← CVS Health Interview Insights

CVS Health·Data Scientist·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

CVS Health data scientist interview with a meaty causal inference case study around flu-shot campaign targeting. The whole thing was essentially one long open-ended problem covering modeling, evaluation, policy design, and online testing. Pretty intense for a single session.

Questions Asked (5)

Q1

You have historical campaign logs with randomized holdouts. How would you design a treatment effect modeling approach to decide who to contact via SMS or Email for an upcoming flu-shot campaign, given demographic features, past vaccination history, engagement signals, and known treatment assignment probabilities?

Data ModelingTechnical Trade-offsA/B Testing & Experimentation
Author's notes

This is where I spent most of my mental energy.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a causal inference task with multi-treatment options (SMS, Email, Control) and heterogeneous treatment effects. Propose a modeling approach that leverages the randomized holdouts to estimate conditional average treatment effects (CATE), then use those estimates to optimize contact decisions under business constraints.

Pro tip: Emphasize that you would validate the model's uplift predictions using a holdout set and potentially run a small pilot before full deployment, and discuss how you'd handle non-compliance or partial exposure.

1. Define the causal estimand and data setup

Clarify that you want to estimate the Conditional Average Treatment Effect (CATE) for each treatment (SMS vs. Control, Email vs. Control) given covariates. Ensure the data includes treatment assignment, outcomes (vaccination), and features; check for randomization validity and overlap.

2. Choose a CATE modeling approach

Select a method that handles multiple treatments and high-dimensional features, such as causal forests, X-learner, or meta-learners (T-learner, S-learner) adapted for multi-arm settings. Consider using inverse probability weighting (IPW) to adjust for known treatment probabilities.

3. Incorporate treatment assignment probabilities

Use the known propensity scores (treatment probabilities) to weight the data or as features in the model, ensuring unbiased effect estimates. This is crucial if the randomization probabilities vary across individuals.

4. Validate and evaluate model performance

Assess model performance using metrics like Qini curve, uplift at various percentiles, and calibration. Use cross-validation or a holdout set to avoid overfitting and ensure the model generalizes.

5. Optimize contact decisions and deploy

Use the estimated CATEs to rank individuals by expected uplift for each channel, then allocate contacts to maximize total effect subject to budget or capacity constraints. Consider a pilot to validate before full rollout.

Key Points to Mention

  • Causal inference vs. predictive modeling: focus on estimating treatment effects, not just outcomes.
  • Handling multiple treatments: use multi-arm CATE methods or pairwise comparisons.
  • Leveraging randomized holdouts: use them to validate causal assumptions and estimate effects.
  • Propensity score adjustment: incorporate known treatment probabilities to correct for confounding.
  • Business constraints: optimize under budget, capacity, or channel-specific costs.
  • Evaluation metrics: Qini curve, uplift, and calibration for model selection.

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

Q2

How would you evaluate uplift models offline? Walk through the metrics you'd use and how you'd factor in channel costs like SMS at $0.02 versus Email at $0.001.

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

Qini curves and AUUC came to mind fast, that part was fine.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the business objective and the need for uplift modeling to optimize incremental impact per channel. Then outline the offline evaluation metrics (e.g., Qini, AUUC) and how to incorporate costs by computing incremental profit curves and cost-sensitive metrics. Finally, discuss validation strategies and practical considerations for deployment.

Pro tip: Emphasize that uplift models should be evaluated on their ability to rank individuals by incremental effect, not just overall accuracy, and that cost differences can drastically change the optimal targeting threshold—so always tie metrics back to ROI.

1. Clarify the goal and setup

Explain that uplift modeling aims to estimate the incremental effect of a treatment (e.g., SMS vs. Email) on an outcome. Offline evaluation requires a holdout set with treatment and control groups, and the goal is to rank individuals by predicted uplift to maximize incremental profit.

