Latent Class Models for Evaluating the Accuracy of Census Counts
Biemer, P. P. (2004, August). Latent Class Models for Evaluating the Accuracy of Census Counts. Presented at , .
In this paper we present a method adjusting the US decennialcensus day counts using a system of four lists (Census, PES, MER,ARL). The primary concern is developing a method that can detecterroneous enumerations that are present in the data. Failure to accountfor erroneous enumerations in the lists will result in overestimationof the census day residents. We use four lists that are indicators ofresidency for obtaining estimates of erroneous enumerations presentlists. The method is similar to detecting heterogeneity in acapture-recapture experiment. The erroneous enumerations are assumed tocome from a >non-resident=population that is rostered on the fourlists with different probabilities then the actual census day>resident= population. Using these four lists, Latent Class Analysisand the software package lEM we were able to fit several reasonablemodels that reflect likely census populations. We discuss theassumptions required for the inference to be valid and indicate therobustness of these models to the violation of these assumptions. Theresults from a simulation study will be presented that show when ourmethod is reliable and situations in which it fails.