pith. sign in

arxiv: 2605.01114 · v1 · submitted 2026-05-01 · 📊 stat.ME

A formal approach to variable selection in difference-in-differences

Pith reviewed 2026-05-09 18:27 UTC · model grok-4.3

classification 📊 stat.ME
keywords difference-in-differencesparallel trendscausal inferencegraphical modelsvariable selectionDAGconditional identification
0
0 comments X

The pith

Graphical criteria identify the minimal covariates needed to justify conditional parallel trends in difference-in-differences.

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

The paper develops a method to choose which covariates to condition on when assuming conditional parallel trends in DiD studies. It uses directed acyclic graphs to represent causal relationships and applies d-separation rules to find the smallest set of variables that make the parallel trends assumption plausible. This approach shows that unconditional parallel trends often fails, and that certain covariates not usually thought of as confounders can still be important to adjust for. The method also addresses how to handle post-treatment variables and treatment-confounder feedback over multiple periods. Finally, it reframes estimation challenges as a mismatch between the identification set and the estimator's adjustment set, providing ways to align them.

Core claim

We propose a formal approach to select the variables that support conditional parallel trends based on graphical criteria. We show that the parallel trends assumption is rarely justified without conditioning on covariates, and that unconditional and conditional parallel trends can conflict with one another. We also demonstrate that a time-invariant covariate with a time-invariant effect on the outcome, which might not ordinarily be considered a confounder in DiD, may be a useful conditioning variable. We clarify that adjustment for a post-treatment covariate depends on what causes that covariate to change. Extending our framework to multiple time periods, we distinguish between treatment and

What carries the argument

Directed acyclic graph (DAG) of the data-generating process, with d-separation and backdoor criteria to determine the minimal adjustment set for conditional parallel trends.

If this is right

  • Parallel trends rarely holds unconditionally, so covariate adjustment is typically required.
  • Unconditional and conditional parallel trends assumptions can contradict each other.
  • Time-invariant covariates can still be necessary for identification even if they have time-invariant effects.
  • Post-treatment covariates require careful consideration based on their causes.
  • Estimation procedures must match the identification adjustment set to avoid bias.

Where Pith is reading between the lines

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

  • Researchers could apply this method to existing DiD studies to check if their chosen covariates are sufficient or minimal.
  • This graphical approach might extend to other identifying assumptions in panel data methods beyond DiD.
  • The alignment of identification and estimation sets could improve robustness in applied work using popular DiD estimators.
  • Testing the sensitivity to different assumed DAGs would be a useful extension for practitioners.

Load-bearing premise

The underlying causal structure among all relevant variables can be correctly represented as a directed acyclic graph.

What would settle it

A real-world DiD application where the graphically selected covariates are adjusted for but the conditional parallel trends assumption is violated, or where the assumption holds without the selected covariates.

Figures

Figures reproduced from arXiv: 2605.01114 by Daniela Rodrigues, Laura A. Hatfield.

