Event-Study Designs for Discrete Outcomes under Transition Independence

We develop a new identification strategy for average treatment effects on the treated (ATT) in panel data with discrete outcomes. Standard difference-in-differences (DiD) relies on parallel trends, which is frequently violated in categorical settings due to mean reversion, out-of-bounds counterfactuals, and ill-defined trends for multi-category outcomes. We propose an alternative identification strategy with transition independence: absent treatment, transition dynamics conditional on pre-treatment outcomes are identical between control and treated groups. To capture unobserved heterogeneity, we introduce a latent-type Markov structure delivering type-specific and aggregate treatment effects from short panels. Three empirical applications yield ATT estimates substantially different from conventional DiD.

💡 Research Summary

This paper tackles a fundamental problem in the evaluation of policies that affect discrete (categorical) outcomes using panel data. Traditional difference‑in‑differences (DiD) methods rely on a parallel‑trends assumption that the untreated potential outcomes of treated and control units would evolve in the same way. When outcomes are bounded and take on a finite set of states—such as employment status, patenting activity, or disability‑related labor‑force participation—parallel trends is often implausible. The authors identify three specific failures of the parallel‑trends framework for discrete outcomes: (1) mean‑reversion bias because treated and control groups start from different state distributions, (2) the generation of counterfactual means that fall outside the logical bounds of probabilities (e.g., negative complaint rates), and (3) the lack of a coherent notion of a single “trend” for multi‑category outcomes, which makes it impossible to compare joint distributions over time.

To overcome these issues, the authors introduce transition independence as a new identifying assumption. Transition independence states that, in the absence of treatment, the conditional transition probabilities of moving from one state to another—given the pre‑treatment state history—are identical for treated and control units. This assumption can be directly tested with pre‑treatment data and respects the bounded, categorical nature of the outcomes because it works with transition matrices rather than level means.

Two practical challenges remain. First, unobserved heterogeneity may affect both treatment assignment and transition dynamics, violating transition independence at the population level. Second, the number of possible pre‑treatment histories grows exponentially with the number of pre‑treatment periods, creating limited‑support problems. The paper resolves both challenges by (i) introducing a latent‑type finite mixture model, where each unit belongs to an unobserved type that has its own Markov transition matrix, and (ii) imposing a low‑order Markov restriction (e.g., first‑order) so that only the most recent outcomes condition the transition probabilities.

Under this combined “latent‑type Markov” framework, the authors prove identification of type‑specific average treatment effects on the treated (LTATTs) and of the aggregate ATT as a weighted average of LTATTs, where the weights are the posterior probabilities of each latent type among the treated. Importantly, identification requires only short panels (as few as two or three periods), making the approach feasible for many applied settings.

A further contribution is the flow decomposition of treatment effects. Because the model works with state‑to‑state transition probabilities, the overall ATT for a given outcome state can be decomposed into inflow (transitions into the state) and outflow (transitions out of the state) components. This decomposition reveals the specific channels through which a policy operates—information that standard DiD, which compares level differences, cannot provide.

The authors illustrate the methodology with three empirical applications.

1. Dodd‑Frank Act – Using complaint‑rate data, conventional DiD predicts a post‑reform increase in complaints, while the transition‑independence estimator predicts a decrease. Moreover, the DiD counterfactual falls below zero, violating probability logic.

2. Norway’s 2003 Patent Reform – University inventors (treated) had roughly twice the pre‑reform patenting rate of non‑university inventors (control). DiD attributes a 4.5 % decline in patenting to the reform, driven by mean‑reversion bias. The transition‑based estimator finds no statistically significant change, suggesting the DiD estimate is severely biased.

3. Americans with Disabilities Act (ADA) – Monthly labor‑force status from the SIPP panel shows that the negative employment effect of the ADA operates primarily through increased transitions from employment directly to out‑of‑labor‑force status. The flow decomposition isolates this outflow channel, while DiD fails to detect any significant effect.

The paper situates its contribution within two strands of the literature: (a) matching on pre‑treatment outcomes, and (b) DiD extensions that allow for nonlinear dynamics. It improves on prior work by providing a formal identification argument for discrete outcomes, incorporating latent heterogeneity without requiring long panels, and delivering a transparent mechanism analysis via flow decomposition.

In sum, the study demonstrates that for discrete outcomes the parallel‑trends assumption is often untenable, and that transition independence combined with a latent‑type Markov structure offers a robust, data‑driven alternative. The approach yields consistent ATT estimates even with short panels, respects the bounded nature of categorical data, and uncovers the underlying transition channels through which policies exert their effects.