betaFeedback

WorksMathematics

Causal Record Linkage

Abstract

Combining observations from multiple data sources creates a valuable resource for research. In causal inference, record linkage methods grant researchers access to the covariates needed for unbiased estimation and offer the means to explore (long-term) outcomes. This paper examines how record linkage conflicts with the assumptions required for identifying causal effects. Our investigation reveals that linkage errors primarily threaten consistency in the causal framework, leading to attenuation bias in the inference due to discrepancies in the linked data. In attempting to address this bias by being more stringent on the linkage, sampling bias inadvertently emerges from a covariate shift between the baseline and the linked samples, as profiles with more distinctive information are more likely to be recovered. Despite these challenges, we identify settings in which record linkage and causal inference can be soundly combined, and show how the effect estimated on linked data can be generalised, with linkage decisions informed accordingly. We further identify when existing and novel solutions support valid causal inference on linked data, and when inference should instead be treated with caution or even abandoned. We illustrate the challenges raised and the potential solutions using real data from a study of women's health and a simulation study.

Introduction

The extent to which reliable causal inference can be achieved on linked data is ambiguous, as linkage processes can compromise the validity of common causal assumptions. In causal inference, baseline information and exposure may come from a prior study while outcomes are collected later. A lack of unique identifiers is a major obstacle towards combining individual records across heterogeneous data sets, and record linkage methods address this by connecting information on the same individuals using partially identifying variables (e.g. birth year, postal code, ethnicity), thereby offering access to broader sets of variables across wider populations and extended time periods. When integrating data through record linkage, it is important to assess the implications on the causal framework and to establish guidelines for selecting records to ensure reliable causal estimates.

In this paper, we consider the causal record linkage task as integrating two data sources to construct a study population defined by their intersection, in order to perform causal inference. Record linkage relies on pre-treatment linking variables ZZ, observed in both sources, to recover complete profiles by linking pre-treatment covariates XX and binary treatment TT from one source with outcome YY from another, inferred to correspond to the same entities. Statistical inference is then conducted on the conditional distribution of Y∣T=t,XY \mid T = t, X for t∈{0,1}t \in\{0, 1\}, for the imperfect selected set of complete profiles built through record linkage, either jointly [81, 40, 39] or via a two-stage approach [94, 93, 59].

There are two main paradigms in record linkage: the Fellegi-Sunter model and the hierarchical entity resolution framework [31, 86, 61] (we illustrate both along this article). All methods essentially output a bimodal distribution of linkage scores, expressing the amount of evidence that two observations belong to the same individual, taking values in [0,1][0, 1] [50, 89, 28, 27, 75]. The modes distinguish obviously distinct profiles (low scores) from similar ones (high scores). A cutoff λ∈[0.5,1)\lambda\in[0.5, 1) is applied to these scores as a decision rule, classifying pairs to yield a valid set of linked records [86, 79]. The choice of this cutoff depends on the goal of the analysis, ideally minimising the false discoveries or tailored to the inference objective [94].

Linking variables ZZ are often demographics [47, 92, 28, 69] and may therefore be causally relevant (e.g. sex, age, socioeconomic status, geographic region). For that reason, linking variables may be involved in the inference, and thus in the treatment effect heterogeneity. We deviate from the literature [94, 40] that ignored the overlap between ZZ and XX, and hence did not investigate the issues induced by record linkage on causal effect estimation [26, 56]. Our framework extends the literature on the implications of record linkage on statistical inference [55, 87, 26, 90] to causal inference. The conditions under which causal inference can be credibly performed on linked data in such a realistic setting have not been examined, yet as we shall demonstrate in this paper, covariate dependent linkage errors interfere with the causal framework and can introduce noteworthy biases in the causal effect estimation.

Causal inference typically rests on three core assumptions [67, 49, 45, 58]: consistency between observed and potential outcomes, positive probability of exposure for all profiles and conditional exchangeability between treatment groups. We explain how record linkage errors that depend on the covariates affect these assumptions. Falsely linked records corrupt consistency, threatening the internal validity of the linked sample. Besides, record linkage preferentially selects distinctive profiles, which shifts the covariate distribution away from the original population, limiting the external validity of causal conclusions. We quantify the resulting biases through a decomposition into attenuation and sampling bias, and establish conditions under which the effect in the original population is identifiable and estimable from the linked sample.

Motivated by the peculiar linkage sample selection mechanism, the main methodological contributions of this paper are to formalise the task of causal inference on linked data in Section 2, to discuss the qualitative implications on effect identifiability in Section 3 and to address the transportability of findings despite the selection. We quantify the linkage induced biases in Section 4 and propose solutions for estimating treatment effects with linked data in Section 5. In Section 6 we provide diagnostic tools for causal record linkage reliability. We illustrate this work with real data from an American study of women’s health through Section 7, and with a simulation study in Section 8. The results presented in this paper are independent of the record linkage method used.

Causal Record Linkage and Sample Selection Bias

The growing availability of observational data has driven advances in causal inference, yet data access remains restricted [65]. Record linkage combines pseudonymised data across sources, enabling the study of novel causal questions. In this section we present the record linkage setting alongside the causal inference problem at stake, before examining the impact of the linkage decisions on the causal roadmap.

Record linkage considers two overlapping samples from the same source population P\mathcal{P}, aiming to construct a study population defined by their intersection: the set of (true) links i.e. pairs of records belonging to the same individuals. The baseline data set A\mathcal{A} gathers baseline characteristics CC and binary treatment TT, while the follow-up data set B\mathcal{B} gathers outcome YY.

Assumption 1. A\mathcal{A} and B\mathcal{B} are overlapping random samples from P\mathcal{P}.

Implicitly, P\mathcal{P} must not be too large relative to A\mathcal{A} and B\mathcal{B}, ensuring a sufficiently likely overlap between samples. The intersection of data sources can then be regarded as a random sample from the collected data [86, 61, 59, 75], and any unit in A\mathcal{A} is expected to be followed up in B\mathcal{B}. Without unique identifiers, observations are linked based on pre-treatment variables ZZ, common to both sources and possibly overlapping with CC.

Assumption 2. The key assumptions for causal identification hold in the source population: consistency, positivity, conditional exchangeability. [49, 45]

We expand on these in § 3. Consequently, all units in A\mathcal{A} are eligible for treatment; we denote the propensity score as a function of the covariates: π(X)=P(T=1∣X)\pi(X) = \mathbb{P}(T = 1 \mid X). Throughout the paper, index ii refers to information from A\mathcal{A}, jj to information from B\mathcal{B}, and NN indexed by a set denotes its size. The observed data are given by:

A={(Ti,Zi,Ci)∣i=1,…,NA},B={(Zj,Yj)∣j=1,…,NB}.\mathcal{A} = \left\{(T_i, Z_i, C_i) \mid i = 1,\ldots,N_{\mathcal{A}}\right\}, \qquad\mathcal{B} = \left\{(Z_j, Y_j) \mid j = 1,\ldots,N_{\mathcal{B}}\right\}.

Sample selection

Notwithstanding the opportunity to address novel causal questions with linked data, record linkage restricts the sample to a selected population whose covariate distribution likely differs from that of the source population [44, 26, 70]. This selection stems from imperfect identifying information, resulting in missed links between entities present in both sources and falsely linked records from distinct entities. Being more stringent by raising the cutoff λ\lambda reduces the latter but increases the former. In practice, the true links may be difficult to distinguish, as indicated by the superposition of linkage scores in Figure 1, obtained with two state-of-the-art algorithms on the application of § 7. Links recovered with high confidence tend to involve individuals with particularly distinctive identifying information, yielding a sample biased towards atypical profiles: record linkage acts as a selection mechanism.

Stacked linkage scores for all records pairs of the SWAN data obtained from BRL and FlexRL

Figure 1. Stacked linkage scores for all records pairs of the SWAN data, obtained from two different methods, BRL [79] (left) and FlexRL [75] (right). Scores (in pale) express the amount of evidence that observations belong to the same entities. To limit contamination from falsely linked data, we select pairs with scores beyond λ∈[0.5,1)\lambda\in[0.5, 1), hoping to mostly detect true links (with scores in dark). Lower λ\lambda yields more falsely linked and higher λ\lambda more missed links.

Related works have considered record linkage for causal inference when identifying information is independent of causally relevant variables, assuming that linkage does not interfere with the causal framework, or restricting inference to homogeneous treatment effects [94, 40]. Since the linking variables may overlap with those driving treatment assignment and outcome, we allow them to be causally relevant and consider XX as the union of ZZ and CC, extending the literature towards effect heterogeneity and exploring how linkage errors interact with the causal framework.

More broadly, this connects to the data fusion literature, where transportability across domains may introduce biases [76, 85, 18, 11, 22, 19]. The structural causal framework [41, 66] formalises such population differences via a selection variable SS, reflecting implicit or explicit conditioning on study enrolment, affecting the distribution of variables of interest [10, 12, 68, 11, 19] We borrow this machinery and introduce a graphical causal model on Section 2.2 to understand how common causal assumptions interact with record linkage in Section 3.

Graphical causal model

Let SS denote the linkage estimator, with Si,j=1S_{i,j}=1 meaning that pair (i,j)(i,j) has a linkage score above λ\lambda and is declared ‘linked’. Inclusion in the analysis depends on linkage, which occurs differentially with respect to XX, making SS a descendant of XX on Figure 2. In other words, linkage errors depend on covariates (atypical profiles are less likely to be missed or falsely linked). Yet, because linkage is not informative about the variables of interest in the analysis, SS has no descendants.

Causal graph with selection and causal graph for linked pairs

Figure 2. Selection diagrams depicting differences between source population contained in the data and linked population obtained with record linkage (indicated by SS). The selection process is made on the linking variables ZZ included in XX, hence SS descends from XX.

We represent factual and counterfactual (observed and potential) outcome distributions using single world intervention graphs (SWIGs) [72]. For unit i∈Ai \in\mathcal{A}, the factual outcome is Yi=fY(Xi,Ti,εYi)Y_i=f_Y(X_i,T_i,\varepsilon_{Y_i}) (analogously for j∈Bj \in\mathcal{B}), and for every unit in P\mathcal{P}, the counterfactual outcome is Y(x,t)=fY(x,t,εY)Y(x,t)=f_Y(x,t,\varepsilon_Y) for any x,t∈{0,1}x,t \in\{0,1\}, sometimes denoted Y(t)Y(t), although substituting YjY_j for Yi(t)Y_i(t) is warranted only when records ii and jj coincide on both treatment and covariates.

Figure 2a depicts the causal structure for the effect of TT on YY confounded by XX, valid for i.i.d. units in P\mathcal{P} and for true links (subscripts omitted). Since linkage and causal mechanisms are intertwined, causal pathways must be considered at the level of linked pairs SS; Figure 2b unravels this representation to make relationships explicit across both data sources, where units become pairs of linked entities forming an i.i.d. sample of selection biased data from P\mathcal{P}.

Independencies can be tested via d-separation (directional flow of information): a path is blocked if it contains a collider that is not conditioned on (nor on any of its descendants), or a non-collider that is conditioned on; two variables are d-separated (by a set), and hence (conditionally) independent, if all paths between them are blocked [66, 83, 67, 64, 32]

On Figure 2b, the independence between pairs breaks once we condition on Si,jS_{i,j}. Indeed, if records ii and jj refer to different individuals, they are independent observations: Xi⊥Yj(t)X_i \perp Y_j(t) for t∈{0,1}t \in\{0,1\}, since the only path between them, {Xi→Si,j←Xj→Yj(t)}\{X_i \rightarrow S_{i,j} \leftarrow X_j \rightarrow Y_j(t)\}, is blocked by the unconditioned collider Si,jS_{i,j}. Once linked however, spurious dependencies emerge: Xi̸ ⁣⊥Yj(t)∣Si,j=1X_i \not\!\perp Y_j(t) \mid S_{i,j}=1. By involving covariates in the linkage process, the latter becomes prone to confirmation bias, as estimation indirectly favours pairs whose covariates align with the relationship between XX and YY under study. Understanding these implications is important to appropriately conduct inference on the potentially imperfect selected set of linked data:

S={(Xi,Ti,Yj)∣Si,j=1,i=1,…,NA,j=1,…,NB}.\mathcal{S} = \{(X_i,T_i,Y_j) \mid S_{i,j}=1, i=1,\ldots,N_A, j=1,\ldots,N_B\}.

Estimand

An estimand defines the intended target of inference, but feasible estimands are constrained by the study design, estimation procedure, data and, critically here, by the selection mechanism induced by record linkage. We characterise the target of record linkage, and its locus in our framework, to establish what the causal estimand can feasibly aim at, given the sample selection and causal model developed above.

Let Δ\Delta denote the true link indicator, target of record linkage, with Δi,j=1\Delta_{i,j}=1 if records ii and jj correspond to the same individual. Among NS=∑i,jSi,jN_S=\sum_{i,j}S_{i,j} linked entities selected for causal inference, true positives TP=NSΔ=∑i,jSi,jΔi,j\mathrm{TP}=N_S\Delta=\sum_{i,j}S_{i,j}\Delta_{i,j} are formed of records referring to the same individual, satisfying Ti=TjT_i=T_j and Xi=XjX_i=X_j, while false positives FP=NS∇=∑i,jSi,j(1−Δi,j)\mathrm{FP}=N_S\nabla=\sum_{i,j}S_{i,j}(1-\Delta_{i,j}) arise from records with similar profiles pertaining to distinct individuals. The remaining true links are missed as false negatives, FN=∑i,j(1−Si,j)Δi,j\mathrm{FN}=\sum_{i,j}(1-S_{i,j})\Delta_{i,j}. These yield linkage quality metrics: False Discovery Proportion (FDP) FPFP+TP=NS∇NS\frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TP}}=\frac{N_S\nabla}{N_S} and sensitivity TPFN+TP\frac{\mathrm{TP}}{\mathrm{FN}+\mathrm{TP}}.

Under Assumption 1, the distributions of variables of interest are identical across true links (Δi,j=1\Delta_{i,j}=1) and non-links (Δi,j=0\Delta_{i,j}=0), which translates into (X,T,Y)⊥ ⁣ ⁣ ⁣⊥Δ(X,T,Y)\perp\!\!\!\perp\Delta. More generally, the latent Δ\Delta plays no role in generating dependence among (X,T,Y,S)(X,T,Y,S), it is therefore omitted from our causal graphical model.

Assumption 3. Δ\Delta is disconnected from the system: Δ⊥ ⁣ ⁣ ⁣⊥V∣W\Delta\perp\!\!\!\perp V\mid W for any V,W⊆{X,T,Y,S}V,W\subseteq\{X,T,Y,S\}.

Under Assumptions 1 and 3, TP and FP share the same distribution of variables of interest, (X,T,Y)⊥ ⁣ ⁣ ⁣⊥Δ∣S(X,T,Y)\perp\!\!\!\perp\Delta\mid S. These independencies imply X⊥ ⁣ ⁣ ⁣⊥ΔX\perp\!\!\!\perp\Delta and X⊥ ⁣ ⁣ ⁣⊥Δ∣SX\perp\!\!\!\perp\Delta\mid S, yet X̸ ⁣⊥ ⁣ ⁣ ⁣⊥SX\not\!\perp\!\!\!\perp S, with implications for the causal estimand.

Remark 1. Under Assumption 1, XiX_i and XjX_j share the same marginal distribution and, registration processes of A\mathcal{A} and B\mathcal{B} are comparable (same patterns of mistakes and missing values), implying for any xx that P(Si,j=1∣Xi=x)=P(Si,j=1∣Xj=x)\mathbb{P}(S_{i,j}=1\mid X_i=x)=\mathbb{P}(S_{i,j}=1\mid X_j=x), meaning in this context that Xi∣Si,j∼Xj∣Si,jX_i\mid S_{i,j}\sim X_j\mid S_{i,j}.

Linkage errors (FP and FN) bias inference [55, 87, 26, 90] and threaten causal validity as Section 3 will show. Falsely linked records undermine the internal validity of the linked data which precludes estimands targeting a specific sample (e.g. individual treatment effect or average treatment effect among true links) and confines inference to estimands targeting the source population. Yet this population is only partially observed through the linked data, which raises concerns about their external validity.

