What are the assumptions of logistic regression, and how do you verify them?

Example Answer

For binary logistic regression, I assume a Bernoulli outcome and model its log-odds as a linear combination of the specified features. That does not require each raw feature to have a straight-line relationship with probability; I can include splines, transformations, and interactions. I also need an appropriate treatment of dependence. Repeated users or clustered observations may require a different model or cluster-aware inference.

I would check the feature matrix for redundant columns and inspect separation. With complete separation, an unpenalized maximum-likelihood coefficient may not have a finite estimate. Predictors and residuals do not need to be normally distributed, and “no outliers” is not a blanket assumption. I would investigate influential records rather than delete them automatically. On held-out data, I would check calibration, log loss, and performance across relevant groups, then revisit functional form or regularization where needed.

Why this works

It focuses on assumptions and failures that matter to the binary model, avoiding unrelated linear-regression diagnostics and arbitrary universal cutoffs.

Follow-up to prepare for

How would you detect and handle complete separation, and what changes when the target has more than two categories?