DEMOGRAPHY© Volume 20, Number 4 November 1983 SOME MODELS OF AGREEMENT AND DISAGREEMENT IN REPEATED MEASUREMENTS OF OCCUPATION Robert M. Hauser Michael P. Massagli Center for Demography and Ecology, Department of Sociology, The University of Wisconsin- Madison, Madison, Wisconsin 53706 Abstract-Measures used by the U.S. Bureau of the Census to assay misclassi- fication and correct marginal distributions-the net difference rate and the index of inconsistency-may produce misleading results and do not fully use information about inconsistency in repeated measurements. We show how multiplicative models of classifications of repeated measurements can be used to locate sources of inconsistency in marginal classifications, sources of discrepancies in classification, and differences between catego- ries in levels of agreement and disagreement. The models are illustrated with an occupational classification from the 1970 CPS-Census Match. Some of our results differ from those reported earlier by the Bureau of the Census. One of the procedures used to evalu- ate the U.S. Census of Population in 1970 was the analysis of agreement and disagreement in repeated measurements of an item for the same person. For example, person records were matched between the Census and the March 1970 Current Population Survey; some of these data are reanalyzed here. Other sources of repeated measurements in- cluded a postenumeration survey and independent checks of institutional rec- ords (U.S. Bureau of the Census, 1975; 1977). The methods illustrated here may also be applied to those data. Typically, in analyzing repeated mea- surements, the Bureau of the Census uses undesirable and outmoded indexes of net (marginal) differences in classifica- tion and of response inconsistency. These may lead to erroneous inferences about patterns and sources of response error and thus they provide little help in improving measurement. In the present study, we illustrate strategies for modeling repeated mea- surements that yield improved measures 449 of net classification discrepancies and that locate specific sources of agreement and disagreement. The reason for the improvement is that our measures of error and of agreement are obtained from explicit models of repeated measure- ment, while the Bureau's measures are based upon traditional ad hoc strategies of index construction. We believe that our analysis begins to address some needs long felt by demographers (Bogue and Murphy, 1964). Though suggested in the literature several years ago (Bishop et aI., 1975, chap. 11), there have been few applications of the present methods to the study of response variability (see Clogg, 1983; Taylor, 1976). The 1970 CPS-Census Match provides data for persons who were enumerated as members of households in both the 1970 Census of Population (20 percent sample) and the March 1970 Current Population Survey. While the match yielded published cross-classifications of many population characteristics, the pre- sent analysis focuses on the classifica- tion of employed persons by major occu- Downloaded from http://read.dukeupress.edu/demography/article-pdf/20/4/449/905594/449hauser.pdf by guest on 05 November 2021