Regression to the mean and changes in risk behavior following study enrollment in a cohort of U.S. women at risk for HIV

Abstract

Purpose: Reductions in risk behaviors are common following enrollment in human immunodeficiency virus (HIV) prevention studies. We develop methods to quantify the proportion of change in risk behaviors that can be attributed to regression to the mean versus study participation and other factors.

Methods: A novel model that incorporates both regression to the mean and study participation effects is developed for binary measures. The model is used to estimate the proportion of change in the prevalence of "unprotected sex in the past 6 months" that can be attributed to study participation versus regression to the mean in a longitudinal cohort of women at risk for HIV infection who were recruited from ten U.S. communities with high rates of HIV and poverty. HIV risk behaviors were evaluated using audio computer-assisted self-interviews at baseline and every 6 months for up to 12 months.

Results: The prevalence of "unprotected sex in the past 6 months" declined from 96% at baseline to 77% at 12 months. However, this change could be almost completely explained by regression to the mean.

Conclusions: Analyses that examine changes over time in cohorts selected for high- or low- risk behaviors should account for regression to the mean effects.

Keywords: HIV; Logistic regression; Risk behavior; Statistical model.