Throughout this article, the potential outcome mean and conditional mean under intervention T=tT=t are denoted μt=E[Y(t)]\mu_t=\mathbb{E}[Y(t)] and μt(X)=E[Y(t)∣X]\mu_t(X)=\mathbb{E}[Y(t)|X] respectively. In contrast, the observed outcome under treatment t∈{0,1}t\in\{0,1\} is denoted Y∣T=tY|T=t. We focus on the Average Treatment Effect (ATE) in the baseline population:

τA=EA[Y(1)−Y(0)]=EA[E[Y(1)−Y(0)∣X]]=μ1A−μ0A=E[μ1A(X)−μ0A(X)].(1)\tau_A=\mathbb{E}_A[Y(1)-Y(0)]=\mathbb{E}_A[\mathbb{E}[Y(1)-Y(0)|X]]=\mu_{1A}-\mu_{0A}=\mathbb{E}[\mu_{1A}(X)-\mu_{0A}(X)]. \tag*{(1)}

The emphasis on the covariate domain in the expectations echoes some considerations regarding potential outcomes to be expanded in Section 3.1. We will now relate the target τA\tau_A to the observed data through the sample ATE among linked units denoted τS\tau_S:

τS=1NS∑ℓ=1NSYℓ(1)−Yℓ(0).(2)\tau_S=\frac{1}{N_S}\sum_{\ell=1}^{N_S}Y_\ell(1)-Y_\ell(0). \tag*{(2)}

Implications for effect identifiability

Record linkage affects causal identifiability through linkage errors and a covariate-dependent selection. Primarily, it interacts with the standard identifiability conditions for internal validity [49, 45, 58]. Falsely linked records corrupt consistency and missed links endanger positivity in finite samples while conditional exchangeability is not compromised. Ultimately, its preferential selection for distinctive profiles imposes to control for the shifted confounders between the baseline population A\mathcal{A} and the linked sample S\mathcal{S} for external validity. These threats, summarised in Table 1, govern the two main components of the bias we develop later in Section 4:

ValidityConditionThreatBiasSolutions
ConsistencyFalsely linked recordsAttenuationFDP vanishes under Assumption 4 or through Z=XZ=X (Remark 2)
InternalPositivityMissed links (finite sample only)ExtrapolationPreferential selection protects atypical profiles under Assumption 5
Conditional exchangeabilityNone
ExternalSS-admissibilityDefinition 1 condition (ii), no access to XjX_jSelectionControl of XjX_j under Assumption 4 or through Remark 2
Covariate shift between A\mathcal{A} and S\mathcal{S}SampleTransport formula under Assumption 5

Table 1. Impact of record linkage on causal identification.

E[τ^S]−τA=E[τ^S]−τS⏟attenuation biasfrom inconsistencies+τS−τA⏟sampling biasfrom covariate shift.\mathbb{E}[\widehat{\tau}_S]-\tau_A=\underbrace{\mathbb{E}[\widehat{\tau}_S]-\tau_S}_{\text{attenuation bias}\\\text{from inconsistencies}}+\underbrace{\tau_S-\tau_A}_{\text{sampling bias}\\\text{from covariate shift}}.

Consistency

Consistency ensures that the observed outcome coincides with the potential outcome under the intervention:

T=t⇒Y=Y(t), thus Y=TY(1)+(1−T)Y(0).T=t \Rightarrow Y=Y(t),\ \text{thus }Y=TY(1)+(1-T)Y(0).

This entails that the treatment assignment mechanism does not affect the potential outcome response [8, 49]. We adopt the weaker distributional consistency [34, 73, 35], requiring only that the observed outcome be equivalent to the potential outcome in distribution. It suffices for the ATE we target and is better suited to linkage, where the observed outcome may belong to a different unit than the one whose potential outcome we target.

The violation of consistency occurs when linked units potential outcomes do not correspond. Indeed, recovering the ATE we would have obtained had we intervened and had we correctly linked the data requires the linked units potential outcomes to share the same distribution. This is plausible if Assumption 2 holds and if Ti=Tj=tT_i=T_j=t, Xi=Xj=xX_i=X_j=x since then Yi=Yi(x,t)∼Yj(x,t)=YjY_i=Y_i(x,t)\sim Y_j(x,t)=Y_j. Thereby, the outcome observed in B\mathcal{B} would reflect the underlying potential outcome from A\mathcal{A}, so distributional consistency, Yj∼Yi(x,t)Y_j\sim Y_i(x,t), would transfer across the linkage. When records belong to distinct individuals who may differ in treatment group and covariates, this correspondence is disrupted since the units likely have different potential outcomes, breaking consistency.

To preview how the violation materialises, consider treatment mismatch on the conditional outcome in the treated. Let TjT_j be the original treatment indicator and TiT_i the treatment indicator after linkage of ii and jj. By Assumption 2 the observed outcome is Yj=Yj(Tj)Y_j=Y_j(T_j), meaning

{Yj=Yj(Ti)∼Yi(Ti)under correct linkage,Yj=Yj(1−Ti)∼Yi(1−Ti)otherwise.\begin{cases} Y_j=Y_j(T_i)\sim Y_i(T_i) & \text{under correct linkage},\\ Y_j=Y_j(1-T_i)\sim Y_i(1-T_i) & \text{otherwise}. \end{cases}

The average conditional outcome in the treated is then:

E[Yj∣Ti=1,X]=E[Yj∣Ti=1,Tj=1,X]⋅P(Ti=Tj∣Ti=1,X)+E[Yj∣Ti=1,Tj=0,X]⋅P(Ti≠Tj∣Ti=1,X)=E[Yj(1)∣X]⋅P(Ti=Tj∣Ti=1,X)+E[Yj(0)∣X]⋅P(Ti≠Tj∣Ti=1,X)=E[Yi(1)∣X]⋅P(Ti=Tj∣Ti=1,X)+E[Yi(0)∣X]⋅P(Ti≠Tj∣Ti=1,X),\begin{aligned} \mathbb{E}[Y_j\mid T_i=1,X] &=\mathbb{E}[Y_j\mid T_i=1,T_j=1,X]\cdot\mathbb{P}(T_i=T_j\mid T_i=1,X)\\ &\quad+\mathbb{E}[Y_j\mid T_i=1,T_j=0,X]\cdot\mathbb{P}(T_i\ne T_j\mid T_i=1,X)\\ &=\mathbb{E}[Y_j(1)\mid X]\cdot\mathbb{P}(T_i=T_j\mid T_i=1,X)+\mathbb{E}[Y_j(0)\mid X]\cdot\mathbb{P}(T_i\ne T_j\mid T_i=1,X)\\ &=\mathbb{E}[Y_i(1)\mid X]\cdot\mathbb{P}(T_i=T_j\mid T_i=1,X)+\mathbb{E}[Y_i(0)\mid X]\cdot\mathbb{P}(T_i\ne T_j\mid T_i=1,X), \end{aligned}

so distributional consistency only holds for the share of linked pairs with Ti=TjT_i=T_j.

Attenuation bias results from the treatment and covariate inconsistencies and takes different forms depending on the estimator used. Treatment mismatch affects all estimators, since observations contribute to inference on the opposite group [94]. Covariate mismatch affects estimators that condition on covariates. For instance, the difference in means targets fY(Xj,t,ϵYj)f_Y(X_j,t,\epsilon_{Y_j}), comparing YjY_j across levels of TiT_i; it is exposed to treatment mismatch. The outcome regression targets fY(x,t,ϵYj)f_Y(x,t,\epsilon_{Y_j}), evaluating μ^Ti=t(Xi=x)\hat{\mu}_{T_i=t}(X_i=x); it is exposed to treatment and covariate mismatch. While inconsistencies threaten causal identification on linked data, their impact can be quantified, as we show in § 4 and further in Appendix Section A.2.2.

Consistency can therefore only be restored through conditions on the linkage, providing a way to derive theoretical guarantees. An ideal case is when false discoveries vanish, achievable by raising the linkage decision threshold λ\lambda, albeit at the cost of more missed links.

Assumption 4. E[FDP]→NS→∞0\mathbb{E}[\mathrm{FDP}] \xrightarrow{N_S\to\infty} 0.

An alternative is to link on all confounders and tolerate residual error, not always feasible in real applications.

Remark 2. When the pre-treatment linking variables ZZ form a valid finite sample adjustment set, X=ZX = Z are observed in A\mathcal{A} and in B\mathcal{B} and linkage performs ‘matching’. Since TT descends from XX, linking on informative covariates raises the probability that Xi,XjX_i,X_j and Ti,TjT_i,T_j coincide, yielding more accurate links, hence supporting distributional consistency.

In many practical settings, it suffices to obtain a causal estimate up to a bounded error, which Section 4 quantifies.

Positivity

Positivity requires that the probability of receiving any value of treatment conditional on covariates is bounded away from 0 and 1, essential for the joint distribution over all treatment values and all covariate profiles to be well defined:

0<P(T=t∣X=x)<1 for all t∈{0,1} and for all x.0 < \mathbb{P}(T = t \mid X = x) < 1 \text{ for all } t \in\{0,1\} \text{ and for all } x.

Positivity transfers from the source population to the linked data under Assumptions 1 and 2. Indeed, the selection mechanism is d-separated from the treatment by the covariates: PS(T=t∣X=x)=PA(T=t∣X=x,S=1)=PA(T=t∣X=x)\mathbb{P}_{S}(T = t \mid X = x) = \mathbb{P}_{A}(T = t \mid X = x, S = 1) = \mathbb{P}_{A}(T = t \mid X = x). Note that identification remains restricted to the covariate and treatment combinations supported in the observed data. For example, if only certain ethnicities participate from baseline to follow-up, or if only younger participants receive treatment and are followed up, positivity is foregone from the outset and biases emerge [26].

In finite samples however, record linkage causes the sample to shrink, hampering the joint conditioning on treatment and confounders in S\mathcal{S} needed to transport the effect towards A\mathcal{A}: PS(T=t,X=x)=PS(T=t∣X=x)PS(X=x)\mathbb{P}_{S}(T = t, X = x) = \mathbb{P}_{S}(T = t \mid X = x)\mathbb{P}_{S}(X = x). The data may suffer random non-positivity: PS(T=t∣X=x)=0\mathbb{P}_{S}(T = t \mid X = x) = 0 for some t∈{0,1}t \in\{0,1\} and some xx. The covariate space depletion may limit the overlap of supports and induce a covariate shift between distributions of S\mathcal{S} and A\mathcal{A}: PS(X=x)≠PA(X=x)\mathbb{P}_{S}(X = x) \ne\mathbb{P}_{A}(X = x) for some xx. For example, our analysis may exclude untreated ‘Hispanic’ women (non-positivity) or ‘White’ women aged 30–35 from a population spanning 25–45 (limited overlap), these profiles being too common to be confidently linked. These empirical limitations primarily stem from missed links and worsen with the linkage decision rule stringency.

These gaps are benign in theory because linkage selection happens preferentially in the frail areas of the covariate space, which protects atypical profiles more at risk of lacking positivity. Methods for generalised treatment effects bridge these gaps by interpolation to transport the causal contrast from S\mathcal{S} to A\mathcal{A} at the cost of untestable modelling assumptions. Under suitable structure on the region of positivity and overlap violations, extrapolation bias can be avoided:

RNS={x∈X; PA(X=x)>0, ∀t∈{0,1} PA(T=t∣X=x)∈(0,1) and min⁡(NS(x),NS0(x),NS1(x))=0}.R_{N_{\mathcal{S}}} = \left\{x \in\mathcal{X};\ \mathbb{P}_{A}(X = x) > 0,\ \forall t \in\{0,1\}\ \mathbb{P}_{A}(T = t \mid X = x) \in(0,1)\text{ and }\min(N_{\mathcal{S}}(x),N_{\mathcal{S}0}(x),N_{\mathcal{S}1}(x)) = 0\right\}.

Assumption 5. RNS⊂int⁡XR_{N_{\mathcal{S}}} \subset\operatorname{int}\mathcal{X} and P(RNS)→NS→∞0\mathbb{P}(R_{N_{\mathcal{S}}}) \xrightarrow[N_{\mathcal{S}}\to\infty]{} 0, where int⁡X\operatorname{int}\mathcal{X} refers to the interior of the covariate space.

That is, violations are finite sample artefacts, internal to the covariate space, supporting transportability and estimation via interpolation [91, 45, 60].

Conditional exchangeability

The conditional distribution of the potential outcomes under treatment t∈{0,1}t \in\{0,1\} is the same among both treatment groups {T=0}\{T = 0\} and {T=1}\{T = 1\} given X=xX = x, [49, 45]:

Y(t)⊥ ⁣ ⁣ ⁣⊥T∣X for t∈{0,1}.Y(t) \mathrel{\perp\!\!\!\perp} T \mid X \text{ for } t \in\{0,1\}.

This implies mean exchangeability over treatment groups, E[Y(t)∣X=x]=E[Y(t)∣T=1,X=x]=E[Y(t)∣T=0,X=x]\mathbb{E}[Y(t)\mid X = x] = \mathbb{E}[Y(t)\mid T = 1, X = x] = \mathbb{E}[Y(t)\mid T = 0, X = x]. In other terms, this condition states that there is no unmeasured confounding.

Linkage does not interfere with conditional exchangeability. Under Assumption 2, the counterfactuals are independent of treatment assignment given the observed covariates in the source population. This condition is directly encoded in Figure 2a since Y(t)Y(t) and TT are d-separated by XX. With record linkage, linked pairs become the exchangeable units at stake, and this conditional independence extends to both correctly and incorrectly linked units. Indeed TiT_i and Yj(t)Y_j(t) are d-separated by {Xi,Si,j}\{X_i, S_{i,j}\} since the only path connecting them in Figure 2b is blocked: it contains a non-collider XiX_i that is conditioned on [83, 67].

Intuitively, suppose that, given BMI, hormone levels, cardiovascular measures, medication, and lifestyle, counterfactual bone densities do not depend on whether individuals are assigned to exercise or not. Then, if we were to randomly relabel who is assigned to exercise in the population, the potential outcomes Y(0)Y(0), Y(1)Y(1) of the relabelled individuals—what would have happened under either exercise regime—would remain unchanged. This relabelling thought experiment extends to linked data: the potential outcomes of individuals in B\mathcal{B} linked to individuals in A\mathcal{A} are not affected by which treatment label they end up matched to.

S-admissibility

The effect in the source population is recoverable from the linked sample as long as the follow-up covariates XjX_j can be controlled. The conditions above discussed the internal validity of S\mathcal{S}. Moreover, the external validity is at stake because the covariate distribution in S\mathcal{S} differs from that in A\mathcal{A}. Consequently, identifying τA\tau_{\mathcal{A}} requires not only that the common causal assumptions hold in S\mathcal{S}, but also that the effect be recoverable from selection bias. We thus examine the causal record linkage problem through the lens of data fusion, where the key challenge is to recover τA\tau_{\mathcal{A}} despite restricting inference to reliably linked pairs.

Causal identification under selection relies on the ‘selection backdoor criterion’ [12, 11, 52].

Definition 1 (selection backdoor criterion). Given a causal system augmented with the sampling mechanism SS, where treatment TT has no other descendant than outcome YY, a set of confounders XX satisfies the selection backdoor criterion if

  1. Y(t)⊥ ⁣ ⁣ ⁣⊥T∣XY(t) \perp\!\!\!\perp T\mid X,

  2. Y(t)⊥ ⁣ ⁣ ⁣⊥S∣X,TY(t) \perp\!\!\!\perp S\mid X,T,

  3. {X,T,Y}\{X,T,Y\} are observed in the selection biased data such that S=1S = 1,

  4. XX is observed at the population level.

We say, equivalently, that there is no unmeasured sample selection bias [52]; that XX is SS-backdoor admissible (or SS-admissible) [68]; that p(Y(t))p(Y(t)) is SS-recoverable from selection biased data i.e. expressible in terms of the distribution under selection SS [12]; or that sample selection bias can be corrected [46].

In the causal record linkage system of Figure 2b, condition (ii) from Definition 1 fails without further assumptions, as Yj(t)⊥ ⁣ ⁣ ⁣⊥Si,j∣Xi,TiY_j(t) \perp\!\!\!\perp S_{i,j}\mid X_i,T_i; conditioning on XjX_j would instead satisfy it. XjX_j is accessible either when the FDP vanishes, as linked units then refer to the same individuals giving Xi=XjX_i = X_j (under Assumption 4 up to a small finite sample error) or when all covariates are linking variables, so Zj=XjZ_j = X_j is observed (Remark 2). In practice, inference should be treated with caution, or abandoned, when the FDP is not controlled. Further, the point of Remark 2 requires careful discussion in application, although in finite samples, valid and asymptotically optimal adjustment sets may differ [78], and ZZ may proxy XX.

