Alternative methods for analyzing cross-sectional studies of common outcomes: A guide for healthcare professionals

Abstract

Objectives

This study compared logistic, Poisson, and log-binomial regression models for estimating prevalence ratios (PRs) in cross-sectional studies with common outcomes, using hypertension prevalence as an applied example. The objective was to identify the most reliable method and reduce misinterpretation when outcome prevalence is high.

Methods

A cross-sectional analysis was conducted on 2022 patient records from King Khalid University Hospital. Hypertension was the primary outcome, aspirin use the exposure, and diabetes mellitus (DM) the confounder. Statistical models included the Mantel–Haenszel prevalence ratio (MHPR, reference), logistic regression, Poisson regression with or without standard error corrections, and log-binomial regression. The MHPR was compared with PRs and 95% confidence intervals (CIs), and percentage changes were used to quantify deviations. Analyses were performed in STATA 17.

Results

The dataset included 43,789 patients. Hypertension prevalence was high (44.7%), aspirin use was reported in 38.6%, and DM in 52.3%. Logistic regression produced inflated estimates, with an unadjusted OR of 4.26 versus MHPR 2.11. After adjusting for DM, the OR declined to 3.78 but still overestimated the association by 110% relative to the MHPR. The Poisson model had the smallest deviation with respect to the adjusted MHPR (0.67% higher), whereas the log-binomial model showed a 2.28% lower value toward the null. Logistic regression yielded a much wider confidence interval (3.74% higher) than the MHPR, whereas Poisson models showed narrower estimated CIs, and the robust and jackknife procedures resulted in minimal differences (0.04% higher).

Conclusion

Logistic regression introduced substantial bias in cross-sectional data with common outcomes. Poisson regression provided more accurate estimates, particularly with robust or jackknife standard errors, while log-binomial regression was valid but prone to convergence issues. Poisson regression with robust standard errors or jackknife standard errors is preferred, to produce reliable PR estimation while avoiding misinterpretation in health research.