Respondent behavior and data quality aspects in panel surveys : four empirical contributions

Abstract

The four essays of this dissertation provide a number of new and unique insights on diverse issues of respondent behavior and aspects of data quality in surveys. The first paper examines the determinants of item nonresponse on several questions of households' wealth, households' income, and respondents' income. It firstly provides empirical evidence that the mechanisms behind item nonresponses and don't know statements differ. The item nonresponse intensity is found to be item specific. The interactions between the respondent and the interviewer and the interview situation are evaluated and it is found that the gender of both, respondent and interviewer, and the age difference between interviewer and respondent have an influence on the occurrence of nonresponse. The second and third paper show that the correlation of item nonresponse with subsequent unit nonresponse is not necessarily positive and linear. The analysis shows a negative correlation of item nonresponse with the newly introduced category of (wealth-) questionnaire nonresponse. With respect to subsequent unit nonresponse it is shown that the correlation pattern with the INR rate is nonlinear. It obeys an inverse-U-shaped pattern, which is explained by simultaneous drop-out of two types of respondents: those with low INR propensity and those with high INR propensity. Finally, the fourth study examines the quality of income data provided by the respondents with respect to rounding. It finds that rounding does not occur at random, but is explicable by cost/benefit considerations of the respondent. The magnitude of rounding is also correlated with the income figure and autocorrelated. This provides evidence that the rounding error is likely to harm estimates of empirical studies with rounded data. From a methodological point of view, this study contributes to the check of ordinality of discrete outcomes of a variable and the adequacy of ordered regression models. All four papers of this dissertation contribute to the understanding of the social interaction processes which occur during a survey interview. The benefit of the insights provided in this dissertation is threefold: first, the findings enable survey institutions to advance the data collection process, in order to reduce data deficiencies and increase the informational value of the survey: We have shown that pairing interviewers based on gender and age may reduce income INR, Face-to-face interviews are beneficial for reducing INR and UNR, and experienced interviewers improve the quality of the collected data. Second, the results could support the sophistication of imputation procedures for missing or misreported data. Since we have shown that the mechanisms behind item nonresponses and don't know statements originate from different response processes, the origin of missing statements should be considered by imputation methods. Third, the methods employed in the studies may improve researchers' ability to rigorously deal with misreports in his or her own empirical analyses, e.g. to use selection models with pre-interview data as instruments. Nonetheless, further research is needed to derive more concrete advise on how to design a survey study to increase the quality of the data collected. Since the approach underlying the studies herein is of empirical nature, the results of these studies are restricted to observable and surveyed characteristics of respondents, interviewers and the interview situation. It is likely that this reflects only part of the story, since a lot of possible determinants may be unobservable, nonmeasurable or not surveyed. This opens avenues of qualitative research in the disciplines of e.g. sociology and psychology.