In this paper faking of personality questionnaires - here under consideration the Myers-Briggs type indicator - is deliberately analyzed by methods of item response theory. This questionnaire is based on the typology of CG. Jung and contains four scales: "Extraversion vs. Introversion", "Thinking vs. Feeling", "Perceiving vs. Judging", and "Sensing vs. Intuition". Administration was either computerized or carried out with pencil and paper, responses had to be given either according a dichotomous or some continuous item response format (analogue scale). The sample consists of 186 volunteers and 129 testees who had to stand the questionnaire in a selection situation. The data were evaluated by the (dichotomous) Rasch model. For this the analogue scale was dichotomized. Model checks applied according to Andersen's likelihood ratio test and some parameter-free "exact test". Analyses of all data together resulted in a relatively small pool of homogeneous items. Separate analyses of both the groups of volunteers and selection testees resulted partly in more proper scales' item pools: For testees in the selection situation a relatively large number of items were deleted because of violating presuppositions of the model. This is interpreted as a faking effect.
Key words: faking, item response theory, computerized testing, Myers-Briggs type indicator (MBTI)
Faking personality questionnaires is meanwhile a multiply established phenomenon (cf. Kubinger, 2001, and Seiwald, 2001, both in this issue), mainly based however, on studies using some social-desirability instructions. In this paper, therefore, another approach was used: Volunteers were opposed to testees who had to stand the test in some selection context. Therefore a certain psychometric approach was used, that is, analysis according to the Rasch model was used in order to disclose whether or not the condition of volunteer or selection does influence the dimensionality of personality questionnaires' scales and the homogenity of their items, respectively.
Analyses according to the well-known Rasch model (cf Rasch, 1980, or Fischer, 1974) are, at the end, based on the following presupposition: If any identifiable sub-group of the sampled testees exists which may be assumed to show some distinct, or at least another answering behavior than other testees, this behavior causes non-homogeneity of a scale's items. As a consequence, Andersen's (1973) likelihood ratio test should discover this fact: Division of the sample into those testees who may be hypothesized to fake according to a certain known variable (external criterion) and into those who may not - according to the same variable -, would lead to a significant test statistic. That is, the data would not fit the model. Nevertheless, analyzing each of the resulting two sub-groups by the Rasch model on its own might offer (almost) homogeneous scale's items, at least for the sub-group where the answering seems true. Though, even the faking sub-group may fit the model if faking good occurs for every testee alike! Speaking technically, the so-called local stochastic independence would then still exist. Hence, some new parameter-free, exact test for model checking (exact testing by the computer programme T-Rasch; Ponocny & Ponocny-Seliger, 1999) applies as well.
2. Survey instrument
Based on the experiences of Kamer (1993) a slightly revised version of the German edition of the Myers-Briggs type indicator (MBTI; Bents & Blank, 1991) was used as survey instrument. This questionnaire refers to the well-known typology by C.G. Jung (1921, 1986).
There is no complete formulation of this typology (cf. Fisseni, 1984), although Jung worked on it repeatedly in the course of his life. When, in the end, the English edition of the questionnaire was established (Myers, 1962), Jung's contemplations were summarized but also changed and broadened.
All in all the MBTI has four bipolar scales. …