Poisson versus logistic regression in a descriptive epidemiologic analysis of data from a Birth Defects Registry

Abstract

The objective of this study was to compare results from four statistical approaches to describe the epidemiology of a birth defect. Cases of tracheoesophageal fistula and/or esophageal atresia were selected from the Texas Birth Defects Registry, and live births from Texas birth certificate data. All were delivered in 1999 through 2009. Data on infant and maternal characteristics were taken from the vital record. Four approaches were compared: a cohort design using Poisson regression (PR) and three case-control analyses using logistic regression (LR) with 10 controls per case (LR10), four controls per case (LR4), or affected controls having other birth defects. Analyses were conducted for all 11 years and for just 1 year (2009). Using all 11 years, there were 869 cases of tracheoesophageal fistula and/or esophageal atresia, and results for PR, LR10, and LR4 were similar. Results using only 2009 were more divergent: PR yielded more statistically significant global tests (type III analyses) and narrower 95% confidence intervals, followed closely by LR10. For both time periods, results using affected controls were different from those using PR and LR. For these descriptive epidemiologic analyses, PR yielded the most precise estimates. However, LR provided similar estimates, and using 10 controls per case yielded precision almost as good as PR.