Addressing Design Bias Due to Instrumental Variables in Survey Experiments: Considering Violations of the Exclusion Restriction

This study proposed a strategy for causal identification in survey experiments by leveraging heteroscedasticity induced through information provision (Z). Here, Z affects attitudes (D), which in turn influences behavior (Y). While Z might traditionally serve as an instrumental variable (IV), a direct effect from Z to the outcome variable (Y) would invalidate it as an IV due to exclusion restriction violations. Building on prior methods, such as Lewbel’s IV approach, which utilize naturally occurring heteroscedasticity in observational data, this study explored the artificial induction of heteroscedasticity within survey experiments. We developed a new unbiased estimator within linear structural model and conducted simulations to confirm that this new estimator outperforms conventional IV estimators. We further addressed the issue of endogenous non-compliance, and discussed that under the homogeneous function assumption, manipulating information ambiguity can yield a consistent bias-corrected estimator. This approach can enhance the possibility of causal identification by survey experiments and extend heteroscedasticity-based methodologies to complex compliance contexts, offering new tools for robust experimental design.