A dichotomous (2-category) outcome variable is often encountered in biomedical research, and Multiple Logistic Regression is often deployed for the analysis of such data. As Logistic Regression estimates the Odds Ratio (OR) as an effect measure, it is only suitable for case-control studies. For cross-sectional and time-to-event studies, the Prevalence Ratio and Cumulative Incidence Ratio can be estimated and easily interpreted. The logistic regression will produce the OR which is difficult to interpret in these studies. In this report, we reviewed 3 alternative multivariate statistical models to replace Logistic Regression for the analysis of data from cross-sectional and time-to-event studies, viz, Modified Cox Proportional Hazard Regression Model, Log-Binomial Regression Model and Poisson Regression Model incorporating the Robust Sandwich Variance. Although none of the models is without flaws, we conclude the last model is the most viable. A numeric example is given to compare the statistical results obtained from all 4 models.
A dichotomous (2-category) outcome variable is indeed ubiquitous in biomedical research enquiries. Here are some examples: (a) a cross-sectional study to compare the prevalence (proportion) of obesity among the adult males and females in Singapore. (b) a clinical trial to compare ethnic differences in the 1-year survival among patients with metastatic non-small cell lung cancer treated with Gefitinib (Iressa). (c) an epidemiologic study to compare the 2-year mortality rate of lung cancer between those who continue to smoke and those who quit after diagnosis.
This article is available only as a PDF. Please click on “Download PDF” to view the full article.