Causal Estimation of Treatment Effects From Observational Health Care Data Using Machine Learning Methods

Abstract

AbstractThe econometrics literature has generally approached problems of causal inference from the perspective of obtaining an unbiased estimate of a parameter in a structural equation model. This requires strong assumptions about the functional form of the model and data distributions. As described in Chapter 3, there is a rapidly growing literature that has used machine learning to estimate causal effects. Machine learning models generally require far fewer assumptions. Traditionally, the identification of causal effects in econometric models rests on theoretically justified controls for observed and unobserved confounders. The high dimensionality of many datasets offers the potential for using machine learning to uncover potential instruments and expand the set of observable controls. Health care is an example of high dimensional data where there are many causal inference problems of interest. Epidemiologists have generally approached such problems using propensity score matching or inverse probability treatment weighting within a potential outcomes framework. This approach still focuses on the estimation of a parameter in a structural model. A more recent method, known as doubly robust estimation, uses mean differences in predictions versus their counterfactual that have been updated by exposure probabilities. Targeted maximum likelihood estimators (TMLE) optimize these methods. TMLE methods are not, inherently, machine learning methods. However, because the treatment effect estimator is based on mean differences in individual predictions of outcomes for those treated versus the counterfactual, super learning machine learning approaches have superior performance relative to traditional methods. In this chapter, we begin with the same assumption of selection of observable variables within a potential outcomes framework. We briefly review the estimation of treatment effects using inverse probability treatment weights and doubly robust estimators. These sections provide the building blocks for the discussion of TMLE methods and their estimation using super learner methods. Finally, we consider the extension of the TMLE estimator to include instrumental variables in order to control for bias from unobserved variables correlated with both treatment and outcomes.