Selection of Double Machine Learning for Causal Effect Estimation under Weak Identification

Estimating the causal effect of the factor of physician supply on cancer mortality with DoubleML PLIV is complicated by the spatial instruments weakness and the point estimate sensitivity to hyperparameters choice under their default settings. Using a panel of 83 Russian regions for 2012–2019 (663 observations), 15 configurations of a learner, the number of folds, and the control composition were tested; the selection was implemented under an objective function penalising the dispersion across 50 sampling splits and auxilary function approximation quality. Configurations were validated on a simulation calibrated (F ≈ 5) with a known true effect for good measure. It was shown that we can prefer wrongly an undereducated model, whereas with the penalty for undereducation, random forest on real data is firmly in the lead; in the independent simulation, this advantage is affirmed only by contrast with gradient boosting, while it is statistically indistinguishable from lasso and SVM ranger. The results yield a practical recommendation for causal machine learning in regional healthcare with weak instruments.