Leveraging Machine Learning to Estimate Effect Modification
Sociologists are often interested in estimating and testing whether some causal effect varies by a modifier of interest. The conventional regression estimator for effect modification is inflexible in functional form and prone to misspecification bias. Machine Learning (ML) algorithms can aid the estimation of effect modification in observational studies by controlling for confounders in a highly flexible, automated, yet principled way. Therefore, leveraging ML for effect modification helps reduce misspecification bias and enhance the credibility of causal identification. We introduce a novel estimator that estimates effect modification in a familiar regression framework after using ML algorithms to fit nuisance components of the model. We show that this estimator is more flexible than the conventional regression model while more efficient and suitable for theory-driven sociological research than other ML-based methods. We use the new estimator to study the modification in the effect of a college degree on adult family income by gender and family income in adolescence in the United States. Along these two dimensions, the benefits of a college degree are rather equally distributed.