How Survey-to-Survey Imputation Can Fail

Newhouse, D. ; Shivakumaran, S. ; Takamatsu, S. ; Yoshida, N.

World Bank Group, Washington, DC

关键词: AVERAGE WAGES;    BIASES;    CALCULATION;    CHANGES IN POVERTY;    CONFIDENCE INTERVALS;   

DOI  :  10.1596/1813-9450-6961
RP-ID  :  WPS6961

学科分类：社会科学、人文和艺术（综合）

来源: World Bank Open Knowledge Repository

【 摘 要 】

This paper proposes diagnostics toassess the accuracy of survey-to-survey imputation methodsand applies them to examine why imputing from the HouseholdIncome and Expenditure Survey into the Labor Force Surveyfails to accurately project poverty trends in Sri Lankabetween 2006 and 2009. Survey-to-survey imputation methodsrely on two key assumptions: (i) that the questions in thetwo surveys are asked in a consistent way and (ii) thatcommon variables of the two surveys explain a large share ofthe intertemporal change in household expenditure andpoverty. In addition, differences in sampling design canlead validation tests to underestimate the accuracy ofsurvey-to-survey predictions. In Sri Lanka, the causes offailure differ across sectors. In the urban sector, theprimary culprit is differences between the two surveys inthe design of the questionnaire. In the rural and estatesectors, the set of common variables in the prediction modeldoes not adequately capture changes in poverty. The paperconcludes that in Sri Lanka, survey-to-survey imputationbetween the Household Income and Expenditure Survey and theLabor Force Survey cannot produce accurate poverty estimatesunless the Labor Force Survey adds additional questions onassets and is redesigned to use a questionnaire that iscompatible with the Household Income and Expenditure Survey.Alternatively, a new welfare-tracking survey that satisfiesthese conditions could be established.