One common approach for predicting rare events in the LOGISTIC procedure is to build a model that disproportionately over-re presents those cases with an event occurring (e.g. a 50-50 event/non- event

Question

Options

• AAll parameter estimates are biased.
• BOnly the intercept estimate is biased.
• COnly the non-intercept parameter estimates are biased.
• DSensitivity estimates are biased.

• A
  3% (1)
• B
  84% (27)
• C
  9% (3)
• D
  3% (1)

Explanation

When you oversample rare events to create an artificial 50/50 split, you've distorted the true prevalence of the event in the population. The intercept in logistic regression captures the baseline log-odds of the event occurring, so it directly encodes that prevalence - an inflated event rate in your sample produces an inflated (biased) intercept. The slope coefficients, however, estimate the relationships between predictors and the outcome, and those relationships are not affected by changing the proportion of events in the sample, leaving them unbiased.

Why the distractors are wrong:

• A is wrong because the bias is isolated to the intercept; slopes remain valid.
• C has it exactly backwards - slopes are the unbiased estimates, not the intercept.
• D is wrong because sensitivity (the rate of correctly predicting events) actually improves with oversampling; it is not the parameter that becomes biased.

Memory tip: Think "intercept = baseline rate." The intercept is the only parameter that cares how many events exist in your data. Slopes only care about which direction and how strongly predictors relate to the outcome - information that oversampling doesn't distort.

No community discussion yet for this question.