Intervention effects estimated from nonrandomized intervention studies are plagued by biases, yet social or structural intervention studies are rarely randomized. There are underutilized statistical methods available to mitigate biases due to self-selection, missing data, and confounding in longitudinal, observational data permitting estimation of causal effects. We demonstrate the use of Inverse Probability Weighting (IPW) to evaluate the effect of participating in a combined clinical and social sexually transmitted infection/human immunodeficiency virus prevention intervention on reduction of incident chlamydia and gonorrhea infections among sex workers in Brazil.
We demonstrate the step-by-step use of IPW, including presentation of the theoretical background, data set up, model selection for weighting, application of weights, estimation of effects using varied modeling procedures, and discussion of assumptions for use of IPW.
A total of 420 sex workers contributed 840 data points on incident chlamydia and gonorrhea infection. Participators were compared with nonparticipators following application of inverse probability weights to correct for differences in covariate patterns between exposure groups and between those who remained in the intervention and those who were lost-to-follow-up. Estimators using 4 model selection procedures provided estimates of intervention effect between odds ratio 0.43 (95% CI, 0.22–0.85) and 0.53 (95% CI, 0.26–1.1).
After correcting for selection bias, loss-to-follow-up, and confounding, our analysis suggests a protective effect of participating in the intervention. Evaluations of behavioral, social, and multilevel interventions to prevent sexually transmitted infection can benefit by introduction of weighting methods such as IPW.
Inverse probability weighting methods are demonstrated step-by-step to evaluate a combined clinical and social sexually transmitted infection/human immunodeficiency virus intervention study with sex workers in Brazil.
From the *Division of Epidemiology, University of California, Berkeley, CA; †Center for Aids Prevention Studies, University of California, San Francisco, CA; and ‡Division of Biostatistics, University of California, Berkeley, CA
The authors gratefully acknowledge Juan Díaz and Magda Chinaglia, investigators on the Encontros study, Angela Donini, the study coordinator, and the study advisory committee and institutional partners (The Population Council, the Brazilian Ministry of Health, Pathfinder do Brasil, Rede Brasileira de Prostitutas, and the state and municipal Secretary of Health) for their dedication to the project that produced these data. We thank Adriana Pinho for data cleaning and Jennifer Ahern and Art Reingold for providing comments on this manuscript.
Data collection for the Encontros project was supported by the Population Council and the Ministry of Health in Brazil. The first author received support from the Fogarty AIDS International Training and Research Program (AITRP) (grant 1 D43 TW00003 to S.A.L.) at the School of Public Health, University of California, Berkeley, while drafting this manuscript.
Correspondence: Sheri A. Lippman, PhD, 50 Beale St, suite 1200, San Francisco CA 94105. E-mail: LippmanS@globalhealth.ucsf.edu.
Received for publication September 18, 2009, and accepted February 2, 2010.