Characterizing Measurement Error in the German Socio-Economic Panel Using Linked Survey and Administrative Data

TL;DR

This study links German survey and administrative data to analyze measurement error in labor earnings, revealing non-random sample selection, non-classical error patterns, and implications for bias in earnings regressions.

Contribution

It provides a detailed characterization of measurement error in German earnings data using linked survey and administrative sources, highlighting non-random selection and non-classical error.

Findings

Survey participation is non-random and based on observables.

Measurement error shows underreporting, autocorrelation, and correlation with true earnings.

Reliability ratios above 0.94 suggest small attenuation bias in simple regressions.

Abstract

This paper exploits the linkage of German administrative social security data (German: Integrierte Erwerbsbiografien) and survey data from the socio-economic panel (Sozio-\"okonomisches Panel, SOEP) for the characterization of measurement error in metrics quantifying individual-specific labor earnings in Germany. We find that survey participants' decision whether to consent to linkage is non-random based on observables. In that sense, the studied sample does not constitute a random sample of SOEP. Further, measurement error is not classical and differential: We observe underreporting of income on average, autocorrelation, and non-zero correlation with the true signal and other observable characteristics. In levels, calculated reliability ratios above 0.94 hint at a relatively small attenuation bias in simple linear univariate regressions with earnings as the explanatory variable. For one-period changes in income the bias from measurement error is exacerbated.