Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: st: getting realitve risk from proportional odds ratio

From   Nick Cox <>
To   "''" <>
Subject   RE: st: getting realitve risk from proportional odds ratio
Date   Wed, 8 Dec 2010 17:41:15 +0000

-mlogit- as such knows nothing of any ordering. However, if your response is genuinely ordinal, then presumably the results will echo this and you can draw upon what you know in interpretation. 

As very recently aired on this list, -constraint- is for linear constraints, and inequalities do not qualify as linear. (For example, a = b defines a line, but a > b defines a half-plane.) 


Bontempo, Daniel E

Thanks Maarten -

I thought the there would be this kind of problem.

I am intrigued by the mlogit suggestion. Would I lose the ordered nature of my dv? The levels are not nominal, but increasing. Could I use the constraint option to specify the order?

Maarten buis

--- On Tue, 7/12/10, Bontempo, Daniel E wrote:
> I have used "ologit dv iv, or" where my dv is 0 to 5
> reported limitations on activities of daily living. (With
> such a restricted range of counts, a Poisson model did not
> seem right.) I requested the odds ratio output format.
> However there seems to be a lot of advice on reporting
> relative risk (RR) instead of OR. I find many formulas that
> calculate RR from OR and rate information for the reference
> group (i.e., iv=0). But here I am unsure if what works for
> OR can work for proportional OR. Since the OR is for a one
> level increase in the DV, what rate can I use in the formula.

I very much doubt whether the odds ratios in an ordered logit
can be meaningfully transformed to risk ratios. The idea behind
the ordered logit is that there are effects on the odds of getting
0 versus more, 0 or 1 versus more, 0, 1, 2 versus more, etc, and
that all these effects are constrained to be equal. That is why
it can give you one effect (odds ratio) for each variable. The 
formulas for transforming OR to RR are approximate, and I don't 
think they play well with the constraint implicit in the ordered 
logit model. My guess would be that you would get 5 (number of 
categories -1) different RR for each variable, so that would 
defeat the very purpose of the proportional odds assumption. So 
either I would use -ologit- and interpret results in terms of odds 
ratios, or I would use -mlogit- and interpret results in terms of 
risk ratios.

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index