Model with categorical * continuous variable interactions in JASP
Hi JASP team,
I have run a frequentist moderated regression with two binary categorical predictors (group, sex), four continuous predictors (e.g., IQ), and interactions between the group categorical variable and each of the other predictors (i.e., group*sex, group*IQ etc). All variables mean centred / effect coded.
I’m now looking to run a Bayesian equivalent analysis in JASP, to obtain inclusion Bayes factors for each of the predictors and interaction terms.
I was intending to use Bayesian ANCOVA for appropriate uninformative priors, categorical variables as fixed factors, continuous as covariates, and manually entering the interaction terms in the model options. I would report the BFincl across matched models. Is this the best approach? I’m unsure of the appropriateness of including interactions between factors and covariates in ANCOVA, particularly following assumptions of homogeneity of regression slopes. I noticed while the effects table computes, I have received errors in the estimates table for the interaction terms (NaN for Mean and SD; 0 for credible intervals). I wasn’t intending on reporting the estimates table so this is not a problem, but are the results from the rest of the analysis valid? Is there an alternative method I should use?
Many thanks for your help.
Comments
Hi Emily,
Not on the JASP team but when I'm interested in Categorical*Continuous interactions from JASP's Bayesian ANCOVA, I tend to use the model comparisons table first, i.e. the main table. I like to set it to report the best model on top, and to report BF01 against this best model. So I can find the interaction(s) which is best-supported by my data, and I know how much better it is than the other interactions.
If I need to or want to comment on each individual BFincl, I tend to phrase it as follows: "Prior to data analysis, the assumed probability of inclusion for this interaction was (insert). Following data analysis, the probability increased/decreased to (insert). Models with (interaction name) were (insert BFincl) times better than models without this interaction term."
I only have one paper in the works where homogeneity of regression slopes isn't given. It's a study of spontaneous speech/language in patients vs. healthy controls, where all of the controls are of course not speech-impaired and most of the patients are dysarthric to various degrees. What I've done here - as the homogeneity of regression slopes is unrealistic to expect - is, I've incorporated the interaction into the null model. That way, I compare my other working hypotheses/models to the hypothesis that dysarthria's confounding effect affects the language outcomes.
Best of luck!!
Yes, I agree with eniseg2. Sorry about my tardiness, I am a little busy at the moment -- hope to be back up to speed soon.
Cheers,
E.J.