Essays on Measurement Error and Nonresponse
Abstract: Essay 1 deals with measurement errors. Matching survey data with administrative records provides unique opportunities to evaluate the statistical properties of respondents' answers. This paper focuses primarily on home ownership and wealth variables. For all analyzed variables we find that the classical assumption of no correlation between measurement error and true value is violated. In an empirical application explaining the housing tenure choice, LPM and Probit models are estimated using both survey and register data on gross financial wealth.Essay 2 (with N. Anders Klevmarken) also analyzes measurement errors. Survey measures of wealth are error prone with a relatively large error variance. The errors are not uncorrelated with the true values but tend to have a negative correlation, which implies that wealthy people tend to under-report and less wealthy to over-report their responses. There is no general tendency of survey data to under-estimate mean wealth with the exception of the last percentile. The under-estimate of the wealth of the very rich is however not due to under-reporting but rather to selective nonresponse. Using simple models this paper discusses consequences of error prone wealth data.Essay 3 (with N. Anders Klevmarken) studies the problem of nonresponse in survey data. Using rich register data to analyze response behavior in a survey on health and economic standard, a model to explain contact and participation probabilities is estimated. Previous attempts to build such models have been constrained by the very limited information available in the sampling frames. One main result is that both probabilities are lower among respondents out of the labor market, who are immigrants and on benefits.Essay 4 analyzes different approaches to adjust for nonresponse bias. When a survey response mechanism depends on the variable of interest measured within the same survey and observed for only part of the sample, the situation is one of nonignorable nonresponse. If the nonresponse is ignored it will most likely generate significant bias in the estimates of the model parameters of interest. To solve this, one option is the joint modelling of the response mechanism and the variable of interest. Another option is to calibrate each observation with weights constructed from auxiliary data. In an application where earnings equations are estimated these approaches are all applied and compared with reference estimates. These reference estimates are based on a large data set without any nonresponse.
This dissertation MIGHT be available in PDF-format. Check this page to see if it is available for download.