Skip to main content
Login | Suomeksi | På svenska | In English

Browsing by Subject "CFA"

Sort by: Order: Results:

  • Mattsson, Markus (2015)
    In this Master's thesis I examine the measurement invariance of the Driver Behavior Questionnaire (DBQ), the perhaps most widely used questionnaire instrument in traffic psychology, across samples of Finnish and Irish young drivers (18 - 25 years of age). The DBQ was developed in the beginning of the 1990s based on principal component analyses. The questionnaire was originally based on a well-tested theory in cognitive ergonomics (the Generic Error Modeling System, GEMS), but in the research that has ensued, the item pool and the factor structure has been determined in an exploratory fashion. This has resulted in an abundance of DBQ versions, which comprise anything from nine to over one hundred items and from one to seven factors. Further, in research articles based on the DBQ, it is a common practice to calculate sum or average scores and compare them across subgroups of respondents. The 28-item version of questionnaire, which is currently perhaps most widely used, is thought to measure two, three or four latent variables. In this thesis I use confirmatory factor analysis and, specifically, analysis of measurement invariance to examine which of the three alternative factor structures functions as the most fitting description of the responses of Finnish and Irish young drivers. The analysis of measurement invariance is based on fitting a series of increasingly restrictive models to data. At each stage of the analysis, an increasing set of parameters are constrained to equality across the samples under comparison. In case the constrained model does not fit the data worse than the unconstrained model, the constrained model can be applied in all (in this thesis both) data sets. The models that are fit to data are, in order: 1) The configural model in which only the number of factors is constrained, 2) the weak invariance model, in which factor loadings are constrained to equality, 3) the strong invariance model, in which also the intercept terms of each item are constrained to equality and 4) the strict invariance model, in which also the error terms of each item are constrained to equality. In addition, models of partial invariance are applied. In these models, only some of the constraints related to each stage of the analysis are preserved. In addition to comparing the models statistically, their fit to data is examined using various descriptive statistics and graphical representations. As a central result I propose that the four-factor model offers the best fit to both data sets, even though the model needs to be modified in an exploratory mode of analysis to ensure sufficient fit to data. Further analyses show that two of the four factors are different in nature in the two samples and that only in the Irish data set do all of the items load on the factors they are expected to. On the other hand, the analysis of the other two factors shows that the items that load on them are interpreted essentially similarly in the two samples and that weak invariance can be assumed on their part. In addition, partial strong invariance can be assumed in the case of one factor, even though even then the values of most of the intercept terms need to be freely estimated in the two data sets. As a conclusion I suggest that, in contrast to the prevailing practice, comparing sum scores based on DBQ factors is dubious and that comparing latent variables scores may be justified only in the case of one factor out of four. As a practical recommendation, I suggest that the factor structure of the DBQ be further developed based on theories of cognitive ergonomics and cognitive psychology and that invariance analyses be performed as a matter of routine before carrying out comparisons of groups based on results of factor analyses.