Intermediate
Open
Pro
VIF, Multicollinearity and What It Actually Breaks
A churn-driver regression includes logins_7d, logins_30d,
sessions_30d, plan_price and tenure_months. The output:
| Feature | Coef | SE | t | VIF |
|---|---|---|---|---|
| logins_7d | 0.42 | 0.31 | 1.35 | 14.2 |
| logins_30d | −0.11 | 0.09 | −1.22 | 18.7 |
| sessions_30d | 0.05 | 0.06 | 0.83 | 9.8 |
| plan_price | −0.020 | 0.004 | −5.0 | 1.3 |
| tenure_months | −0.008 | 0.002 | −4.0 | 1.1 |
Overall F-test p < 0.0001, R² = 0.38.
- Interpret the VIF of 18.7 concretely. What is R^2 of
logins_30dregressed on the other four features? - The product manager concludes "login activity doesn't matter, only price and tenure do." Explain why that conclusion is not supported.
- Propose two remedies for the collinear block and state, for each, what you gain and what you give up. Would you bother if the model's only purpose were next-month churn prediction?
Share this question