Competencies as dependent variables in regression models.

Autor/in	Rohwer, Götz
Institution	Leibniz-Institut für Bildungsverläufe
Titel	Competencies as dependent variables in regression models.
Quelle	Bamberg: Leibniz Institute for Educational Trajectories (LIfBi) (2014), 23 S. PDF als Volltext kostenfreie Datei Link als defekt meldenVerfügbarkeit
Reihe	NEPS working paper. 33
Beigaben	Literaturangaben
Sprache	englisch
Dokumenttyp	online; Monographie; Graue Literatur
Schlagwörter	Rasch-Modell; Skalierung; Test; Schüler; Skalierung; Test; Variable; Regressionsanalyse; Rasch-Modell; Schuljahr 05; Schüler; Regressionsanalyse; Mathematische Kompetenz; Variable; NEPS (National Educational Panel Study)
Abstract	This paper considers methods to define dependent variables representing results of competence tests; as an example I refer to NEPS data on math competencies of 5th grade pupils. The simplest and easily comprehensible method is to use the number of correct responses in a competence test as values of a quantitative dependent variable in a regression model. Instead of simply using the number of correct responses one can define weighted versions which could take into account that items might have different importance for the competence that the test is intended to measure. However, I show that it is easily misleading to think of such weights as 'item difficulties' which can be derived from proportions of wrong responses. Instead of using these simple approaches to the construction of a dependent variable, one can start from a probabilistic framework. As an example, I consider a Rasch model that allows one to construct a variable representing latent competencies which can subsequently be used as a dependent variable in regression models. I argue that this approach has two disadvantages, compared with using a simple summary index. The Rasch model introduces a nonlinear metric which is difficult to understand and therefore makes it difficult to interpret effects of explanatory variables. Moreover, the Rasch model employs a notion of "item difficulties" which are derived from the distribution of competencies of the persons participating in the test. I then discuss the proposal to use so-called plausible values for the construction of dependent variables. I distinguish between versions with and without conditioning variables. I show that using plausible values, when derived from models including conditioning variables, entail striking forms of statistical discrimination, and propose that this approach should not be used for sociological analyses. Finally, I briefly consider models which avoid a reference to latent competencies and instead directly relate the observable response patterns to values of explanatory variables. While attractive at first sight, this approach has the drawback that such models must be supplemented by a procedure for aggregating item-specific probabilities. (Orig.).
Erfasst von	DIPF \| Leibniz-Institut für Bildungsforschung und Bildungsinformation, Frankfurt am Main
Update	2020/2