OBJECTIVES:

In outcomes research, we deal with rare cases frequently. If medications is just approved building a sample size on the claims datasets, which has 4-6 months lag may take some time and you had to work with small numbers for the initial studies. For certain cases, the diagnosis is real so even have a researcher have enough coverage, can not get enough sample size due to characteristics of the disease. Using a Department of Defense dataset, we analyzed a rare events and proposed a model that mitigates the bias. In particular, we analyzed the probability of using IVIg Therapy for a rare skin condition called Bullous pemphigoid.

METHODS: Since standard logit regressions are biased in rare event, we applied two correction methods. The first was Bayesian correction which involved correcting observed fraction in the sample with the fraction from population. We estimated adjusted constant coefficient in the logit model. For the second involved weightining proportions of ones and zeros in the sample to equal the true proportions in the population using non-parametric tests. Kalmogorov-Smirnov two sample test and The Mann-Whitney U tests be used on predicted probabilities of logit regression to see whether differences exists.

RESULTS:

To apply the methodology, we constructed a retrospective cohort study to estimate IVIg Therapy for Bullous pemphigoid. 2,041 patients are identified from Department of Defense data. IVIg therapy were rare (1.02%). The estimated probability was 0.71% for standard models, 0.98% for the prior correction method and 1.03% for the weighted mechanism. The predicted event probabilities were significantly different.

CONCLUSIONS: Standard logistic regression are proven to underestimate probabilities with rare events. Two correction method that we examined mitigate the bias associated with rare events. The predicted event probabilities adjusted for rare event bias were significantly different from the unadjusted ones.