Abstract:For a binary outcome Y, generated by a simple threshold crossing mod& with a single exogenous normally distributed explanatory variable X, the OLS estimator of the coefficient on X in a linear probability model is a consistent estimator of the average partial effect of X. Even in this very simple setting, we show that when allowing for X to be endogenously determined, the 2SLS estimator, using a normally distributed instrumental variable Z does not identify the same causal parameter. It instead estimates the average partial effect of Z, scaled by the coefficient on Z in the linear first-stage model for X, denoted gamma(1), or equivalently, it estimates the average partial effect of the population predicted value of X, Z gamma(1). These causal parameters can differ substantially as we show for the normal Probit model, which implies that care has to be taken when interpreting 2SLS estimation results in a linear probability model. Under joint normality of the error terms, IV Probit maximum likelihood estimation does identify the average partial effect of X. The two-step control function procedure of Rivers and Vuong can also estimate this causal parameter consistently, but a double averaging is needed, one over the distribution of the first-stage error V and one over the distribution of X. If instead a single averaging is performed over the joint distribution of X and V, then the same causal parameter is estimated as the one estimated by the 2SLS estimator in the linear probability model. The 2SLS estimator is a consistent estimator when the average partial effect is equal to 0, and the standard Wald test for this hypothesis has correct size under strong instrument asymptotics. We show that, in general, the standard weak instrument first-stage F-test interpretations do not apply in this setting.

Interpreting TSLS Estimators in Information Provision Experiments

The causal interpretation of estimated associations in regression models

Inference on LATEs with covariates

Causal inference under over-simplified longitudinal causal models

Overidentification in Shift-Share Designs

The interpetation of 2SLS with a continuous instrument: A weighted LATE representation

Causal Interpretation of Estimands Defined by Exposure Mappings

Stein-like estimators for causal mediation analysis in randomized trials

Binary Outcomes, OLS, 2SLS and IV Probit

Learning causal effects from many randomized experiments using regularized instrumental variables

Causal Inference for Human-Language Model Collaboration

C-Learner: Constrained Learning for Causal Inference and Semiparametric Statistics

Information Equivalence in Survey Experiments

A Framework for Eliciting, Incorporating, and Disciplining Identification Beliefs in Linear Models

On the implied weights of linear regression for causal inference

Program Evaluation and Causal Inference with High-Dimensional Data

Causal Interpretation of Regressions With Ranks

Inferring Effect Ordering Without Causal Effect Estimation

Valid Causal Inference with (Some) Invalid Instruments

Causal mediation analysis: From simple to more robust strategies for estimation of marginal natural (in)direct effects

Bayesian Predictive Synthesis For Causal Inference With TSCS Data: A Multilevel State-Space Factor Model With Hierarchical Shrinkage Priors