Beyond condition (ii)(ii), the conditions from Definition 1 hold by definition of the causal record linkage task, guaranteeing SS-recoverability of the effect under control over XjX_j, but not its estimation in practice: the linkage selection shrinks the covariate space relative to A\mathcal{A}, inducing the covariate shift, and finite sample artefacts may induce positivity and overlap violations. These concerns remain manageable, as they occur at random and internally to the covariate space, as formalised by Assumption 5, which supports estimation of the transport formula between the selected data and the baseline.

Theorem 1. In the causal record linkage context represented in Figure 2, under assumptions 1, 2, 4, 5, τA=EA[τS(X)]\tau_A = \mathbb{E}_A[\tau_S(X)] is identified and estimable from the available data.

In the above, τA\tau_A is the estimand of Equation (1) and τS\tau_S is the sample ATE of Equation (2); τS(X)\tau_S(X) is the sample conditional ATE in the selected linked units. The proof of Theorem 1 is deferred to the Appendix Section A.2. Note that Assumption 4 on the FDP is met by stricter linkage while Assumption 5 needs the linked sample to be as large as possible; both may not hold in practice but the error may be negligible, especially when more covariates are linking variables—see the earlier discussion in § 3.1.

To conclude, ensuring identification requires imposing stringency on the linkage, trading consistency, and correspondingly attenuation bias, against the covariate shift and its sampling bias. If the FDP vanishes (Assumption 4), consistency and SS-admissibility are restored, and the covariate shift is corrected by the transport formula under Assumption 5 so τA\tau_A is recoverable. § 4 quantifies the bias trade-off, and § 5 describes the tools available for estimating the transport formula. The simulations of § 8 will later let us observe concretely how these biases interplay with the effect estimation.

Bias decomposition

Record linkage may be interpreted as a matching procedure for estimating treatment effects from multiple data sources. Like a nearest neighbour estimator, it links each baseline unit to a follow-up unit with similar values of pre-treatment linking variables Z⊆XZ \subseteq X, thereby approximating the potential outcome Yi(t)Y_i(t) of a unit ii with the observed outcome Yj∣Tj=tY_j \mid T_j=t of a potentially distinct unit jj. In this Section, we examine the implications of this procedure, contrasting estimators built on linked data to the targeted estimand. Especially, we quantify in Theorem 2 how falsely linked data bias causal estimates.

We distinguish different standard consistent ATE estimators: difference in means, propensity score stratification, inverse probability weighting, matching and outcome regression (compared in the application and simulations § 7 and § 8, and developed in Appendix Section A.1); from non-adjusted methods, through indirect adjustment via the propensity score, to outcome regression controlling for covariates. This section focuses, as a simplified case, on the bias of the difference in means estimator relative to the average causal contrast in the baseline population. Such estimator may be seen as an outcome regression estimator with omitted variables [62], and while extending this analysis formally to the remaining estimators is left out for future work, we hope this development helps clarify the biases at stake when estimating causal effects from linked data.

The difference in means ATE estimator computed on the selected linked data is described as

τ^S=1NS1∑i,jSi,jTiYj−1NS0∑i,jSi,j(1−Ti)Yj.(3)\widehat{\tau}_S = \frac{1}{N_{S1}} \sum_{i,j} S_{i,j}T_iY_j - \frac{1}{N_{S0}} \sum_{i,j} S_{i,j}(1-T_i)Y_j. \tag*{(3)}

We compare τ^S\widehat{\tau}_S with our target, the ATE in the baseline population from Equation (1), to study the bias of the causal record linkage procedure:

E[τ^S]−τA=E[ζ^S]+E[ξ^S],\mathbb{E}[\widehat{\tau}_S] - \tau_A = \mathbb{E}[\widehat{\zeta}_S] + \mathbb{E}[\widehat{\xi}_S],

with ζ^NS\widehat{\zeta}_{NS} the sampling bias and ξ^NS\widehat{\xi}_{NS} the attenuation bias, defined as follows:

ζ^NS=1NS1∑i,j{Si,jTiμ1(Xi)}−1NS0∑i,j{Si,j(1−Ti)μ0(Xi)}−τA,\widehat{\zeta}_{NS} = \frac{1}{N_{S1}} \sum_{i,j} \{S_{i,j}T_i\mu_1(X_i)\} - \frac{1}{N_{S0}} \sum_{i,j} \{S_{i,j}(1-T_i)\mu_0(X_i)\} - \tau_A,
ξ^NS=1NS1∑i,j{Si,jTi(Yj−μ1(Xj))}−1NS0∑i,j{Si,j(1−Ti)(Yj−μ0(Xj))}+1NS0∑i,j{Si,j(1−Ti)(μ0(Xi)−μ0(Xj))}−1NS1∑i,j{Si,jTi(μ1(Xi)−μ1(Xj))}.\begin{aligned} \widehat{\xi}_{NS} ={}& \frac{1}{N_{S1}} \sum_{i,j} \{S_{i,j}T_i(Y_j-\mu_1(X_j))\} - \frac{1}{N_{S0}} \sum_{i,j} \{S_{i,j}(1-T_i)(Y_j-\mu_0(X_j))\} \\ &+ \frac{1}{N_{S0}} \sum_{i,j} \{S_{i,j}(1-T_i)(\mu_0(X_i)-\mu_0(X_j))\} - \frac{1}{N_{S1}} \sum_{i,j} \{S_{i,j}T_i(\mu_1(X_i)-\mu_1(X_j))\}. \end{aligned}

The sampling bias captures the difference in ATE between the selected linked sample SS and the baseline population A\mathcal{A} due to missed links (FN). It results from the covariate shift (brought up in § 3.2 and developed in § 3.4) between distributions in SS and in A\mathcal{A} intensifying with linkage stringency.

The attenuation bias (announced in § 3.1) is due to falsely linked records (FP) and emerges via two mechanisms. The first mechanism implies a contradiction in the causal effect from mismatches in treatment groups: Yj∣Tj=1−t,XjY_j \mid T_j = 1-t, X_j contrasted with μt(Xj)\mu_t(X_j). The magnitude of this part of the bias is the opposite of the causal effect targeted and it appears in proportion to the share of FP pairing units from different treatment groups [93].

The second mechanism arises from discrepancies between covariates inducing differences in linked units potential outcomes: μt(Xj)\mu_t(X_j) contrasted with μt(Xi)\mu_t(X_i). This part of the bias is analogous to the conditional bias of matching estimators due to matching discrepancy [1, 3, 4, 5]. It is a manifestation of coefficient dilution arising from errors in independent variables [82, 14, 33, 87, 90].

The impact of consistency violations on estimation, that takes shape as attenuation bias, depends on the ATE estimator used, and on whether the potential outcome it targets is defined at the level of the covariates or in its aggregated form, as discussed in § 3.1. To frame this impact in the context of the difference in means estimator, we introduce additional assumptions.

When using record linkage for causal inference, we presume to be able to recover a suitable sample. Although all observations are assumed eligible for treatment, not all profiles are being selected. We therefore require that record linkage recover a positive fraction of both treated and control observations.

Assumption 6. ∀t∈{0,1}, E[NSt]>0\forall t \in\{0,1\},\ \mathbb{E}[N_{St}] > 0.

Alternatively we could link treated and untreated subsets from A\mathcal{A} to B\mathcal{B} separately and discard pairs involving records in B\mathcal{B} linked to both treatment groups.

We also require that linkage errors do not systematically confound treatment groups (Ti=t,Tj=1−tT_i=t,T_j=1-t), i.e. the majority of falsely linked pairs should connect entities who received the same treatment (Ti=t,Tj=tT_i=t,T_j=t).

Assumption 7. ∀t∈{0,1}, E[NS∣Ti=t,Tj=1−t]≤E[NS∣Ti=t,Tj=t]\forall t \in\{0,1\},\ \mathbb{E}[N_{S\mid T_i=t,T_j=1-t}] \le\mathbb{E}[N_{S\mid T_i=t,T_j=t}].

Finally we need a consistent estimation method.

Assumption 8. The ATE estimator used is consistent.

The difference in means estimator is consistent for the linked sample ATE when the treatment is randomly assigned. Other estimators may additionally require correctly specified models for the propensity score or for the outcome, for instance.

Theorem 2. In the causal record linkage context represented in Figure 2, under assumptions 1, 2, 3, 6, 7, 8, the causal effect estimation suffers from attenuation bias due to falsely linked data: E[τ^S]=τS(1−E[FDP])\mathbb{E}[\widehat{\tau}_S] = \tau_S(1-\mathbb{E}[FDP]).

With Theorems 1 and 2, we have characterised the gap between the estimator and the estimand in two steps. The ATE estimator computed on linked data, although converging towards the true ATE on the selected sample, is hampered by attenuation bias (Theorem 2). This ATE on the selected sample differs from the targeted ATE on the baseline population due to sampling bias induced by selection (Theorem 1). For instance, if individuals with extreme BMI are more confidently linked, the effect of physical activity on bone density in the selected sample may differ from the targeted effect over the whole baseline population, which spans a broader range of BMI values. As a result, recovering the causal effect in the baseline population requires eliminating attenuation bias and transporting the effect over samples.

Consider the implications of these results. The trade-off between inconsistencies and covariate shift in the causal identification transposes into a trade-off between attenuation and sampling bias in the inference: a balance between falsely linked records and missed links, FDP and sensitivity, bias and variance. If the linked data are mostly true links, sampling bias drives our estimator. If falsely linked data link records from different treatment groups, our estimator accounts for adversarial contributions to the causal effect and, if they link records within the same treatment group, our estimator behaves like a one-to-one matching estimator; the main concern in that case becomes ‘how similar are the potential outcomes of the falsely linked records?’

Attenuation bias dominates when the FDP is non-negligible, while sampling bias dominates when the decision rule is too strict. When the linked sample is large enough to support causal inference, this suggests favouring a relatively stringent decision rule so as to render attenuation bias tolerable, while the resulting sampling bias could still be corrected for under the assumptions that make Definition 1 valid. Note that, were the population unchanged from baseline to follow-up (NA=NB=∑i,jΔi,jN_A = N_B = \sum_{i,j} \Delta_{i,j}), the record linkage task would be easier [29, 28, 51, 74], the expected potential outcomes in A\mathcal{A} and B\mathcal{B} would coincide on average so the bias would only stem from opposite contributions, which would also be less likely.

Estimation

The remaining obstacle to estimating τA\tau_A on linked data is the covariate shift between S\mathcal{S} and A\mathcal{A}, which is handled at the estimation stage. Unbiased causal effect estimation on linked data hinges on a stringent record linkage decision rule, which ensures consistency and effect recoverability from the selected data S\mathcal{S} (§ 3). Linkage thereby favours atypical profiles and the set S\mathcal{S} of observations from A\mathcal{A} augmented with outcome information from B\mathcal{B} is not representative. Even with negligible falsely linked records, a standard estimator applied to S\mathcal{S} therefore targets τS\tau_{\mathcal{S}} rather than τA\tau_A (§ 4). We present three families of generalised treatment effect (g)-methods that correct this shift.

G-methods cope with this issue by exploiting the representativeness of A\mathcal{A}, at the cost of untestable interpolation [45, 22, 19]. We consider three families of estimators, used in the simulations of § 8 and in the real data application of § 7. The g-formula [76] models conditional outcomes and marginalises over the targeted population. Weighting [85, 18, 42] uses propensity score modelling to transform observations and reflect the targeted population characteristics. Doubly robust approaches combine both and remain consistent if either model is correctly specified.

The two ingredients for reliable causal effect estimation on linked data are formalised with Theorems 1 and 2. The generalised treatment effect estimators can correct the sampling bias and recover τA\tau_A from S\mathcal{S} despite the selection induced by linkage. Though, they cannot correct the attenuation bias due to falsely linked records. Nevertheless, when linkage is imperfect, matching similar profiles remains adequate to identify causal effects, or at least their direction as we illustrate in § 7. The intuition being that falsely linked individual records may still create an instance for which the observed outcome is a reasonable potential outcome.

G-formula estimators

G-formula estimators use outcome modelling to generalise findings from observations at hand towards the targeted population [76, 21, 22, 19]. To wit, the potential outcome distributions are learnt from the selected linked data and applied on the baseline population. The classical g-formula treatment effect estimator τ^G\widehat{\tau}_{G} is defined as

1NA∑i=1NA(μ^1,S(xi)−μ^0,S(xi))(4)\frac{1}{N_A}\sum_{i=1}^{N_A}\left(\widehat{\mu}_{1,S}(x_i)-\widehat{\mu}_{0,S}(x_i)\right) \tag*{(4)}

where μ^t,S(x)\widehat{\mu}_{t,S}(x) is fitted using the selected linked data for treatment group tt. When Assumption 5 necessary to generalise the effect from SS to AA is not reasonable, causal identification is confined to regions where positivity holds. We can correct this estimator with a Prediction Powered Inference (PPI) term accounting for the difference between modelled potential and observed outcomes [38]. The use of such estimator for generalised causal contrast is fairly novel [16]. We define the PPI g-formula estimator τ^GPPI\widehat{\tau}_{GPPI} as:

1NA∑i=1NAω^i(μ^1,S(xi)−μ^0,S(xi))+1NS∑ℓ=1NStℓ(yℓ−ω^ℓ⋅μ^1,S(xℓ))−(1−tℓ)(yℓ−ω^ℓ⋅μ^0,S(xℓ)).(5)\frac{1}{N_A}\sum_{i=1}^{N_A}\widehat{\omega}_i\left(\widehat{\mu}_{1,S}(x_i)-\widehat{\mu}_{0,S}(x_i)\right)+\frac{1}{N_S}\sum_{\ell=1}^{N_S}t_\ell\left(y_\ell-\widehat{\omega}_\ell\cdot\widehat{\mu}_{1,S}(x_\ell)\right)-(1-t_\ell)\left(y_\ell-\widehat{\omega}_\ell\cdot\widehat{\mu}_{0,S}(x_\ell)\right). \tag*{(5)}

Individual weights ω^i\widehat{\omega}_i are included to obtain a more efficient estimator, with lower variance; they control the contribution of the estimated potential outcomes into estimation, adapting to the accuracy between potential and observed outcomes of the predictive model used—see [7, 38, 80] for further details.

Weighting estimators

Weighting estimators calibrate the observations from the selected linked data to represent the baseline population. They transport evidence from one population to the other by accounting for the covariate distribution shift between populations with weights ω\omega. The evidence in the selected linked population is derived from inverse probability weighting using propensity score estimation. As a consequence of their definition, weighting estimators might be sensitive to the lack of random positivity [77, 17, 45]. The expected conditional outcome in treatment group tt is identified as:

EA[ES[Y∣T=t,X=x]]=ES[ES[Y∣T=t,X=x]ω]=ES[ES[Y1{T=t}∣X=x]PS(T=t∣X=x)ω].\mathbb{E}_{A}\left[\mathbb{E}_{S}[Y\mid T=t,\mathbf{X}=x]\right]=\mathbb{E}_{S}\left[\mathbb{E}_{S}[Y\mid T=t,\mathbf{X}=x]\omega\right]=\mathbb{E}_{S}\left[\frac{\mathbb{E}_{S}[Y\mathbb{1}_{\{T=t\}}\mid\mathbf{X}=x]}{\mathbb{P}_{S}(T=t\mid\mathbf{X}=x)}\omega\right].

We then define the IPSW estimator τ^IPSW\widehat{\tau}_{IPSW} and the EB estimator τ^EB\widehat{\tau}_{EB} for the treatment effect with:

1NS∑ℓ=1NSω^ℓyℓ(tℓπ^(xℓ)−1−tℓ1−π^(xℓ)).(6)\frac{1}{N_S}\sum_{\ell=1}^{N_S}\widehat{\omega}_\ell y_\ell\left(\frac{t_\ell}{\widehat{\pi}(x_\ell)}-\frac{1-t_\ell}{1-\widehat{\pi}(x_\ell)}\right). \tag*{(6)}

These estimators differ in their weighting scheme. Entropy Balancing (EB) or Calibration Weighting (CW) [42, 19, 36] constructs balancing weights by minimising the negative entropy subject to balancing and normalising constraints, aligning the moments of the selected sample SS with those of the baseline population A\mathcal{A}. Inverse Probability of Sampling Weighting (IPSW) [85, 18, 19] instead weights observations by the covariate density ratio PA(X=x)/PS(X=x)\mathbb{P}_{\mathcal{A}}(X=x)/\mathbb{P}_{\mathcal{S}}(X=x) which can be developed into PA(S=1)/PA(S=1∣X)\mathbb{P}_{\mathcal{A}}(S=1)/\mathbb{P}_{\mathcal{A}}(S=1\mid X) [18, 19]. We estimate these sampling weights using a logistic regression.

