On logistic regression analysis of dichotomized responses

Area of Science:

• Biostatistics
• Statistical Modeling

Background:

• Treatment effect estimation is crucial in clinical studies.
• Logistic regression is commonly used, but its performance with baseline adjustments needs clarification.

Purpose of the Study:

• To evaluate the properties of treatment effect estimates using odds ratios from logistic regression, adjusting for baseline values.
• To compare the statistical power and accuracy of different treatment effect metrics.

Main Methods:

• Analysis of treatment effect estimates from logistic regression models.
• Comparison of adjusted versus unadjusted analyses for odds ratios.
• Evaluation of risk difference as an alternative metric.

Main Results:

• Adjusting for baseline values increases the estimated treatment effect and its standard error, leading to greater statistical power.
• The adjusted odds ratio may not accurately represent the true effect due to baseline value dependency.
• The risk difference derived from logistic regression approximates the true risk difference.

Conclusions:

• Adjusted logistic regression offers increased power for detecting treatment effects.
• Risk difference is a more reliable metric than odds ratio when adjusting for baseline values in this context.
• Different treatment effect metrics show comparable statistical power at the baseline mean.