tively; for public and private transfers, 12 percent and 21 percent, respectively; and for property income, 24 percent and 32 percent, respectively (Jabine, King, and Petroni, 1990: Table 10.8; see also Citro and Kalton, 1993: Tables 3-4, 3-5 for comparisons of nonresponse rates for such specific income sources as AFDC and SSI).

The imputation process maximizes the available sample size for analysis from a survey by providing filled-in records for respondents whose records would otherwise have to be discarded if key analytical variables were missing. However, the process can introduce error. No definitive evaluation has been conducted of the imputation procedures used in the March CPS or SIPP; however, available evidence suggests that the procedures are a source of error and could be improved.

The Census Bureau currently applies very complex procedures, which it refers to as statistical matches, to impute values in the March CPS for whole groups of variables, such as income and employment-related items. The records are classified by a number of characteristics, and the record that is the best match is selected as the "donor" to supply the missing values to the record requiring imputation (the "host"). The Census Bureau's statistical matching procedures have, over the years, replaced somewhat less complex "hot-deck" imputations for more and more items. In the hot-deck method, the data records are arrayed by geographic area and processed sequentially, and the reported values are used to update matrices of characteristics. A record with a missing item has the most recently updated value assigned from the appropriate matrix. Hot-deck methods are largely used for imputation in SIPP.

David et al. (1986) compared the Census Bureau's imputations of earnings in the March CPS with a regression-based imputation—using data from the Internal Revenue Service from a 1981 exact-match CPS-IRS file as the measure of truth—and found that the CPS methods performed quite well in reproducing the overall shape of the earnings distribution. However, they and other analysts have determined that the CPS imputations are less successful for small groups, such as minorities and specific occupations (Coder, no date: Lillard, Smith, and Welch, 1986). Coder (1991), in an exact match of the March 1986 CPS with IRS records for married couples with earnings, found that records with imputations for CPS earnings contributed significantly to the overall underestimate of wages and salaries in the CPS in comparison with the IRS tax returns. Thus, while mean CPS earnings in cases with no imputations were 98 percent of mean IRS earnings, mean CPS earnings in cases with imputations were only 89 percent of mean IRS earnings. Also, while 95 percent of cases with no imputations had CPS earnings within 1 decile of IRS earnings, only 66 percent of cases with imputations were in this close agreement.

The available evidence suggests that the SIPP imputation procedures could also be improved. Several studies have focused on the population eligible for assistance programs and have identified problems because the current proce-



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement