Pith. sign in

REVIEW 4 major objections 5 minor 36 references

State-policy heterogeneity analyses should bound state-specific effects instead of estimating conditional averages, because bounds can determine a policy's sign for individual states even when point identification fails.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A difference-in-differences bounding method is proposed that reports ranges for state-specific policy effects, and it recovers effect signs more reliably than conditional average effect estimates in simulations.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection State-specific effect bounds are a genuinely useful shift of target, but the data-informed tau undercalibrates in the paper's own simulations and the application reads more certainty than the method supports. the 4 major comments →

arxiv 2602.08643 v2 pith:R3WD5XOZ submitted 2026-02-09 stat.ME stat.AP

State policy heterogeneity analyses: considerations and proposals

Max Rubinstein , Megan S. Schuler , Elizabeth A. Stuart , Bradley D. Stein , Max Griswold , Elizabeth M. Stone , Beth Ann Griffin This is my paper
classification stat.ME stat.AP MSC 62D20
keywords state policy evaluationtreatment effect heterogeneitystate-specific treatment effectsconditional average treatment effectspartial identificationdifference-in-differencessensitivity analysistreatment coarsening
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper is trying to establish that in state-policy studies, the standard target of heterogeneity analyses—the conditional average treatment effect (CATE)—is usually not the quantity policymakers need, and that researchers should instead bound the state-specific treatment effect (ITE): the effect of the policy on a particular state. The authors argue that the CATE is comparatively easy to identify but only supports associational statements ('effects are larger in states with…'), while the ITE is policy-relevant but not point-identified without strong assumptions. Their proposal is a difference-in-differences bounding framework in which each state's effect is reported as an interval, with the interval width governed by an explicit sensitivity parameter that can be informed by pre-treatment prediction errors. In simulations, these ITE bounds determine the sign of the true state effect more reliably than CATE confidence intervals do. The payoff is concrete: applied to Medicaid expansion and high-volume buprenorphine prescribing, the bounds point to a negative effect in Illinois and a positive effect in New Mexico, despite conventional average-effect estimates being near zero and statistically insignificant.

Core claim

The central claim is that when the data cannot point-identify a state-level policy effect, the inferential goal should shift from estimating a conditional average to bounding the individual (state-specific) effect. Under a bounded-counterfactual-error assumption—the error in predicting what each state would have done without (or with) the policy is no larger than a chosen τ_i—each state's ITE lies in an interval centered on a unit-level difference-in-differences estimate. The paper develops this interval, shows how pre-treatment residuals can inform τ_i, and proves that untreated states generally require larger sensitivity parameters than treated states (Z1 > Z0). Simulations show that the r

What carries the argument

The key machinery is the unit-level difference-in-differences estimator ˆψ^{DiD}_{i,m} combined with Assumption 2 (bounded counterfactual error), which converts an unverifiable point-identifying assumption into an interval ψ_{i,m} ∈ [ˆψ^{DiD}_{i,m} − τ_i, ˆψ^{DiD}_{i,m} + τ_i]. The width τ_i is the sensitivity parameter; the paper proposes setting τ_i = Z∥˜ε_i∥ using pre-treatment residuals, with Lemma 2 and Proposition 6 showing how the residual norm can be scaled for treated versus untreated units. Treatment coarsening enters through a two-stage randomization framework that interprets coarsened DiD estimates as stochastic interventions over policy versions, and the paper offers three ways

Load-bearing premise

The load-bearing premise is that the pre-treatment residuals, scaled by a constant, genuinely cover the size of the error in predicting each state's no-policy (or with-policy) outcome; the paper notes this is not guaranteed, and its own simulations show that for control states the data-informed bounds cover the true effect only about 57–59% of the time at the chosen Z.

What would settle it

A falsifying experiment would simulate states where the post-treatment counterfactual error is, by construction, larger or more skewed than the pre-treatment residual norm used for τ; if the data-informed bounds cover the true ITE at rates well below the intended level even at the recommended Z, the calibration assumption fails. In the application, re-estimating the Illinois and New Mexico bounds with a longer pre-period and a different comparison pool would provide a direct check: if either state's interval flips sign or contains zero under a moderate Z, the headline application conclusion wo

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can report 'the effect in state i lies between L and U' and interpret it as the causal effect for that state, rather than an average over states with similar characteristics.
  • In the paper's simulations, the ITE bounds determine the sign of the true state effect more reliably than CATE confidence intervals, especially for treated states.
  • Pre-treatment residuals can be used to calibrate sensitivity parameters, and varying the multiplier Z gives a tipping-point analysis that is directly interpretable on the scale of observed pre-period deviations.
  • When treatment is coarsened, untreated-state bounds require larger sensitivity parameters; the paper formalizes why and gives three concrete strategies (known M(1), union of bounds, conservative τ).
  • In the application, the bounds identify Illinois as having a negative and New Mexico as having a positive Medicaid-expansion effect on high-volume buprenorphine prescribing, consistent across eight specifications.