Figure 1
Figure 1. Figure 1: Causal diagram in which equi-confounding by the unobserved V0 is denoted by the equal magnitude ∗ on the edges to the outcome at both time periods. To identify the average affect of A1 on Y1, we must block the backdoor path, A1 ← V0 ∗−→ Y1. However, this path cannot be blocked because V0 is not measured. This challenge motivates a DiD identification strategy, in which we leverage pre-treatment outcome meas… view at source ↗
Figure 2
Figure 2. Figure 2: Causal diagram in which equi-confounding is represented by the absence of an arrow from the unmeasured confounder, V0, to the change in the outcome over time, ∆Y . In this representation, equi-confounding is encoded as a single missing edge from V0 to ∆Y , rather than an equality constraint across two backdoor paths. By inspection, there are no backdoor paths from A1 to ∆Y . As we will show in Section 5, t… view at source ↗
Figure 3
Figure 3. Figure 3: Same model as view at source ↗
Figure 4
Figure 4. Figure 4: Scenario with a baseline covariate W0 having a time-varying relationship to outcomes. (a) Natural representation. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. 9 view at source ↗
Figure 5
Figure 5. Figure 5: Scenario with a time-invariant covariate, W0, and a time-varying covariate, Zt, that evolves due to a shock, Q0. (a) Natural representation. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. Although there are backdoor paths from A1 to ∆Y in the compact representation, the time symmetry of the confounding ensures perfect offsetting in DiD such t… view at source ↗
Figure 6
Figure 6. Figure 6: Scenario with a time-invariant covariate W0 and a time-varying covariate Zt. (a) Natural representation. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. In this case, unconditional parallel trends no longer holds, σ(∆Y, A1) = a + gijk − gih ̸= a, which means we should resort to conditional parallel trends. By applying the backdoor criterion to… view at source ↗
Figure 7
Figure 7. Figure 7: Scenario in which the time-varying covariate Wt’s evolution is governed by an auto-regressive process. (a) Natural diagram. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. This scenario illustrates a case in which adjustment for the post-treatment covariate is not required for identification, but can be included without adverse consequences. 5… view at source ↗
Figure 8
Figure 8. Figure 8: Scenario in which the time-varying covariate Wt’s evolution is governed by observed confounder Z0 and an auto-regressive process. (a) Natural diagram. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. 5.3.3 Unmeasured confounder In view at source ↗
Figure 9
Figure 9. Figure 9: Scenario in which the time-varying covariate Wt’s evolution is governed by the unobserved confounder V0 and an auto-regressive process. (a) Natural diagram. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. This scenario demonstrates that adjustment for the post-treatment covariate can be the only available route to identification when unobserve… view at source ↗
Figure 10
Figure 10. Figure 10: Scenario in which the time-varying covariate Wt’s evolution depends on treatment and an auto￾regressive process. (a) Natural diagram. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. This scenario illustrates the case in which the post-treatment covariate must not be included in the adjustment set, as doing so would block a causal path from tr… view at source ↗
Figure 11
Figure 11. Figure 11: Scenario in which a static treatment may have time-varying effects. (a) Natural diagram. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. From the compact representation in Panel (b), we see that {W0} is a minimally sufficient adjustment set to identify the effect of treatment on both ∆Y1 and ∆Y2. Of note, across all scenarios in this section,… view at source ↗
Figure 12
Figure 12. Figure 12: Scenario with a dynamic treatment regime and treatment-confounder feedback via W1. (a) Natural diagram. (b) Compact representation. Illustrative linear model coefficients are shown as lowercase letters on each edge. The key takeaway from these two scenarios is that the distinction between the type of treatment and the rollout strategy has direct consequences for covariate adjustment: treatment-confounder … view at source ↗
Figure 13
Figure 13. Figure 13 view at source ↗
read the original abstract

Difference-in-differences (DiD) identification relies mainly on a parallel trends assumption about untreated potential outcomes. Researchers often relax this assumption by assuming conditional parallel trends within units with the same covariate values. However, the process of selecting which covariates to include in this assumption is often \emph{ad hoc}. We propose a formal approach to select the variables that support conditional parallel trends based on graphical criteria. We show that the parallel trends assumption is rarely justified without conditioning on covariates, and that unconditional and conditional parallel trends can conflict with one another. We also demonstrate that a time-invariant covariate with a time-invariant effect on the outcome, which might not ordinarily be considered a confounder in DiD, may be a useful conditioning variable. We clarify that adjustment for a post-treatment covariate depends on what causes that covariate to change. Extending our framework to multiple time periods, we distinguish between treatment type and rollout strategy and examine the problem of treatment-confounder feedback. On the estimation side, we argue that the difficulty of incorporating covariates in DiD, often framed as an estimator problem, is more accurately understood as a misalignment between the adjustment set used by the estimator and the adjustment set required for identification. This misalignment affects several popular estimation procedures, and resolving it requires not a change of estimator, but a change in how covariates enter the estimation procedure. We show how to achieve this alignment for all estimators we evaluate.

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.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a formal graphical approach, based on directed acyclic graphs and d-separation, for selecting the minimal set of covariates that justify a conditional parallel trends assumption in difference-in-differences designs. It argues that unconditional parallel trends is rarely justified without covariates, that unconditional and conditional parallel trends can conflict, that time-invariant covariates with time-invariant effects can be useful conditioning variables, that post-treatment covariate adjustment depends on the causes of change in the covariate, and extends the framework to multiple periods while distinguishing treatment type from rollout strategy and addressing treatment-confounder feedback. On the estimation side, it reframes difficulties with covariates as misalignment between the identification adjustment set and the estimator's adjustment set, and shows how to achieve alignment for the estimators considered.

