|
Abstract:
|
Multiple Systems Estimation has become a standard method in the area of human rights statistics, especially in the attempt to quantify victims of human trafficking and modern slavery. Data in these areas tend to show specific characteristics, for example low or zero counts in many of the cells of the data tables, and this needs to be taken into account for proper inference. In addition, the choice between a very large number of possible models can materially affect the estimates, as well as being burdensome computationally. A standard approach, even if not fully valid in the case of sparse data, is to use the Bayes Information Criterion (BIC) to choose between all the possible models. Another is a forward stepwise model choice strategy. Whatever method is used, a bootstrap approach can give confidence intervals take account of model choice, rather than being calculated conditional on one particular model. Careful computational approaches, and, in the case of the BIC, modifications to the method itself, make it possible to speed up computation times by as much as orders of magnitude. This facilitates the use of the methods and assessments of their relative performance.
|