Where Pith is reading between the lines

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

  • A natural extension the paper leaves implicit is applying the same residual-calibration bounding to staggered adoption and dynamic effects; the logic should carry over, but the counterfactual error bound would need to account for treatment timing.
  • The paper's own simulation shows data-informed bounds covering control-state ITEs only about 57–59% of the time at Z=2, suggesting practitioners should treat Z as a sensitivity knob tilted toward larger values for untreated states, not as a calibrated confidence level.
  • Because CATE confidence intervals target a projected average rather than individual states, published state-policy heterogeneity findings based on interaction terms may overstate confidence in which states benefit; reanalysis with these bounds could change policy rankings.
  • The bounding approach could be combined with placebo-covariate balance checks or Bayesian hierarchical models as an additional robustness overlay, though the paper mentions Bayesian models only in passing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper addresses a recurring problem in state-level policy heterogeneity analyses: standard estimands such as CATEs and their linear projections are often not aligned with policy questions about what happened or would happen in a specific state. It formalizes the distinction between ITEs, CATEs, and CDEs, and shows how treatment coarsening further complicates interpretation. The main proposal is to abandon point identification and instead bound state-specific ITEs under a bounded counterfactual error assumption (Assumption 2), with sensitivity parameters informed by pre-treatment residuals. Formal results include worst-case bounds under bounded errors (Proposition 5), a probabilistic MSE-based lemma (Lemma 2), and a normality-based scaling result for control-state errors (Proposition 6). A simulation compares ITE bounds with CATE confidence intervals, and an application to Medicaid expansion and high-volume buprenorphine prescribing reports robust-looking negative bounds for Illinois and positive bounds for New Mexico across 6 of 8 specifications.

Significance. The conceptual contribution is genuinely useful. Distinguishing ITE, CATE, and CDE, and making the case that CATEs are often associational rather than causal for state-specific questions, is an important clarification for applied work. The bounding framework is transparent and interpretable: when the true tau* is known, the interval contains the ITE by construction, and the oracle version gives correct coverage. The paper is also unusually open about limitations: Section 4.2 explicitly concedes that no pre-data rule guarantees tau_i >= tau_i^star, and Table 4 reports the consequences. The availability of code for simulations and application is a concrete strength. However, the paper's headline claims—that the method can "more reliably determine the sign of the ITEs than CATE estimates" and that the Illinois/New Mexico findings are strongly suggested—depend on calibration choices that the paper does not deliver. The framework is defensible as a sensitivity analysis, but not as a calibrated procedure with the stated coverage/validity implications.

