OPL: Stata module for optimal policy learning and multi-action optimal policy learning

OPL is a package for learning optimal policies from data for empirical welfare maximization. Specifically, OPL allows to find "treatment assignment rules" that maximize the overall welfare, defined as the sum of the policy effects estimated over all the policy beneficiaries. OPL learns the optimal policy empirically, i.e. based on data and observations obtained from previous (same or similar) implemented policies. OPL carries out empirical welfare maximization within three policy classes: (i) Threshold-based; (ii) Linear-combination; and (iii) Decision-tree. Empirical welfare maximization requires the estimation of the Conditional Average Treatment Effect (CATE) of the past policy. Currently, OPL estimates CATE via linear and non-linear Regression Adjustment (RA), allowing for the target outcome to be continuous, binary, count, or fractional. The treatment variable of reference must be binary 0/1. opl_ma_fb implements first-best Optimal Policy Learning (OPL) algorithm to estimate the best treatment assignment given an outcome and a set of observed covariates and treatment effects. It allows for different risk preferences in decision-making (i.e., risk-neutral, risk-averse linear, risk-averse quadratic). This command uses linear regression for estimating nuisance conditional means. opl_ma_vf estimates the value-function for multi-action Optimal Policy Learning via three different methods: 1. Regression Adjustment (RA): estimates expected outcomes for each action using regression models. 2. Inverse Probability Weighting (IPW): uses estimated propensity scores to reweigh observations. 3. Doubly Robust (DR): combines RA and IPW for a more robust estimator.