Abstract
With data from a pretest-posttest design with a nonequivalent control group, researchers have two main options for identifying the causal impact of a treatment: covariate adjustment (CA) or difference-in-differences (DID). Each strategy rests on strong assumptions: the adjustment criterion (unconfoundedness) for CA and the common trend assumption for DID. In practice, neither of these assumptions is typically fully satisfied, resulting in biased effect estimates. To address uncertainty about potential assumption violations, we propose reliability-based bounds for the average treatment effect (ATE) that rely on a linear structural causal model and the reliability of the outcome's pretest measure with respect to an unobserved confounder. Unlike existing bracketing approaches, the reliability-based bounds do not require the stationarity assumption and remain valid even when the pretest and posttest measures are on different scales. Using simulated data, we demonstrate that the proposed bounds outperform existing methods by covering the true ATE even when stationarity is violated, while often producing narrower bounds when stationarity holds. The bounds also remain informative under moderate violations of linearity and the single-unobserved-confounder assumption. An empirical example illustrates how the bounds are established based on subject-matter knowledge and that they tend to cover the benchmark effect estimate from a corresponding randomized experiment.