arxiv: 2605.05984 · v1 · submitted 2026-05-07 · 📊 stat.ME

Recognition: unknown

Separable Effects in Four-Arm and Two-Arm Designs

Chan Park, Youmi Suk

Authors on Pith no claims yet

Pith reviewed 2026-05-08 07:37 UTC · model grok-4.3

classification 📊 stat.ME

keywords separable effectscausal inferencemediation analysisfour-arm designtwo-arm designfalsification testsefficient influence functionstreatment decomposition

0 comments

The pith

A framework supplies distinct identification and estimation strategies for separable effects of treatment components depending on whether the data come from four-arm designs with independent assignment or two-arm designs with bundled assign

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a general framework to analyze separable effects, which examine the causal impact of each component of a treatment after decomposing the overall treatment. In four-arm designs the components can be assigned independently, permitting direct identification and estimation of each component's effect. In two-arm designs the components are strictly bundled, so identification requires additional assumptions that the framework makes explicit and testable. The authors derive tailored identification results and efficient influence function estimators with machine learning and cross-fitting for both designs, then add two falsification tests that use four-arm data to check the key assumptions needed for two-arm analyses. This distinction matters because many interventions in education, medicine, and policy are delivered as bundles, yet the ability to isolate component effects can lead to more precise mechanism understanding and better-targeted recommendations.

Core claim

The central claim is that separable effects can be identified and estimated with distinct strategies according to the design: four-arm data, in which treatment components are assigned independently, allow direct identification without the stronger assumptions required when components are bundled in two-arm data. For the bundled case the framework further supplies two falsification tests that leverage four-arm data to assess the identifying assumptions of consistency, positivity, and no unmeasured confounding for the component effects. Estimation in both cases proceeds via efficient influence functions combined with machine learning and cross-fitting, and the methods are illustrated on both a

What carries the argument

The general framework that supplies separate identification and estimation procedures for four-arm designs (independent component assignment) versus two-arm designs (bundled component assignment), implemented through efficient influence function estimators with machine learning and cross-fitting, together with falsification tests that compare results across the two designs.

If this is right

When four-arm data are available, separable effects of each component can be identified without the stronger assumptions required for bundled data.
The two falsification tests provide a practical check on whether two-arm data can be used reliably for separable-effects analysis.
Efficient influence function estimators with cross-fitting remain consistent even when machine learning is used to estimate nuisance functions.
Application to data such as the National Assessment of Educational Progress yields mechanism-specific insights into interventions like extended time accommodations.
Simulation results confirm that the proposed estimators achieve good performance under the stated assumptions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Collecting four-arm data when feasible could improve the reliability of mechanism analysis for many bundled interventions beyond education.
The falsification tests might be adapted to other settings where both bundled and unbundled data sources exist for the same intervention.
The framework could be extended to time-varying components or more than two components to handle richer longitudinal or multi-factor designs.
Matching the analytic strategy to the exact data-generating design reduces the risk of misattributing effects to the wrong component.

Load-bearing premise

The consistency, positivity, and no unmeasured confounding assumptions for the component effects hold in the two-arm design so that four-arm data can be used to test them.

What would settle it

If the effect estimates obtained by applying the two-arm identification strategy to a population also observed under a four-arm design differ substantially from the estimates obtained directly from the four-arm data, this would show that the identification assumptions for the two-arm design are violated.

Figures

Figures reproduced from arXiv: 2605.05984 by Chan Park, Youmi Suk.

**Figure 1.** Figure 1: Directed acyclic graph illustrating the decomposition of the extended time accommoda view at source ↗

**Figure 2.** Figure 2: Directed acyclic graph (DAG) and single-world intervention graph (SWIG) for treatment view at source ↗

