Modeling censored-in-the-middle condom use data

Many HIV prevention programs seek to reduce the risk of infection through increases in condom use. Condom use is often expressed as a proportion: the number of sexual contacts in which a condom is used divided by the total number of sexual contacts. The distribution of this proportion has several undesirable characteristics-the principal one is bimodality. Bimodality results from excessive numbers of 0% and 100% responses, creating distributions that are censored-in-the-middle. The purpose of this paper is to show how censored condom use data can be usefully modeled using Tobit regression. Tobit regression first transforms observed variable scores into latent variable scores, scores on an unobserved, hypothetical condom use variable, and then the latent variable scores are modeled using one or more explanatory variables. Data from the National Institute on Drug Abuse Cooperative Agreement for HIV/AIDS Community-Based Outreach/Intervention Research program for frequency of condom use and number of drug injecting sex partners were used to illustrate the method. We found that for every additional drug injecting sex partner, the probability of using condoms decreased by approximately 1%.