Evaluation of log-rank tests for infrequent observations from a multi-state process, with application to HPV vaccine...

Genital infection by human papillomavirus (HPV) is a common sexually transmitted disease, with over 25 per cent prevalence among young women in the US. Infections are usually without symptoms and transient (or reversible), but a small proportion of infections persist and are believed to be responsible for nearly all cervical cancers and precursor lesions such as cervical intraepithelial neoplasia (CIN). Therefore, successful vaccines against persistent HPV infections could have a great impact in preventing cervical cancers. In trials being planned, ongoing, and recently completed, a log-rank or a similar test may be employed to assess a vaccine effect in comparison to placebo, with an infection 'event' defined to capture persistent but not transient infections. However, it is not clear how best to define such an event, because (1) diagnostic tests cannot distinguish a persistent from a transient infection, (2) participants are only examined periodically, and (3) there can be misclassification errors in the detection of infections. This paper evaluates several definitions of persistent infection that are based on periodically observed infection statuses by postulating a multi-state model for persistent and transient infections. The type I error and the power of tests on vaccine efficacy based on these operational definitions are then examined under various scenarios of how a vaccine might affect the infection-disease process. We find that none of the candidates performs satisfactorily, thus raising concerns that clinical trials based only on infection endpoints will not be reliable.