Significance. If the graphical criteria are shown to correctly encode and imply the counterfactual equality required by conditional parallel trends, the paper would provide a principled, non-ad-hoc method for covariate selection in DiD that could improve identification transparency and validity in applied work. The multi-period extensions and the reframing of estimation issues as alignment problems rather than estimator choice are practically relevant contributions that could influence how researchers specify and implement DiD models.

major comments (3)
  1. [§3] §3 (graphical criteria for conditional parallel trends): the central claim that d-separation on the proposed DAG identifies the minimal adjustment set supporting conditional parallel trends requires an explicit derivation showing how the graph encodes the counterfactual processes E[Y(0)_t | X, G] and the no-anticipation assumption; without this, standard d-separation on observed variables does not automatically guarantee the required equality of expected untreated potential outcomes across groups.
  2. [§4.2] §4.2 (conflict between unconditional and conditional PT): the demonstration that these two assumptions can conflict is load-bearing for the motivation to use graphical selection; a concrete counter-example or theorem (with the specific DAG and potential-outcome equations) is needed to show when the conflict arises and whether the proposed minimal adjustment set resolves it without additional parametric restrictions.
  3. [§6] §6 (multi-period extension and treatment-confounder feedback): the distinction between treatment type and rollout strategy is important, but the graphical criterion for handling feedback must be shown to preserve identification of the target parameter; if the feedback loop is not fully blocked by the selected covariates, the conditional PT may fail in later periods.
minor comments (2)
  1. The abstract states that 'adjustment for a post-treatment covariate depends on what causes that covariate to change,' but the main text should include a small table or figure contrasting the two cases (e.g., covariate change caused by treatment vs. by an exogenous factor) to improve clarity.
  2. Notation for potential outcomes and time indices is introduced gradually; a single consolidated notation table at the beginning of §2 would help readers track the distinction between observed and counterfactual quantities.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which identify opportunities to strengthen the rigor of our graphical framework for covariate selection in difference-in-differences. We address each major comment below and outline the revisions we will undertake.

read point-by-point responses
  1. Referee: [§3] §3 (graphical criteria for conditional parallel trends): the central claim that d-separation on the proposed DAG identifies the minimal adjustment set supporting conditional parallel trends requires an explicit derivation showing how the graph encodes the counterfactual processes E[Y(0)_t | X, G] and the no-anticipation assumption; without this, standard d-separation on observed variables does not automatically guarantee the required equality of expected untreated potential outcomes across groups.

    Authors: We agree that an explicit derivation is required to link d-separation directly to the counterfactual equality E[Y(0)_t | X, G] under the no-anticipation assumption. In the revised manuscript we will add a dedicated subsection (or appendix) that derives this connection step by step: we will specify how the DAG encodes the potential-outcome processes, show that the no-anticipation assumption corresponds to the absence of certain directed paths, and prove that d-separation on the minimal adjustment set implies the required conditional independence of untreated potential outcomes across groups. revision: yes

  2. Referee: [§4.2] §4.2 (conflict between unconditional and conditional PT): the demonstration that these two assumptions can conflict is load-bearing for the motivation to use graphical selection; a concrete counter-example or theorem (with the specific DAG and potential-outcome equations) is needed to show when the conflict arises and whether the proposed minimal adjustment set resolves it without additional parametric restrictions.

    Authors: We acknowledge the need for a fully specified counter-example. The revision will include a concrete DAG together with the associated potential-outcome equations that exhibit a conflict between unconditional and conditional parallel trends. We will then apply our graphical criterion to recover the minimal adjustment set and verify that it restores identification of the target parameter without imposing extra parametric restrictions beyond those already stated in the paper. revision: yes

  3. Referee: [§6] §6 (multi-period extension and treatment-confounder feedback): the distinction between treatment type and rollout strategy is important, but the graphical criterion for handling feedback must be shown to preserve identification of the target parameter; if the feedback loop is not fully blocked by the selected covariates, the conditional PT may fail in later periods.

    Authors: We agree that preservation of identification must be shown explicitly when treatment-confounder feedback is present. In the revised Section 6 we will provide a formal argument (supported by path-blocking analysis) demonstrating that the covariates selected by our graphical criterion close all feedback paths that would otherwise violate conditional parallel trends in subsequent periods. We will also add a brief numerical illustration confirming that the target parameter remains identified under the proposed adjustment set. revision: yes

Circularity Check

0 steps flagged

No circularity: standard graphical criteria applied to DiD identification

full rationale

The paper's derivation applies established d-separation and adjustment-set criteria from graphical causal models to the conditional parallel trends assumption in DiD. This is an extension of existing tools rather than a self-referential construction; the parallel trends statement is not redefined in terms of the selected covariates, no parameters are fitted to the target result and then relabeled as predictions, and no load-bearing uniqueness theorems or ansatzes are imported solely via self-citation. The central claim remains an independent mapping from DAG structure to covariate sets, with the identification result following from the graphical rules rather than reducing to the paper's own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on standard causal-graph assumptions and the mapping of parallel trends to conditional independence statements in a DAG; no new free parameters, invented entities, or ad-hoc axioms are introduced in the abstract.

