Answers are grounded in this paper and cite their sources.
In German surveys, educational attainment is measured with two questionnaire items, the first referring to the highest general school-leaving qualification, and the second referring to the highest vocational qualification or higher education degree obtained. For analysis, information from both items is usually combined into one variable. There is currently no standard way for doing so, hindering comparisons across studies and the aggregation of research data across data sources. This paper presents the results of a large-scale validation of several candidates for a German standard education variable. The candidates are based on the official International Standard Classification of Education (ISCED) and the academic CASMIN education coding scheme originating in comparative social stratification research, which are both commonly used in Germany even in non-comparative research. The validation uses ALLBUS 2018 data and a data-driven selection of 157 linear validation (dependent) variables. Candidate standard education variables that retain a higher relative explanatory power across validation variables than the most detailed education variable in linear multiple regression models are regarded as more valid than candidates showing larger losses of explanatory power. Results clearly point to the superiority of CASMIN compared to ISCED when predicting a wide range of outcomes. Only for a few dependent variables, ISCED performs better, and only if it is measured in a more detailed fashion than the official way by accounting for specificities of the German educational system. Most survey data sets, however, only offer a derived ISCED variable reflecting main levels. The commonly used aggregation into three broad education levels loses, on average, half the explanatory power of the detailed benchmark, irrespective of whether it is derived from ISCED or CASMIN. Thus, studies controlling for education with just three categories may risk non-measured education effects to bias effects of other variables in the model.
No takes yet. Share an insight, caveat, or question.
Schneider et al. (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: