Interpret a Streaming Regression Without Overstating Causality
Interpret a Streaming Regression Without Overstating Causality
You are provided with an Ordinary Least Squares (OLS) regression model designed to explain a user's streaming hours over the last 28 days:
streaming_hours_l28d = beta_0 + beta_tv * primary_device_tv + beta_mobile * primary_device_mobile + beta_tenure * tenure + beta_titles * number_of_titles + error
primary_device_tv and primary_device_mobile are mutually exclusive indicator variables. The omitted reference group consists of users whose primary device is neither TV nor mobile. The regression output includes each coefficient, its standard error, and a 95% confidence interval.
A product manager has interpreted this as a "causal-inference regression."
Your task is to:
- Explain the model and its components to the product manager in clear, non-technical language.
- Evaluate the product manager's proposal: "Because the estimated coefficient on
primary_device_tvis positive, the product should nudge users to make a TV their primary device in order to increase streaming hours."
Constraints & Assumptions:
- The data are observational; no random assignment or natural experiment is implied by the regression output alone.
- All time and count units must be confirmed.
number_of_titlesmay be measured within the same 28-day window as the outcome, affecting its temporal role.- A positive coefficient and a confidence interval excluding zero indicate a conditional association, not necessarily a causal effect.
- You may propose additional data collection or experiments.
Clarifying Questions to Ask:
- How is "primary device" assigned, and when is it measured relative to the outcome?
- Is
number_of_titlesa baseline preference, a behavior during the outcome window, or something else? - Which variables were selected before seeing the results, and what important common causes are available?
- Is the business decision about sending a nudge, changing actual viewing device, or increasing future streaming hours?
- Can users be randomized at the account or household level, and can exposure and device switching be logged?
AI interview prep
Core dilemma
Distinguishing correlation from causation in observational data and designing experiments to establish causal relationships for product decisions.
3 tradeoffs and 3 stress tests locked
Interview prep
Unlock the interviewer X-Ray rubric, tradeoff matrix, stress tests, clarifying questions, and key concepts.
Interpret a Streaming Regression Without Overstating Causality
Interpret a Streaming Regression Without Overstating Causality
You are provided with an Ordinary Least Squares (OLS) regression model designed to explain a user's streaming hours over the last 28 days:
streaming_hours_l28d = beta_0 + beta_tv * primary_device_tv + beta_mobile * primary_device_mobile + beta_tenure * tenure + beta_titles * number_of_titles + error
primary_device_tv and primary_device_mobile are mutually exclusive indicator variables. The omitted reference group consists of users whose primary device is neither TV nor mobile. The regression output includes each coefficient, its standard error, and a 95% confidence interval.
A product manager has interpreted this as a "causal-inference regression."
Your task is to:
- Explain the model and its components to the product manager in clear, non-technical language.
- Evaluate the product manager's proposal: "Because the estimated coefficient on
primary_device_tvis positive, the product should nudge users to make a TV their primary device in order to increase streaming hours."
Constraints & Assumptions:
- The data are observational; no random assignment or natural experiment is implied by the regression output alone.
- All time and count units must be confirmed.
number_of_titlesmay be measured within the same 28-day window as the outcome, affecting its temporal role.- A positive coefficient and a confidence interval excluding zero indicate a conditional association, not necessarily a causal effect.
- You may propose additional data collection or experiments.
Clarifying Questions to Ask:
- How is "primary device" assigned, and when is it measured relative to the outcome?
- Is
number_of_titlesa baseline preference, a behavior during the outcome window, or something else? - Which variables were selected before seeing the results, and what important common causes are available?
- Is the business decision about sending a nudge, changing actual viewing device, or increasing future streaming hours?
- Can users be randomized at the account or household level, and can exposure and device switching be logged?