major comments (4)
  1. [Section 4.2 and Table 4] The central claim rests on the data-informed sensitivity parameter tau_i^m = Z ||epsilon_i||. The paper itself states in Section 4.2 that "none of these approaches guarantee tau_i^m >= tau_i^star", so the bounds are not guaranteed to contain the ITE. Table 4 quantifies the problem: at Z=2, coverage for treated units is 0.833–0.842 and for control units only 0.570–0.587. These are far below any nominal confidence level. Thus the abstract's claim that the bounds "can more reliably determine the sign of the ITEs than CATE estimates" and the application's "strongly suggest" language overstate what the simulations establish. The paper should either calibrate Z (e.g., by reporting coverage over a grid of Z and choosing a value that meets a target) or explicitly reframe the intervals as pure sensitivity intervals with no coverage guarantee. This is not a presentation issue; it is load-bearing f
  2. [Proposition 6, Remark 11, and Section 4.4] The formal support for control-state bounds requires assumptions that are not met in the simulation or application. Proposition 6 assumes normality of the errors and known mean/scale parameters alpha and gamma, plus inequality (11); none of these are verifiable from pre-period data. More importantly, Remark 11 says that for untreated units one should generally choose Z1 > Z0, but the simulation (Section 4.4) and the application (Section 5.1) set Z1 = Z0 = 2. The simulation's poor control-state coverage (0.54–0.59) is, in part, a direct consequence of not following the paper's own guidance. This matters because the control-state bounds are the weakest link. The authors should either implement a calibration that respects Remark 11 and report coverage, or explain why equal Z1 and Z0 is a deliberate choice and temper the conclusions accordingly.
  3. [Section 4.4 and Table 4] The simulation's headline comparison is between ITE bounds and CATE confidence intervals, but the CATE estimator targets a linear projection of a coarsened CATE, not the ITE. It is unsurprising that intervals for one estimand have poor coverage and sign recovery for a different estimand. The claim "bounding state-specific effects can more reliably determine the sign of the ITEs than CATE estimates" is therefore a comparison of incomparable objects. A fairer assessment would compare the ITE bounds against an estimator that also targets the ITE, or at least report the CATE comparison with an explicit caveat that the two methods answer different questions. The paper does note this caveat, but the abstract and Section 6 still present the comparison as evidence for the bounding method.
  4. [Section 5.1 and Table 3] The application's robustness claim is summarized only as counts of "6 of 8" specifications for Illinois and New Mexico. Table 3 does not report which specification produced which sign, the values of Z used, or how the alternative specifications differ in assumptions. Since the eight specifications are not independent—they share the same data, the same pre-treatment residual construction, and the same choice Z1=Z0=2—the count "6 of 8" overstates the evidential weight. The authors should provide a full specification-by-specification table and show how the classification changes as Z is varied, rather than only reporting a binary count.
minor comments (5)
  1. [Definition 5] The coarsened ITE is denoted tilde(psi), whereas Definition 1 uses psi for the ITE; the notation should be consistent to avoid confusion.
  2. [Section 3.4] "Cohon's d" should be "Cohen's d".
  3. [Section 1] "implemeted" is a typo for "implemented".
  4. [Section 4.2] The paragraph introducing Z^m would benefit from stating explicitly that Z=0 corresponds to exact parallel trends and that larger Z does not make the interval a confidence interval; the current wording is close but could be sharper given the paper's emphasis on transparency.
  5. [Appendix A.2.2] Equation (11), used in the text and in the skeptical summary, does not have a visible equation number in the manuscript. Please number all key equations referenced in the text.

Circularity Check

0 steps flagged

No circular reasoning: the ITE bounds are an explicit sensitivity analysis, not a fitted prediction.

full rationale

The derivation of the ITE bounds is an explicit, assumption-driven sensitivity analysis rather than a fitted prediction. Assumption 2 states |Ŷ_i,T(m)−Y_i,T(m)| ≤ τ_i^m, and equation (2) then places ψ_i,m in [ψ̂_i,m^DiD − τ̂_i^m, ψ̂_i,m^DiD + τ̂_i^m]. The paper immediately acknowledges that choosing τ = τ⋆ makes Assumption 2 hold by definition and that the open question is how to choose τ. This is a transparent tautology, not a hidden reduction: the bounds are not fitted to the target ITE, and no guarantee is claimed for data-informed τ. The simulation uses a fixed Z=2, reports control-state coverage of 0.57–0.59 in Table 4, and explicitly disclaims formal statistical guarantees, so the method is not presented as a calibrated predictor. The application's Illinois/New Mexico statements are robustness counts over eight sensitivity specifications and remain conditional on an unvalidated τ choice; that is a validity limitation, not circularity. Load-bearing support comes from external references (Manski & Pepper, Rambachan & Roth, VanderWeele & Hernán), and the authors' self-citations are contextual—prior estimates of Medicaid expansion effects and high-volume prescriber concentration—rather than used to force the paper's conclusions. No step in the claimed derivation chain reduces to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 8 axioms · 1 invented entities

The central method depends on user-chosen sensitivity parameters (Z, norm) and unverifiable assumptions about counterfactual errors. The invented entity M_i(1) is a theoretical construct needed to make coarsened ITEs well-defined, with no independent empirical evidence.