We estimate the propensity score with a logistic regression, on the baseline A\mathcal{A}: πA(X)=PA(T=1∣X)\pi_{\mathcal{A}}(X)=\mathbb{P}_{\mathcal{A}}(T=1\mid X), and on the restricted selected subset S\mathcal{S}: πS(X)=PS(T=1∣X)\pi_{\mathcal{S}}(X)=\mathbb{P}_{\mathcal{S}}(T=1\mid X). As noted in § 3.2 these propensity scores should coincide: PS(T∣X)=PA(T∣X)\mathbb{P}_{\mathcal{S}}(T\mid X)=\mathbb{P}_{\mathcal{A}}(T\mid X), although they tend to differ empirically. According to the original work, we should use πS(X)\pi_{\mathcal{S}}(X) as we generalise from population S\mathcal{S} [85, 18, 19]. In the simulation study, using the propensity score estimated in the more stable baseline population yields better estimates. In the real data application, following the original idea yields more accurate estimates.

Doubly robust estimators

Doubly robust versions of these estimators exist [9, 57, 19]. They combine modelling of the outcome Equation (4) and of the treatment probability Equation (6) to form the Augmented IPSW τ^AIPSW\widehat{\tau}_{\mathrm{AIPSW}} and the Augmented EB τ^AEB\widehat{\tau}_{\mathrm{AEB}} estimators:

1NA∑i=1NAμ^1,S(xi)−μ^0,S(xi)+1NS∑ℓ=1NSω^ℓ(tℓ(yℓ−μ^1,S(xℓ))π^(xℓ)−(1−tℓ)(yℓ−μ^0,S(xℓ))1−π^(xℓ)).(7)\frac{1}{N_{\mathcal{A}}}\sum_{i=1}^{N_{\mathcal{A}}}\widehat{\mu}_{1,\mathcal{S}}(x_i)-\widehat{\mu}_{0,\mathcal{S}}(x_i)+\frac{1}{N_{\mathcal{S}}}\sum_{\ell=1}^{N_{\mathcal{S}}}\widehat{\omega}_{\ell}\left(\frac{t_{\ell}(y_{\ell}-\widehat{\mu}_{1,\mathcal{S}}(x_{\ell}))}{\widehat{\pi}(x_{\ell})}-\frac{(1-t_{\ell})(y_{\ell}-\widehat{\mu}_{0,\mathcal{S}}(x_{\ell}))}{1-\widehat{\pi}(x_{\ell})}\right). \tag*{(7)}

These estimators are consistent when one of the models (outcome or propensity score) is correctly specified.

All three families of estimators can address the covariate shift between A\mathcal{A} and S\mathcal{S} that gives rise to sampling bias. However, they require sufficiently large treatment arms in S\mathcal{S} to fit the outcome or propensity models. Moreover, the gg-formula relies on a correctly specified outcome model and risks extrapolating beyond the linked support. Weighting estimators require overlap between the supports of A\mathcal{A} and S\mathcal{S} and can become unstable as these supports diverge. Doubly robust estimators combine both approaches and remain consistent if either model is correctly specified.

As with the assumptions underlying Theorem 1, these requirements are in tension with the linkage stringency needed to eliminate attenuation bias, which shrinks S\mathcal{S} and increases its divergence from A\mathcal{A}. Estimator performance therefore depends on the application. Under substantial covariate shift, the simulations (§ 8) show that weighting fails while outcome modelling recovers the effect. In contrast, in the application (§ 7), all estimators exhibit biases of similar magnitude. The assumptions underlying each estimator thus inform the linkage decisions discussed in the next section.

Guiding linkage decisions

Reliable causal effect estimation with linked data is contingent on identifiability of the causal contrast in the linked data, recoverability of the experimental distribution from linkage sampling bias, and generalisability of the estimation towards the intended target of inference. The linked data selection thus calls for complementary evaluation measures to assess record linkage quality and validity of the causal framework. Having established the main condition under which the selection backdoor criterion holds: Assumption 4, allowing us to recover the causal effect in theory, we now provide guidelines for jointly navigating record linkage and causal framework in order to estimate the causal effect in practice.

Before record linkage, it is worth gauging how hard the task is. A naive linkage approach—matching records on exact agreement of identifying information—reveals the discriminative power of the linking variables, indicating how much ambiguity the linkage procedure will face.

Step 1—Record linkage across decision thresholds λ\lambda Run the record linkage procedure and apply a grid of cutoffs λ∈[0.5,1)\lambda\in[0.5,1) to the potential pairs of records in the output. This will shed light on the trade-offs selection induces (linkage quality versus linked sample size versus representativeness of the baseline).

Step 2—Examine the linkage output Record linkage performance may be estimated without unique identifiers [13, 15, 84, 43, 29, 71, 23, 24, 25, 56, 74]; here we focus on controlling the estimated FDP. The default output (λ=0.5\lambda= 0.5) gives the largest linked set and the most reliable FDP estimate. Estimating the FDP across thresholds from 0.5 up to the highest linkage score observed shows the potential gains from stricter linkage (but FDP estimation is destabilised by smaller samples). The estimated FDP bounds the attenuation bias of the linked sample ATE estimator and helps assess Assumption 4: controlling it ensures consistency in the linked data and validity of the selection backdoor criterion. One should also track the linked sample size, overall and within treatment subgroups, as a function of λ\lambda to assess Assumption 6, since stricter linkage trades sample size for linkage quality. What constitutes a reasonable FDP should be judged per application.

Step 3—Scrutinise the resulting linked data The linked sample should remain faithful to the baseline population. Generalisability and effect estimation remain possible when the analytic support is smaller than that of the target population, provided the lack of positivity and overlap is at random rather than structural [88, 19], and violations are internal [91, 60]. However, theory only guarantees identification in regions of the covariate space shared by the linked sample SS and the baseline population AA, and the sampling bias resulting from an increasingly small and atypical sample of data linked confidently may dominate the estimator. Deviations between SS and AA may be assessed with the Maximum Mean Discrepancy (MMD) between covariate distributions, the Intersection over Union (IoU) between variable supports, the Standardised Mean Difference (SMD) between variables. In addition, the potential positivity and overlap violations in the linked sample SS could be examined to assess Assumption 5. Covariate distributions could be compared across treatment status {T=1}\{T = 1\} versus {T=0}\{T = 0\} and samples AA versus SS.

Step 4—Decide whether to proceed with causal inference Based on the joint behaviour of estimated FDP, deviations between SS and AA and sample size, determine whether the linked data supports valid inference or whether inference should be abandoned. As a practical rule, restrict candidate cutoffs to those retaining a minimum viable sample per treatment arm (e.g. in the application of § 7 we target at least 100 linked units per group), then select the strictest λ\lambda at the ‘elbow’ of the MMD curve—the point where linked and baseline samples begin to diverge—while monitoring FDP.

Step 5—Apply generalised estimators to the validated linked sample Once the linked sample SS is built, compute standard estimators and g-methods. Although the principle of separating design and analysis cautions against letting design choices inform estimation, comparing results across thresholds λ\lambda may reveal the shifting balance between attenuation bias (from falsely linked pairs) and sampling bias (from missed links). This trade-off however, may be ambiguous as standard error has many roots and, in the causal record linkage setting where linking variables and covariates are intertwined, there is no evidence that increased variance, or deviations from the main effect, reflects linkage error.

Following these guidelines, we illustrate the selection of the cutoff λ\lambda for our study of women’s health application on Figure 3, running the linkage procedure across a grid λ∈[0.5,1)\lambda\in[0.5,1) (Step 1) and tracking sample size on the upper x-axis. We first examine the linkage output (Step 2) via the estimated FDP^∈[0,1]\widehat{\mathrm{FDP}} \in[0,1], computed from linkage scores [29, 74], providing an empirical order of magnitude for the attenuation bias characterised in Theorem 2. As the linked sample shrinks, the estimation becomes less reliable, but remains useful for guiding selection and determining whether inference should be abandoned. We then scrutinise the linked data (Step 3) using the Maximum Mean Discrepancy, MMD ∈[0,1]\in[0, 1] [37], measuring joint distributional differences between SS and AA in a Gaussian RBF kernel space, providing an empirical counterpart to the sampling bias derived in § 4. We complement this with the Intersection over Union (IoU) of propensity score supports and covariate Standardised Mean Differences (SMD), and inspect

Complementary metrics for BRL and FlexRL across linkage thresholds, including MMD, FDP, Sensitivity, IoU, propensity score support, SMD measures, and BMI and Age distributions

Figure 3. Complementary metrics illustrating the trade-off between minimisation of linkage errors and maximisation of resemblance between linked and baseline samples, for record linkage on the SWHAN data (BRL left, FlexRL right). Black lines indicate means and grey bands standard deviations across 100 procedures.

histograms of some causally relevant variables (lower panels of Figure 3). We additionally report the true FDP and sensitivity.

Based on these diagnostics, we select the cutoff λ=0.82\lambda= 0.82, pointed out on the upper panel of Figure 3, where linked and baseline samples begin to diverge, controlling FDP and retaining enough observations per treatment arm (Step 4). The proportion of treated units remains stable across thresholds λ\lambda, consistent with S⊥ ⁣ ⁣ ⁣⊥T∣XS \perp\!\!\!\perp T \mid X. At very strict cutoffs, the IoU becomes unstable, signalling a potentially irreversible covariate shift under which all estimators become dominated by sampling bias.

Navigating record linkage performance and representativeness of the linked sample relative to the baseline, subject to the limited number of linked observations is important for consistency to be tenable and for the effect to be recoverable. On Figure 4 and Table 2 in the following section, we present the estimation results for the effect of physical activity on two physiological outcomes.

Effect of physical activity on hip bone mineral density and dehydroepiandrosterone sulfate hormone levels

Figure 4. Effect of physical activity on hip bone mineral density (HP-BMD) (top, no effect, average ATE on true links is 0) and on dehydroepiandrosterone sulfate (DHA-S) hormone levels (bottom, negative effect, average ATE on true links is −18-18) over 100 procedures, after record linkage through BRL (median indicated by a square, left point) and FlexRL (median indicated by a triangle, right point), on the SWAN data. We compare ATE and g-ATE estimators on selected linked data S\mathcal{S} and on baseline data A\mathcal{A} respectively.

Physical activity on HP-BMDPhysical activity on DHA-S
estimatorBRLFlexRLestimatorBRLFlexRL
ATE on S\mathcal{S}: diff. in means−0.04-0.04 (0.01)−0.01-0.01 (0.03)diff. in means51.85 (9.25)51.34 (20.06)
YY-regression0.01 (0.01)−0.01-0.01 (0.02)YY-regression9.06 (7.76)7.35 (14.53)
π\pi-stratification−0.04-0.04 (0.02)−0.03-0.03 (0.03)π\pi-stratification79.11 (11.47)71.39 (20.13)
IPW0.27 (0.10)0.28 (0.27)IPW7.03 (7.24)5.45 (13.89)
matching0.02 (0.01)0.01 (0.02)matching5.89 (16.20)3.72 (18.06)
gg-ATE on A\mathcal{A}: DR0.07 (0.09)0.00 (0.05)DR5.96 (8.53)5.68 (17.85)
gg-formula0.11 (0.05)0.00 (0.08)gg-formula5.96 (8.53)5.68 (17.85)
PPI gg-formula−0.02-0.02 (0.01)−0.03-0.03 (0.05)PPI gg-formula2.11 (10.44)−3.08-3.08 (17.63)
IPSW−0.07-0.07 (0.03)−0.02-0.02 (0.03)IPSW15.36 (9.28)8.84 (17.48)
EB−0.01-0.01 (0.01)0.00 (0.01)EB16.51 (3.34)14.42 (6.32)

Table 2. Bias mean and standard deviation over a 100 procedures of evaluating the effect of physical activity on HP-BMD (left) and on DHA-S (right) for BRL and FlexRL on the SWAN data. ATE and g-ATE estimators are derived from selected linked data S\mathcal{S} and baseline population A\mathcal{A} respectively.

SWAN application results

The SWAN data come from a longitudinal Study of Women’s Health Across the Nation in the United States and are distributed by the Inter-university Consortium for Political and Social Research (ICPSR). We build the baseline data set A\mathcal{A} using the baseline (1996-1997) and visit 1 (1997-1999), and we use visit 2 (1998-2000) as the follow-up data set B\mathcal{B}. File A\mathcal{A} contains 1,393 records, file B\mathcal{B} 1,836 and 1,194 entities are observed in both files (they form the set of true links). As an illustration, we consider the effect of initiating physical activity on two outcome variables: the hip bone mineral density (HP-BMD) and the (rescaled) dehydroepiandrosterone sulfate (DHA-S) hormone, among pre-menopausal women. In the baseline and in the true links 22% of the units are treated.

These data are of particular interest for this work because they are openly available, suited for causal inference, fitting with the setting we tackle and they contain unique identifiers that allow to evaluate the causal record linkage procedure and discuss methodology. We work with ethnicity, age, discretised height and BMI, and a truncated portion of the unique identifier present in both data sets, as linking variables. Exact matching on all these linking variables yields an FDP of 0.3 and a sensitivity of 0.5. Based on the literature and on an exploratory analysis of the true links, we include baseline HP-BMD and DHA-S levels, ethnicity, BMI, age as covariates in the causal analysis [48, 54, 6, 30, 53, 63].

After reviewing the available software for record linkage we select BRL and FlexRL [79, 75] for evaluation, as they are state-of-the-art methods available in R representing the two main paradigms in record linkage: the Fellegi-Sunter model and the hierarchical entity resolution framework [31, 86] and they are able to link records on this application. Following guidelines from § 6 we run 100 procedures for both algorithms with the MMD elbow detection yielding λ=0.82\lambda= 0.82, on average achieving an FDP of 0.1 and retaining approximately 500 linked observations—see Figure 3.

Figure 4 presents the distributions of the estimated effect of physical activity on both outcomes, HP-BMD and DHA-S, for both record linkage algorithms, BRL and FlexRL. For each outcome, we compare standard ATE estimators on the selected linked data S\mathcal{S} (cited in § 4) with generalised treatment effect estimators, g-ATE, on the baseline population A\mathcal{A} (specified in § 5) to ATE in the true links; corresponding average biases are reported in Table 2. Across linkage algorithms, the biases of all estimators are of similar magnitude.

There is no effect of physical activity on HP-BMD. All estimators tally around zero; correctly and incorrectly linked pairs contribute similarly in absence of signal. In contrast, there is a negative effect of physical activity on DHA-S. On Table 2 we see that the ‘matching’ ATE estimator and ‘PPI g-formula’ g-ATE estimator yield the most consistent estimates for both record linkage algorithms. We observe that FlexRL is more noisy and less conservative than BRL, which is expected [75 a].

Interestingly, weighting g-methods (IPSW, EB) are not particularly sensitive to positivity and overlap violations in this application. All estimators seem robust to the proportion of false discoveries in the linked data, suggesting some tolerance to deviations from Assumption 4. However, we noticed that all estimators become dominated by the sampling bias with an overly strict linkage decision rule. The true effect average is around −18-18 which we approach with λ=0.82\lambda= 0.82. With λ=0.95\lambda= 0.95 estimators estimate an effect of +15+15; we also see that the MMD deviates from the null with the increasing stringency of the decision rule on Figure 3 and that the IoU of propensity score support shrinks.

Overall, this application illustrates the practical difficulties of causal record linkage: linkage quality and linked observations after selection are limited. We complement the application with a simulation study demonstrating the practical utility of our methodology in the following section.

Simulations

To better understand the estimation behaviour in the causal record linkage framework, we perform a simulation study. Files A\mathcal{A} and B\mathcal{B} gather respectively 500 and 700 patient records, including 400 true links created by randomly duplicating records between sources. We generate five independent variables X\mathbf{X} from multinomial distributions. We define the propensity score as π(X)=P(T=1∣X)=Λ(γX1+γX2+γX4)\pi(\mathbf{X}) = \mathbb{P}(T = 1|\mathbf{X}) = \Lambda(\gamma X_1 + \gamma X_2 + \gamma X_4), where Λ(⋅)\Lambda(\cdot) is the logistic distribution function. We define the outcome as Y(t)=αtX1+βX2+εY(t) = \alpha tX_1 + \beta X_2 + \varepsilon incorporating effect heterogeneity through an interaction with X1X_1. While C=(X1,X2,X4)\mathbf{C} = (X_1, X_2, X_4) are involved in the causal framework, the record linkage model uses Z=(X1,X2,X3,X5)\mathbf{Z} = (X_1, X_2, X_3, X_5) to link the data sources. Variables X\mathbf{X} have a decreasing distribution so that higher values are rarer; we do so to illustrate that, although noise in the linked data is expected to shrink the effect towards zero, sampling bias can inflate it if the profiles record linkage preferentially selects are associated with a larger causal effect. Conversely, when selected profiles are associated with a smaller causal effect, sampling bias reinforces the attenuation bias; this is what we observe empirically for the null effect of physical activity on HP-BMD in § 7.

