Estimate heterogeneous treatment effects with causal ML

Q: Estimate heterogeneous treatment effects with causal ML

This question evaluates a candidate's competency in causal machine learning, heterogeneous treatment effect estimation, off-policy evaluation, and fairness-aware policy design using observational data.

Q: How do I approach Machine Learning interview questions?

Machine Learning questions require understanding of core concepts and practice. PracHub provides solutions with explanations to help you master machine learning interviews.

Q: What difficulty level is this interview question?

This is a hard difficulty Machine Learning question, commonly asked during Technical Screen rounds at TikTok.

Q: What role is this question designed for?

This question is commonly asked for Data Scientist candidates at TikTok during technical interviews.

Question

Context

You are given large-scale, logged observational data from an always-on promotion. Each record contains features X (user/context), a binary treatment A (promotion shown or not), and an outcome Y (e.g., revenue in the next 7 days). Treatment assignment was not randomized and may depend on X.

Goal: Learn heterogeneous treatment effects (HTE) to target users and maximize incremental revenue with a safe, fair, and evaluable deployment plan.

Assumptions (state and justify any deviations):

Binary treatment A ∈ {0,1}. Outcome Y is measured over a fixed horizon after exposure and is not affected by future actions (no carryover beyond window). No interference across users. Unconfoundedness given observed X is assumed plausible.
If promotion has a known cost c per exposure, the policy should maximize net incremental revenue (uplift minus cost).

Task

Propose a causal ML approach that addresses the following:

(a) Choose among T-/S-/X-/DR-learners for HTE, and justify your choice using bias–variance and ignorability considerations.

(b) Specify modeling choices for nuisance functions and HTE, and how you will use cross-fitting to reduce overfitting and nuisance bias.

(c) Describe how you will handle positivity/overlap violations and extreme propensities.

(d) Define evaluation metrics (policy value estimation, uplift/Qini, PEHE proxy) and explain how to implement off-policy evaluation with IPW and doubly robust estimators.

(e) Propose fairness constraints (e.g., demographic parity of treatment) and explain how they affect the learned policy.

(f) Outline how you would A/B test the learned policy safely before full deployment.

Estimate heterogeneous treatment effects with causal ML

Quick Overview

Estimate heterogeneous treatment effects with causal ML

Context

Task

Write your answer