free parameters (4)
  • Sensitivity parameter Z^m = Z=2 in simulation; Z in {1, 1.5, 2} in application
    Scales pre-treatment residuals to set tau; user-chosen rather than estimated. Central to bound width and validity.
  • Norm choice for pre-treatment residuals = MAE in simulation; L-infinity in application
    The choice of ||epsilon|| changes tau and therefore the reported bounds; multiple norms are permitted.
  • Rurality threshold in application = States with more than 25% of buprenorphine prescribers in rural communities
    Defines a binary effect modifier; the threshold is an analyst choice and affects stratification.
  • Mean/variance scaling for untreated-state error (alpha, gamma) = User-specified; examples alpha=1, gamma=1 or alpha=gamma=0
    Proposition 6 uses alpha and gamma to relate unobservable counterfactual error to pre-treatment residuals under normality; they are not estimated from data.
axioms (8)
  • domain assumption SUTVA, consistency, and no anticipatory effects
    Invoked in Section 3.1 and Assumptions 3.a/4.b so that observed outcomes equal potential outcomes under observed treatment.
  • domain assumption A single well-defined control condition M=0
    Needed to define Y_i(0) and the ITE; the paper acknowledges in a footnote that in reality the control regime may also be versioned.
  • domain assumption Parallel trends (Assumptions 1, 5.a, 5.b)
    Used to form the starting DiD estimate Yhat_i,T(m); motivates the construction in Section 4.1 and the propositions in Appendix A.1.2.
  • domain assumption Bounded counterfactual error (Assumption 2)
    The core bounding assumption: |Yhat_i,T(m) - Y_i,T(m)| <= tau_i. It is guaranteed for the unobserved true tau*, but not for any data-informed choice.
  • domain assumption Pre-treatment residuals are informative about post-treatment counterfactual error
    Used in Section 4.2, Remark 9, and Proposition 6 to justify tau = Z||epsilon_i|| and the MSE bound C = omega||epsilon_i||^2.
  • ad hoc to paper Normality and mean/variance scaling of errors (Proposition 6)
    Invoked only to derive Z1 = C_alpha_gamma Z0 for untreated states; not required for the core bounds under known tau.
  • domain assumption Exclusion restriction: A affects Y only through M
    Assumption 4.a, needed to define Y(A=1)=Y(M(1)) under treatment coarsening.
  • domain assumption Positivity of treatment assignment
    Assumptions 3.c/4.d; required for the CATE identification expressions and weights.
invented entities (1)
  • M_i(1): counterfactual policy version assignment for untreated states no independent evidence
    purpose: Defines the coarsened ITE psi_{i,M_i(1)} and coarsened CATE for untreated states under treatment coarsening.
    Latent assignment variable; identified only under ignorability assumptions 5.c/5.d or user choices; no direct falsifiable handle outside the paper.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of State policy heterogeneity analyses: considerations and proposals." pith.science (2026). https://pith.science/paper/R3WD5XOZ

@misc{pith2026260208643,
  author       = {Pith},
  title        = {Pith review of: State policy heterogeneity analyses: considerations and proposals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R3WD5XOZ}},
  note         = {Machine review of arXiv:2602.08643}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

State-level policy studies often conduct heterogeneity analyses that quantify how treatment effects vary across state characteristics. These analyses may be used to inform state-specific policy decisions, or to infer how the effect of a policy changes in combination with other state characteristics. However, in state-level settings with varied contexts and policy landscapes, multiple versions of similar policies, and differential policy implementation, the causal quantities targeted by these analyses may not align with the inferential goals. This paper clarifies these issues by distinguishing several causal estimands relevant to heterogeneity analyses in state-policy settings, including state-specific treatment effects (ITE), conditional average treatment effects (CATE), and controlled direct effects (CDE). We argue that the CATE is often the easiest to identify and estimate, but may not be the most policy relevant target of inference. Moreover, the widespread practice of coarsening distinct policies or implementations into a single indicator further complicates the interpretation of these analyses. Motivated by these limitations, we propose bounding ITEs as an alternative inferential goal, yielding ranges for each state's policy effect under explicit assumptions that quantify deviations from the ideal identifying conditions. These bounds target a well-defined and policy-relevant quantity, the effect for specific states. We develop this approach within a difference-in-differences framework and discuss how sensitivity parameters may be informed using pre-treatment data. Through simulations we demonstrate that bounding state-specific effects can more reliably determine the sign of the ITEs than CATE estimates. We then illustrate this method to examine the effect of the Affordable Care Act Medicaid expansion on high-volume buprenorphine prescribing.

Figures

Figures reproduced from arXiv: 2602.08643 by Beth Ann Griffin, Bradley D. Stein, Elizabeth A. Stuart, Elizabeth M. Stone, Max Griswold, Max Rubinstein, Megan S. Schuler.