**Figure 3.** Figure 3: Two examples where H0(i) and H0(ii) are not equivalent to (12) and (13), respectively. The dotted nodes represent unmeasured variables. However, if A3 holds, the null H0(i) is equivalent to: E[MaY =1,aM (G) | X(G) = x] = E[MaY =0,aM (G) | X(G) = x] . (12) In this case, rejecting H0(i) does provide valid evidence against AY -partial isolation. Thus, the validity of this test depends on the plausibility of A… view at source ↗

read the original abstract

Robins and Richardson (2010) reformulated mediation analysis by decomposing treatments into multiple components and examining separable effects of each component. While this approach is increasingly popular, existing work has analyzed ``two-arm'' data, where components are strictly bundled and manipulated simultaneously. However, in practice, four-arm data where components are assigned independently are often available. For example, testing accommodations might strictly bundle extra time with a separate session or allow them to be assigned separately. To address this distinction, we propose a general framework for analyzing separable effects in four-arm and two-arm designs. This framework provides distinct identification and estimation strategies for each design. For estimation, we utilize efficient influence function estimators coupled with machine learning and cross-fitting techniques. Additionally, we introduce two falsification tests for key identification assumptions required in the two-arm design by leveraging four-arm data. We investigate the performance of the proposed estimators via a simulation study and demonstrate their application by studying the effect of extended time accommodations using data from the National Assessment of Educational Progress. Ultimately, this separable effects analysis enables practitioners to clearly communicate underlying mechanisms and derive informative policy recommendations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives separate identification and estimation paths for separable effects in four-arm designs versus two-arm ones, plus falsification tests that use four-arm data to check two-arm assumptions, but the tests need unstated comparability between the two designs.

read the letter

The main thing to know is that this paper sets up distinct identification and estimation strategies for separable effects depending on whether the data come from a four-arm design, where treatment components can be assigned independently, or the more common two-arm design where they are bundled together. It also proposes two falsification tests that leverage four-arm data to probe the key assumptions needed for two-arm analyses, such as consistency, positivity, and no unmeasured confounding for the separate components. They back the methods with efficient influence function estimators that incorporate machine learning and cross-fitting, run a simulation study, and apply everything to National Assessment of Educational Progress data on test accommodations like extra time and separate sessions. That application shows how the framework can separate mechanisms in a real policy setting. The work builds directly on Robins and Richardson without circularity, and the simulation results give some evidence that the estimators behave reasonably in finite samples. The softer spot is the falsification tests. These only hold if the four-arm and two-arm populations share the same potential outcome distributions and confounding structures. The paper does not provide guidance on checking or ensuring that comparability, so differences in who selects into each design or in unmeasured factors could make the tests unreliable or misleading. That is the main practical limitation. This paper is for causal inference researchers who already work with mediation or mechanism questions in randomized experiments and who sometimes have access to multi-arm data. A reader looking for concrete ways to handle bundled treatments would get usable tools and an example to follow. It deserves a serious referee because the core distinctions are grounded in existing theory, the estimation approach is standard and reproducible, and they include both simulations and a real-data demonstration. I would recommend sending it for peer review, with reviewers asked to examine the comparability assumption behind the falsification tests.

Referee Report

2 major / 2 minor

Summary. The paper proposes a general framework for analyzing separable effects of treatment components in four-arm designs (where components are assigned independently) and two-arm designs (where components are bundled). It develops distinct identification results and efficient influence function (EIF) estimators using machine learning and cross-fitting for each design, introduces two falsification tests for the key identifying assumptions (consistency, positivity, no unmeasured confounding for components) in the two-arm case by leveraging four-arm data, validates the estimators via simulation, and applies the methods to National Assessment of Educational Progress data on extended time accommodations.

Significance. If the identification and falsification results hold, the work offers a concrete advance over existing separable-effects analyses (Robins and Richardson 2010) by explicitly handling the common practical distinction between bundled and unbundled designs. The use of EIF estimators with cross-fitting, the simulation study, and the real-data demonstration are strengths that support practical utility in policy settings such as education testing. The falsification tests, if valid, would be a notable contribution for assessing the credibility of two-arm analyses.

major comments (2)

[Identification and falsification tests sections] The two falsification tests for the two-arm design's identifying assumptions (consistency, positivity, and no unmeasured confounding for components) rely on an implicit exchangeability or comparability assumption between the four-arm and two-arm populations/mechanisms. This is not stated as an explicit identifying assumption in the identification section, nor is it addressed in the falsification-test derivation or the simulation design; without it the tests cannot reliably detect violations in the two-arm setting.
[Identification results] The distinct identification strategies for the four-arm versus two-arm designs are presented as internally consistent under standard causal assumptions, but the manuscript does not provide a formal proof or explicit statement showing that the two-arm identification reduces to the four-arm case when the bundling is removed (or vice versa). This comparison would strengthen the claim that the framework is unified.

