Statistical models of relationships among variables often assume simple, linear relations such as "the higher the work demands for employees, the more exhaustion results" or "the better the support by supervisors, the less exhaustion".

In more and more areas of research, however, more complex relationships among variables, so-called moderator or interaction effects, are assumed.

Moderator Effects

A moderator effect is present if the relation between two variables varies depending on the value of a third variable. A stress research example would be the assumption that high work demands lead to exhaustion only if there is a lack of necessary resources. Sufficient resources provided, high demands can be fulfilled without exhaustion (Figure 1).

Schematic diagram of a moderator effect between work demands (x-axis, horizontal), resources, and exhaustion (y-axis, vertical) using separate regression lines: If resources are lacking, increasing demands lead to exhaustion (regression line has positive slope). If sufficient resources are being provided, no relation between demands and exhaustion would be present (horizonal regression line).

Figure 1. Schematic diagram of a moderator effect between work demands, resources, and exhaustion.

Statistically speaking, the value of a variable which is to be predicted (criterion; e.g. exhaustion) depends on specific combinations of values of variables used for prediction (predictors; e.g. work demands and resources).

Such effects have considerable practical importance, as e.g. in many jobs it may be impossible to influence outcomes such as exhaustion by decreasing work demands (for economical reasons, or due to the work itself, such as in emergency and rescue services or the military), but it may still be possible to influence resources.

Analysis of Moderator Effects

Empirical tests for assumed moderator effects involve certain statistical problems which are specific for moderator effects or aggravated in their presence (especially reliability and distribution problems). From a content perspective, variable conceptualization and operationalization problems may also lead to difficulties in testing for moderator effects.

Furthermore, suboptimal statistical methods using manifest variables (without taking measurement error into account) are frequently employed, as well as methods resulting in distortion and loss of information (such as artificial dichotomization of samples). In light of these problems, empirical results when testing for moderator effects are often ambiguous and there still is considerable uncertainty regarding the existence and the extent of assumed moderator effects in many content areas.

Research project

The present research project aimes at contributing to methodologically more sound statistical testing procedures for moderator effects, and at shedding light on ambiguous empirical results. To this end, methodological analyses will be conducted allowing to compare the practical applicability of both well-established testing methods for moderator effects as well as current methods using latent variable frameworks.

In addition, those methods appearing to be most promising and least problematic in a given context will be used for comparative analyses of data sets from work and organizational psychology. On the one hand, this will be data sets where moderator effects are to be expected given the specific combination and operationalization of variables. On the other hand, this will be data sets where such effects would not be expected, in order to support and specify theoretical assumptions about moderator effects in a methodologically profound way.

It is intended to contribute to an increased precision of theories and to an increased use of more advanced testing procedures regarding moderator effects in work and organizational psychology and other application areas.