Hello all,
I have two related questions about modeling with imbalanced binary outcomes in logistic regression.
1. Rare events vs class imbalance
In King and Zeng’s “Logistic Regression in Rare Events Data” (https://gking.harvard.edu/files/0s.pdf), the authors discuss bias in logistic regression when the event of interest is very rare. My question is: should rare events be viewed simply as an extreme case of class imbalance, or is “rare-event bias” a distinct statistical problem beyond ordinary class imbalance?
2. Improving predictive performance with low default rates
I am working on a personal project to predict loan default using logistic regression. This is not a research paper, so my primary goal is predictive performance. In my dataset, the default rate is about 2 percent to 3 percent. I am considering approaches to improve predictive performance under this degree of imbalance.
One idea is to use weighted logistic regression. However, I am unsure whether this is appropriate. Since logistic regression is estimated by maximum likelihood, adding weights effectively changes the likelihood contribution of the minority class, which alters the underlying probability model. My concern is that weighting may distort the estimated probabilities rather than improving prediction in a principled way.
Given a default rate in the 2–3 percent range, what methods are recommended in Stata to enhance predictive performance without compromising the interpretation of predicted probabilities? Are weighting, rare-events logistic regression (e.g., firthlogit or King and Zeng’s correction), or resampling approaches appropriate here?
Thank you.
I have two related questions about modeling with imbalanced binary outcomes in logistic regression.
1. Rare events vs class imbalance
In King and Zeng’s “Logistic Regression in Rare Events Data” (https://gking.harvard.edu/files/0s.pdf), the authors discuss bias in logistic regression when the event of interest is very rare. My question is: should rare events be viewed simply as an extreme case of class imbalance, or is “rare-event bias” a distinct statistical problem beyond ordinary class imbalance?
2. Improving predictive performance with low default rates
I am working on a personal project to predict loan default using logistic regression. This is not a research paper, so my primary goal is predictive performance. In my dataset, the default rate is about 2 percent to 3 percent. I am considering approaches to improve predictive performance under this degree of imbalance.
One idea is to use weighted logistic regression. However, I am unsure whether this is appropriate. Since logistic regression is estimated by maximum likelihood, adding weights effectively changes the likelihood contribution of the minority class, which alters the underlying probability model. My concern is that weighting may distort the estimated probabilities rather than improving prediction in a principled way.
Given a default rate in the 2–3 percent range, what methods are recommended in Stata to enhance predictive performance without compromising the interpretation of predicted probabilities? Are weighting, rare-events logistic regression (e.g., firthlogit or King and Zeng’s correction), or resampling approaches appropriate here?
Thank you.
