← Dropbox Interview Insights

Dropbox·Data Scientist·Technical Phone Screen·Senior

Senior
Jul 2026

Summary

Dropbox data scientist interview with a deep experimentation design question focused on their file recommender system. Single question but it covered basically every dimension of experiment design you can think of, from randomization unit to alpha spending to Simpson's paradox. Dense.

Questions Asked (1)

Q1

Design an online experiment to evaluate a file recommender system in production across multiple organizations, where users within the same org can influence each other's behavior. Cover: primary metrics, secondary metrics, guardrails, randomization unit, clustering and ICC assumptions for sample size, variance reduction, sequential monitoring, ramp schedule, novelty/carryover controls, logging for exposure and attribution, a difference-in-differences fallback if full randomization isn't possible, and stop/ship criteria including how to handle Simpson's paradox across tenants and roles.

A/B Testing & ExperimentationProduct Analytics & MetricsSystem Design
Author's notes

This was a lot.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the experiment lifecycle: design (metrics, randomization, sample size), execution (monitoring, ramp, logging), and decision (stop/ship, Simpson's paradox). Emphasize how network effects and multi-tenancy complicate standard A/B testing and propose concrete solutions like cluster randomization, variance reduction, and difference-in-differences. Conclude with robust decision criteria that account for heterogeneity across tenants and roles.

Pro tip: Proactively address interference by proposing a cluster-randomized design with org as the unit, and pre-specify how you'll detect and resolve Simpson's paradox using tenant- and role-level subgroup analyses.

1. Define Metrics and Randomization

Choose primary (e.g., file engagement rate), secondary (e.g., click-through, session time), and guardrail (e.g., latency, error rate) metrics. Select randomization unit (org or user cluster) to handle interference.

2. Plan Sample Size and Variance Reduction

Estimate sample size using cluster randomization with ICC assumptions; apply variance reduction techniques like CUPED or stratification by org size/role.

3. Implement Monitoring and Ramp Schedule

Set up sequential monitoring with alpha spending, control for novelty/carryover via washout periods, and design a gradual ramp schedule (e.g., 1% → 5% → 50%).

4. Ensure Robust Logging and Attribution

Log exposures, user interactions, and cross-user influences; attribute outcomes correctly by tracking org-level and user-level events.

5. Define Stop/Ship Criteria and Handle Simpson's Paradox

Pre-register decision rules based on primary metric lift and guardrails; analyze subgroups (tenant, role) to detect and resolve Simpson's paradox before shipping.

Key Points to Mention

  • Cluster randomization by organization to account for interference and ICC
  • Variance reduction techniques like CUPED or stratification
  • Sequential monitoring with alpha spending and ramp schedule
  • Logging exposure and attribution for network effects
  • Difference-in-differences as fallback when randomization isn't feasible
  • Simpson's paradox detection via tenant- and role-level subgroup analysis

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.