Evaluating Regression Imputation Models: An Example from the Services Sectors Portion of the Economic Census

• Activity Number: 401
• Type: Topic Contributed
• Date/Time: Thursday, August 12, 2004 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Quatracia Williams*+
• Companies: U.S. Census Bureau
• Keywords: imputation ; multicolinearity ; multiple regression ; mean absolute error ; mean absolute deviation

Abstract

The Economic Census uses a variety of statistical models for item imputation. Currently, the services sectors portion of the Economic Census relies on industry-average-ratio imputation as its primary statistical imputation model, developing parameters from no-intercept simple linear regression models using weighted least squares estimation to compensate for heteroscedasticity. Our prior research (using 1997 data) showed improved predictions for these Economic Census sectors using multiple regression models. The high correlation between candidate covariates introduces the model-fitting issue of multicolinearity. The presence of multicollinearity and heteroscedasticity renders the "traditional" regression evaluation diagnostics inappropriate. We present a model selection procedure that uses alternative robust statistics and relies primarily on cross-validation. In theory, multiple regression models constructed from multicolinear data (with unequal error variances) can be used with other datasets for prediction, provided that the prediction region does not change. This paper evaluates our methods using data from the 1997 and 2002 Economic Censuses.