When Dynamic Panel Regressions Misread Treatment Effects

A coefficient is not automatically a causal effect

A country’s economic performance can influence its political choices, while those choices may affect future performance. Similarly, a firm’s productivity can influence whether it exports, and exporting may change its subsequent productivity. These are examples motivating dynamic panel methods, which follow multiple units over time and account for persistent outcomes and unit-specific differences.

In the theoretical preprint Heterogeneous Treatment Effects via Dynamic Panel Data Methods, Philip Marx, Elie Tamer and Xun Tang ask what a treatment coefficient means when effects differ across units and depend on earlier treatment. Their framework allows treatment to start, stop and recur. It excludes spillovers between units and effects of future treatments on present outcomes.

The central warning concerns a particular first-differenced instrumental-variable estimand: the population quantity targeted by the specified regression procedure. A familiar regression form does not, by itself, make its treatment coefficient an interpretable average causal effect.

Why accounting for history may not suffice

The authors first examine a baseline period followed by two treatment periods. Their regression relates changes in outcomes to changes in treatment and a lagged outcome change. Its instruments are conditional predictions of those regressors given the baseline outcome and first-period treatment. This benchmark need not coincide with a regression using only raw lagged variables as instruments.

They consider sequential exchangeability: after conditioning on observed treatment and outcome history, current treatment carries no additional information about the relevant present and future potential outcomes—or their means, in the weaker version. With suitable overlap in treatment probabilities, this assumption already supports established methods for identifying causal effects.

But it does not ensure that this IV coefficient averages those effects appropriately.

The authors decompose the coefficient into a nonnegatively weighted average of current effects and two additional terms. One reflects departures from those desirable weights. The other reflects outcome trends that would occur without current treatment, potentially retaining consequences of earlier treatment. Because lagged outcomes themselves contain earlier treatment responses, heterogeneity in those responses also affects the regression’s weighting.

Consequently, the coefficient can be negative even when every current treatment effect is the same strictly positive number.

What restores an interpretable average?

The paper characterizes restrictions under which this problem disappears. Restrictions on how earlier mean trends and effects vary with baseline outcomes remove trend contamination and make the treatment-effect weights sum to one. Those weights can still be negative when current effects vary.

Under full sequential exchangeability, an additional restriction—that the mean current effect for each prior treatment history does not vary with baseline and preceding potential outcomes—guarantees the paper’s squared-weight, nonnegative averaging formula.

The necessity claims have an important qualification: they concern guarantees across compatible distributions or treatment-assignment rules, holding specified parts of the model fixed. Failure of a restriction does not mean every particular application fails. However, the authors show that predictable effect heterogeneity can combine with a weak first stage to make the coefficient arbitrarily positive or negative across compatible assignment rules.

Identification without sequential exchangeability

A separate result takes a different route. The authors impose a first-order autoregressive structure on outcomes without current treatment, while allowing treatment gains to vary across units, dates and histories.

Under conditional mean-zero restrictions on outcome shocks, an additional restriction on treatment gains, and a full-rank instrument condition, a history-conditioned first-differenced IV regression identifies the average current effect among units treated for the first time, conditional on their baseline outcome. The result applies from period three onward.

The gain restriction requires that, within this first-time-treated group and baseline condition, fixed effects and past shocks provide no further information about the mean gain. It does not require treatment choices to ignore those latent factors. This is identification under alternative structural assumptions, not assumption-free inference. Formal asymptotic analysis of the proposed flexible estimation approach lies beyond the paper’s identification results.

Equal pre-trends can hide later bias

The authors also distinguish sequential exchangeability from parallel trends, which concerns comparisons of untreated outcome trends across treatment groups. Randomization conditional on observed history supports sequential exchangeability, but does not generally imply parallel trends when later assignment responds to earlier outcomes.

Their pre-trend example makes the distinction concrete. With persistent untreated outcomes, equal baseline and next-period variances, and treatment selected using the sum of those two outcomes plus independent noise, treated and untreated groups have equal mean pre-treatment changes.

Nevertheless, selection partly reflects the latest temporary shock. Its subsequent disappearance creates unequal untreated trends after treatment begins. A pre-trend test then has no population discrepancy to detect, even though parallel trends fails. The lesson is specific: equal earlier trends cannot, on their own, validate a later comparison when treatment responds to recent outcomes.