Screening for non-linearity in binary logistic regression can be achieved by visualizing:

Question

Options

• AA scatter plot of binary response versus a predictor variable.
• BA trend plot of empirical logit versus a predictor variable.
• CA logistic regression plot of predicted probability values versus a predictor variable.
• DA box plot of the odds ratio values versus a predictor variable.

• B
  92% (34)
• C
  3% (1)
• D
  5% (2)

Explanation

Visualizing the empirical logit (log of observed odds) against a predictor reveals whether the log-odds relationship is linear - the core assumption of logistic regression - making a trend plot of empirical logit the correct diagnostic tool. A scatter plot of the raw binary response (0s and 1s) against a predictor (A) is useless for detecting non-linearity since the discrete values obscure any underlying pattern. A plot of predicted probabilities from the fitted model (C) reflects what the model already assumes, so it cannot independently diagnose violations of its own linearity assumption. A box plot of odds ratios (D) is not a standard diagnostic and odds ratios are single summary statistics, not continuous values you'd plot over a predictor.

Memory tip: Think "logit linearity check" - logistic regression assumes a linear relationship on the logit scale, so the diagnostic must also operate on that scale. If the empirical logit trend is curved, you have non-linearity.

No community discussion yet for this question.