Survey ordering and the measurement of welfare

Abstract Economic policy and research rely on the accurate measurement of welfare. In nearly all instances, measuring welfare requires collecting data via long household surveys. If survey response patterns change over the course of a survey to introduce measurement error, this measurement error can be either classical (i.e., changing distributions, leading to noise) or non-classical (i.e., changing expectations, leading to bias). We embed an experiment in a survey by randomly assigning a questionnaire with either the assets module near the beginning of the survey or the assets module at the end of the survey, delaying enumeration of assets by about 60 minutes. We find no evidence in the full sample that survey ordering introduces differential response patterns, either in the number of reported assets or the reported value of those assets. In exploratory analysis of heterogeneity, we find evidence of non-classical measurement error due to survey ordering within sub-samples of respondents who (i) are from larger households or (ii) have low levels of education. Our experimental design can be generalized to serve as an ex-post test of data quality with respect to questionnaire length.