minor comments (2)

[Notation and setup] Notation for potential outcomes and counterfactuals is introduced without a consolidated table or glossary; readers must track component-specific notation across sections.
[Simulation study] The simulation study reports performance metrics but does not include sensitivity checks to the strength of the comparability assumption between designs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment in detail below and indicate the revisions we will make to strengthen the paper.

read point-by-point responses

Referee: [Identification and falsification tests sections] The two falsification tests for the two-arm design's identifying assumptions (consistency, positivity, and no unmeasured confounding for components) rely on an implicit exchangeability or comparability assumption between the four-arm and two-arm populations/mechanisms. This is not stated as an explicit identifying assumption in the identification section, nor is it addressed in the falsification-test derivation or the simulation design; without it the tests cannot reliably detect violations in the two-arm setting.

Authors: We agree with the referee that the falsification tests rely on an exchangeability assumption between the four-arm and two-arm data-generating mechanisms, which was not explicitly stated. This assumption is necessary for the tests to be valid in detecting violations specific to the two-arm setting. In the revised manuscript, we will explicitly state this comparability assumption in the identification section, incorporate it into the derivation of the falsification tests, and ensure the simulation design accounts for it by simulating comparable populations. This will make the conditions under which the tests are reliable clear to readers. revision: yes
Referee: [Identification results] The distinct identification strategies for the four-arm versus two-arm designs are presented as internally consistent under standard causal assumptions, but the manuscript does not provide a formal proof or explicit statement showing that the two-arm identification reduces to the four-arm case when the bundling is removed (or vice versa). This comparison would strengthen the claim that the framework is unified.

Authors: We appreciate this suggestion to clarify the connection between the two identification strategies. While the strategies are tailored to their respective designs, we will add an explicit statement and a brief proof sketch in the revised manuscript demonstrating that the two-arm identification result reduces to the four-arm result in the absence of bundling (i.e., when the treatment components are assigned independently). This will be placed in the identification section or as a remark to highlight the unified framework. revision: yes

Circularity Check

0 steps flagged

Framework extends external mediation literature with design-specific identification and falsification strategies that remain independent of fitted values or self-referential definitions

full rationale

The paper's core contributions—distinct identification/estimation strategies for four-arm versus two-arm designs, efficient influence function estimators with cross-fitting, and two falsification tests leveraging four-arm data to check two-arm assumptions—derive from standard causal identification results (e.g., Robins and Richardson 2010) plus explicit positivity/consistency/no-unmeasured-confounding assumptions that are stated separately from the target quantities. No equation reduces a claimed prediction or test statistic to a fitted parameter by construction, no uniqueness theorem is imported from the authors' prior work, and the falsification tests are defined via observable contrasts between designs rather than tautological re-labeling of inputs. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on extending standard causal identification assumptions to four-arm and two-arm contexts; no explicit free parameters, new entities, or ad-hoc axioms are described.

axioms (1)

domain assumption Standard causal assumptions including consistency, positivity, and no unmeasured confounding for treatment components in mediation settings.
Invoked implicitly as the foundation for identification strategies in both designs, consistent with the referenced Robins and Richardson (2010) work.

pith-pipeline@v0.9.0 · 5487 in / 1398 out tokens · 56192 ms · 2026-05-08T07:37:06.743445+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

300 extracted references · 13 canonical work pages

[1]

Robins and Thomas S

James M. Robins and Thomas S. Richardson , title =. Causality and Psychopathology: Finding the Determinants of Disorders and their Cures , editor =. 2010 , pages =

2010
[2]

Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence , pages =

Pearl, Judea , title =. Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence , pages =. 2001 , publisher =

2001
[3]

and Greenland, Sander , title =

Robins, James M. and Greenland, Sander , title =. Epidemiology , year =
[4]

and Richardson, Thomas S

Robins, James M. and Richardson, Thomas S. and Shpitser, Ilya , title =. Probabilistic and Causal Inference: The Works of. 2022 , pages =

2022
[5]

arXiv , year =

Modeling Interference Via Symmetric Treatment Decomposition , author =. arXiv , year =
[6]

Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes , volume=

Pearl, Judea , year=. Does Obesity Shorten Life? Or is it the Soda? On Non-manipulable Causes , volume=. Journal of Causal Inference , publisher=. doi:10.1515/jci-2018-2001 , number=

work page doi:10.1515/jci-2018-2001 2018
[7]

