Doubly robust methods – let’s get more!
So far, we’ve discussed a broad range of methods that can be used to estimate causal effects. In earlier chapters, we discussed linear regression; in later chapters, we discussed matching, propensity score weighting, and the meta-learner framework. The latter allowed us to go beyond the limitations of linear regression by plugging arbitrary machine learning models as base learners. Meta-learners turned out to be very flexible as they offer all the benefits of contemporary machine learning.
Do we need another thing?
As we learned, propensity scores alone can be used to deconfound the data (as in propensity score weighting; note that this only holds for observed confounding). The same is true for regression models, where we can deconfound the data by simply controlling for the right variable(s) by including them in the regression formula. Propensity score models are sometimes referred to as treatment models (as they aim...