axioms (2)
  • standard math Causal relationships among variables can be represented by a directed acyclic graph (DAG).
    Invoked to apply d-separation and other graphical criteria for identifying adjustment sets.
  • domain assumption The (conditional) parallel trends assumption corresponds to specific conditional independence relations in the DAG.
    Central link between the DiD identifying assumption and the graphical selection procedure.

pith-pipeline@v0.9.0 · 5544 in / 1465 out tokens · 30838 ms · 2026-05-09T18:27:51.420041+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages

  1. [1]

    The Estimation of Causal Effects by Difference-in-Difference Methods

    Lechner M. The Estimation of Causal Effects by Difference-in-Difference Methods. Foundations and Trends in Econometrics. 2011;4(3):165-224

  2. [2]

    What’s Trending in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature

    Roth J, Sant’Anna PHC, Bilinski A, Poe J. What’s Trending in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature. Journal of Econometrics. 2023;235(2):2218-44

  3. [3]

    Nothing to See Here? A non-inferiority approach to evaluation of parallel trends in difference-in-differences

    Bilinski A, Hatfield LA. Nothing to See Here? A non-inferiority approach to evaluation of parallel trends in difference-in-differences. Statistics in Medicine. 2026 Feb;45(3-5)

  4. [4]

    Semiparametric Difference-in-Differences Estimators

    Abadie A. Semiparametric Difference-in-Differences Estimators. Review of Economic Studies. 2005;72:1- 19

  5. [5]

    Confounding and Regression Adjustment in Difference-in-Differences Studies

    Zeldow B, Hatfield LA. Confounding and Regression Adjustment in Difference-in-Differences Studies. Health Services Research. 2021;56:932-41

  6. [6]

    Causality: Models, Reasoning, and Inference

    Pearl J. Causality: Models, Reasoning, and Inference. 2nd ed. New York: Cambridge University Press; 2000

  7. [7]

    Causal Inference in Statistics: A Primer

    Pearl J, Glymour M, Jewell NP. Causal Inference in Statistics: A Primer. Chichester, West Sussex: Wiley; 2016

  8. [8]

    Causal Knowledge as a Prerequisite for Con- founding Evaluation: An Application to Birth Defects Epidemiology

    Hern´ an MA, Hern´ andez-D´ ıaz S, Werler MM, Mitchell AA. Causal Knowledge as a Prerequisite for Con- founding Evaluation: An Application to Birth Defects Epidemiology. American Journal of Epidemiology. 2002;155(2):176-84

  9. [9]

    Use of Directed Acyclic Graphs (DAGs) to Identify Confounders in Applied Health Research: Review and Recommendations

    Tennant PWG, Murray EJ, Arnold KF, Berrie L, Fox MP, Gadd SC, et al. Use of Directed Acyclic Graphs (DAGs) to Identify Confounders in Applied Health Research: Review and Recommendations. International Journal of Epidemiology. 2021;50(2):620-32

  10. [10]

    A Crash Course in Good and Bad Controls

    Cinelli C, Forney A, Pearl J. A Crash Course in Good and Bad Controls. Sociological Methods & Research. 2024;53(3):1071-104

  11. [11]

    Overlap in Observational Studies with High-Dimensional Covariates

    D’Amour A, Ding P, Feller A, Lei L, Sekhon J. Overlap in Observational Studies with High-Dimensional Covariates. Journal of Econometrics. 2021;221(2):644-54

  12. [12]

    To Adjust or Not to Adjust? Sensitivity Analysis of M-Bias and Butterfly-Bias

    Ding P, Miratrix LW. To Adjust or Not to Adjust? Sensitivity Analysis of M-Bias and Butterfly-Bias. Journal of Causal Inference. 2015;3(1):41-57

  13. [13]

    Identification Without Exogeneity Under Equiconfounding in Linear Recursive Structural Systems

    Chalak K. Identification Without Exogeneity Under Equiconfounding in Linear Recursive Structural Systems. In: Chen X, Swanson NR, editors. Recent Advances and Future Directions in Causality, Prediction, and Specification Analysis. New York, NY: Springer New York; 2013. p. 27-55. 25

  14. [14]

