Propose a multi-touch attribution model and its data pipeline
Tests your grasp of attribution models and their data engineering needs. Propose a rule-based model (e.g., time-decay), outline the data pipeline for it, and acknowledge privacy-driven signal loss. A red flag is ignoring the challenge of identity resolution.
What's really being asked
This question tests your ability to weigh the trade-offs between model complexity, implementation cost, and business value. The interviewer wants to see if you understand that perfect attribution is impossible due to modern signal loss (iOS 14.5, cookie deprecation) and that the core challenge is often data governance and identity resolution, not the specific algorithm. It assesses your practical data systems design skills and business pragmatism.
The full answer
A strong answer addresses four key areas in order. First, model selection and justification: Acknowledge the long sales cycle and propose a rule-based model like Time-Decay or a Position-Based (U-shaped) model. Justify this as a pragmatic starting point because it's interpretable, auditable, and doesn't have the high data requirements of ML models. Second, data pipeline architecture: Describe the key stages of ingesting data from all sources (CRM, ad platforms, web analytics), transforming it with a heavy focus on identity resolution to stitch together user touchpoints, and applying the model logic. Third, parameters and limitations: Discuss the critical importance of setting an appropriate attribution window (e.g., 90-180 days for B2B), noting a short window can shift credit by over 20%. Explicitly mention the reality of signal loss, which has reduced MTA coverage to 30-60% of 2020 levels. Fourth, an evolution path: Briefly mention advanced models like Markov Chains or Shapley Values as a future step, but only after the foundational data pipeline is solid and data volume exceeds 2,000+ conversions per month.
The mistakes people make
A common mistake is jumping straight to a complex, black-box ML model without discussing the high data requirements (2,000+ conversions/month), interpretability trade-offs, or the massive engineering effort required. Another red flag is describing a model without explaining the data processing needed to support it; the question has two parts. Ignoring the impact of modern privacy changes like iOS 14.5, cookie deprecation, or walled gardens shows a lack of current, real-world knowledge. Finally, the biggest red flag is failing to mention identity resolution, as this is the hardest technical part of the problem and ignoring it suggests a superficial understanding.
What usually comes next
Expect questions like: How would you handle offline touchpoints, like a sales call? How would you validate that your model is better than last-touch? What specific identity resolution techniques would you use (e.g., deterministic vs. probabilistic)? What would the schema look like for a touchpoints table in your data warehouse?
A concrete example
For a 90-day sales cycle, I'd propose a Time-Decay model with a 90-day lookback window and a 7-day half-life. The pipeline would ingest events from Segment, Google Ads, and Salesforce. We'd build a user journey table in our data warehouse, keyed on a resolved user_id. An hourly dbt job would stitch anonymous IDs to known user IDs based on email capture events. The final attribution model would be another dbt model that processes these journeys, assigns credit, and outputs a table showing credited revenue per channel per day.
Interview question
When initially proposing a multi-touch attribution (MTA) model, what is the primary pragmatic reason to recommend a rule-based model like Time-Decay over a complex machine learning model?
- a.Rule-based models can incorporate offline touchpoints more easily than ML models.
- b.Complex ML models are only effective for very short sales cycles and high conversion volumes.
- c.Rule-based models are inherently more resistant to privacy-driven signal loss.
- d.They offer a better balance of interpretability, auditability, and lower initial data/engineering requirements.Correct
Why? this is the answer
The card emphasizes that rule-based models are a pragmatic starting point due to their interpretability, auditability, and lower data/engineering requirements, especially compared to ML models that need high data volumes. Distractor A is incorrect because signal loss is a general challenge for all attribution models, not a specific advantage of rule-based ones.
Just read this? Test yourself on what you have been reading.
Read the original → improvado.io
- #analytics
- #data engineering
- #attribution
- #metrics
You just looked this up. Could you explain it out loud?
That is the part interviews actually test. Tezvyn takes questions like this one and gives you what the interviewer is really checking, the answer that lands, and the mistake that ends the conversation, in the four minutes before your next meeting.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.
We are hiring for this. Open roles that interview on analytics — each one lists the topics its interview covers.
See open roles