Figure 1
Figure 1. Figure 1: Causal estimands highlights that a common target of inference – the OLS projection of the coarsened CATE – may yield little insight about the causal relationships between a policy change and an effect modifier even in a simple setting. Appendix B provides a more general but simple mathematical illustration showing these results using linear models, again illustrating why the CATE, CDE, and ITE diverge. 3.4… view at source ↗
Figure 2
Figure 2. Figure 2: Simulation results data-informed bounds. However, the performance gap remains between treated and control units. This underscores that the method’s performance depends on how well τ 1−Ai i is chosen. We again emphasize, however, that the CATE intervals target a projected CATE rather than the ITE, so it is not surprising this method does not target the ITE well. Finally, additional results in Appendix D sho… view at source ↗
Figure 3
Figure 3. Figure 3: Simulation results pansion—which, starting in 2014, allowed states to receive federal funding to expand Medicaid eligibility to 138% of the federal poverty line—was not associated with high-volume prescribing [30]. By contrast, manda￾tory PDMP laws, which require prescribers to check the PDMP before prescribing restricted substances such as opioids, were associated with a decrease in high-volume prescribin… view at source ↗
Figure 4
Figure 4. Figure 4: State-specific effect bounds 5.1 State-specific effects We now instead take the approach that our goal is not to estimate average effects, CATEs, or CDEs, but to bound the state-specific effects. We use the difference-in-differences style approach outlined above, constructing the state-specific effect counterfactual estimates using the pre-treatment outcomes for that state plus the term given by the trends… view at source ↗

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

36 extracted references · 2 linked inside Pith

  1. [1]

    Adaptive interventions in child and adolescent mental health.Journal of Clinical Child & Adolescent Psychology, 45(4):383–395, 2016

    Daniel Almirall and Andrea Chronis-Tuscano. Adaptive interventions in child and adolescent mental health.Journal of Clinical Child & Adolescent Psychology, 45(4):383–395, 2016

    2016

  2. [2]

    Princeton university press, 2009

    Joshua D Angrist and J¨ orn-Steffen Pischke.Mostly harmless econometrics: An empiricist’s companion. Princeton university press, 2009

    2009

  3. [3]

    Autoregressive models for panel data causal inference with application to state-level opioid policies.arXiv preprint arXiv:2408.09012, 2024

    Joseph Antonelli, Max Rubinstein, Denis Agniel, Rosanna Smart, Elizabeth Stuart, Matthew Cefalu, Terry Schell, Joshua Eagan, Elizabeth Stone, Max Griswold, et al. Autoregressive models for panel data causal inference with application to state-level opioid policies.arXiv preprint arXiv:2408.09012, 2024

    arXiv 2024

  4. [4]

    Models as approximations i.Statistical Science, 34(4):523–544, 2019

    Andreas Buja, Lawrence Brown, Richard Berk, Edward George, Emil Pitkin, Mikhail Traskin, Kai Zhang, and Linda Zhao. Models as approximations i.Statistical Science, 34(4):523–544, 2019. 27

    2019

  5. [5]

    Difference-in-differences with multiple time periods.Jour- nal of econometrics, 225(2):200–230, 2021

    Brantly Callaway and Pedro HC Sant’Anna. Difference-in-differences with multiple time periods.Jour- nal of econometrics, 225(2):200–230, 2021

    2021

  6. [6]

    Double/debiased machine learning for treatment and structural parameters, 2018

    Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. Double/debiased machine learning for treatment and structural parameters, 2018

    2018

  7. [7]

    All medicaid expansions are not created equal: the geography and targeting of the affordable care act

    Craig Garthwaite, John A Graves, Tal Gross, Zeynal Karaca, Victoria R Marone, and Matthew J Notowidigdo. All medicaid expansions are not created equal: the geography and targeting of the affordable care act. Technical report, National Bureau of Economic Research, 2019

    2019

  8. [8]

    Cambridge university press, 2007

    Andrew Gelman and Jennifer Hill.Data analysis using regression and multilevel/hierarchical models. Cambridge university press, 2007

    2007

  9. [9]

    Phillip Heiler. Heterogeneous treatment effect bounds under sample selection with an application to the effects of social media on political polarization.Journal of Econometrics, 244(1):105856, 2024

    2024

  10. [10]

    Effect or treatment heterogeneity? policy evaluation with aggregated and disaggregated treatments

