Hello,
I am in the process of planning my analysis of Lighthouse Studio MaxDiff collected data. My design meets the requirement to run a hierarchical bayes (e.g. each item is shown 4+ times per person) and extract individual level preference share scores for each item (13) in my study. My understanding is that these scores can be used for post-hoc cluster analysis (although often latent class analysis is conducted instead).
Would post-hoc modelling of the scores using an ANOVA be a valid approach?
The model would be individualscore(as extracted from the hierarchical bays and transformed to scores) ~ item*demographics.
The model would generate the mean preference share with confidence intervals for every item*demographic (e.g. binary age, under 50 or over 50 would lead to 26 means in my case). I would then use AIC to identify one or more ANOVA models with different combinations of demographics that fit the data well without too many parameters. I am using this data to get a predicted score for demographic subgroups, with demographic information from them in a separate dataset.
I have only seen a few mentions of this sort of approach in the literature (and mostly people conduct post-hoc cluster analysis using k-means for example).
The only issue I could potentially forsee is that there isnt a confidence interval around the individual level scores (from variation in how they score rather than the overall population), which ideally I would propagate into the ANOVA model using boostrapping for example. This could potentially be complex to disentangle as missing data on items at the individual level for is leveraged from the item population mean.
Thanks for your help,
Richard