relative risk ratio from meologit

[Jennifer May]

Hi all, is it possible to generate/calculate adjusted relative risk ratios from a multilevel mixed-effects ordered logistic regression (meologit)? If so, how would I do this please? I've had a google and search on here but haven't managed to find the answer.

Any help appreciated.

Thanks,
Jen

[Bruce Weaver]

Hello Jennifer. I have to wonder if your question is motivated by prior experience with the -mlogit- command, where the rrr option is used to "[r]eport results as relative-risk ratios". If your background is in epidemiology & medical research, you likely learned that a relative risk is another term for risk ratio, and that both of them = the risk of some outcome in one group divided by the risk of that outcome in another (independent) group. But that's not what the relative-risk in RRR is. Believe it or not, relative-risk is just another name for odds. And therefore, the RRR is just an odds ratio. Because I found the RRR label very confusing when I first used -mlogit-, I cobbled together some notes for myself after I figured things out. Perhaps you'll find them helpful. You can view them here:

• https://www.researchgate.net/publica...Risk_Ratio_RRR

So, if my guess about what motivated your question was right, you just need to use the or option for -melogit-. If my guess was wrong, you probably need to provide more info about what you want to do.

Cheers,
Bruce

[Jennifer May]

Hi Bruce, thank you for your reply. Does the same concept apply for meologit? I have an ordered categorical outcome, trial arm as the main fixed effect and a random effect for site, so I'm running a multilevel mixed-effects ordered logistic regression (meologit). The SAP asks for results presented as adjusted relative risk ratios for intervention versus control, but meologit produces odds ratios. Are these ORs the same as RRR in this context? Or do I need to compute the RRR from the model? Thank you for your help!

[Maarten Buis]

I suspect that the SAP (whatever/whoever that may be) mean odds ratios, but we cannot be sure. Statistical terminology is not as standardized as we would like, especially across disciplines. The only way to find out is to have a discussion with that person.

If my suspicion is correct, than you can get odds ratios from a meologit model, just add the or option.

Even though I am on record on this list as a strong proponent for reporting odds ratios, I have not found an effective way of talking about effects in terms of odds ratios from an ordered logit model. I know what they mean, but I have not been able to devise a sentence less than 30 words long that accurately describes it in substantive terms and where the audience has not lost track of what I want to say half way through that sentence.

[Bruce Weaver]

I suspect that the SAP (whatever/whoever that may be) mean odds ratios, but we cannot be sure. Statistical terminology is not as standardized as we would like, especially across disciplines. The only way to find out is to have a discussion with that person.

If my suspicion is correct, than you can get odds ratios from a meologit model, just add the or option.

Hi Jennifer. I completely agree with Maarten.

PS- What does SAP stand for?

[Joseph Coveney]

What does SAP stand for?

Statistical Analysis Plan.

[Joseph Coveney]

I have an ordered categorical outcome, trial arm as the main fixed effect and a random effect for site . . . The SAP asks for results presented as adjusted relative risk ratios for intervention versus control . . .

Did the SAP actually prescribe fitting a hierarchical / mixed-effects ordered-logistic regression model to these data?

If this is an industrial product-development clinical study, then it's not considered fiduciarily prudent to hold the financial investment hostage to such a statistical analysis method's risk of nonconvergence.

More commonly, an SAP will specify criteria by which to dichotomize the ordered categories into what amounts to therapeutic-success and therapeutic-failure cases, and then use a noniterative (e.g., Mantel-Haenszel) technique in order to accommodate the "effect for site" and report "results presented as adjusted relative risk—i.e., adjusted risk ratio—for intervention versus control".

If so, then the SAP will be expecting something like the following (or the SAS equivalent)

cs out trt, by(cid)

where out is the dichotomized outcome variable, trt is the "intervention versus control" variable and cid is clinic (site) ID.

Otherwise, as Maarten suggests above, you might want to check back with whoever wrote (or signed-off on) the SAP for clarification, because as both he and Bruce mention "relative risk ratio" doesn't make sense in the context of ordered-categorical response variables.