    Phillip Heiler and Michael Knaus. Effect or treatment heterogeneity? policy evaluation with aggregated and disaggregated treatments. 2022

    2022

  11. [11]

    A causal framework for evaluating drivers of policy effect heterogeneity using difference-in-differences.Health Services and Outcomes Research Methodology, pages 1–22, 2025

    Gary Hettinger, Youjin Lee, and Nandita Mitra. A causal framework for evaluating drivers of policy effect heterogeneity using difference-in-differences.Health Services and Outcomes Research Methodology, pages 1–22, 2025

    2025

  12. [12]

    Jennifer L. Hill. Bayesian nonparametric modeling for causal inference.Journal of Computational and Graphical Statistics, 20(1):217–240, 2011

    2011

  13. [13]

    Statistics and causal inference.Journal of the American statistical Association, 81(396):945–960, 1986

    Paul W Holland. Statistics and causal inference.Journal of the American statistical Association, 81(396):945–960, 1986

    1986

  14. [14]

    Nonparametric estimation of average treatment effects under exogeneity: A review

    Guido W Imbens. Nonparametric estimation of average treatment effects under exogeneity: A review. Review of Economics and statistics, 86(1):4–29, 2004

    2004

  15. [15]

    Semiparametric doubly robust targeted double machine learning: a review.arXiv preprint arXiv:2203.06469, 2022

    Edward H Kennedy. Semiparametric doubly robust targeted double machine learning: a review.arXiv preprint arXiv:2203.06469, 2022

    Pith/arXiv arXiv 2022

  16. [16]

    Wild bootstrap inference for wildly different cluster sizes

    James G MacKinnon and Matthew D Webb. Wild bootstrap inference for wildly different cluster sizes. Journal of Applied Econometrics, 32(2):233–254, 2017

    2017

  17. [17]

    The wild bootstrap for few (treated) clusters.The Econometrics Journal, 21(2):114–135, 2018

    James G MacKinnon and Matthew D Webb. The wild bootstrap for few (treated) clusters.The Econometrics Journal, 21(2):114–135, 2018. 28

    2018

  18. [18]

    Randomization inference for difference-in-differences with few treated clusters.Journal of Econometrics, 218(2):435–450, 2020

    James G MacKinnon and Matthew D Webb. Randomization inference for difference-in-differences with few treated clusters.Journal of Econometrics, 218(2):435–450, 2020

    2020

  19. [19]

    How do right-to-carry laws affect crime rates? coping with ambiguity using bounded-variation assumptions.Review of Economics and Statistics, 100(2):232–244, 2018

    Charles F Manski and John V Pepper. How do right-to-carry laws affect crime rates? coping with ambiguity using bounded-variation assumptions.Review of Economics and Statistics, 100(2):232–244, 2018

    2018

  20. [20]

    Emma E McGinty, Nicholas J Seewald, Sachini Bandara, Magdalena Cerd´ a, Gail L Daumit, Matthew D Eisenberg, Beth Ann Griffin, Tak Igusa, John W Jackson, Alene Kennedy-Hendricks, et al. Scaling interventions to manage chronic disease: innovative methods at the intersection of health policy research and implementation science.Prevention Science, 25(Suppl 1)...

    2024

  21. [21]

    Bayesian hierarchical models.Jama, 320(22):2365–2366, 2018

    Anna E McGlothlin and Kert Viele. Bayesian hierarchical models.Jama, 320(22):2365–2366, 2018

    2018

  22. [22]

    Direct and indirect effects

    Judea Pearl. Direct and indirect effects. InProbabilistic and causal inference: the works of Judea Pearl, pages 373–392. 2022

    2022

  23. [23]

    A more credible approach to parallel trends.Review of Economic Studies, 90(5):2555–2591, 2023

    Ashesh Rambachan and Jonathan Roth. A more credible approach to parallel trends.Review of Economic Studies, 90(5):2555–2591, 2023

    2023

  24. [24]

    Heterogeneous interventional effects with multiple mediators: Semiparametric and nonparametric approaches.Journal of Causal Inference, 11(1):20220070, 2023

    Max Rubinstein, Zach Branson, and Edward H Kennedy. Heterogeneous interventional effects with multiple mediators: Semiparametric and nonparametric approaches.Journal of Causal Inference, 11(1):20220070, 2023

    2023

  25. [25]

