Compute ITT, TOT, and LATE with noncompliance
Company: Netflix
Role: Data Scientist
Category: Analytics & Experimentation
Difficulty: medium
Interview Round: Onsite
In the same personalization experiment, not everyone assigned to treatment actually receives personalization (noncompliance).
You are given user-level columns:
- `Z` (0/1): randomized assignment (instrument)
- `D` (0/1): actually received personalization (treatment received)
- `Y` (float): `minutes_streamed`
Tasks:
1) Define and compute the **Intention-to-Treat (ITT)** effect of assignment on minutes streamed.
2) Define what people often mean by **Treatment-on-the-Treated (TOT)** in this setting, and compute it.
3) Using an IV interpretation, compute the **LATE** for compliers and explain what assumptions are required.
4) Conceptually classify users into compliance types (compliers, always-takers, never-takers, defiers) and state which group LATE pertains to.
Overview: This question evaluates understanding of causal inference and experimental analysis, specifically estimation and interpretation of ITT, TOT, and LATE under noncompliance along with classification of compliance types (compliers, always-takers, never-takers, defiers).
Read the full Netflix Data Scientist interview experience this question came from