BEST advice

classic Classic list List threaded Threaded
4 messages Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

BEST advice

wikipeterson
Hi John,

I am working on a comparison of genders with regard to motivational constructs.  I am hoping to get some advice as to how to best model the data.

1) Should I use your BEST software and run separate comparisons for each construct or use a hierarchical model?

2) What distribution should I use to describe the data? The constructs are measured for each student as averages of sets of questions on a survey using a five point Likert-type scale, so they can range from 1 to 5. The data are skewed toward lower values, but in a the handbook citing a validation study, some of the constructs had positive skewness.

3) Rather than using broad normal distributions, I’m thinking that the hyper-priors could be uniform ranging from 1 to 5 for the group means, and I could use uniform(0,2) for the group standard deviations since that would give equal credence across the range of possibilities for data that ranges from 1 to 5. Does that make sense?

Thanks,
Steve
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: BEST advice

wikipeterson
I suppose another way to ask questions 2 and 3 is, is there anything wrong with this based on the post-predictive check?


Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: BEST advice

John K. Kruschke
Administrator

The normal (or t) distribution does not appear to be a good description of your data distributions.

I do not (yet) have any models for sums of ordinal responses. Meanwhile, you might try using a model of censored data, as exemplified here: http://doingbayesiandataanalysis.blogspot.com/2012/01/complete-example-of-right-censoring-in.html
Sorry I don't have a ready-to-use model for your  data, but I hope this helps...
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: BEST advice

wikipeterson
I had an idea. For describing my data where each observation is an average for a set of Likert scales (1-5), would it make sense to subtract 1 and divide everything by four so that they are on a 0-1 scale and use a beta likelihood?
Loading...