Imputation strategies for missing values in post-enumeration surveys

To estimate census undercount, a post-enumeration survey (PES) is taken, and an attempt is made to find a matching census record for each individual in the PES; the rate of successful matching provides an estimate of census coverage. Undercount estimation is performed within poststrata defined by geographic, demographic, and housing characteristics, X. Portions of X are missing for some individuals due to survey nonresponse; moreover, a match status Y cannot be determined for all individuals. A procedure is needed for imputing the missing values of X and Y. This paper reviews the imputation methods used in the 1986 Test of Adjustment Related Operations (Schenker 1988) and proposes two alternative model-based methods: (1) a maximum-likelihood contingency-table estimation procedure that ignores the missing-data mechanism; and (2) a new Bayesian contingency table estimation procedure that does not ignore the missing-data mechanism. The first method is computationally simpler, but the second is preferred on conceptual and scientific grounds.