Relaxing Assumptions, Improving Inference: Integrating Machine Learning and the Linear Regression

Valid inference in an observational study requires a correct control specification, but a correct specification is never known. I introduce a method that constructs a control vector from the observed data that, when included in a linear regression, adjusts for several forms of bias. These include nonlinearities and interactions in the background covariates, biases induced by heterogeneous treatment effects, and specific forms of interference. The first is new to political science; the latter two are original contributions. I incorporate random effects, a set of diagnostics, and robust standard errors. With additional assumptions, the estimates allow for causal inference on both binary and continuous treatment variables. In total, the model provides a flexible means to adjust for biases commonly encountered in our data, makes minimal assumptions, returns efficient estimates, and can be implemented through publicly available software.

Attached PDF:

Attributes

License CC BY
DOI https://doi.org/10.1017/S0003055422001022
Type Journal Articles
C-ID POLS 160 - Methods