Comparison of statistical analysis methods for object case best-worst scaling.
Cheung KL., Mayer S., Simon J., de Vries H., Evers SMAA., Kremer IEH., Hiligsmann M.
Aims: Different methods have been used to analyze "object case" best-worst scaling (BWS). This study aims to compare the most common statistical analysis methods for object case BWS (i.e. the count analysis, multinomial logit, mixed logit, latent class analysis, and hierarchical Bayes estimation) and to analyze their potential advantages and limitations based on an applied example. Methods: Data were analyzed using the five analysis methods. Ranking results were compared among the methods, and methods that take respondent heterogeneity into account were presented specifically. A BWS object case survey with 22 factors was used as a case study, tested among 136 policy-makers and HTA experts from the Netherlands, Germany, France, and the UK to assess the most important barriers to HTA usage. Results: Overall, the five statistical methods yielded similar rankings, particularly in the extreme ends. Latent class analysis identified five clusters and the mixed logit model revealed significant preference heterogeneity for all, with the exception of three factors. Limitations: The variety of software used to analyze BWS data may affect the results. Moreover, this study focuses solely on the comparison of different analysis methods for the BWS object case. Conclusions: The most common statistical methods provide similar rankings of the factors. Therefore, for main preference elicitation, count analysis may be considered as a valid and simple first-choice approach. However, the latent class and mixed logit models reveal additional information: identifying latent segments and/or recognizing respondent heterogeneity.