Intraclass Correlation Coefficients (ICCs) From Longitudinal Group-Randomized Trials of HIV/STI/Pregnancy Prevention Programs and Effects of Covariate Adjustment

Introduction: Group-randomized trials (GRTs) are one of the most rigorous scientific methods for evaluating the effectiveness of prevention programs delivered in groups (e.g., schools, community-based organizations). Efficiently sizing GRTs with the desired power and precision to assess intervention effects requires estimates of the intraclass correlation coefficient (ICC) – the degree to which outcomes of individuals clustered within groups are correlated. ICC estimates can vary widely depending on outcome, population, and setting, and small changes in ICCs can have large effects on needed sample size. Furthermore, using regression adjustment for covariates and repeated measures analyses can substantially reduce ICCs and therefore needed sample size. The purpose of this study is to add to the small body of existing literature on ICCs in the area of youth sexual risk-taking behaviors and to examine the impact of adjustment for certain covariates on ICCs and sample size requirements.

Methods: The present study applied multilevel regression analyses with and without repeated measures, and with and without various covariate adjustments, to existing data from three federally-funded school-based GRTs of HIV/STI/pregnancy prevention in youth to obtain ICC estimates under a variety of scenarios/conditions. Covariates used for adjustment were related to the outcome variable at the individual level (e.g., student age and GPA, parent education level, indicators of acculturation, living situation and probation status).

Results: ICCs for frequency of unprotected sex in past 3 months, number of unprotected partners, and condom/birth control use at last sex ranged from <0.000 – 0.06, with adjustment for covariates and repeated measurements reducing the ICC in some, but not all, cases. Assuming an average of 4 classes of 25 students per school, the observed ranges of ICCs correspond to needed sample sizes of 8 – 32 schools to detect with 80% power and 0.05 significance the typically small effect sizes (d=0.2) seen in behavioral interventions.

Conclusions: These data provide further evidence that using the appropriate adjusted ICC estimate at the design stage can have great impact on estimates of the number of schools to recruit and therefore where resources are directed in costly GRTs. We found that in most cases the ICC was reduced when adjusting for covariates, implying fewer schools would be needed. However, in some cases adjusting for covariates increased the ICC, further supporting the need for careful consideration of inclusion criteria for covariates at both the design and analysis stages.