When Do Treatment Changes Identify Causal Effects?

Abstract: This paper clarifies the identifying assumptions underlying causal inference based on treatment changes rather than treatment levels, and their relationship to conventional identification strategies. We characterize two distinct structural models, with non-nested identifying assumptions, under which treatment-change identification is valid conditional on observed covariates. We demonstrate that the identifying assumptions relying on treatment changes are generally not nested with those of methods relying on treatment levels, such as selection-on-observables strategies that control for past outcomes, treatments, and covariates, or difference-in-differences approaches that difference outcomes rather than treatments over time. We show, however, that under a random-walk restriction on the treatment process, conditioning on treatment changes is equivalent to conditioning on treatment levels given lagged treatment. This and other equivalence results motivate overidentification tests by jointly considering methods based on treatment levels and changes. Beyond these tests, the non-nesting results carry a structural double robustness implication: an estimator that differences both the outcome and the treatment over time, such as two-way fixed effects regression, remains consistent if either the treatment-change assumption or the parallel-trends assumption holds, without requiring both simultaneously. We characterize the causal models consistent with each method, investigate finite-sample behavior in a simulation study, and present an empirical application to cigarette demand.

• The paper demonstrates that treatment-change-based identification is valid under two distinct structural models that are not nested within levels-based approaches.
• It shows that the non-nesting of assumptions allows for Hausman-type overidentification tests and establishes TWFE estimators' double robustness under dynamic conditions.
• Simulation studies and an empirical application on cigarette demand validate the theoretical insights while highlighting practical differences in policy-relevant estimates.

Identification of Causal Effects via Treatment Changes

Overview

The paper "When Do Treatment Changes Identify Causal Effects?" (2606.02234) provides a rigorous characterization of when causal effects can be consistently estimated by exploiting treatment changes (rather than static treatment levels) in panel or longitudinal data. The key contributions lie in the comparison of identification strategies based on treatment changes, traditional level-based identification (including selection-on-observables and difference-in-differences, DiD), and conditions for joint validity, overidentification testing, and double robustness properties.

Structural Models for Treatment-Change Identification

Two non-nested structural models are identified under which conditioning on treatment changes, given covariate histories, achieves causal identification:

• Model A: Requires additive, time-invariant unobserved confounders ("fixed effects") in the treatment equation and exclusion of dynamic treatment effects (i.e., current outcomes are unaffected by past treatments and confounders are additively separable). Here, differencing the treatment over time removes fixed effects, and identification is valid if contemporaneous changes are exogenous conditional on past covariates.
• Model B: Imposes a random-walk structure on the treatment process; the treatment at time $T$ equals treatment at $T-1$ plus a serially uncorrelated innovation. This stronger restriction grants identification even when dynamic treatment effects are present because treatment changes (the innovation) are orthogonal to all past confounding variables conditional on observed covariates.

The paper emphasizes that neither model nests the other—allowing for fundamental differences in the causal structures under which treatment-change identification holds.

Non-Nesting with Level-Based Strategies

The central novel claim is that the assumptions enabling identification through treatment changes are not nested within those that underpin identification via treatment levels (i.e., levels-based selection-on-observables or DiD). In particular:

• Treatment changes may be exogenous even if level-based assumptions (such as conditional independence of levels given observed history) fail, and vice versa.
• The random-walk treatment process (Model B) is the critical scenario in which strategies based on treatment changes and those based on levels become equivalent: conditioning on the innovation is sufficient for both approaches. Outside this scenario, the two differ fundamentally and can be exploited to construct overidentification tests.

Furthermore, the paper formally details the divergence of their assumptions using a suite of structural examples and directed acyclic graphs (DAGs), showing, for instance, how dynamic treatment effects and feedback from treatments to covariates violate identifiability via treatment changes but not via levels.

Overidentification Testing and Double Robustness

Given that treatment-change-based and level-based strategies are generally non-nested but coincide under special structural conditions (e.g., the treatment follows a random walk), the paper justifies Hausman-type overidentification tests. These tests compare estimates from both approaches: significant divergence indicates failure of at least one set of identifying assumptions.

A foundational implication is a structural double robustness (DR) result for two-way fixed effects (TWFE) estimators. TWFE remains consistent for the effect of treatment if either the assumptions for identification via treatment changes or the parallel-trends (DiD) assumption hold, but does not require both. This is established for settings in which both outcomes and treatments are differenced, complementing recent results on double robustness in event-study literature.

Simulation and Empirical Illustration

Simulation studies demonstrate each estimator's (CIA-∆D, CIA-D, CIA-∆Y, TWFE) finite-sample performance in settings where only one, multiple, or none of the necessary identification assumptions hold. The outcomes confirm the precise non-nesting of strategies, as well as the DR property of TWFE.

An empirical application to state-level panel data on cigarette demand shows the divergent estimates across identification strategies. Notably, the treatment-change-based estimator yields substantially larger price elasticity (in absolute terms) compared to the more tightly clustered DiD-, selection-on-observables-, and TWFE-based estimates. Hausman tests consistently reject the equivalence of treatment-change against level-based strategies, with the latter group in mutual agreement—providing strong evidence against the joint validity of the assumptions behind both approaches.

Nonparametric causal forest estimators broadly corroborate the OLS patterns and highlight the need for careful structural justification when employing treatment-change-based designs.

Implications and Future Directions

The analysis clarifies that exploiting quasi-randomness in treatment changes is valid only under explicit, testable structural conditions, and that these should not be naively conflated or substituted for standard levels-based identification strategies. Empirical researchers must critically evaluate and explicitly state the structure of their identifying assumptions when relying on treatment changes, especially with dynamic or policy-relevant treatments.

The formal equivalence (under a random walk) and non-nesting enable robust sensitivity analysis and overidentification testing in empirical applications. The double robustness of TWFE offers a practical safeguard in the presence of uncertainty about the data-generating process, though it does not substitute for detailed structural scrutiny.

Potential future research directions include:

• Extending identification analysis to settings with multiple or endogenous treatment switching, staggered adoption, or heterogeneous dynamic effects.
• Developing systematic diagnostics for the empirical plausibility of random walk or exclusion-type structural assumptions.
• Tactical integration of doubly robust estimation in complex panel or high-dimensional settings, leveraging machine learning.

Conclusion

This paper provides a rigorous structural foundation for identification using treatment changes, demonstrates the broad non-nesting with existing level-based approaches, and develops principled inferential diagnostics for cases where both strategies could be considered. The results establish clear guidance and necessary caution for researchers choosing between, or combining, these strategies in causal inference with longitudinal data.