We analyse a scenario with good linkage quality (FDP after selection below 0.05); we obtained similar results with more difficult scenarios (FDP after selection below 0.2). Linking variables have a relatively high discriminative strength with 7 unique values each, and we do not induce any registration error. Exact matching on all linking variables yields an FDP of 0.5 and a sensitivity of 1. This simulation scenario results in a default FDP of 0.15 for both methods BRLBRL and FlexRLFlexRL linking 290 and 330 records respectively, and an optimised FDP of 0.03 for both methods BRLBRL and FlexRLFlexRL linking 110 records and 130 records respectively. The BRLBRL method is known to be more conservative [75 a], and that is indeed what we observe.

The simulation study illustrates the discussions from § 3 and § 4. Using the default linkage threshold λ=0.5\lambda= 0.5, ATE estimators exhibit attenuation bias from falsely linked data due to covariate discrepancies and adversarial contributions—see the upper Figure 5. Falsely linked data yield larger variance than correctly linked data.

Three-panel boxplot simulation study comparing BRL and FlexRL record linkage methods

Figure 5. ATE of the simulation study over 100 procedures, after record linkage through BRL (square, left point) and FlexRL (triangle, right point). On the top, estimators evaluated on TP and FP (of the same or of opposite treatment group) for λ=0.5\lambda= 0.5. On the middle, evolution of outcome regression estimator (YY-regression) with linkage stringency (assessed by λ\lambda, the resulting FDP and the number of linked observations NSN_S). On the bottom, ATE estimators evaluated on true links and selected linked data SS with λ=0.95\lambda= 0.95 and, gg-ATE estimators evaluated on baseline data AA.

The trade-off between attenuation bias and sampling bias is illustrated for the outcome regression estimator in the middle panel of Figure 5; we observed the same behaviour for all estimators. These results illustrate that deviations from the main effect or increasing variance may not reflect inconsistencies in the linked data but rather may arise from sampling bias; highlighting limitations of approaches aiming at minimising the estimated effect variance [94]. In the middle panel of Figure 5, sampling bias induces larger variance and deviations from the main effect despite an FDP lower than 0.03.

Optimising MMD and controlling the estimated FDP yields λ=0.95\lambda= 0.95 for BRLBRL and FlexRLFlexRL. We report in Table 3 the average biases of ATE estimators on selected linked data S\mathcal{S} and g-ATE estimators on baseline population A\mathcal{A}. While the estimation from rigorously selected linked data (middle of the lower panel on Figure 5) is biased,

estimatorBRLBRLFlexRLFlexRL
ATE on S\mathcal{S}diff. in means1.27 (0.63)2.00 (0.60)
ATE on S\mathcal{S}YY-regression1.26 (0.52)1.81 (0.51)
ATE on S\mathcal{S}π\pi-stratification1.31 (0.73)1.84 (0.60)
ATE on S\mathcal{S}IPW1.42 (1.07)2.19 (1.21)
ATE on S\mathcal{S}matching1.28 (0.51)1.79 (0.54)
g-ATE on A\mathcal{A}DR0.10 (0.29)0.18 (0.37)
g-ATE on A\mathcal{A}g-formula0.10 (0.29)0.19 (0.35)
g-ATE on A\mathcal{A}PPI g-formula-0.63 (0.47)-0.89 (0.53)
g-ATE on A\mathcal{A}IPSW——
g-ATE on A\mathcal{A}EB——

Table 3. Bias mean and standard deviation over a 100 procedures of the simulation study for BRLBRL and FlexRLFlexRL. ATE and g-ATE estimators are obtained on selected linked data and baseline population respectively.

g-methods relying on outcome modelling (right of the lower panel on Figure 5) recover the true effect. Weighting methods though are affected by the lack of positivity and overlap between the covariate distributions of SS and A\mathcal{A} and are not reliable.

Conclusion and Discussion

This paper frames causal inference on linked data as a selection problem that primarily threatens consistency. Under a negligible false discovery proportion or a valid linking variable adjustment set, the selection backdoor criterion ensures effect identifiability and transportability to the target. We further highlight the balance between linkage reliability and sample representativeness. To address the bias we advocate for the use of g-methods, whose assumptions offer different strategies for causal inference and can be challenged using generalised additive models or sensitivity analysis. We provide a principled framework for selecting linked data and estimating causal effects while accounting for the trade-offs inherent in record linkage and generalisability.

When record linkage is imperfect, it remains capable of matching similar profiles and allows the treatment effect to be recovered with g-methods, provided that the linking variables overlap sufficiently with the confounders. Reliable causal estimation thus remains possible despite linkage errors to some extent. This connects record linkage to matching, though unlike matching methods with an adaptable number of matches, record linkage as we studied it does not benefit from building clusters of comparable entities, as it aims to recover the one-to-one assignment between entities across two data files. Moreover, record linkage performance improves with the dimensionality of linking variables, whereas matching discrepancy worsens with the dimensionality of covariates. Bridging these two paradigms further might be a promising direction for future work on causal record linkage.

The selection problem we investigated is not unique to causal inference and would appear for any statistical inference on linked data. More broadly, the causal record linkage framework is an extreme instance of the classical problem of learning under a non-ignorable selection mechanism that shifts the observed (‘training’) distribution away from the target population. Our case is particularly severe because record linkage selection targets a narrow, atypical subpopulation; were the selected population representative of the majority of the baseline observations, the selection mechanism would remain non-ignorable, but the covariate distributions from which we learn and to which we generalise would overlap far more.

Uncertainty is not addressed in this article. Horvitz-Thompson or robust variance estimators cannot account for the uncertainty arising from the entire process (data sampling, linkage, propensity score and treatment effect estimation), and variance estimators derived under simple random sampling are inappropriate because linkage is informed by covariates. Ideally, we would propagate uncertainty via bootstrapping. However, as with matching methods [2], the bootstrap is not valid for causal record linkage because it cannot preserve the distributions of links and non-links while maintaining independence across samples. Moreover, it cannot ensure adequate representation of treated and control units given the limited number of linked observations available in each sample. Therefore, we only report point estimates in the applications and leave the quantification of uncertainty throughout the causal record linkage process for future research. Notably, the argument for what to do in practice may change when uncertainty quantification is desired, trading bias for variance reduction.

Explainability and diagnostic tools are difficult to develop because causal relationships are only observed for linked records, and only partial information is available from non-linked records (covariates in the baseline and outcomes in the follow-up). We therefore proposed diagnostic measures to help scientists understand how atypical the linked population is, report on the precariousness of the estimated treatment effect across the space of confounders, and clarify whether uncertainty stems from data scarcity or reliance on modelling. With this essay on causal record linkage, we hope to have provided a set of concepts that can anchor future methodologies for causal inference on linked data.

Appendix

Standard average treatment effect estimators

difference in means:

1NS1∑i,jSi,jTiYj−1NS0∑i,jSi,j(1−Ti)Yj\frac{1}{N_{S1}}\sum_{i,j} S_{i,j}T_iY_j-\frac{1}{N_{S0}}\sum_{i,j}S_{i,j}(1-T_i)Y_j

regression estimator:

1NS∑i,jSi,j(μ^1(Xi)−μ^0(Xi)),\frac{1}{N_S}\sum_{i,j}S_{i,j}\left(\widehat{\mu}_1(X_i)-\widehat{\mu}_0(X_i)\right),

where μ^t(Xi)\widehat{\mu}_t(X_i) estimates E[Yj∣Ti=t,Xi]\mathbb{E}[Y_j\mid T_i=t,X_i],

propensity score stratification:

∑stratum s∈S(1NS1s∑{i,j}∈sTiYj−1NS0s∑{i,j}∈s(1−Ti)Yj)NSsNS,\sum_{\substack{\text{stratum }s\in S}}\left(\frac{1}{N_{S1}^{s}}\sum_{\{i,j\}\in s}T_iY_j-\frac{1}{N_{S0}^{s}}\sum_{\{i,j\}\in s}(1-T_i)Y_j\right)\frac{N_S^{s}}{N_S},

inverse probability weighting:

1NS∑i,jSi,j(TiYjπ^(Xi)−(1−Ti)Yj1−π^(Xi)),\frac{1}{N_S}\sum_{i,j}S_{i,j}\left(\frac{T_iY_j}{\widehat{\pi}(X_i)}-\frac{(1-T_i)Y_j}{1-\widehat{\pi}(X_i)}\right),

matching:

1NS∑i,j(2Ti−1)(Yj−1M∑i′∈JM(i)Yj′)Si,j.\frac{1}{N_S}\sum_{i,j}(2T_i-1)\left(Y_j-\frac{1}{M}\sum_{i'\in\mathcal{J}_M(i)}Y_{j'}\right)S_{i,j}.

Proof of Theorems

We adopt the shorthand S△S\triangle for TP and S∇S\nabla for FP (with S∇AS\nabla_A, S∇BS\nabla_B specifying conditioning on XiX_i or XjX_j, and S∇tt′S\nabla_{tt'} disambiguating falsely linked pairs with Ti=tT_i=t and Tj=t′T_j=t'; no such distinction is needed for TP). Additionally, potential outcomes in SS might refer to those of records from A\mathcal{A} or from B\mathcal{B}, we sometimes index these quantities with SAS_A or SBS_B. We refer the reader to § 2 for other notations.

Objectives We consider the causal record linkage system presented in § 2 and represented in Figure 2. Our estimand, the target of inference, is the Average Treatment Effect (ATE) in the baseline population (Equation (1) in the article)

τA=EA[Y(1)−Y(0)]=μ1A−μ0A.\tau_A=\mathbb{E}_A[Y(1)-Y(0)]=\mu_{1A}-\mu_{0A}.

Let τ^S=1NS1∑i,jSi,jTiYj−1NS0∑i,jSi,j(1−Ti)Yj\widehat{\tau}_S=\frac{1}{N_{S1}}\sum_{i,j}S_{i,j}T_iY_j-\frac{1}{N_{S0}}\sum_{i,j}S_{i,j}(1-T_i)Y_j denote the difference in means estimator (Equation (3) in the article), computed on the selected linked data. The bias of the causal record linkage procedure is thus:

E[τ^S]−τA=E[ζ^NS]+E[ξ^NS].\mathbb{E}[\widehat{\tau}_S]-\tau_A=\mathbb{E}[\widehat{\zeta}_{NS}]+\mathbb{E}[\widehat{\xi}_{NS}].

It decomposes into the sampling bias ζ^NS\widehat{\zeta}_{NS} and the attenuation bias ξ^NS\widehat{\xi}_{NS}. The sampling bias is due to the record linkage induced selection mechanism and results from the missed links while the attenuation bias is due to the falsely linked pairs, which contaminates the signal with opposite contributions owing to the mismatch in treatment groups of some falsely linked pairs and, with effect dilution owing to the discrepancies between potential outcomes of all falsely linked pairs. We remind the reader of the biases definitions given in § 4:

ζ^NS=1NS1∑i,j{Si,jTiμ1(Xi)}−1NS0∑i,j{Si,j(1−Ti)μ0(Xi)}−τA,\widehat{\zeta}_{NS}=\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}T_i\mu_1(X_i)\}-\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-T_i)\mu_0(X_i)\}-\tau_A,
ξ^NS=1NS1∑i,j{Si,jTi(Yj−μ1(Xj))}−1NS0∑i,j{Si,j(1−Ti)(Yj−μ0(Xj))}+1NS0∑i,j{Si,j(1−Ti)(μ0(Xi)−μ0(Xj))}−1NS1∑i,j{Si,jTi(μ1(Xi)−μ1(Xj))}.\begin{aligned} \widehat{\xi}_{NS}={}&\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}T_i(Y_j-\mu_1(X_j))\}-\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-T_i)(Y_j-\mu_0(X_j))\}\\ &+\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-T_i)(\mu_0(X_i)-\mu_0(X_j))\}-\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}T_i(\mu_1(X_i)-\mu_1(X_j))\}. \end{aligned}

Note that μt(Xi)\mu_t(X_i) does not depend on the actual treatment the unit received under conditional exchangeability (see § 3.3, which holds in the selected data given XiX_i for both correctly and incorrectly linked pairs). The same property applies to μt(Xj)\mu_t(X_j) given XjX_j. However in practice, we cannot condition on XjX_j which is unobserved unless either Assumption 4 or the point of Remark 2 holds.

The sampling bias is the difference in ATE between data set A and its subset S

Under assumption 6,

E[ζ^NS]=E[1NS1∑i,j{Si,jTiμ1(Xi)}−1NS0∑i,j{Si,j(1−Ti)μ0(Xi)}−τA]=E[1NS1∑i,jSi,jTiμ1(Xi)]−E[1NS0∑i,jSi,j(1−Ti)μ0(Xi)]−τA=μ1SA−μ0SA−(μ1A−μ0A)=τSA−τA.\begin{aligned} \mathbb{E}[\widehat{\zeta}_{NS}] = \mathbb{E}\left[\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}T_i\mu_1(X_i)\}-\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-T_i)\mu_0(X_i)\}-\tau_A\right] \\ &= \mathbb{E}\left[\frac{1}{N_{S1}}\sum_{i,j}S_{i,j}T_i\mu_1(X_i)\right]-\mathbb{E}\left[\frac{1}{N_{S0}}\sum_{i,j}S_{i,j}(1-T_i)\mu_0(X_i)\right]-\tau_A \\ &=\mu_{1S_A}-\mu_{0S_A}-(\mu_{1A}-\mu_{0A}) \\ &=\tau_{S_A}-\tau_A. \end{aligned}

The attenuation bias reflects the impact of falsely linked data on the causal inference

Falsely linked data impact causal inference, with adversarial contributions from linking records that pertain to different treatment groups and, with discrepancies from linking records that differ in covariates:

ξ^NS=1NS1∑i,j{Si,jTi(Yj−μ1(Xj))}−1NS0∑i,j{Si,j(1−Ti)(Yj−μ0(Xj))}(8)\widehat{\xi}_{NS}=\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}T_i(Y_j-\mu_1(X_j))\}-\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-T_i)(Y_j-\mu_0(X_j))\} \tag*{(8)}
+1NS0∑i,j{Si,j(1−Ti)(μ0(Xi)−μ0(Xj))}−1NS1∑i,j{Si,jTi(μ1(Xi)−μ1(Xj))}(9)+\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-T_i)(\mu_0(X_i)-\mu_0(X_j))\}-\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}T_i(\mu_1(X_i)-\mu_1(X_j))\} \tag*{(9)}
=1NS1∑i,j{Si,jΔi,jTi(Yj−μ1(Xj))}(10)=\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}\Delta_{i,j}T_i(Y_j-\mu_1(X_j))\} \tag*{(10)}
−1NS0∑i,j{Si,jΔi,j(1−Ti)(Yj−μ0(Xj))}(11)-\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}\Delta_{i,j}(1-T_i)(Y_j-\mu_0(X_j))\} \tag*{(11)}
+1NS1∑i,j{Si,j(1−Δi,j)Ti(Yj−μ1(Xj))}(12)+\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}(1-\Delta_{i,j})T_i(Y_j-\mu_1(X_j))\} \tag*{(12)}
−1NS0∑i,j{Si,j(1−Δi,j)(1−Ti)(Yj−μ0(Xj))}(13)-\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-\Delta_{i,j})(1-T_i)(Y_j-\mu_0(X_j))\} \tag*{(13)}
+1NS0∑i,j{Si,jΔi,j(1−Ti)(μ0(Xi)−μ0(Xj))}(14)+\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}\Delta_{i,j}(1-T_i)(\mu_0(X_i)-\mu_0(X_j))\} \tag*{(14)}
−1NS1∑i,j{Si,jΔi,jTi(μ1(Xi)−μ1(Xj))}(15)-\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}\Delta_{i,j}T_i(\mu_1(X_i)-\mu_1(X_j))\} \tag*{(15)}
+1NS0∑i,j{Si,j(1−Δi,j)(1−Ti)(μ0(Xi)−μ0(Xj))}(16)+\frac{1}{N_{S0}}\sum_{i,j}\{S_{i,j}(1-\Delta_{i,j})(1-T_i)(\mu_0(X_i)-\mu_0(X_j))\} \tag*{(16)}
−1NS1∑i,j{Si,j(1−Δi,j)Ti(μ1(Xi)−μ1(Xj))}.(17)-\frac{1}{N_{S1}}\sum_{i,j}\{S_{i,j}(1-\Delta_{i,j})T_i(\mu_1(X_i)-\mu_1(X_j))\}. \tag*{(17)}

Proof of Theorems

Components (8) and (9) separate into TP/FP in (9)/(10) and (11)/(12) respectively. By consistency in the source population (Assumption 2), there is no difference between observed and potential outcomes in TP, and component (9) vanishes. Likewise, since TP refer to the same individuals, their potential outcomes in A\mathcal{A} and in B\mathcal{B} coincide, and component (11) vanishes. In contrast, FP between records from different treatment groups induce opposite contributions (10) and, all FP induce bias through the discrepancy between their baseline and follow-up potential outcomes (12).