Does obesity shorten life? The importance of well-defined interventions to answer causal questions , volume=

Hernán, M A and Taubman, S L , year=. Does obesity shorten life? The importance of well-defined interventions to answer causal questions , volume=. International Journal of Obesity , publisher=. doi:10.1038/ijo.2008.82 , number=

work page doi:10.1038/ijo.2008.82 2008
[8]

Journal of the Royal Statistical Society Series B: Statistical Methodology , pages =

Park, Chan and Stensrud, Mats J and Tchetgen Tchetgen, Eric J , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , pages =
[9]

2026 , journal=

Identifying Causes of Test Unfairness: Manipulability and Separability , author=. 2026 , journal=. doi:10.48550/arXiv.2601.13449 , url=

work page doi:10.48550/arxiv.2601.13449 2026
[10]

2025 , journal=

Addressing an extreme positivity violation to distinguish the causal effects of surgery and anesthesia via separable effects , author=. 2025 , journal=. doi:10.48550/arXiv.2504.01171 , url=

work page doi:10.48550/arxiv.2504.01171 2025
[11]

Evidence Factors in Fuzzy Regression Discontinuity Designs with Sequential Treatment Assignments , volume=

Lee, Youjin and Suk, Youmi , year=. Evidence Factors in Fuzzy Regression Discontinuity Designs with Sequential Treatment Assignments , volume=. Psychometrika , publisher=. doi:10.1017/psy.2025.10033 , number=

work page doi:10.1017/psy.2025.10033 2025
[12]

Psychometrika , author=

Causal Mediation and Functional Outcome Analysis with Process Data , DOI=. Psychometrika , author=. 2026 , pages=

2026
[13]

, title =

Robins, James M. , title =. Mathematical Modelling , year =
[14]

and Robins, James M

Richardson, Thomas S. and Robins, James M. , title =. 2013 , number =

2013
[15]

Section 9

On the application of probability theory to agricultural experiments: Essay on principles. Section 9. , author=. Statistical Science , volume=. 1923 , note=

1923
[16]

M. S. Bartlett , journal =. Measles Periodicity and Community Size , volume =
[17]

and Wahba, Grace

Kimeldorf, George S. and Wahba, Grace. A Correspondence Between Bayesian Estimation on Stochastic Processes and Smoothing by Splines. Annals of Mathematical Statistics. 1970

1970
[18]

, author=

Estimating causal effects of treatments in randomized and nonrandomized studies. , author=. Journal of Educational Psychology , volume=. 1974 , publisher=

1974
[19]

Rubin , journal =

Donald B. Rubin , journal =. Inference and Missing Data , volume =
[20]

Bayesian Inference for Causal Effects: The Role of Randomization

Rubin, Donald B. Bayesian Inference for Causal Effects: The Role of Randomization. The Annals of Statistics
[21]

Dawid, A. P. , title =. Journal of the Royal Statistical Society: Series B (Methodological) , volume =
[22]

A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity , volume =

Halbert White , journal =. A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity , volume =
[23]

Large Sample Properties of Generalized Method of Moments Estimators , volume =

Lars Peter Hansen , journal =. Large Sample Properties of Generalized Method of Moments Estimators , volume =
[24]

Rosenbaum and Donald B

Paul R. Rosenbaum and Donald B. Rubin , journal =. The Central Role of the Propensity Score in Observational Studies for Causal Effects , volume =
[25]

LaLonde , journal =

Robert J. LaLonde , journal =. Evaluating the Econometric Evaluations of Training Programs with Experimental Data , volume =
[26]

1986 , author =

A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect , journal =. 1986 , author =

1986
[27]

On Asymptotically Efficient Estimation in Semiparametric Models , volume =

Anton Schick , journal =. On Asymptotically Efficient Estimation in Semiparametric Models , volume =
[28]

1986 , author =

Meta-analysis in clinical trials , journal =. 1986 , author =

1986
[29]

and Rosenbaum, Paul R

Holland, Paul W. and Rosenbaum, Paul R. Conditional Association and Unidimensionality in Monotone Latent Variable Models. Annals of Statistics. 1986

1986
[30]

Zeger , journal =

Kung-Yee Liang and Scott L. Zeger , journal =. Longitudinal Data Analysis Using Generalized Linear Models , volume =
[31]

