A Residual Prediction Test for the Well-Specification of Linear Instrumental Variable Models (2506.12771v1)

Abstract: The linear instrumental variable (IV) model is widely applied in observational studies. The corresponding assumptions are critical for valid causal inference, and hence, it is important to have tools to assess the model's well-specification. The classical Sargan-Hansen J-test is limited to the overidentified setting, where the number of instruments is larger than the number of endogenous variables. Here, we propose a novel and simple test for the well-specification of the linear IV model under the assumption that the structural error is mean independent of the instruments. Importantly, assuming mean independence allows the construction of such a test even in the just-identified setting. We use the idea of residual prediction tests: if the residuals from two-stage least squares can be predicted from the instruments better than randomly, this signals misspecification. We construct a test statistic based on sample splitting and a user-chosen machine learning method. We show asymptotic type I error control. Furthermore, by relying on machine learning tools, our test has good power for detecting alternatives from a broad class of scenarios. We also address heteroskedasticity- and cluster-robust inference. The test is implemented in the R package RPIV and in the ivmodels software package for Python.