    Exploiting Equality Constraints in Causal Inference

    Zhang C, Cinelli C, Chen B, Pearl J. Exploiting Equality Constraints in Causal Inference. In: Proceed- ings of the 24th International Conference on Artificial Intelligence and Statistics (AISTATS). vol. 130. San Diego, California, USA; 2021

  15. [15]

    On Negative Outcome Con- trol of Unobserved Confounding as a Generalization of Difference-in-Differences

    Sofer T, Richardson DB, Colicino E, Schwartz J, Tchetgen Tchetgen EJ. On Negative Outcome Con- trol of Unobserved Confounding as a Generalization of Difference-in-Differences. Statistical Science. 2016;31(3)

  16. [16]

    Assumption Trade-Offs When Choosing Identification Strategies for Pre-Post Treatment Effect Estimation: An Illustration of a Community-Based Intervention in Madagascar

    Weber AM, van der Laan MJ, Petersen ML. Assumption Trade-Offs When Choosing Identification Strategies for Pre-Post Treatment Effect Estimation: An Illustration of a Community-Based Intervention in Madagascar. Journal of Causal Inference. 2015;3(1):109-30

  17. [17]

    Gain Scores Revisited: A Graphical Models Perspective

    Kim Y, Steiner PM. Gain Scores Revisited: A Graphical Models Perspective. Sociological Methods & Research. 2021;50(3):1353-75

  18. [18]

    Using Causal Diagrams to Assess Parallel Trends in Difference-in- Differences Studies

    Renson A, Dukes O, Shahn Z. Using Causal Diagrams to Assess Parallel Trends in Difference-in- Differences Studies. arXiv; 2025. ArXiv:2505.03526 [stat]

  19. [19]

    Selection and Parallel Trends

    Ghanem D, Sant’Anna PHC, W¨ uthrich K. Selection and Parallel Trends. arXiv; 2025

  20. [20]

    Difference-in-Differences When Parallel Trends Holds Conditional on Covariates

    Caetano C, Callaway B. Difference-in-Differences When Parallel Trends Holds Conditional on Covariates. arXiv; 2024

  21. [21]

    Controlling Time-Varying Confounding in Difference-in-Differences Studies Using the Time- Varying Treatments Framework

    Myint L. Controlling Time-Varying Confounding in Difference-in-Differences Studies Using the Time- Varying Treatments Framework. Health Services and Outcomes Research Methodology. 2023

  22. [22]

    Identifying and Estimating Effects of Sustained Interventions under Parallel Trends Assumptions

    Renson A, Hudgens M, Keil A, Zivich P, Aiello A. Identifying and Estimating Effects of Sustained Interventions under Parallel Trends Assumptions. Biometrics. 2023;79(4):2998-3009

  23. [23]

    Structural Nested Mean Models Under Parallel Trends Assumptions

    Shahn Z, Dukes O, Richardson D, Tchetgen ET, Robins J. Structural Nested Mean Models Under Parallel Trends Assumptions. arXiv; 2022. Available from:http://arxiv.org/abs/2204.10291

  24. [24]

    When Should We Use Unit Fixed Effects Regression Models for Causal Inference with Longitudinal Data? American Journal of Political Science

    Imai K, Kim IS. When Should We Use Unit Fixed Effects Regression Models for Causal Inference with Longitudinal Data? American Journal of Political Science. 2019;63(2):467-90

  25. [25]

    Should We Combine Difference in Differences with Con- ditioning on Pre-Treatment Outcomes? Toulouse School of Economics

    Chab´ e-Ferret S. Should We Combine Difference in Differences with Con- ditioning on Pre-Treatment Outcomes? Toulouse School of Economics

  26. [26]

    Available from:https://www.tse-fr.eu/publications/ should-we-combine-difference-differences-conditioning-pre-treatment-outcomes

    17-824. Available from:https://www.tse-fr.eu/publications/ should-we-combine-difference-differences-conditioning-pre-treatment-outcomes

  27. [27]

    Matching and Regression to the Mean in Difference-in-Differences Analysis

    Daw JR, Hatfield LA. Matching and Regression to the Mean in Difference-in-Differences Analysis. Health Services Research. 2018;53(6):4138-56. 26

  28. [28]

    Well-Balanced or Too Matchy–Matchy? The Controversy over Matching in Difference-in- Differences

    Ryan AM. Well-Balanced or Too Matchy–Matchy? The Controversy over Matching in Difference-in- Differences. Health Services Research. 2018;53(6):4106-10

