How Would YOU Approach This (Statistically)?

[clip.clop]

I've already got a fairly solid (i.e., if I end up going with it I don't think too many people would bat an eye at it) plan for the analysis of these data with respect to the research question. But, I've got this annoying gut feeling it could be better and this is a situation where I'm REALLY not (power-wise) rich enough to forgo variance for the sake of ease. And I don't want to skew your comments, so I'm not telling what my plan is right now. So, what test(s) would you perform to approach this/these research question(s)?

The ultimate aim of this entire project (not just this one paper) is to figure out how we might predict a certain specific quality of life outcome (Z), if at all, given the ways in which most of this (specific) subfield currently collects data pertinent to Z.

Unless otherwise noted, all variables are obtained via self-report, Likert scale (where possible) items on a questionnaire completed pre-treatment and post-treatment. Note: the pre-treatment dataset is quite famished on some variables--that's a whole other issue worth discussing, so we're trying to not let it be the focus of this paper to any unnecessary degree. Plus, the ultimate interest here relies heavily on what happens post-treatment. As counter-intuitive as that might seem, just trust me.

The variables we're looking at in terms of how they independently and/or collectively affect or predict Z are:

Psychological
Depression (via BDI)
Anxiety (via STAI)
GAF (not Likert or self-report--duh)
Body image

Sociodemographic
Gender
Age
Education
Race
Partnership status
Employment status

There are also several relevant physical variables to consider, but I'm reluctant to include them because they're all self-report and I hardly have them all in the pre-tx dataset.

Any thoughts are very much appreciated.

(Obviously I've generalized things a bit for obvious reasons.)

---

Members don't see this ad.

 

[PsyDr]

We need more information about z. Is it binary?

 

[clip.clop]

No; self-report on a 5 point Likert scale of agreement.

FWIW, this is archival data, so that should answer a lot of questions. On the flip side, we have a bit of a needle in a hay stack here, so even if the data are limiting it's still super publishable. Which is why I want to make sure I get it right.

 

[Ollie123]

We have nowhere near enough information to tell you how to approach this. We have no idea how correlated your IVs are and whether multicollinearity is going to be an issue. No idea if you are interested in changes in the IVs or raw scores post-tx. No idea what the distribution of your IVs and DVs is like and whether it may be necessary to address floor/ceiling effects or transform anything. I can't tell if your reference to being "famished" on some variables means you have an enormous amount of missing data or just that the variables aren't measured in optimal ways. If its missing data (but not an unreasonable amount), use MI to handle it since multiple variables with high missingness will tank your N fast. If not then you probably just need to drop the variables with missing data. The obvious answer that any first year grad student would give you is some form of regression. I assume that's not the answer you are looking for because you could ask anyone else wandering around the lab/department. Cumulative logit/probit ordinal regression is likely the most conservative choice, though most in this field would just treat the 1-5 agreement scale as continuous and most journals are unlikely to take issue with it.

Not sure if that's helpful or not. I don't feel like I'm saying anything an introductory stats class wouldn't cover, so I'm guessing there is some complicating piece to this that wasn't being included. You seem very focused on the pre vs. post treatment thing and what the variables actually represent. Statistically, it shouldn't make the slightest bit of difference but the fact that you are focusing on it makes me think there is some reason it does that you didn't mention...