    Max Rubinstein, Amelia Haviland, and David Choi. Balancing weights for region-level analysis: The effect of medicaid expansion on the uninsurance rate among states that did not expand medicaid.The Annals of Applied Statistics, 17(2):1469–1490, 2023

    2023

  26. [26]

    dabblers

    Brendan Saloner, Barbara Andraka Christou, Adam J Gordon, and Bradley D Stein. Article commen- tary: It will end in tiers: A strategy to include “dabblers” in the buprenorphine workforce after the x-waiver.Substance Abuse, 42(2):153–157, 2021

    2021

  27. [27]

    Growing importance of high-volume buprenorphine prescribers in oud treatment: 2009–2018

    Megan S Schuler, Andrew W Dick, Adam J Gordon, Brendan Saloner, Rose Kerber, and Bradley D Stein. Growing importance of high-volume buprenorphine prescribers in oud treatment: 2009–2018. Drug and alcohol dependence, 259:111290, 2024

    2009

  28. [28]

    Schuler, Beth Ann Griffin, Magdalena Cerd´ a, Emma E

    Megan S. Schuler, Beth Ann Griffin, Magdalena Cerd´ a, Emma E. McGinty, and Elizabeth A. Stuart. Methodological challenges and proposed solutions for evaluating opioid policy effectiveness.Health Services and Outcomes Research Methodology, 21:21–41, 2021. 29

    2021

  29. [29]

    Schuler, Sara E

    Megan S. Schuler, Sara E. Heins, Rosanna Smart, Beth Ann Griffin, David Powell, Elizabeth A. Stuart, Bryce Pardo, Sierra Smucker, Stephen W. Patrick, Rosalie Liccardo Pacula, and Bradley D. Stein. The state of the science in opioid policy research.Drug and Alcohol Dependence, 214:108137, 2020

    2020

  30. [30]

    High-volume buprenorphine prescribers: examining state policy contexts.Drug and Alcohol Dependence Reports, page 100406, 2026

    Megan S Schuler, Flora Sheng, Brendan Saloner, Adam J Gordon, and Bradley D Stein. High-volume buprenorphine prescribers: examining state policy contexts.Drug and Alcohol Dependence Reports, page 100406, 2026

    2026

  31. [31]

    In- vestigating the complexity of naloxone distribution: Which policies matter for pharmacies and potential recipients.Journal of health economics, 97:102917, 2024

    Rosanna Smart, David Powell, Rosalie Liccardo Pacula, Evan Peet, Rahi Abouk, and Corey S Davis. In- vestigating the complexity of naloxone distribution: Which policies matter for pharmacies and potential recipients.Journal of health economics, 97:102917, 2024

    2024

  32. [32]

    Bradley D Stein, Rachel K Landis, Flora Sheng, Brendan Saloner, Adam J Gordon, Mark Sorbero, and Andrew W Dick. Buprenorphine treatment episodes during the first year of covid: a retrospective examination of treatment initiation and retention.Journal of general internal medicine, 38(3):733–737, 2023

    2023

  33. [33]

    Causal inference under multiple versions of treatment

    Tyler J VanderWeele and Miguel A Hernan. Causal inference under multiple versions of treatment. Journal of causal inference, 1(1):1–20, 2013

    2013

  34. [34]

    conditional average treatment effects

    Brian G Vegetabile. On the distinction between “conditional average treatment effects” (cate) and “individual treatment effects” (ite) under ignorability assumptions.arXiv preprint arXiv:2108.04939, 2021. 30 A Theoretic results A.1 Identification results: treatment coarsening We consider causal identification in the presence of treatment coarsening from a...

    Pith/arXiv arXiv 2021

  35. [35]

    selection- on-observables

    To motivate Proposition 1, we make the following assumptions: Assumption 3.a(Consistency). Y= kX m=0 I(M=m)Y(m). Consistency states that the observed values ofYis equal to the potential outcome under the respective treatment. Assumption 3.b(Ignorability). Y(m)⊥M|X, m= 0, . . . , k. Ignorability states thatMis effectively randomized with respect toY(m) con...

  36. [36]

    fundamental problem of causal inference:

    We consider the following model of state-level potential outcomes with respect to both a multi-valued interventionM, some effect modifying covariate of interest covariateX, and a second effect modifierU, where, for simplicity, we have that (X i, Ui) iid ∼ N  0,   σ2 x σxu σxu σ2 u    . The potential outcomesY i(m, x) are then given by Yi(0, x)...

    2013

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.