Measurement Errors in Recall Food Expenditure Data

Household expenditure data is an important input into the study of consumption and savings behaviour and of living standards and inequality. Because it is collected in many surveys, food expenditure data has formed the basis of much work in these areas. Recently, there has been considerable interest in properties of different ways of collecting expenditure information. It has also been suggested that measurement error in expenditure data seriously affects empirical work based on such data. The Canadian Food Expenditure Survey asks respondents to first estimate their household's food expenditures and then record food expenditures in a diary for two weeks. This unique experiment allows us to compare recall and diary based expenditure data collected from the same individuals. Under the assumption that the diary measures are "true" food consumption, this allows us to observe errors in measures of recall food consumption directly, and to study the properties of those errors. Under this assumption, measurement errors in recall food consumption data appear to be substantial, and they do not have many of the properties of classical measurement error. In particular, they are neither uncorrelated with true consumption nor conditionally homoscedastic. In addition, they are not well approximated by either a normal or log normal distribution. We also show evidence that diary measures are themselves imperfect, suffering for example, from "diary exhaustion". This suggests alternative interpretations for the differences between recall and diary consumption measures. Finally, we compare estimates of income and household size elasticities of per capita food consumption based on the two kinds of expenditure data and, in contrast to some previous work, find little difference between the two.