data management query

[Rose Matthews]

I have 4 questions in a questionnaire (Knee Injury and Osteoarthritis Outcome Score - KOOS), that together (the 4 questions) all relate to a specific function
of the knee (lets call this F1). Each of the 4 questions are ordinal scores from 1(worst outcome) to 5 (best outcome)

I want to test how F1 compares with the new treatment compared to the control.

I'm at cross roads, as seen in previous posts I have tried latent analysis and now more recently have been recommended Confirmatory Factor analysis which personally I don't think makes sense.

1. Am I complicating this too much, trying to do something cool?

2. Should I just perform an ordinal regression determining how F1 compares with new treatment to the control
The only problem is I don't know how I would manage my data for F1, considering I have 4 ordinal variables (q1-q4) with a potential score of 1-5

* Example generated by -dataex-. For more info, type help dataex
clear
input float(q1 q2 q3 q4 procedure) str5 ptid
5 5 2 3 1 "12A"  
3 1 2 1 0 "147A" 
5 3 1 3 1 "1987D"
4 3 1 5 1 "1098G"
1 5 3 2 1 "379S" 
1 1 1 2 0 "918D" 
end

[Clyde Schechter]

Well, compared to what is usually seen in the medical literature, your planned approaches are rather sophisticated. I wouldn't say you are overthinking it, although I do wonder if you will end up confusing the audience of your ultimate presentation or publication of this work--most clinical folk would be overwhelmed by the mention of a latent variable.

That said, you don't show the code you used for your "latent analysis," nor what you might do for confirmatory factor analysis. The latent variable analysis that comes to my mind for this kind of problem looks something like:

gsem (q1 q2 q3 q4 <- F1, ologit) (F1 <- i.procedure, regress)

The thing to note here is that the first equation in that model is a confirmatory factor analysis, and you regress the latent factor on the procedure variable. I'm not sure what you were planning to do for confirmatory factor analysis. I have certainly seen people do a simple confirmatory factor analysis by itself and then use predicted values of the latent factor as variables the outcome variable for a separate regression against a procedure variable. I've even done it myself. But it's really the wrong way to go. It's a misuse of latent variables to "reify" them in that way. If you want a non-latent function variable, then I would just do some (weighted) linear combination of the values of q1 q2 q3 and q4 and use that. (Yes, this is problematic, too, in that it treats the ordinal variables as if it were interval level data. But if the response scale is roughly "equally spaced" (whatever that means in your context) it's a decent approximation.

So, I would go with this "latent analysis" approach for a publication or presentations to statistically savvy audiences. For presentations to clinical people, I would supplement, or even replace that heavily with tabular output of the effects of the procedure on each of the q* responses so that they will have a fighting chance of understanding what happened.

Finally, all of these approaches have their strengths and weaknesses to trade off. It would not surprise me if other Forum members respond disagreeing with my advice. It's not cut and dried.

[Rose Matthews]

Thanks really informative, well with regards to the model's I've tried... I've tried several and had problems as seen in my posts here:

https://www.statalist.org/forums/for...s-not-feasible
https://www.statalist.org/forums/for...several-option

Howver, I've made some further progress today by finding a published method to calculate K1 that is
k1 = (q1+q2+q3+q4)*5
Which shows 100(best) 0 (worst outcome)

As you may imaging this was published in a clinical paper by clinicians,
However, I thought this was too simplistic. I'll try out Post 2 and look forward to see other people's thoughts