Assumptions and Diagnostics in Logistic Regression
What you'll learn: How to verify the key assumptions underlying logistic regression and check whether your model is behaving properly.
Why Assumptions Matter
Unlike linear regression which assumes normality of residuals, logistic regression has different—and often fewer—strict assumptions. But violating the assumptions it does have can still produce misleading results about the relationship between predictors and your binary outcome.
Core Assumptions of Logistic Regression
1. Linearity in the Log-Odds
The relationship between predictors and the log-odds (logit) must be linear. If age predicts disease risk, the logit of disease probability should change linearly with age—not the probability itself. A curved relationship in log-odds space means logistic regression won't fit well.
Check it: Plot log-odds against each continuous predictor; look for straight-line patterns, not curves.
2. Independence of Observations
Each observation must be independent—one person's outcome shouldn't influence another's. Think: patient records are independent; repeated measures on the same patient are not.
Check it: Review your data collection process. Clustered or repeated-measures data needs special GLM techniques.
3. Absence of Multicollinearity
Just like multiple linear regression, predictors shouldn't be highly correlated with each other. If height (cm) and height (inches) are both predictors, coefficients become unstable and uninterpretable.
Check it: Calculate variance inflation factors (VIF) or examine correlation matrices among predictors.
Common Diagnostics
Beyond assumptions, check for influential outliers using deviance residuals and leverage measures. Large residuals suggest observations the model fits poorly; high leverage points have unusual predictor combinations.
Key Takeaway: Logistic regression requires linearity in log-odds (not probability), independent observations, and low multicollinearity—verify these before trusting your model's coefficients and predictions.