Invited Commentary: The Seedy Side of Causal Effect Estimation with Machine Learning.

Machine learning algorithms have become an increasingly popular tool for data analysis, including causal effect estimation. Once a causal parameter is expressed in terms of the observed data (i.e., identified), the next task is selecting an appropriate estimator. Suppose a sufficient adjustment set is W. In the case where W takes only a few different values, one may be able to nonparametrically estimate the outcome process (i.e., the nuisance function for the g-formula) or the propensity score (i.e., the nuisance function for the inverse probability weighting estimator). However, W is often high dimensional in epidemiology, so nonparametric estimators may no longer be feasible. This issue has been referred to as the curse of dimensionality. Instead, one can use a model that “borrows” information from observations with “nearby” values of W. However, this borrowing does not come freely. The validity of the results now relies on the correct specification of the model for the nuisance function. As parametric models specified by epidemiologists can be relatively simplistic, flexible data-adaptive machine learning algorithms are thought to better meet model specification assumptions. However, the use of machine learning brings complications.