2. Choose appropriate offline metrics

Use ranking-based metrics like Qini coefficient, AUUC (Area Under the Uplift Curve), or the uplift curve itself. These measure how well the model identifies individuals with positive incremental response. Also consider calibration of uplift predictions if needed.

3. Incorporate channel costs

Convert uplift into expected incremental profit by multiplying predicted uplift by the value of the outcome (e.g., conversion value) and subtracting the treatment cost. For SMS at $0.02 and Email at $0.001, compute the net incremental profit per individual and use it to build a profit curve.

4. Optimize targeting threshold

Determine the optimal threshold for targeting based on the profit curve: target individuals where expected incremental profit > 0. Compare thresholds across channels to decide which channel to use for each segment, considering cost differences.

5. Validate and iterate

Use cross-validation or bootstrapping to assess metric stability. Simulate deployment by applying the model to a holdout set and computing realized incremental profit. Discuss potential biases (e.g., selection bias) and the need for ongoing monitoring.

Key Points to Mention

  • Qini coefficient and AUUC as standard uplift evaluation metrics
  • Uplift curve and its interpretation for ranking quality
  • Cost-sensitive evaluation: incremental profit = uplift * value - cost
  • Channel cost differences: SMS $0.02 vs Email $0.001, and their impact on targeting
  • Optimal targeting threshold based on expected profit
  • Validation techniques: cross-validation, bootstrapping, and holdout testing

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

Q3

Given a budget that allows contacting at most 40% of eligible customers, how would you rank customers and assign them to a channel?

Roadmap PrioritizationTechnical Trade-offsProduct Strategy
Author's notes

Argmax over channel-specific uplift minus cost, pretty standard.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a constrained optimization: maximize total expected value (e.g., incremental profit or retention) subject to a 40% contact limit. Then propose a two-stage approach: first rank customers by predicted uplift or value, then assign the top 40% to the most cost-effective channel based on channel-specific response models and constraints.

Pro tip: Emphasize that the ranking should be based on incremental impact (uplift) rather than pure propensity, and that channel assignment should consider both cost and capacity—this shows you understand causal inference and operational realities.

1. Define objective and constraints

Clarify the business goal (e.g., maximize profit, minimize churn) and the exact constraint: at most 40% of eligible customers can be contacted. Also note any channel-specific constraints (budget, capacity).

2. Build predictive models

Develop models to estimate each customer's expected response or uplift for each channel, along with the expected cost and value. Use historical data and causal methods (e.g., uplift modeling) to avoid selection bias.

3. Rank customers by expected value

For each customer, compute a score that combines predicted uplift and value (e.g., expected incremental profit). Rank customers in descending order of this score.

4. Assign channels optimally

For the top 40% of customers, assign each to the channel that maximizes expected net value (uplift minus cost) while respecting channel capacities. This can be formulated as an assignment problem and solved with linear programming or greedy heuristics.

5. Validate and iterate

Test the strategy via holdout groups or A/B tests, measure actual incremental lift, and refine models and constraints based on results.

Key Points to Mention

  • Uplift modeling vs. propensity modeling: focus on incremental impact.
  • Cost-sensitive ranking: incorporate channel costs and customer value.
  • Channel capacity and budget constraints beyond the 40% limit.
  • Use of optimization techniques (e.g., linear programming) for assignment.
  • Validation through randomized experiments to measure true lift.
  • Scalability and deployment considerations in a production environment.

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

Q4

How would you design an online validation test to compare model-based targeting against uniform random targeting? What success metrics and stopping rules would you define?

A/B Testing & ExperimentationProduct Analytics & Metrics
Author's notes

I went with a gated rollout framing, start small, compare vaccination rates adjusted for baseline differences, use sequential testing with a pre-specified stopping rule for harm.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and defining the target population, then outline a randomized controlled experiment where users are split between model-based targeting and uniform random targeting. Specify primary and secondary success metrics, along with statistical stopping rules that balance early detection with false positive control.

