Policy Targeting from Causal Inference to Production

Context

You completed a causal-inference project estimating the effect of a binary marketing treatment (e.g., a targeted offer) on a business outcome (e.g., customer spend or profit) using observational and/or experimental data. You now need to productionize a targeting policy that treats customers who are most likely to benefit, subject to operational and fairness guardrails.

Assume:

• Treatment T ∈ {0,1} delivered at time t.
• Outcome Y measured after treatment (e.g., 90-day profit or conversion).
• Feature vector X with potential confounders (e.g., demographics, historical spend, engagement, credit/risk, channel, time/seasonality, eligibility).

Tasks

1. Causal graph and estimands
   • Formulate a DAG for treatment → outcome including key confounders. Justify conditional independence assumptions (e.g., ignorability, SUTVA, positivity).
   • State your estimands: ATE and ITE/uplift (CATE).
2. Method choice and overlap
   • Choose and justify a method to estimate heterogeneous treatment effects (e.g., doubly robust learner, causal forest, uplift gradient boosting). Explain pros/cons.
   • Describe how you will check overlap/positivity and how you would handle violations.
3. Training and validation
   • Describe sample splitting and cross-fitting for nuisance models (propensity and outcome models).
   • Explain how to tune hyperparameters without biasing effect estimates.
   • Describe how you would use policy risk/off-policy evaluation (IPW/DR) to compare targeting policies.
4. Diagnostics
   • Produce uplift curves and Qini coefficients; discuss calibration and interpretation.
   • Describe sensitivity analysis for unobserved confounding (e.g., Rosenbaum bounds) and balance checks.
5. Deployment
   • Translate ITEs into a treatment policy with operational guardrails (budget, eligibility, risk), fairness constraints across cohorts, and post-deployment monitoring.