Investigating factor structure of scores on the outcome questionnaire using factor mixture modeling
The Outcome Questionnaire (OQ-45; Lambert et al., 1996) has been widely employed as a psychotherapy outcome monitoring measure following research findings that support various aspects of its validity and sensitivity to change. Despite its broad usage in both clinical and research settings, some of its psychometric properties are not definite. The three subscales of the OQ-45 are designed to measure three distinct, but related, aspects of psychological functioning. However, neither the one- nor three-factor models have been supported by previous research. Likewise, the results of the current study supported neither of those factor structures. It was suspected that heterogeneity in data might have led to the lack of the confirmatory factor analysis model fit. Therefore, factor mixture modeling (FMM), a combination of confirmatory factor analysis and latent class analysis, was employed to investigate potential heterogeneity of the data. Among the series of factor mixture models with varying numbers of classes that were fitted, the two-class, unconditional FMM based on the revised three-factor solution was decided to best describe the data under analysis. Although three covariates of clinical status, sex, and race were selected as known sources of heterogeneity and incorporated into the FMMs (i.e., conditional model), the findings were contradictory to expectations. The implications of these findings in counseling were discussed in terms of aggregating OQ-45 scores and its score interpretation. Furthermore, this study demonstrates the process involved and dilemmas encountered in choosing the best fitting FMM. There is currently no criterion for assessing individual model fit. Instead, models’ fit are compared using various information criteria (IC). And, as was found in the current study, these ICs are frequently contradictory. Thus, the process of identifying the best fitting model cannot rest solely on fit indices but must also depend on interpretation of models and consideration of the ultimate use of the results. In the current study, consideration of transition matrices and the pattern of latent means across classes contributed as much to model selection as fit index interpretation.