analysis of ordinal outcome variable when some people have "not applicable"

[Rich Goldstein]

I have a situation where the outcome (response, dependent) variable is basically ordinal but, for some people, the question is not applicable (NA)
and that creates an issue; specifically (note that this is one example of 3), people with a spinal cord injury (SCI) were asked how
frequently their spasms were painful (5 point scale from "never" to "very often"); however, some people never had spasms and thus this
is not applicable (NA ranges from about 1% to about 7% across the 3 items of interest). I have thought of
several possible ways of analyzing this and will use more than 1. However, I know of no literature on this.

My questions, thus, are (1) any literature on this? (2)n any thoughts on
the various types of analyses listed below? (3) any lit comparing
the various types of analyses below? (4) any other analytic suggestions?

Various type of analysis that I can conceive of (listed in no particular order); note that "ordinal logit" should be read to include partial
PO models as per Peterson and Harrell:

a. ordinal logit ignoring the NAs (i.e., drop these observations)
b. multinomial logit using NA as a 6th category
c. including the NAs with the never group and using ordinal logit
d. Using Anderson's stereotype model with 2 dimensions at least at the start - slogit
e. expanding "d." by using 3 (or more) dimensions
f. something analogous to a two-part model where the first part
is whether there are spasms (binary logit) and the second is an ordinal logit

I will also posted this on the Data Methods Forum (if I can login!)

[Stephen Jenkins]

How about (g) -heckoprobit-? Analogous to your (f) two-part model. Selection step (whether answer rather than respond NA) + model step (Spasm code response| respond). Obviously, route (g) involves assumptions you might not want to make, but all routes involve them ...

[Rich Goldstein]

thank you Stephen Jenkins - I will look into that as I certainly think it will be simpler than my (f) - as you say, and as is always the case, any analysis has underlying assumptions; this is one of the reasons that I will do at least 2 estimates, being sure that the assumptions are at least somewhat different

[Felix Bittmann]

I just read about twopm in another thread, which is probably what you mean under f.) https://journals.sagepub.com/doi/pdf...867X1501500102

[Rich Goldstein]

sort of, thank you; however, "twopm is designed to estimate models in which the positive outcome is continuous. It does not deal with discrete or count outcome" (from the help file), and that is not my situation which is why my language in (f) above is not as clear as it could

[daniel klein]

What's the research question? What's the theoretical model?

[Rich Goldstein]

sorry about that - my clients are a bit confused/conflicted about the research question but here is a first cut - clinicians "know" that certain clinical data is important to know with respect to the question of the frequency of pain in spasms (e.g., completeness of injury, level of injury, time since injury, etc.); but there is apparently no literature re: the impact of demographics (e.g., gender, age) and this is what they want to investigate; so, one way to look at what the question is: does the addition of demographics to clinical predictors allow a "better" (more informative?) model - my clients have no theoretical model at this point and, while it appears clear that there is an underlying continuous variable (at least for those for whom the question is applicable), they prefer to look the issue in categories as that is how they deal with patients - hope this helps

[daniel klein]

Thanks. I think the "best" model ultimately depends on the precise research question and goals. If your clients are primarily interested in patients who suffer pain, excluding those who don't experience spasms might be reasonable. If your clients wish to predict painful spasms based on demographics across the patient population, combining patients without spasms and those with painless ones might suffice. If your clients want to know how demographics correlate differently with experiencing spasms and pain intensity, you probably want a two-part model. If your clients seek to develop interventions and you believe there is some selection process, you need to account for that. Obviously, those are just general remarks that may not help much with your specific questions.

[Richard Williams]

Is eoprobit with the select option worth considering? I’ve never used the eregress commands but they look interesting.

[Rich Goldstein]

thank you for this suggestion; the problem I see is that, although all these people are being treated in some way, I know nothing about their treatment, or their treatment status, so I'm not sure that this would be helpful; it is, however, good to know about the command

[Richard Williams]

thank you for this suggestion; the problem I see is that, although all these people are being treated in some way, I know nothing about their treatment, or their treatment status, so I'm not sure that this would be helpful; it is, however, good to know about the command

I would think eoprobit would be used to implement your option F, if in fact you want to pursue that option. I’m not sure if it would be better or worse than heckoprobit or whether it is a good option at all in your case.

[Rich Goldstein]

thanks - while waiting for the final data, I will continue to look into eoprobit

[Richard Williams]

You might also look at this great paper by Mize, Dian, and Long:

https://journals.sagepub.com/doi/abs...81175019852763

Among other things, when there are multiple plausible models (e.g. mlogit and ologit) it lets you assess how much difference the choice of model makes.

I might also be tempted to treat no spasms as category 1 and then run ologit.

[Richard Williams]

That is, 0 = No Spasm, 1 = Spasm but never any pain, 2 = Occasional Pain, etc.

[Rich Goldstein]

Richard Williams - thanks for the suggestion and the citation!