δ Q 0 , m ( τ ) = ( F 1 0 , m ) − 1 ( τ ) − ( F 0 0 , m ) − 1 ( τ ) , and Δ E 0 , m ( τ ) = e τ ( F 1 0 , m ) − e τ ( F 0 0 , m ) \Delta_E^{0,m}(\tau) = e_\tau(F_1^{0,m})-e_\tau(F_0^{0,m}) Δ E 0 , m ( τ ) = e τ ( F 1 0 , m ) − e τ ( F 0 0 , m ) , where e τ ( F ) e_\tau(F) e τ ( F ) denotes the τ \tau τ -expectile of F F F . With finite first moments, Δ E 0 , m ( 0.5 ) = Δ μ 0 , m \Delta_E^{0,m}(0.5)=\Delta_\mu^{0,m} Δ E 0 , m ( 0.5 ) = Δ μ 0 , m . QTE and the Expectile Treatment Effect (ETE) are functionals of the marginal cut-off distributions, rather than averages of conditional effects or direct QR and ER fits. Their map conditions for weak convergence and bootstrap validity are discussed in Remark 3.4 and Proposition S.4.1.
Within ∣ R ∣ ≤ 0.11 |R| \le0.11 ∣ R ∣ ≤ 0.11 , separate DR fits on each side include a linear vote-margin term and controls X X X . Predictions at R = 0 R=0 R = 0 are averaged over the same pooled within-bandwidth covariate sample. This approximates H 0 m H_0^m H 0 m as the bandwidth shrinks; at the fixed bandwidth, it remains a finite-sample standardisation. Standard CR uses observation weights; Z-SACR uses inverse municipality-size weights in both fitting and covariate averaging.11 The fitted marginal cut-off distributions are monotonically rearranged before inversion or integration [3 ] .
Each of the 499 municipality-level bootstrap draws repeats the side-specific DR fits, pooled-covariate standardisation, rearrangement, and distribution-to-functional transformations. Pointwise intervals use bootstrap SDs; uniform bands and uniform homogeneity tests use normalised bootstrap IQRs. Section S.6.4 describes the grids and integration. These distribution-derived results illustrate implementation; their full-pipeline asymptotic validity is not established by the direct score-process results here.12
Figure S.1. Local RD DTE path at the vote-margin cut-off: Z-SACR (left) and Standard CR (right)
Figure S.2. Local RD QTE path at the vote-margin cut-off: Z-SACR (left) and Standard CR (right)
Figure S.3. Local RD ETE path at the vote-margin cut-off: Z-SACR (left) and Standard CR (right)
Table S.7. Local RD treatment-effect summaries at the vote-margin cut-off
Notes: All effects are obtained from the two separately fitted cut-off CDFs; inference uses 499 municipality-level bootstrap draws. Zero path excluded means that the zero function is outside the simultaneous band. The homogeneity column gives the bootstrap p p p -value for a constant treatment-effect path. Positive DTE values mean a higher probability of scoring at or below the stated threshold under political turnover.
Figures S.1 to S.3 and Table S.7 report the distributional results. Both DTE paths are positive throughout the central grid, but neither simultaneous band is positive throughout, so the bands do not establish first-order stochastic dominance. Both QTE and ETE paths are negative. None of the homogeneity tests rejects a constant path at 5%, although the Z-SACR DTE test has p = 0.0942 p = 0.0942 p = 0.0942 .
Integrating the fitted cut-off DTE gives a DR-implied mean effect of − 0.0672 -0.0672 − 0.0672 for Z-SACR, with a pointwise 95% bootstrap confidence interval of [− 0.1131 -0.1131 − 0.1131 , − 0.0214 -0.0214 − 0.0214 ]. The corresponding conventional effect is − 0.0568 -0.0568 − 0.0568 , with an interval of [− 0.0991 -0.0991 − 0.0991 , − 0.0145 -0.0145 − 0.0145 ]. By construction, these integrals equal the fitted expectile differences at τ = 0.5 \tau= 0.5 τ = 0.5 to numerical precision, providing an internal check on the common cut-off-distribution construction.
Distributional grid construction
The grid construction fits a normal distribution and two- and three-component Gaussian mixtures to the outcome data and computes their implied curvature envelopes, i.e., the absolute diagonal element of the Hessian matrix of the corresponding function, on a dense grid. For each target, it uses the largest resulting grid size. The grid rule described in Section S.5.2 gives target-specific sizes of 29, 92, and 87 for DTE, QTE, and ETE; their maximum, 92, is used for all three reported paths. The reported evaluation ranges cover the central 0.1–0.9 region. A separate internal DR grid of 92 points spans the empirical 0.1st–99.9th percentiles, on which each cut-off CDF after rearrangement is linearly interpolated for quantile inversion and expectile integration. Remaining lower- and upper-tail masses are assigned to the corresponding endpoints as a maintained tail specification.
Cluster size has two distinct dimensions. Conditional noninformativeness means that the conditional outcome distribution, or the particular attribute of interest, does not vary with cluster size (given covariates) [9 , 13 ] . Size negligibility concerns asymptotic influence: under unrestricted within-cluster dependence, max g N g 2 / N → 0 \max_g N_g^2/N \to0 max g N g 2 / N → 0 is a benchmark for conventional Gaussian inference. Its failure defines the ULC regime considered in this paper, but does not imply failure of Gaussian inference under every dependence structure. The following classification separates these questions.
Table 5.
Conditional noninformativeness does not guarantee invariance of marginal targets, since cluster size may remain associated with covariates. Conversely, distributional informativeness does not rule out a common zero for a particular score. As Proposition 2.1 establishes, a common cluster-level score zero preserves the structural zero under deterministic reweighting; identification additionally requires separation, and inference requires the remaining assumptions. SACR limits size-driven influence by using cluster-average scores, with its inferential conclusions supplied by § 3 . Without a common zero, equal cluster representation can define a different meaningful target, but the present inference results cannot automatically be invoked for it. An aggregate zero alone does not ensure consistent estimation of the sampling covariance, although the coincident analytical and multiplier-bootstrap score covariance estimators can support conservative pointwise inference for the weighting-specific target; see Remark 3.3. There is no general efficiency ranking between observation-weighted and cluster-balanced estimators.