  29. [29]

    Benefits and Costs of Matching Prior to a Difference in Differences Analysis When Parallel Trends Does Not Hold

    Ham DW, Miratrix L. Benefits and Costs of Matching Prior to a Difference in Differences Analysis When Parallel Trends Does Not Hold. arXiv; 2024. Available from:http://arxiv.org/abs/2205.08644

  30. [30]

    Correlation and Causation

    Wright S. Correlation and Causation. Journal of Agricultural Research. 1921;20(7):557-85

  31. [31]

    Graphical Tools for Linear Path Models; 2018

    Chen B, Pearl J, Kline R. Graphical Tools for Linear Path Models; 2018. 469

  32. [32]

    Using Propensity Scores in Difference-in-Differences Models to Estimate the Effects of a Policy Change

    Stuart EA, Huskamp HA, Duckworth K, Simmons J, Song Z, Chernew ME, et al. Using Propensity Scores in Difference-in-Differences Models to Estimate the Effects of a Policy Change. Health Services and Outcomes Research Methodology. 2014;14(4):166-82

  33. [33]

    A New Approach to Causal Inference in Mortality Studies with a Sustained Exposure Period—Application to Control of the Healthy Worker Survivor Effect

    Robins JM. 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. 1986;7(9-12):1393-512

  34. [34]

    Marginal Structural Models and Causal Inference in Epidemiol- ogy

    Robins JM, Hern´ an MA, Brumback B. Marginal Structural Models and Causal Inference in Epidemiol- ogy. Epidemiology (Cambridge, Mass). 2000;11(5):550-60

  35. [35]

    Unfinished Business: Paid Family Leave in California and the Future of U.S

    Milkman R, Appelbaum E. Unfinished Business: Paid Family Leave in California and the Future of U.S. Work-Family Policy. Ithaca: ILR Press, an imprint of Cornell University Press; 2013

  36. [36]

    Paid Family and Medical Leave Programs: State Pathways and Design Options

    Glynn SJ, Bradley AL, Veghte BW. Paid Family and Medical Leave Programs: State Pathways and Design Options. Washington, DC: National Academy of Social Insurance; 2017

  37. [37]

    Two-Way Fixed Effects Estimators with Heterogeneous Treat- ment Effects

    De Chaisemartin C, D’Haultfœuille X. Two-Way Fixed Effects Estimators with Heterogeneous Treat- ment Effects. American Economic Review. 2020;110(9):2964-96

  38. [38]

    Matching As An Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme

    Heckman J, Ichimura H, Todd P. Matching As An Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme. The Review of Economic Studies. 1997;64(4):605-54

  39. [39]

    Characterizing Selection Bias Using Experimental Data

    Heckman J, Ichimura H, Smith J, Todd P. Characterizing Selection Bias Using Experimental Data. Econometrica. 1998;66(5):1017-98

  40. [40]

    Doubly Robust Difference-in-Differences Estimators

    Sant’Anna PHC, Zhao J. Doubly Robust Difference-in-Differences Estimators. Journal of Econometrics. 2020;219(1):101-22

  41. [41]

    Did: Treatment Effects with Multiple Periods and Groups; 2022

    Callaway B, Sant’Anna PHC. Did: Treatment Effects with Multiple Periods and Groups; 2022. Available from:https://cloud.r-project.org/web/packages/did/index.html

  42. [42]

    Estimators DID with Multiple Groups and Periods [Package]; 2026

    Quispe A, Ciccia D, Knau F, M´ elitine M, Sow D, Zhang S, et al.. Estimators DID with Multiple Groups and Periods [Package]; 2026. Available from:https://github.com/Credible-Answers/did_ multiplegt. 27

  43. [43]

    Difference in Differences with Time-Varying Covariates

    Caetano C, Callaway B, Payne S, Rodrigues HS. Difference in Differences with Time-Varying Covariates. arXiv; 2024. Available from:http://arxiv.org/abs/2202.02903

  44. [44]

    Modern Applied Statistics with S

    Venables WN, Ripley BD. Modern Applied Statistics with S. 4th ed. New York: Springer; 2002. 28 A Identification via Wright’s rules and Cram´ er’s covariance for- mula in the context of linear SCMs Scenario 5.2(Figure 6) Pairwise covariances via Wright’s rules:σ(∆Y, A 1) =a+gijk−gih σ(∆Y, W0) =ga−ih+ijk σ(∆Y, Z0) =iga−h+jk σ(A1, W0) =g σ(A1, Z0) =gi Substi...