Holland , title =

Paul W. Holland , title =. Journal of the American Statistical Association , volume =
[32]

Rubin , title =

Donald B. Rubin , title =. Journal of the American Statistical Association , volume =
[33]

Newey , journal =

Whitney K. Newey , journal =. Semiparametric Efficiency Bounds , volume =
[34]

Robins and Sander Greenland , journal =

James M. Robins and Sander Greenland , journal =. Identifiability and Exchangeability for Direct and Indirect Effects , volume =
[35]

Robins , title =

James M. Robins , title =. Communications in Statistics - Theory and Methods , volume =
[36]

Robins and Andrea Rotnitzky and Lue Ping Zhao , journal =

James M. Robins and Andrea Rotnitzky and Lue Ping Zhao , journal =. Estimation of Regression Coefficients When Some Regressors Are Not Always Observed , volume =
[37]

The Annals of Statistics , number =

van der Vaart, Aad W , title =. The Annals of Statistics , number =
[38]

Statistics in Medicine , volume =

Rotnitzky, Andrea and Robins, James , title =. Statistics in Medicine , volume =
[39]

Causal Inference from Complex Longitudinal Data

Robins, James M. Causal Inference from Complex Longitudinal Data. Latent Variable Modeling and Applications to Causality. 1997

1997
[40]

Heckman and Hidehiko Ichimura and Petra E

James J. Heckman and Hidehiko Ichimura and Petra E. Todd , journal =. Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme , volume =
[41]

Gray and Ron Brookmeyer , journal =

Sarah M. Gray and Ron Brookmeyer , journal =. Estimating a Treatment Effect from Multidimensional Longitudinal Data , volume =
[42]

On the Role of the Propensity Score in Efficient Semiparametric Estimation of Average Treatment Effects , volume =

Jinyong Hahn , journal =. On the Role of the Propensity Score in Efficient Semiparametric Estimation of Average Treatment Effects , volume =
[43]

Statistics in medicine , volume=

Multiple imputation of missing blood pressure covariates in survival analysis , author=. Statistics in medicine , volume=. 1999 , publisher=

1999
[44]

1999 , volume=

IEEE Transactions on Information Theory , title=. 1999 , volume=

1999
[45]

Robins and Sander Greenland and Fu-Chang Hu , title =

James M. Robins and Sander Greenland and Fu-Chang Hu , title =. Journal of the American Statistical Association , volume =
[46]

and Demétrio, Clarice G

Ridout, Martin S. and Demétrio, Clarice G. B. and Firth, David , title =. Biometrics , volume =
[47]

Scharfstein and Andrea Rotnitzky and James M

Daniel O. Scharfstein and Andrea Rotnitzky and James M. Robins , title =. Journal of the American Statistical Association , volume =. 1999 , publisher =

1999
[48]

and Hern

Robins, James M. and Hern. Marginal structural models and causal inference in epidemiology , journal=. 2000 , volume=

2000
[49]

and Rotnitzky, Andrea and Scharfstein, Daniel O

Robins, James M. and Rotnitzky, Andrea and Scharfstein, Daniel O. Sensitivity Analysis for Selection bias and unmeasured Confounding in missing Data and Causal inference models. Statistical Models in Epidemiology, the Environment, and Clinical Trials. 2000

2000
[50]

Gray and Ron Brookmeyer , title =

Sarah M. Gray and Ron Brookmeyer , title =. Journal of the American Statistical Association , volume =
[51]

A Generalized Representer Theorem

Sch \"o lkopf, Bernhard and Herbrich, Ralf and Smola, Alex J. A Generalized Representer Theorem. Computational Learning Theory. 2001

2001
[52]

Carroll , journal =

Xihong Lin and Raymond J. Carroll , journal =. Semiparametric Regression for Clustered Data Using Generalized Estimating Equations , volume =
[53]

, title =

Spiegelhalter, David J. , title =. Statistics in Medicine , volume =
[54]

and Omar, Rumana Z

Turner, Rebecca M. and Omar, Rumana Z. and Thompson, Simon G. , title =. Statistics in Medicine , volume =
[55]

Machine learning , volume=

Random forests , author=. Machine learning , volume=. 2001 , publisher=

2001
[56]