From the difference between observed and potential outcomes of falsely linked records from different treatment groups arises adversarial contributions; for each treatment group from lines (10):

E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}(Yj−μt(Xj))]=E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=t}(Yj−μt(Xj))]+E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}(Yj−μt(Xj))]=E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=t}{E[Yj∣Tj=t,Xj]−μt(Xj)}]+E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}{E[Yj∣Tj=1−t,Xj]−μt(Xj)}]=E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=t}{E[Yj(t)∣Xj]−μt(Xj)}]+E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}{E[Yj(1−t)∣Xj]−μt(Xj)}]=E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=t}(μt(Xj)−μt(Xj))]+E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}(μ1−t(Xj)−μt(Xj))]=E[1NSt∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}(μ1−t(Xj)−μt(Xj))]=E[NS∇t,1−tNSt1NS∇t,1−t∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}(μ1−t(Xj)−μt(Xj))]=E[NS∇t,1−tNSt(μ1−tS∇B−μtS∇B)].\begin{aligned} &\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\left(Y_j-\mu_t(X_j)\right)\right]\\ ={}&\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=t\}}\left(Y_j-\mu_t(X_j)\right)\right]\\ &+\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=1-t\}}\left(Y_j-\mu_t(X_j)\right)\right]\\ ={}&\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=t\}}\left\{\mathbb{E}[Y_j\mid T_j=t,X_j]-\mu_t(X_j)\right\}\right]\\ &+\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=1-t\}}\left\{\mathbb{E}[Y_j\mid T_j=1-t,X_j]-\mu_t(X_j)\right\}\right]\\ ={}&\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=t\}}\left\{\mathbb{E}[Y_j(t)\mid X_j]-\mu_t(X_j)\right\}\right]\\ &+\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=1-t\}}\left\{\mathbb{E}[Y_j(1-t)\mid X_j]-\mu_t(X_j)\right\}\right]\\ ={}&\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=t\}}\left(\mu_t(X_j)-\mu_t(X_j)\right)\right]\\ &+\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=1-t\}}\left(\mu_{1-t}(X_j)-\mu_t(X_j)\right)\right]\\ ={}&\mathbb{E}\left[\frac{1}{N_{St}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=1-t\}}\left(\mu_{1-t}(X_j)-\mu_t(X_j)\right)\right]\\ ={}&\mathbb{E}\left[\frac{N_{S\nabla t,1-t}}{N_{St}}\frac{1}{N_{S\nabla t,1-t}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbb{1}_{\{T_i=t\}}\mathbb{1}_{\{T_j=1-t\}}\left(\mu_{1-t}(X_j)-\mu_t(X_j)\right)\right]\\ ={}&\mathbb{E}\left[\frac{N_{S\nabla t,1-t}}{N_{St}}\left(\mu_{1-t}S\nabla_{\mathcal{B}}-\mu_tS\nabla_{\mathcal{B}}\right)\right]. \end{aligned}

We use the consistency of the ATE estimator (Assumption 8) to replace its empirical average by its expectation, and the consistency in the source population (Assumption 2) to replace observed outcomes by potential outcomes.

Proof of Theorems

Then, component (10) of ξ^NS\widehat{\xi}_{NS} develops into:

E[1NS1∑i,jSi,j(1−Δi,j)1{Ti=1}(Yj−μ1(Xj))]−E[1NS0∑i,jSi,j(1−Δi,j)1{Ti=0}(Yj−μ0(Xj))]=E[NS∇1,0NS1](μ0S∇B−μ1S∇B)−E[NS∇0,1NS0](μ1S∇B−μ0S∇B)=E[NS∇1,0NS1+NS∇0,1NS0](μ0S∇B−μ1S∇B).\begin{aligned} \mathbb{E}\left[\frac{1}{N_{S1}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=1\}}(Y_j-\mu_1(X_j))\right]-\mathbb{E}\left[\frac{1}{N_{S0}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=0\}}(Y_j-\mu_0(X_j))\right] \\ &=\mathbb{E}\left[\frac{N_{S\nabla1,0}}{N_{S1}}\right](\mu_0S\nabla_B-\mu_1S\nabla_B)-\mathbb{E}\left[\frac{N_{S\nabla0,1}}{N_{S0}}\right](\mu_1S\nabla_B-\mu_0S\nabla_B) \\ &=\mathbb{E}\left[\frac{N_{S\nabla1,0}}{N_{S1}}+\frac{N_{S\nabla0,1}}{N_{S0}}\right](\mu_0S\nabla_B-\mu_1S\nabla_B). \end{aligned}

We observe the opposite contributions, reflecting the opposite of the treatment effect. It arises when falsely linked records belong to different treatment groups. This bias involves the potential outcomes of records from the follow-up data that are falsely linked, defined over the XX from B\mathcal{B} which we do not have access to.

From the information discrepancy in falsely linked records arises a difference between potential outcomes from A\mathcal{A} and from B\mathcal{B}; for each treatment group from lines (12):

E[1NS ⁣t∑i,jSi,j(1−Δi,j)1{Ti=t}(μt(Xi)−μt(Xj))]=E[1NS ⁣t∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=t}(μt(Xi)−μt(Xj))]+E[1NS ⁣t∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}(μt(Xi)−μt(Xj))]=E[NS∇t,tNS ⁣t1NS∇t,t∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=t}(μt(Xi)−μt(Xj))]+E[NS∇t,1−tNS ⁣t1NS∇t,1−t∑i,jSi,j(1−Δi,j)1{Ti=t}1{Tj=1−t}(μt(Xi)−μt(Xj))]=E[NS∇t,tNS ⁣t](μtS∇A−μtS∇B)+E[NS∇t,1−tNS ⁣t](μtS∇A−μtS∇B)=E[NS∇t,t+NS∇t,1−tNS ⁣t](μtS∇A−μtS∇B)=E[FDP⁡t](μtS∇A−μtS∇B).\begin{aligned} \mathbb{E}\left[\frac{1}{N_{S\!t}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=t\}}(\mu_t(X_i)-\mu_t(X_j))\right] \\ &=\mathbb{E}\left[\frac{1}{N_{S\!t}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=t\}}\mathbf{1}_{\{T_j=t\}}(\mu_t(X_i)-\mu_t(X_j))\right] \\ &+\mathbb{E}\left[\frac{1}{N_{S\!t}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=t\}}\mathbf{1}_{\{T_j=1-t\}}(\mu_t(X_i)-\mu_t(X_j))\right] \\ &=\mathbb{E}\left[\frac{N_{S\nabla t,t}}{N_{S\!t}}\frac{1}{N_{S\nabla t,t}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=t\}}\mathbf{1}_{\{T_j=t\}}(\mu_t(X_i)-\mu_t(X_j))\right] \\ &+\mathbb{E}\left[\frac{N_{S\nabla t,1-t}}{N_{S\!t}}\frac{1}{N_{S\nabla t,1-t}}\sum_{i,j}S_{i,j}(1-\Delta_{i,j})\mathbf{1}_{\{T_i=t\}}\mathbf{1}_{\{T_j=1-t\}}(\mu_t(X_i)-\mu_t(X_j))\right] \\ &=\mathbb{E}\left[\frac{N_{S\nabla t,t}}{N_{S\!t}}\right](\mu_tS\nabla_A-\mu_tS\nabla_B) \\ &+\mathbb{E}\left[\frac{N_{S\nabla t,1-t}}{N_{S\!t}}\right](\mu_tS\nabla_A-\mu_tS\nabla_B) \\ &=\mathbb{E}\left[\frac{N_{S\nabla t,t}+N_{S\nabla t,1-t}}{N_{S\!t}}\right](\mu_tS\nabla_A-\mu_tS\nabla_B) \\ &=\mathbb{E}\left[\operatorname{FDP}_t\right](\mu_tS\nabla_A-\mu_tS\nabla_B). \end{aligned}

Drawing on results from the matching literature, under Remark 2 and assumptions 9 and 10, this discrepancy is stochastically bounded: (μtS∇A−μtS∇B)=O(1NS ⁣t1/k)(\mu_tS\nabla_A-\mu_tS\nabla_B)=\mathcal{O}\left(\frac{1}{N_{S\!t}^{1/k}}\right).

Assumption 9. The potential outcomes satisfy standard regularity conditions, ensuring smoothness of μt(X)\mu_t(X).

Assumption 10. The covariates X∈RkX\in\mathbb{R}^{k} have compact, convex support, and density bounded away from zero.

The stochastic order of this bias is derived from the proof of theorem 1 in [1].

Proof of Theorems

Then, component (12) of ξ^NS\widehat{\xi}_{NS} develops into:

E[1NS0∑i,jSi,j(1−Δi,j)(1−Ti)(μ0(Xi)−μ0(Xj))]−E[1NS1∑i,jSi,j(1−Δi,j)Ti(μ1(Xi)−μ1(Xj))]=E[FDP0](μ0S∇A−μ0S∇B)−E[FDP1](μ1S∇A−μ1S∇B).\begin{aligned} &\mathbb{E}\left[\frac{1}{N_{S0}}\sum_{i,j} S_{i,j}(1-\Delta_{i,j})(1-T_i)\left(\mu_0(X_i)-\mu_0(X_j)\right)\right] \\ &\quad-\mathbb{E}\left[\frac{1}{N_{S1}}\sum_{i,j} S_{i,j}(1-\Delta_{i,j})T_i\left(\mu_1(X_i)-\mu_1(X_j)\right)\right] \\ &=\mathbb{E}[\mathrm{FDP}_0]\left(\mu_{0S\nabla_A}-\mu_{0S\nabla_B}\right)-\mathbb{E}[\mathrm{FDP}_1]\left(\mu_{1S\nabla_A}-\mu_{1S\nabla_B}\right). \end{aligned}

This result explains the effect dilution due to incorrect information being used. It results from the discrepancy between potential outcomes from A\mathcal{A} and from B\mathcal{B} in any falsely linked pairs.

To summarise the Appendix Section A.2.1 and Section A.2.2, we studied the bias:

E[τ^S]−τA=E[ζ^NS]+E[ξ^NS],\mathbb{E}[\widehat{\tau}_S]-\tau_A=\mathbb{E}[\widehat{\zeta}_{NS}]+\mathbb{E}[\widehat{\xi}_{NS}],

and defined the sampling bias due to missed links:

E[ζ^NS]=μ1SA−μ0SA−(μ1A−μ0A)=τSA−τA.\mathbb{E}[\widehat{\zeta}_{NS}]=\mu_{1S_A}-\mu_{0S_A}-(\mu_{1A}-\mu_{0A})=\tau_{S_A}-\tau_A.

as well as the attenuation bias due to falsely linked data:

E[ξ^NS]=E[NS∇1,0NS1+NS∇0,1NS0](μ0SB−μ1SB)+E[FDP0](μ0SA−μ0SB)−E[FDP1](μ1SA−μ1SB).\mathbb{E}[\widehat{\xi}_{NS}]=\mathbb{E}\left[\frac{N_{S\nabla1,0}}{N_{S1}}+\frac{N_{S\nabla0,1}}{N_{S0}}\right](\mu_{0S_B}-\mu_{1S_B})+\mathbb{E}[\mathrm{FDP}_0](\mu_{0S_A}-\mu_{0S_B})-\mathbb{E}[\mathrm{FDP}_1](\mu_{1S_A}-\mu_{1S_B}).

Note that ζ^NS\widehat{\zeta}_{NS} only depends on information from A\mathcal{A} while ξ^NS\widehat{\xi}_{NS} depends on A\mathcal{A} and on B\mathcal{B}. From assumptions 1 and 3, X⊥ ⁣ ⁣⊥ΔX\perp\!\!\perp\Delta and X⊥ ⁣ ⁣⊥Δ∣SX\perp\!\!\perp\Delta\mid S and the potential outcome means are the same among correctly and incorrectly linked profiles (μtS∇=μtSΔ=μtS\mu_{tS\nabla}=\mu_{tS\Delta}=\mu_{tS}). Besides, X⊥̸ ⁣ ⁣⊥SX\not\perp\!\!\perp S and the impact of linkage errors comes from that dependence. Therefore,

E[ξ^NS]=E[NS∇1,0NS1+NS∇0,1NS0](μ0SB−μ1SB)+E[FDP0](μ0SA−μ0SB)−E[FDP1](μ1SA−μ1SB).\mathbb{E}[\widehat{\xi}_{NS}]=\mathbb{E}\left[\frac{N_{S\nabla1,0}}{N_{S1}}+\frac{N_{S\nabla0,1}}{N_{S0}}\right](\mu_{0S_B}-\mu_{1S_B})+\mathbb{E}[\mathrm{FDP}_0](\mu_{0S_A}-\mu_{0S_B})-\mathbb{E}[\mathrm{FDP}_1](\mu_{1S_A}-\mu_{1S_B}).

Moreover, under assumptions 1 and 7, the proportion of opposite contributions E[NS∇1,0NS1+NS∇0,1NS0]\mathbb{E}\left[\frac{N_{S\nabla1,0}}{N_{S1}}+\frac{N_{S\nabla0,1}}{N_{S0}}\right] is bounded by E[FDP]\mathbb{E}[\mathrm{FDP}], with the bound saturated in the edge case of Assumption 7 (when half of the falsely linked data links records from different treatment groups). We obtain that result developing three steps.

First, Assumption 7 implies

E[NS∇t,1−tNSt]≤12E[FDPt].\mathbb{E}\left[\frac{N_{S\nabla t,1-t}}{N_{St}}\right]\leq\frac{1}{2}\mathbb{E}[\mathrm{FDP}_t].

Second, we notice that

FDP=FPFP+TP=NS∇NS, hence FDPt=NS∇tNSt=NS∇t,t+NS∇t,1−tNSt.\mathrm{FDP}=\frac{FP}{FP+TP}=\frac{N_{S\nabla}}{N_S},\ \text{hence}\ \mathrm{FDP}_t=\frac{N_{S\nabla t}}{N_{St}}=\frac{N_{S\nabla t,t}+N_{S\nabla t,1-t}}{N_{St}}.

Third, for any t∈{0,1}t \in\{0,1\}

E[FDPt]=E[E[FDPt∣X]]=E[E[∑i,jSi,j(1−Δi,j)∑i,jSi,j∣T=t,X]]=E[E[∑i,jSi,j(1−Δi,j)∑i,jSi,j|X]]=E[E[FDP∣X]]=E[FDP]\begin{aligned} \mathbb{E}[\mathrm{FDP}_t] = \mathbb{E}\left[\mathbb{E}[\mathrm{FDP}_t \mid X]\right] = \mathbb{E}\left[\mathbb{E}\left[\left.\frac{\sum_{i,j} S_{i,j}(1-\Delta_{i,j})}{\sum_{i,j} S_{i,j}}\right|T=t,X\right]\right] \\ &= \mathbb{E}\left[\mathbb{E}\left[\frac{\sum_{i,j} S_{i,j}(1-\Delta_{i,j})}{\sum_{i,j} S_{i,j}}\middle|X\right]\right] = \mathbb{E}[\mathbb{E}[\mathrm{FDP}\mid X]] = \mathbb{E}[\mathrm{FDP}] \end{aligned}

since (X,T)⊥ ⁣ ⁣ ⁣⊥Δ(X,T)\perp\!\!\!\perp\Delta by Assumption 1 and since S⊥ ⁣ ⁣ ⁣⊥T∣XS\perp\!\!\!\perp T\mid X.

