This estimator runs into problems with small AP as the Rogan-Gladen conversion is really using expected apparent prevalence. Getting the expected value of the AP distribution is complex and the expected value given a single observation is not in general the ratio of positives / count. The expected apparent prevalence is never less than the specificity but the observed value often is. To deal with this the R-G estimator truncates at zero.