Pro tip: In healthcare, consider ethical and regulatory constraints: ensure the test doesn't disadvantage any group and that metrics align with clinical outcomes, not just engagement. Also, pre-register your analysis plan to avoid p-hacking and build trust with stakeholders.

1. Define Objective and Hypotheses

Clearly state the goal (e.g., increase conversion or engagement) and formulate null and alternative hypotheses for the comparison between model-based and random targeting.

2. Design Experiment and Randomization

Choose the randomization unit (e.g., user, session), ensure proper sample size and power, and set up control (uniform random) and treatment (model-based) groups. Consider stratification if needed.

3. Select Success Metrics

Define primary metric (e.g., conversion rate) and secondary metrics (e.g., revenue, engagement, fairness). Include guardrail metrics to monitor unintended consequences.

4. Establish Stopping Rules

Pre-specify stopping rules: fixed horizon or sequential testing with alpha spending. Include criteria for early stopping for efficacy, futility, or harm, and adjust for multiple comparisons.

5. Analyze and Interpret Results

Use appropriate statistical tests (e.g., t-test, Bayesian methods) to compare groups, check for practical significance, and consider subgroup analyses. Communicate findings with confidence intervals and effect sizes.

Key Points to Mention

  • Randomization unit and sample size calculation to ensure adequate power.
  • Primary and secondary metrics, including business KPIs and guardrail metrics.
  • Statistical stopping rules (e.g., O'Brien-Fleming alpha spending) to control Type I error.
  • Consideration of novelty effects and long-term impact.
  • Ethical and regulatory considerations in healthcare, such as fairness and privacy.
  • Pre-registration of analysis plan to avoid p-hacking and ensure reproducibility.

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

Q5

How would you detect segments where your model is actually causing harm, meaning negative uplift, and what would you do about them in the targeting policy?

Root Cause AnalysisA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

This one surprised me a little.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining negative uplift as a causal effect where treated individuals have worse outcomes than they would have under control, then describe how to estimate it using experimental data or causal inference methods. Explain how to identify segments with negative uplift and outline a targeting policy that excludes or adjusts for these segments, while considering business and ethical implications.

Pro tip: Emphasize that detecting negative uplift requires a causal framework, not just correlation, and that ignoring it can lead to ethical and financial harm. Mention that in healthcare, such as at CVS Health, this is critical for patient safety and regulatory compliance.

1. Define and Measure Negative Uplift

Clearly define negative uplift as a negative causal effect (treatment harms). Use randomized controlled trials (A/B tests) or causal inference methods like uplift modeling to estimate individual treatment effects.

2. Identify Harmful Segments

Analyze the estimated treatment effects to find subgroups where the effect is significantly negative. Use techniques like causal forests, meta-learners, or segmented analysis to detect these segments.

3. Validate and Understand Root Causes

Validate findings with holdout data or additional experiments. Investigate why the model causes harm in these segments (e.g., confounding, data drift, or inherent treatment risk).

4. Adjust Targeting Policy

Modify the targeting policy to exclude harmful segments from treatment or apply alternative interventions. Consider constraints like fairness, business goals, and ethical guidelines.

5. Monitor and Iterate

Continuously monitor the policy's impact, re-evaluate segments, and update the model as new data arrives to prevent future harm.

Key Points to Mention

  • Causal inference vs. correlation: negative uplift is a causal concept, so use methods like randomized experiments or uplift modeling.
  • Uplift modeling techniques: meta-learners (S-, T-, X-learner), causal forests, or two-model approach.
  • Segmentation: use decision trees or clustering on estimated treatment effects to find subgroups with negative uplift.
  • Business and ethical implications: in healthcare, negative uplift could harm patients; must comply with regulations and prioritize safety.
  • Targeting policy adjustments: exclude harmful segments, use alternative treatments, or apply constraints in optimization.
  • Continuous monitoring: set up guardrail metrics and A/B tests to detect and mitigate negative effects over time.

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