As a result, we can write:

E[τ^S]=τA+τSA−τA−τSBE[NS∇1,0NS1+NS∇0,1NS0]+E[FDP0](μ0SA−μ0SB)−E[FDP1](μ1SA−μ1SB)=τSA−τSBE[NS∇1,0NS1+NS∇0,1NS0]+E[FDP](μ0SA−μ0SB)−E[FDP](μ1SA−μ1SB)=τSA−τSBE[NS∇1,0NS1+NS∇0,1NS0]+E[FDP](τSB−τSA)=τSA(1−E[FDP])+τSB(E[FDP]−E[NS∇1,0NS1+NS∇0,1NS0])(18)\begin{aligned} \mathbb{E}[\widehat{\tau}_S] = \tau_A + \tau_{S_A} - \tau_A - \tau_{S_B}\mathbb{E}\left[\frac{N_{S}\nabla_{1,0}}{N_{S1}}+\frac{N_{S}\nabla_{0,1}}{N_{S0}}\right] + \mathbb{E}[\mathrm{FDP}_0](\mu_{0S_A}-\mu_{0S_B})-\mathbb{E}[\mathrm{FDP}_1](\mu_{1S_A}-\mu_{1S_B}) \\ &= \tau_{S_A}-\tau_{S_B}\mathbb{E}\left[\frac{N_{S}\nabla_{1,0}}{N_{S1}}+\frac{N_{S}\nabla_{0,1}}{N_{S0}}\right]+\mathbb{E}[\mathrm{FDP}](\mu_{0S_A}-\mu_{0S_B})-\mathbb{E}[\mathrm{FDP}](\mu_{1S_A}-\mu_{1S_B}) \\ &= \tau_{S_A}-\tau_{S_B}\mathbb{E}\left[\frac{N_{S}\nabla_{1,0}}{N_{S1}}+\frac{N_{S}\nabla_{0,1}}{N_{S0}}\right]+\mathbb{E}[\mathrm{FDP}](\tau_{S_B}-\tau_{S_A}) \\ &= \tau_{S_A}(1-\mathbb{E}[\mathrm{FDP}])+\tau_{S_B}\left(\mathbb{E}[\mathrm{FDP}]-\mathbb{E}\left[\frac{N_{S}\nabla_{1,0}}{N_{S1}}+\frac{N_{S}\nabla_{0,1}}{N_{S0}}\right]\right) \tag*{(18)} \end{aligned}

meaning τ^S\widehat{\tau}_S is attenuated compared to τSA\tau_{S_A} by E[FDP]\mathbb{E}[\mathrm{FDP}], attenuation that is compensated up to E[FDP]\mathbb{E}[\mathrm{FDP}] through the contribution of τSB\tau_{S_B}.

As noted in Remark 1 under Assumption 1, distributions of covariates are the same in the selected records from AA and BB, thus the information discrepancy disappears on average hence for any t∈{0,1}t \in\{0,1\}, μtSA=μtSB\mu_{tS_A}=\mu_{tS_B} and τSA=τSB=τS\tau_{S_A}=\tau_{S_B}=\tau_S. Thereby, Equation (18) becomes

E[τ^S]=τS(1−E[NS∇1,0NS1+NS∇0,1NS0])\mathbb{E}[\widehat{\tau}_S]=\tau_S\left(1-\mathbb{E}\left[\frac{N_{S}\nabla_{1,0}}{N_{S1}}+\frac{N_{S}\nabla_{0,1}}{N_{S0}}\right]\right)

which in the edge case of Assumption 7 (the worst case) writes:

E[τ^S]=τS(1−E[FDP]).\mathbb{E}[\widehat{\tau}_S]=\tau_S(1-\mathbb{E}[\mathrm{FDP}]).

The effect on the baseline population is recoverable from sampling bias

The causal effect is SS-recoverable from the record linkage induced selection mechanism SS presented in Figure 2 under Assumptions 2 and 4. Indeed, point (i)(i) of the selection backdoor criterion Definition 1 holds under Assumption 2 discussed in § 3.3; point (ii)(ii) holds under Assumption 4 or alternatively under the point of Remark 2, we elaborated on that point in § 3.4; point (iii)(iii) and point (iv)(iv) hold by definition of the causal record linkage task, as XX is measured in both, the selected (biased) population and the baseline population.

The potential outcome distribution is SS-recoverable from the selection biased data and the causal effect in the baseline is identified through the following transport formula (see [12, 68, 11]):

pA(Yi(t))=∑xpA(Yi(t)∣Xi=x)pA(Xi=x)(19)p_A(Y_i(t))=\sum_x p_A(Y_i(t)\mid X_i=x)p_A(X_i=x) \tag*{(19)}
=∑xpS(Yi(t)∣Xi=x)pA(Xi=x),(20)=\sum_x p_S(Y_i(t)\mid X_i=x)p_A(X_i=x), \tag*{(20)}

which states that we can identify the causal effect in the baseline population through the effect in the specific selected population SS.

By conditional exchangeability and, by distributional consistency (under Assumption 4 or alternatively under the point of Remark 2), the distribution in the selected sample further develops:

pA(Yi(t))=∑xpS(Yi(t)∣Xi=x)pA(Xi=x)(21)p_A(Y_i(t)) = \sum_x p_S(Y_i(t)\mid X_i=x)p_A(X_i=x) \tag*{(21)}
=∑xpS(Yi(t)∣Ti=t,Xi=x)pA(Xi=x)(22)= \sum_x p_S(Y_i(t)\mid T_i=t, X_i=x)p_A(X_i=x) \tag*{(22)}
=∑xpS(Yj∣Ti=t,Xi=x)pA(Xi=x).(23)= \sum_x p_S(Y_j\mid T_i=t, X_i=x)p_A(X_i=x). \tag*{(23)}

From line (18) we can estimate the causal effect in the baseline population by estimating the effect in the specific selected population SS and average it over XX in the baseline A\mathcal{A}.

Finally, under Assumption 5, an estimator of pS(Yj∣Ti=t,Xi=x)p_S(Y_j\mid T_i=t, X_i=x) over x∈Sx \in S could approximate pA(Yi(t)∣Xi=x)p_A(Y_i(t)\mid X_i=x) over x∈Ax \in\mathcal{A}, even when the covariate distributions in SS and A\mathcal{A} differ. Indeed, the lack of overlap can be encompassed under random internal violations of positivity via interpolation—smoothing over the surrounded region of non-positivity—though extrapolation is ill-advised [91, 60]. This last assumption guarantees that the resultant expression is estimable from the available data [20]. Hence,

τA=EA[τS(X)]\tau_{\mathcal{A}} = \mathbb{E}_{\mathcal{A}}[\tau_S(X)]

may be recovered using methods for generalised treatment effect estimation.

To recover the effect in practice though, we need to eliminate the biases developed in Appendix Section A.2.1 and Section A.2.2 (i.e. we need to ensure Assumption 4), before transporting the effect estimated in SS to A\mathcal{A} as justified here in Appendix Section A.2.3.

Author contributions statement

KR carried out the investigation conceptualised by SP. All authors drew up the simulation framework. KR implemented the estimation methods and conducted the data application and the analysis. She prepared the first manuscript draft, reviewed and edited by MH, MW and SP. All authors read and approved this manuscript.

Acknowledgements

The authors thank the anonymous reviewers for their valuable suggestions.

Data availability

The experiments and the data sets are available on GitHub.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Conflict of interest

None declared.

