A practical guide for multivariate analysis of dichotomous outcomes

Area of Science:

• Biostatistics
• Epidemiology
• Health Research Methods

Background:

• Dichotomous outcome variables are common in biomedical research.
• Multiple Logistic Regression (MLR) is frequently used but estimates Odds Ratios (ORs).
• ORs are difficult to interpret in cross-sectional and time-to-event studies, unlike Prevalence Ratios or Cumulative Incidence Ratios.

Purpose of the Study:

• To review alternative multivariate statistical models to MLR for analyzing dichotomous outcomes.
• To compare the interpretability and viability of different models for specific study designs.
• To identify a suitable replacement for MLR in cross-sectional and time-to-event research.

Main Methods:

• Review of three alternative multivariate statistical models: Modified Cox Proportional Hazard Regression, Log-Binomial Regression, and Poisson Regression with Robust Sandwich Variance.
• Comparison of these models against traditional Logistic Regression.
• Illustrative numeric example to compare statistical results.

Main Results:

• Logistic Regression yields Odds Ratios, which are less interpretable in cross-sectional and time-to-event studies.
• Alternative models like Modified Cox, Log-Binomial, and Poisson Regression offer more interpretable effect measures (e.g., ratios).
• Poisson Regression with robust variance emerged as the most viable option among the reviewed alternatives.

Conclusions:

• Standard Logistic Regression is not optimal for all study types involving dichotomous outcomes.
• Poisson Regression with robust variance provides a more interpretable and viable approach for cross-sectional and time-to-event studies.
• Careful selection of statistical models is crucial for accurate interpretation of effect measures in epidemiological research.