Mon, Jan 23, 2017 at noon:
H. Luke Shaefer
Economists have devoted increasing attention to the magnitude and consequences of measurement error in their data. Most discussions of measurement error are based on the "classical" assumption that errors in measuring a particular variable are uncorrelated with the true value of that variable, the true values of other variables in the model, and any errors in measuring those variables. In this survey, we focus on both the importance of measurement error in standard survey-based economic variables and on the validity of the classical assumption. We begin by summarizing the literature on biases due to measurement error, contrasting the classical assumption and the more general case. We then argue that, while standard methods will not eliminate the bias when measurement errors are not classical, one can often use them to obtain bounds on this bias. Validation studies allow us to assess the magnitude of measurement errors in survey data, and the validity of the classical assumption. In principle, they provide an alternative strategy for reducing or eliminating the bias due to measurement error. We then turn to the work of social psychologists and survey methodologists which identifies the conditions under which measurement error is likely to be important. While there are some important general findings on errors in measuring recall of discrete events, there is less direct guidance on continuous variables such as hourly wages or annual earnings. Finally, we attempt to summarize the validation literature on specific variables: annual earnings, hourly wages, transfer income, assets, hours worked, unemployment, job characteristics like industry, occupation, and union status, health status, health expenditures, and education. In addition to the magnitude of the errors, we also focus on the validity of the classical assumption. Quite often, we find evidence that errors are negatively correlated to true values. The usefulness of validation data in telling us about errors in survey measures can be enhanced if validation data is collected for a random portion of major surveys (rather than, as is usually the case, for a separate convenience sample for which validation data could be obtained relatively easily); if users are more actively involved in the design of validation studies; and if micro data from validation studies can be shared with researchers not involved in the original data collection.