Total Survey Error in Social Surveys : Measurement and Nonresponse Error in Finland and Sweden

This thesis examines Total Survey Error (TSE) in social surveys, specifically focusing on measurement and nonresponse error through three articles that utilize probability sample and registry data collected in Finland and Sweden. This is crucial to study, since survey nonresponse rates are increasing, and repeatedly survey data have been unable to predict crucial societal events such as elections. This thesis contributes by utilizing novel analytical approaches, which are enabled by access to high-quality data. The thesis begins with a literature review of the TSE framework (Groves et al., 2009) and relevant theories and previous research. The Measurement component of TSE relates to how we measure our concepts of interest, and the thesis focuses on measurement error: the deviation between the recorded response and the truth, either due to random (e.g., nonattitudes) or correlated errors (e.g., social desirability). This relates to the psychology of survey response (Tourangeau et al., 2000) and the formulation of public opinion (Zaller, 1992). The Representation component of TSE relates to how well the responding sample can represent the population, and the thesis focuses on nonresponse error: meaning the error that occurs when not all sampled individuals respond to the survey, which is especially detrimental if there are structural differences between respondents and nonrespondents (i.e., nonresponse bias). This relates to response propensity theories such as the social exchange theory (Dillman et al., 2014) and leverage-salience theory (Groves et al., 2000). The literature review results in an analytical framework that identifies key aspects of measurement and nonresponse error, and ways to estimate the errors and, to some extent, counteract the errors. This leads the way to three research articles. In Article 1, we study the occurrence of measurement error when surveying non-factual issue positions. In Article 2, we study the extent of mismatches regarding factual demographics by comparing survey self-reports and registry data. Finally, in Article 3, we study nonresponse trends and risk of bias by looking at nonresponse predictors and the impact of a deteriorating survey climate and cohort replacement. All three articles also provide insights into potential interaction effects that, in turn, affect TSE. We find that measurement error is low when surveying factual aspects such as demographics, while measurement error is moderate when surveying non-factual aspects such as issue positions. Almost all of the measurement error is due to random error, while the correlated errors were marginal. Meanwhile, the increasing nonresponse is almost wholly attributable to a deteriorating survey climate, not cohort replacement. Also, while nonresponse has increased, the risk of nonresponse bias does not appear to have done so. There is also an interplay between measurement and nonresponse error, strengthening the argument for substantive researchers to both estimate and actively counteract both types of error.