Journal of the American Statistical Association , volume =

Peter D Hoff and Adrian E Raftery and Mark S Handcock , title =. Journal of the American Statistical Association , volume =
[57]

and Rubin, Donald B

Frangakis, Constantine E. and Rubin, Donald B. , title =. Biometrics , volume =
[58]

Understanding statistics: statistical issues in psychology, education, and the social sciences , volume=

Partitioning variation in multilevel models , author=. Understanding statistics: statistical issues in psychology, education, and the social sciences , volume=. 2002 , publisher=

2002
[59]

Statistics in Medicine , volume =

Goldstein, Harvey and Browne, William and Rasbash, Jon , title =. Statistics in Medicine , volume =
[60]

Stefanski and Dennis D

Leonard A. Stefanski and Dennis D. Boos , journal =. The Calculus of M-Estimation , volume =
[61]

Newey and James L

Whitney K. Newey and James L. Powell , journal =. Instrumental Variable Estimation of Nonparametric Models , volume =
[62]

Journal of Machine Learning Research , pages =

Meir, Ron and Zhang, Tong , title =. Journal of Machine Learning Research , pages =. 2003 , volume =

2003
[63]

American Economic Review , Volume =

Abadie, Alberto and Gardeazabal, Javier , Title =. American Economic Review , Volume =. 2003 , Pages =

2003
[64]

and Ridder, Geert , title =

Hirano, Keisuke and Imbens, Guido W. and Ridder, Geert , title =. Econometrica , volume =
[65]

and Tangen, Catherine M

Petrylak, Daniel P. and Tangen, Catherine M. and Hussain, Maha H.A. and Lara, Primo N. and Jones, Jeffrey A. and Taplin, Mary Ellen and Burch, Patrick A. and Berry, Donna and Moinpour, Carol and Kohli, Manish and Benson, Mitchell C. and Small, Eric J. and Raghavan, Derek and Crawford, E. David , title =. New England Journal of Medicine , volume =
[66]

Journal of the American Statistical Association , volume =

Peter Hall and Jeff Racine and Qi Li , title =. Journal of the American Statistical Association , volume =
[67]

and Davidian, Marie , title =

Lunceford, Jared K. and Davidian, Marie , title =. Statistics in Medicine , volume =
[68]

Optimal Structural Nested Models for Optimal Sequential Decisions

Robins, James M. Optimal Structural Nested Models for Optimal Sequential Decisions. Proceedings of the Second Seattle Symposium in Biostatistics: Analysis of Correlated Data. 2004

2004
[69]

2005 , author =

Future directions in the treatment of androgen-independent prostate cancer , journal =. 2005 , author =

2005
[70]

The Review of Economic Studies , volume =

Abadie, Alberto , title =. The Review of Economic Studies , volume =
[71]

Journal of Epidemiology & Community Health , volume=

A brief conceptual tutorial of multilevel analysis in social epidemiology: linking the statistical concept of clustering to the idea of contextual phenomenon , author=. Journal of Epidemiology & Community Health , volume=. 2005 , publisher=

2005
[72]

, title =

Bang, Heejung and Robins, James M. , title =. Biometrics , volume =
[73]

censoring

Causal inference through potential outcomes and principal stratification: application to studies with" censoring" due to death , author=. Statistical Science , pages=
[74]

, title =

Athey, Susan and Imbens, Guido W. , title =. Econometrica , volume =
[75]

Journal of the American Statistical Association , volume =

Michael E Sobel , title =. Journal of the American Statistical Association , volume =. 2006 , publisher =

2006
[76]

and Raudenbush, S

Hong, G. and Raudenbush, S. W. Evaluating kindergarten retention policy: A case study of causal inference for multilevel observational data. Journal of the American Statistical Association. 2006

2006
[77]

Instruments for causal inference:

Hern. Instruments for causal inference:. Epidemiology , volume=
[78]

Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

Identifying direct and indirect effects in a non-counterfactual framework , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2007 , publisher=

2007
[79]

and Gilbert, Peter B

Jemiai, Yannis and Rotnitzky, Andrea and Shepherd, Bryan E. and Gilbert, Peter B. , title = ". Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =
[80]

Biometrics , volume=

Discussions on ``Principal Stratification Designs to Estimate Input Data Missing Due to Death'' , author=. Biometrics , volume=

Showing first 80 references.