Validity of Bernoulli census, log-linear, and truncated binomial models for correcting for underestimates in prevalence studies.