On resampling methods for model assessment in penalized and unpenalized logistic regression

TL;DR

This paper evaluates various resampling methods for assessing logistic regression models, revealing biases in leave-one-out crossvalidation and recommending alternative techniques for more accurate performance estimation.

Contribution

It compares the effects of different resampling techniques on model performance metrics in penalized and unpenalized logistic regression, highlighting biases and proposing better methods.

Findings

Leave-one-out crossvalidation biases c-statistics towards zero.

Bias is more severe for ridge regression estimators.

Leave-pair-out and five-fold crossvalidation provide more accurate estimates.

Abstract

Penalized logistic regression methods are frequently used to investigate the relationship between a binary outcome and a set of explanatory variables. The model performance can be assessed by measures such as the concordance statistic (c-statistic), the discrimination slope and the Brier score. Often, data resampling techniques, e.g. crossvalidation, are employed to correct for optimism in these model performance criteria. Especially with small samples or a rare binary outcome variable, leave-one-out crossvalidation is a popular choice. Using simulations and a real data example, we compared the effect of different resampling techniques on the estimation of c-statistics, discrimination slopes and Brier scores for three estimators of logistic regression models, including the maximum likelihood and two maximum penalized-likelihood estimators. Our simulation study confirms earlier studies reporting that leave-one-out crossvalidated c-statistics can be strongly biased towards zero. In addition, our study reveals that this bias is more pronounced for estimators shrinking predicted probabilities towards the observed event rate, such as ridge regression. Leave-one-out crossvalidation also provided pessimistic estimates of the discrimination slope but nearly unbiased estimates of the Brier score. We recommend to use leave-pair-out crossvalidation, five-fold crossvalidation with repetition, the enhanced or the .632+ bootstrap to estimate c-statistics and leave-pair-out or five-fold crossvalidation to estimate discrimination slopes.