Code for Minimax Policy Learning under Unobserved Confounding

ConfoundingRobustPolicy is a wrapper object. The main choices are:

• Optimizer
  • Gradient-based approaches
  • Global optimization
• Uncertainty set and weight subproblem solution
  • This is passed to the gradient-based approach as a function callback for computing the optimal weights.
• Performance Evaluation
  • simulation (known dgp)
  • observational data (pass in propensity scores, evaluate by IPW)
• Type of treatment
  • binary treatment (we reparametrize with respect to the probability of assigning T=1)
  • multiple treatments

The .fit() method is specialized to handle various combinations of the above configurations.

ConfoundingRobustPolicy takes as input:

• baseline_pol (function returning baseline policy)

ConfoundingRobustPolicy.fit() takes as input:

• X data
• T data (integer-coded)
• Y
• q0 nominal propensities
• log gamma series of sensitivity parameters to optimize over (some approaches leverage the nested structure of uncertainty sets
• optimization params a dictionary of optimization parameters
• eval_conf a dictionary of evaluation parameters.

optimization params is a dictionary with the following configuration parameters:

-optimizer: function callback, e.g. get_opt_tree_policy. -pol_opt: name of policy class

• unc_set_type: indicator of uncertainty set type (interval or budgeted)
• opt_params: method-specific parameters (e.g. step size for gradients; tree depth for optimal tree)
• BASELINE_POL: ctrl_p_1_mt, -type:'IPW'

Performance Evaluation

• For multiple treatments: oracle_risk is a function that takes in the n x k matrix of policy assignment probabilities (robust, and baseline), integer-coded treatments. oracle_risk sums over treatment partitions.

Replication code Run methods_test.py to replicate simulation from paper.