References

  1. [1]Alberto Abadie and Guido W. Imbens. Large sample properties of matching estimators for average treatment effects. Econometrica, 74(1):235–267, 2006. doi: https://doi.org/10.1111/j.1468-0262.2006.00655.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1468-0262.2006.00655.x.
  2. [2]Alberto Abadie and Guido W. Imbens. On the failure of the bootstrap for matching estimators. Econometrica, 76(6):1537–1557, 2008. ISSN 00129682, 14680262. URL http://www.jstor.org/stable/40056514.
  3. [3]Alberto Abadie and Guido W. Imbens. Bias-corrected matching estimators for average treatment effects. Journal of Business & Economic Statistics, 29(1):1–11, 2011. doi: 10.1198/jbes.2009.07333. URL https://doi.org/10.1198/jbes.2009.07333.
  4. [4]Alberto Abadie and Guido W. Imbens. A martingale representation for matching estimators. Journal of the American Statistical Association, 107(498):833–843, 2012. doi: 10.1080/01621459.2012.682537. URL https://doi.org/10.1080/01621459.2012.682537.
  5. [5]Alberto Abadie and Guido W. Imbens. Matching on the estimated propensity score. Econometrica, 84(2):781–807, 2016. doi: https://doi.org/10.3982/ECTA11293. URL https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA11293.
  6. [6]Ahmad H Alghadir, Sami A Gabr, and Einas Al-Eisa. Physical activity and lifestyle effects on bone mineral density among young adults: sociodemographic and biochemical analysis. Journal of physical therapy science, 2015. doi: 10.1589/jpts.27.2261.DOI
  7. [7]Anastasios N. Angelopoulos, John C. Duchi, and Tijana Zrnic. Ppi++: Efficient prediction-powered inference, 2024. URL https://arxiv.org/abs/2311.01453.
  8. [8]Joshua D. Angrist, Guido W. Imbens, and Donald B. Rubin. Identification of Causal Effects Using Instrumental Variables. Journal of the American Statistical Association, 91(434):444–455, 1996. ISSN 0162-1459. doi: 10.2307/2291629.DOI
  9. [9]Heejung Bang and James M Robins. Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4):962–973, 2005. doi: 10.1111/j.1541-0420.2005.00377.x.DOI
  10. [10]Elias Bareinboim and Judea Pearl. Controlling selection bias in causal inference. In Neil D. Lawrence and Mark Girolami, editors, Proceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics, volume 22 of Proceedings of Machine Learning Research, pages 100–108, La Palma, Canary Islands, 21–23 Apr 2012. PMLR. URL https://proceedings.mlr.press/v22/bareinboim12.html.
  11. [11]Elias Bareinboim and Judea Pearl. Causal inference and the data-fusion problem. Proceedings of the National Academy of Sciences, 113(27):7345–7352, 2016. doi: 10.1073/pnas.1510507113. URL https://www.pnas.org/doi/abs/10.1073/pnas.1510507113.DOI
  12. [12]Elias Bareinboim, Jin Tian, and Judea Pearl. Recovering from selection bias in causal and statistical inference. Proceedings of the AAAI Conference on Artificial Intelligence, 28(1), Jun. 2014. doi: 10.1609/aaai.v28i1.9074. URL https://ojs.aaai.org/index.php/AAAI/article/view/9074.DOI
  13. [13]Thomas R. Belin and Donald B. Rubin. A method for calibrating false-match rates in record linkage. Journal of the American Statistical Association, 90(430):694–707, 1995. ISSN 01621459, 1537274X. URL http://www.jstor.org/stable/2291082.DOI
  14. [14]Chris M. Bishop. Training with noise is equivalent to tikhonov regularization. Neural Computation, 7(1):108–116, 01 1995. ISSN 0899-7667. doi: 10.1162/neco.1995.7.1.108. URL https://doi.org/10.1162/neco.1995.7.1.108.
  15. [15]Tony Blakely and Clare Salmond. Probabilistic record linkage and a method to calculate the positive predictive value. International Journal of Epidemiology, 31(6):1246–1252, 12 2002. ISSN 0300-5771. doi: 10.1093/ije/31.6.1246. URL https://doi.org/10.1093/ije/31.6.1246.
  16. [16]Riccardo Cadei, Ilker Demirel, Piersilvio De Bartolomeis, Lukas Lindorfer, Sylvia Cremer, Cordelia Schmid, and Francesco Locatello. Prediction-powered causal inferences, 2025. URL https://arxiv.org/abs/2502.06343.
  17. [17]S. R. Cole and M. A. Hernán. Constructing inverse probability weights for marginal structural models. American journal of epidemiology, 168(6):656–664, 2008. doi: https://doi.org/10.1093/aje/kwn164.
  18. [18]Stephen R. Cole and Elizabeth A. Stuart. Generalizing evidence from randomized clinical trials to target populations: The actg 320 trial. American Journal of Epidemiology, 172(1):107–115, 06 2010. ISSN 0002-9262. doi: 10.1093/aje/kwq084. URL https://doi.org/10.1093/aje/kwq084.
  19. [19]Bénédicte Colnet, Imke Mayer, Guanhua Chen, Awa Dieng, Ruohong Li, Gaël Varoquaux, Jean-Philippe Vert, Julie Josse, and Shu Yang. Causal Inference Methods for Combining Randomized Trials and Observational Studies: A Review. Statistical Science, 39(1):165 – 191, 2024. doi: 10.1214/23-STS889. URL https://doi.org/10.1214/23-STS889.
  20. [20]Juan Correa and Elias Bareinboim. Causal effect identification by adjustment under confounding and selection biases. Proceedings of the AAAI Conference on Artificial Intelligence, 31(1), Feb. 2017. doi: 10.1609/aaai.v31i1.11060. URL https://ojs.aaai.org/index.php/AAAI/article/view/11060.DOI
  21. [21]Issa J. Dahabreh, Sarah E. Robertson, Jon A. Steingrimsson, Elizabeth A. Stuart, and Miguel A. Hernan. Extending inferences from a randomized trial to a new target population. Statistics in Medicine, 2019. doi: https://doi.org/10.1002/sim.8426.
  22. [22]Issa J Dahabreh, Sebastien J-P A Haneuse, James M Robins, Sarah E Robertson, Ashley L Buchanan, Elizabeth A Stuart, and Miguel A Hernán. Study designs for extending causal inferences from a randomized trial to a target population. American Journal of Epidemiology, 190(8):1632–1642, 12 2020. ISSN 0002-9262. doi: 10.1093/aje/kwaa270. URL https://doi.org/10.1093/aje/kwaa270.
  23. [23]Abel Dasylva and Arthur Goussanou. Estimating linkage errors under regularity conditions. In Proceedings of the Survey Research Methods Section. American Statistical Association, 2020.
  24. [24]Abel Dasylva and Arthur Goussanou. Estimating the false negatives due to blocking in record linkage. Survey Methodology, 47(2), 2021. ISSN 1492-0921.
  25. [25]Abel Dasylva and Arthur Goussanou. On the consistent estimation of linkage errors without training data. Japanese Journal of Statistics and Data Science, 5, 2022. doi: 10.1007/s42081-022-00153-3.DOI
  26. [26]James C. Doidge and Katie L. Harron. Reflections on modern methods: linkage error bias. International Journal of Epidemiology, 48(6):2050–2060, 10 2019. ISSN 0300-5771. doi: 10.1093/ije/dyz203. URL https://doi.org/10.1093/ije/dyz203.
  27. [27]Ashimiyu B Durojaiye, Lisa L Puett, Scott Levin, Matthew Toerper, Nicolette M McGeorge, Kristen L W Webster, Gurmehar S Deol, Hadi Kharrazi, Harold P Lehmann, and Ayse P Gurses. Linking electronic health record and trauma registry data: Assessing the value of probabilistic linkage. Methods of information in medicine, 2018.
  28. [28]Stacie B. Dusetzina, Seth Tyree, Anne-Marie Meyer, Adrian Meyer, Laura Green, and William R. Carpenter. Linking data for health services research: A framework and instructional guide [internet]. https://www.ncbi.nlm.nih.gov/books/NBK253312/, 2014. URL https://www.ncbi.nlm.nih.gov/books/NBK253312/. An Overview of Record Linkage Methods.
  29. [29]Ted Enamorado, Benjamin Fifield, and Kosuke Imai. Using a probabilistic model to assist merging of large-scale administrative records. American Political Science Review, 113(2):353–371, 2019. doi: 10.1017/S0003055418000783.DOI
  30. [30]Kaoutar Ennour-Idrissi, Elizabeth Maunsell, and Caroline Diorio. Effect of physical activity on sex hormones in women: a systematic review and meta-analysis of randomized controlled trials. Breast Cancer Research, 17, 2015. doi: 10.1186/s13058-015-0647-3.
  31. [31]I. P. Fellegi and A. B. Sunter. A theory for record linkage. Journal of the American Statistical Association, 64 (328):1183–1210, 1969.DOI
  32. [32]Patrick Forré and Joris M. Mooij. A mathematical introduction to causality, 2023. URL https://staff.fnwi.uva.nl/j.m.mooij/articles/causality_lecture_notes_2023.pdf.
  33. [33]Chris Frost and Simon G. Thompson. Correcting for regression dilution bias: Comparison of methods for a single predictor variable. Journal of the Royal Statistical Society. Series A (Statistics in Society), 163(2):173–189, 2000. ISSN 09641998, 1467985X. URL http://www.jstor.org/stable/2680496.DOI
  34. [34]Els Goetghebeur, Saskia le Cessie, Bianca De Stavola, Erica EM Moodie, Ingeborg Waernbaum, and “on behalf of” the topic group Causal Inference (TG7) of the STRATOS initiative. Formulating causal questions and principled statistical answers. Statistics in Medicine, 39(30):4922–4948, 2020. doi: https://doi.org/10.1002/sim.8741. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.8741.
  35. [35]Heyang Gong, Chaochao Lu, and Yu Zhang. Distribution-consistency structural causal models, 2024. URL https://arxiv.org/abs/2401.15911.
  36. [36]Noah Greifer. WeightIt: Weighting for Covariate Balance in Observational Studies, 2025. URL https://ngreifer.github.io/WeightIt/. R package version 1.5.0.
  37. [37]Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test. Journal of Machine Learning Research, 13(25):723–773, 2012. URL http://jmlr.org/papers/v13/gretton12a.html.DOI
  38. [38]Jessica Gronsbell, Jianhui Gao, Yaqi Shi, Zachary R. McCaw, and David Cheng. Another look at inference after prediction, 2025. URL https://arxiv.org/abs/2411.19908.
  39. [39]Sharmistha Guha and Jerome P Reiter. Regression-assisted bayesian record linkage for causal inference in observational studies with covariates spread over two files. Journal of statistical planning and inference, 2024.DOI
  40. [40]Sharmistha Guha, Jerome P. Reiter, and Andrea Mercatanti. Bayesian causal inference with bipartite record linkage. Bayesian Analysis, 17(4):1275 – 1299, 2022. doi: 10.1214/21-BA1297. URL https://doi.org/10.1214/21-BA1297.
  41. [41]Trygve Haavelmo. The statistical implications of a system of simultaneous equations. Econometrica, 11(1):1–12, 1943. ISSN 00129682, 14680262. URL http://www.jstor.org/stable/1905714.DOI
  42. [42]Jens Hainmueller. Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. Political Analysis, 20(1):25–46, 2012. doi: 10.1093/pan/mpr025.DOI
  43. [43]Katie L Harron, James C Doidge, Hannah E Knight, Ruth E Gilbert, Harvey Goldstein, David A Cromwell, and Jan H van der Meulen. A guide to evaluating linkage quality for the analysis of linked data. International Journal of Epidemiology, 46(5):1699–1710, 09 2017. ISSN 0300-5771. doi: 10.1093/ije/dyx177. URL https://doi.org/10.1093/ije/dyx177.
  44. [44]James J. Heckman. Sample selection bias as a specification error. Econometrica, 47(1):153–161, 1979. ISSN 00129682, 14680262. URL http://www.jstor.org/stable/1912352.DOI
  45. [45]Miguel A. Hernan and James M. Robins. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC, 2020.
  46. [46]Miguel A. Hernán, Sonia Hernández-Díaz, and James M. Robins. A structural approach to selection bias. Epidemiology, 2004.pubmed.ncbi.nlm.nih.gov/15308962
  47. [47]Theodore Hershberg, Alan Burstein, and Robert Dockhorn. Record linkage. Historical Methods Newsletter, 9(2-3):137–163, 1976. doi: 10.1080/00182494.1976.10112639. URL https://doi.org/10.1080/00182494.1976.10112639.
  48. [48]Andrew Y.Y. Ho and Annie W.C. Kung. Determinants of peak bone mineral density and bone area in young women. Journal of Bone and Mineral Metabolism, 23, 2005. doi: 10.1007/s00774-005-0630-7.DOI
  49. [49]Guido W. Imbens and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press, 1 edition, April 2015. ISBN 978-0-521-88588-1 978-1-139-02575-1. doi: 10.1017/CBO9781139025751.DOI
  50. [50]M. A. Jaro. Probabilistic linkage of large public health data files. Statistics in Medicine, 14(5–7):491–498, 1995.pubmed.ncbi.nlm.nih.gov/7792443
  51. [51]J. E. Johndrow, K. Lum, and D. B. Dunson. Theoretical limits of microclustering for record linkage. Biometrika, 105(2):431–446, 03 2018. ISSN 0006-3444. doi: 10.1093/biomet/asy003. URL https://doi.org/10.1093/biomet/asy003.
  52. [52]Eben Kenah. A potential outcomes approach to selection bias. Epidemiology, 34(6):865–872, 2023. ISSN 1044-3983. doi: 10.1097/ede.0000000000001660. URL http://dx.doi.org/10.1097/EDE.0000000000001660.
  53. [53]Anna Kopiczko, Monika Łopuszańska Dawid, and Karol Gryko. Bone mineral density in young adults: the influence of vitamin d status, biochemical indicators, physical activity and body composition. Archives of Osteoporosis, 15, 2020. doi: 10.1007/s11657-020-0684-0.DOI
  54. [54]Selahattin Kumru, Recep Ozmerdivenli, Suleyman Aydin, Abdullah Yasar, Nermin Kilic, Cem Parmaksiz, and Suna Aydin. Effects of regular physical exercise on serum leptin and androgen concentrations in young women. The Journal of Men’s Health & Gender, 2(2):218–222, 2005. ISSN 1571-8913. doi: https://doi.org/10.1016/j.jmhg.2005.03.002. URL https://www.sciencedirect.com/science/article/pii/S1571891305000488.
  55. [55]P. Lahiri and Michael D. Larsen. Regression analysis with linked data. Journal of the American Statistical Association, 100(469):222–230, 2005. ISSN 01621459. URL http://www.jstor.org/stable/27590532.DOI
  56. [56]Joseph Lam, Andy Boyd, Robin Linacre, Ruth Blackburn, and Katie Harron. Generating synthetic identifiers to support development and evaluation of data linkage methods. International Journal of Population Data Science, 9(1), 07 2024. doi: 10.23889/ijpds.v9i1.2389.10.23889/ijpds.v7i3.1794. URL https://ijpds.org/article/view/2389.
  57. [57]Dasom Lee, Shu Yang, Lin Dong, Xiaofei Wang, Donglin Zeng, and Jianwen Cai. Improving trial generalizability using observational studies. Biometr ics, 79(2):1213–1225, 12 2021. ISSN 0006-341X. doi: 10.1111/biom.13609. URL https://doi.org/10.1111/biom.13609.
  58. [58]Fan Li, Peng Ding, and Fabrizia Mealli. Bayesian causal inference: a critical review. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 381(2247):20220153, 2023. doi: 10.1098/rsta.2022.0153. URL https://royalsocietypublishing.org/doi/abs/10.1098/rsta.2022.0153.DOI
  59. [59]Shanshan Luo, Yechi Zhang, and Wei Li. Multiply robust estimation of causal effects using linked data, 2023. URL https://arxiv.org/abs/2309.08199.
  60. [60]Maxime Léger, Arthur Chatton, Florent Le Borgne, Romain Pirracchio, Sigismond Lasocki, and Yohann Foucher. Causal inference in case of near-violation of positivity: comparison of methods. Biometrical journal, 2022. doi: 10.1002/bimj.202000323.DOI
  61. [61]N. G. Marchant, B. I. P. Rubinstein, and R. C. Steorts. Bayesian graphical entity resolution using exchangeable random partition priors. Journal of Survey Statistics and Methodology, 11(3):569–596, 2023.DOI
  62. [62]Caleb H Miles, Linda Valeri, and Brent Coull. Measurement error-robust causal inference via constructed instrumental variables. Biometrics, 82(2):ujag057, 06 2026. ISSN 1541-0420. doi: 10.1093/biomtc/ujag057. URL https://doi.org/10.1093/biomtc/ujag057.
  63. [63]Len De Nys, Esther F. Ofosu, Gemma C. Ryde, Jenni Connelly, and Anna C. Whittaker. Physical activity influences cortisol and dehydroepiandrosterone (sulfate) levels in older adults: A systematic review and meta-analysis. Journal of Aging and Physical Activity, 31(2):330 – 351, 2023. doi: 10.1123/japa.2021-0501.DOI
  64. [64]Elizabeth L. Ogburn and Tyler J. VanderWeele. Causal Diagrams for Inference. Statistical Science, 29(4):559 – 578, 2014. doi: 10.1214/14-STS501. URL https://doi.org/10.1214/14-STS501.
  65. [65]Ivan Olier, Yiqiang Zhan, Xiaoyu Liang, and Victor Volovici. Causal inference and observational data. BMC Medical Research Methodology, 23, 2023. ISSN 1471-2288. doi: 10.1186/s12874-023-02058-5.DOI
  66. [66]Judea Pearl. Causal diagrams for empirical research. Biometrika, 82(4):669–688, 1995. ISSN 00063444, 14643510. URL http://www.jstor.org/stable/2337329.
  67. [67]Judea Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, 2 edition, 11 2009. ISBN 9780521895606.
  68. [68]Judea Pearl and Elias Bareinboim. External Validity: From Do-Calculus to Transportability Across Populations. Statistical Science, 29(4):579 – 595, 2014. doi: 10.1214/14-STS486. URL https://doi.org/10.1214/14-STS486.
  69. [69]Sean Randall, Adrian Brown, James Boyd, Rainer Schnell, Christian Borgs, and Anna Ferrante. Sociodemographic differences in linkage error: an examination of four large-scale datasets. BMC Health Services Research, 2018. doi: 10.1186/s12913-018-3495-x.DOI
  70. [70]Kyle Raze, Nicole Perales, and Liana Christin Landivar. Non-random assignment of individual identifiers and selection into linked data: Implications for research. Working papers, Center for Economic Studies, U.S. Census Bureau, 2026. URL https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-06.html.
  71. [71]Dean Resnick and Jana Asher. Measurement of type i and type ii record linkage error. In Using and Linking Administrative and Auxiliary Data, Government Statistics Section. Joint Statistical Meetings, 09 2019. URL https://ww2.amstat.org/meetings/proceedings/2019/data/assets/pdf/1199462.pdf.
  72. [72]Thomas S Richardson and James M Robins. Single world intervention graphs (swigs): A unification of the counterfactual and graphical approaches to causality. Technical report, University of Washington, Department of Statistics, 2013. URL https://csss.uw.edu/files/working-papers/2013/wp128.pdf.
  73. [73]Thomas S. Richardson and James M. Robins. Potential outcome and decision theoretic foundations for statistical causality. Journal of Causal Inference, 2023. doi: https://doi.org/10.1515/jci-2022-0012.
  74. [74]Kayané Robach, Michel H Hof, and Mark A van de Wiel. False discovery estimation in record linkage. Statistics in Medicine, 44(23-24):e70292, 2025a. doi: https://doi.org/10.1002/sim.70292. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.70292.
  75. [75]Kayané Robach, Stéphanie L van der Pas, Mark A van de Wiel, and Michel H Hof. A flexible model for record linkage. Journal of the Royal Statistical Society Series C: Applied Statistics, 74(4):1100–1127, 02 2025b. ISSN 0035-9254. doi: 10.1093/jrsssc/qlaf016. URL https://doi.org/10.1093/jrsssc/qlaf016.
  76. [76]James Robins. A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect. Mathematical Modelling, 7(9):1393–1512, 1986. ISSN 0270-0255. doi: https://doi.org/10.1016/0270-0255(86)90088-6. URL https://www.sciencedirect.com/science/article/pii/0270025586900886.
  77. [77]James M. Robins, Miguel Ángel Hernán, and Babette Brumback. Marginal structural models and causal inference in epidemiology. Epidemiology, 11, 2000.pubmed.ncbi.nlm.nih.gov/10955408
  78. [78]Nadja Rutsch, Sara Magliacane, and Stéphanie van der Pas. Finite sample-optimal adjustment sets in linear gaussian causal models, 2025. URL https://arxiv.org/abs/2506.12869.
  79. [79]M. Sadinle. Bayesian estimation of bipartite matchings for record linkage. Journal of the American Statistical Association, 112(518):600–612, 2017.arxiv.org/abs/1601.06630
  80. [80]Stephen Salerno, Jiacheng Miao, Awan Afiaz, Kentaro Hoffman, Anna Neufeld, Qiongshi Lu, Tyler H McCormick, and Jeffrey T Leek. ipd: an r package for conducting inference on predicted data. Bioinformatics, 41 (2), 2025. doi: 10.1093/bioinformatics/btaf055.DOI
  81. [81]Mingyang Shan, Kali S. Thomas, and Roee Gutman. A multiple imputation procedure for record linkage and causal inference to estimate the effects of home-delivered meals. The Annals of Applied Statistics, 15(1):412 – 436, 2021. doi: 10.1214/20-AOAS1397. URL https://doi.org/10.1214/20-AOAS1397.
  82. [82]C. Spearman. The proof and measurement of association between two things. The American Journal of Psychology, 15(1):72–101, 1904. ISSN 00029556. URL http://www.jstor.org/stable/1412159.
  83. [83]Peter Spirtes, Clark Glymour, and Richard Scheines. Causation, Prediction, and Search. The MIT Press, 01 2001. ISBN 9780262284158. doi: 10.7551/mitpress/1754.001.0001. URL https://doi.org/10.7551/mitpress/1754.001.0001.
  84. [84]John D. Storey and Robert Tibshirani. Statistical significance for genomewide studies. Proceedings of the National Academy of Sciences, 100(16):9440–9445, 2003. doi: 10.1073/pnas.1530509100. URL https://www.pnas.org/doi/abs/10.1073/pnas.1530509100.DOI
  85. [85]Elizabeth A. Stuart, Stephen R. Cole, Catherine P. Bradshaw, and Philip J. Leaf. The use of propensity scores to assess the generalizability of results from randomized trials. Journal of the Royal Statistical Society Series A: Statistics in Society, 174(2):369–386, 11 2010. ISSN 0964-1998. doi: 10.1111/j.1467-985X.2010.00673.x. URL https://doi.org/10.1111/j.1467-985X.2010.00673.x.
  86. [86]A. Tancredi and B. Liseo. A hierarchical Bayesian approach to record linkage and population size problems. The Annals of Applied Statistics, 5(2B), 2011.
  87. [87]Andrea Tancredi and Brunero Liseo. Regression analysis with linked data: problems and possible solutions. Statistica, 75(1):19–35, Jan. 2015. doi: 10.6092/issn.1973-2201/5821. URL https://rivista-statistica.unibo.it/article/view/5821.DOI
  88. [88]Elizabeth Tipton. Improving generalizations from experiments using propensity score subclassification: Assumptions, properties, and contexts. Journal of Educational and Behavioral Statistics, 38(3):239–266, 2013. ISSN 10769986, 19351054. URL http://www.jstor.org/stable/41999424.DOI
  89. [89]M Tromp, A C J Ravelli, N Méray, J B Reitsma, and G J Bonsel. An efficient validation method of probabilistic record linkage including readmissions and twins. Methods of information in medicine, 2008.pubmed.ncbi.nlm.nih.gov/18690369
  90. [90]Zhenbang Wang, Emanuel Ben-David, Guoqing Diao, and Martin Slawski. Regression with linked datasets subject to linkage error. WIREs Computational Statistics, 14(4):e1570, 2022. doi: https://doi.org/10.1002/wics.1570. URL https://wires.onlinelibrary.wiley.com/doi/abs/10.1002/wics.1570.
  91. [91]Daniel Westreich and Stephen R Cole. Invited commentary: positivity in practice. American journal of epidemiology, 2010. doi: 10.1093/aje/kwp436.DOI
  92. [92]William E. Winkler. Chapter 14 - record linkage. In C.R. Rao, editor, Handbook of Statistics, volume 29 of Handbook of Statistics, pages 351–380. Elsevier, 2009. doi: https://doi.org/10.1016/S0169-7161(08)00014-X. URL https://www.sciencedirect.com/science/article/pii/S016971610800014X.
  93. [93]J P H Wortman. Record Linkage Methods with Applications to Causal Inference and Election Voting Data. PhD thesis, Department of Statistical Science Duke University, 2019.
  94. [94]Joan Heck Wortman and Jerome P. Reiter. Simultaneous record linkage and causal inference with propensity score subclassification. Statistics in Medicine, 37(24):3533–3546, 2018. doi: https://doi.org/10.1002/sim.7911. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.7911.

Paper details

vlnc:W_w72ce5yf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents