-64- 
Interviewers. Three interviewers with positive significant effects on visible errors show the 
same for invisible errors. Looking at Table VI, we can see that, compared to the interviewer who 
completed most cases, two interviewers have significant negative effects on both types of errors. The 
rest of the 19 interviewers have significant effects only on one of the errors types, for four the signs 
of the coefficients are reversed across equations. In sum, interviewer effects on the two indicators of 
data quality are consistent for some but not all interviewers. 
Respondents. Respondents' education reduces errors of both types. I don't quite know what 
to make of the different thresholds shown in the two models-for visible errors, the largest difference 
in coefficients was between people graduated Volksschule and middle school, while for invisible 
errors, the jump was between middle school and Fachhochschulreife. 
Life course complexity, namely the fragmentation of chronological sequences into life 
domains, has a direct positive effect on invisible but not on visible errors. When introducing an 
interaction of interview time with life course complexity, however, it turned out that the effect of 
complexity on visible errors is conditional on time. That is, an increase of visible but not invisible 
errors could be and was avoided by taking more time for recalling and probing. 
Age increases both types of errors for the two cohorts considered in this paper. A comparison 
with older cohorts in the Life History Study and in other retrospective surveys might prove 
interesting. The same is true for the effects of gender shown above; in regard to invisible errors, 
female respondents data are better regardless of living situation, while with respect to visible errors, 
single women have more errors than men, and mothers less than other people. 
Have I shown that data quality is best analyzed as a multidimensional concept? Especially with 
respect to interviewer effects, experience and productivity, my results are perhaps less conclusive than 
I would like them to be. On the other hand, with respect to the dynamics of time and volume, as well 
as with respect to respondents attributes, I find interesting differences in effects. It is not always be 
unambiguous what these differences mean; before I proceed to the next section and some remarks 
about limitations in data and design, I report results of the last analysis carried out in this paper, 
attempting to trace the relationship between visible and invisible errors. 
Appendix C, column 1, reports the coefficients for invisible errors with visible errors added 
as an explanatory variable (logged to correct for skewness). If visible and invisible errors were 
generated by basically the same processes, visible errors could be interpreted as a proxy for these 
processes insofar they are not accurately measured by the other variables in the model. On the other