Capture recapture methods using data on cases appearing on overlapping incomplete lists have been widely used for prevalence estimates. Nevertheless, workers have, in general, employed several different methods and there appears no consensus as to which is optimal.The authors, using several different methods, undertook an “empirical” investigation of data in 21 reports available in the literature, for which they could compare the known total in a group with the capture recapture estimate derived from data given on three overlapping lists. Comparisons were undertaken of the accuracy of estimates derived from one of six methods, Bayesian or non-Bayesian.The authors found use of the most complex log-linear model for three sources, the all-two-way interaction model, a method not in general use, generated estimates notably more accurate and with greater coverage of the true value by both calculated and distributional intervals than those generated by more frequently used methods. Moreover, the lower limits on these estimates were “better” (i.e. lower than the true values but closer to them). The upper limits of the estimates generated by all-two-way interaction model, were however, “worse”, very often uselessly large.The authors suggest investigators consider estimates and lower limits derived from the all-two-way interaction model in addition to and in comparison with those derived from any other methods traditionally employed.
©2012 Walter de Gruyter GmbH & Co. KG, Berlin/Boston