Measure Speaker's Impact Using Propensity Score Matching

Quick Overview

Evaluates causal inference for optional smart speaker ownership using propensity score matching and staggered adoption. Strong answers address self-selection, construct matched or weighted controls, validate balance and pre-trends, estimate ATT, and run placebo and sensitivity analyses.

Measure Speaker's Impact Using Propensity Score Matching

Company: Roku

Role: Data Scientist

Category: Analytics & Experimentation

Difficulty: medium

Interview Round: Technical Screen

##### Scenario The company releases an optional smart speaker and wants to understand its effect on user engagement, but users self-select into purchasing. ##### Question Without being able to run a forced A/B test, how would you measure the causal impact of owning the speaker on engagement? Describe your preferred design and why. ##### Hints Discuss quasi-experiments: difference-in-differences, propensity-score matching, instrumental variables, pre-post checks.

Quick Answer: Evaluates causal inference for optional smart speaker ownership using propensity score matching and staggered adoption. Strong answers address self-selection, construct matched or weighted controls, validate balance and pre-trends, estimate ATT, and run placebo and sensitivity analyses.

|Home/Analytics & Experimentation/Roku
Roku logo
Roku
Jul 12, 2025, 6:59 PM
mediumData ScientistTechnical ScreenAnalytics & Experimentation
41
0

Measure a Speaker's Impact Using Propensity Score Matching

A company releases an optional smart speaker. Some users purchase and link the speaker to their account, while others do not. Users self-select into ownership, and you cannot run a forced A/B test. You want to measure the causal impact of owning the speaker on engagement.

Constraints & Assumptions

  • Assume user-level panel data on engagement, speaker activation timing, rich pre-treatment covariates, marketing exposure, devices, geography, and staggered adoption.
  • Self-selection is a major concern.
  • The preferred design should compare treated users with comparable non-treated or not-yet-treated users.
  • Include assumptions, diagnostics, and sensitivity analysis.

Clarifying Questions to Ask Guidance

  • What engagement outcome matters: streaming hours, sessions, purchases, retention, or device usage?
  • Is treatment defined by purchase, activation, linking, or first use?
  • What pre-treatment history and marketing exposure data are available?
  • Are there never-adopters and not-yet-adopters throughout the study period?

Part 1 - Identification Strategy

How would you measure the causal impact of owning the speaker on engagement?

What This Part Should Cover Guidance

  • Propensity score matching or weighting, difference-in-differences with staggered adoption, event study, or combined matched DiD.
  • Estimand such as ATT.
  • Assumptions: conditional exchangeability, overlap, no anticipation, and parallel trends if using DiD.

Part 2 - Construct Control Group

How would you construct the control group?

What This Part Should Cover Guidance

  • Propensity score model using pre-treatment engagement, trends, demographics, geography, devices, marketing exposure, and prior product usage.
  • Matching, weighting, calipers, common support, and balance diagnostics.
  • Avoiding post-treatment variables.

Part 3 - Estimate and Validate

How would you estimate impact and validate the design?

What This Part Should Cover Guidance

  • Outcome model, matched or weighted comparison, DiD/event-study regression, robust standard errors, placebo tests, pre-trend checks, and sensitivity to unobserved confounding.
  • Heterogeneity and duration effects.

What a Strong Answer Covers Guidance

A strong answer recognizes self-selection, constructs comparable controls with pre-treatment information, validates balance and pre-trends, and reports uncertainty and sensitivity rather than treating ownership as randomized.

Follow-up Questions Guidance

  • What if adopters were already increasing engagement before purchase?
  • How would you handle users influenced by a marketing campaign?
  • What if overlap between adopters and non-adopters is poor?
Loading comments...