Improving Information for Social Policy Decisions -- The Uses of Microsimulation Modeling: Volume II, Technical Papers (1991)
Chapter: AFDC-Census Match
AFDC-Census Match
Springs and Beebout (1976) matched 1973 survey data for Aid to Families with Dependent Children (AFDC) families to data from a stratified subsample of the 1970 decennial census public-use sample to augment the AFDC file with information on the nonparticipant population “at risk” for the program. The stratified subsample contained 156,961 household records for 513,113 persons, plus 30,900 “household” records for residents of group quarters. The 1973 AFDC survey is a national sample of families that received an AFDC assistance payment in January 1973. The sample size of the AFDC survey was 33,829 families.
Since the two files represented different years’ data, it was necessary to adjust one of the files to conform with the other. A form of static aging was therefore used on the 1970 public-use sample file, controlling it to totals from the March 1973 CPS. In addition, there was a second-stage adjustment to some state population totals derived from the Census Bureau’s state population estimates. Also, there was an adjustment of income in the 1970 public-use sample so that aggregate income by source would be nearly in agreement with the March 1973 CPS. Finally, the Micro Analysis of Transfers to Households (MATH) model was used to simulate AFDC eligibility on the census file.
Prior to the statistical match, some recoding was performed on both files so that definitions of the match variables were consistent. The first step of the statistical match was a partitioning of the files, with the potential matches limited to those within partitions. The partitions were defined using state, family type, race, and standard metropolitan statistical area (SMSA). Within categories, a distance measure was defined by using the following matching variables: number of persons aged 0–20, eligibility status for AFDC, reason for eligibility, AFDC family type, assistance unit unearned income, assistance unit earned income, mother’s employment status, education of assistance unit head, age of head, and number of persons in household. The definition of a match depended on whether the variable was categorical or continuous, with a tolerance allowed for the matching of continuous variables. In addition, in these cases the distance function was a function of the closeness of the match, so some credit was given for a near match, but not as much as for a match.
The remaining issue was the treatment of the sample weights. In some ways, the procedure for handling differing sample weights was an attempt to mimic the benefits of constrained statistical matching. To quote Springs and Beebout (1976:22): “Therefore, a procedure had to be designed so that records with greatly differing sample weights could be matched while preserving the correct totals and distributions from both files on the merged data set.” Very roughly speaking, the minimum sampling weight of a matched pair was attributed to the matched pair. The remaining residual weight was given to the
