A control-conditional test for assessing HIV vaccine efficacy in small sample preclinical studies

In this paper, we discuss the analysis of data from small sample animal studies intended to evaluate HIV vaccine efficacy. The focus is on the chimpanzee model with HIV infection, a binary outcome, of primary interest. The problem becomes that of testing for a difference in independent binomial proportions, but aspects of the study design call into question the use of standard approaches. As sample sizes may be as small as one or two per group in this context, it is tempting to utilize previous data; such usage, however, carries a high price in terms of additional assumptions. We present a test, referred to as the control-conditional test, which conditions on the control data and assumes (in a manner of Bayesian estimation) only vague prior information. Comparisons are made with Fisher's exact test and an exact unconditional test. The control-conditional test is also generalized to allow the analysis of data from a differential dose design.