calculating relative risk ratio from logistic regression

[Martin Imelda Borg]

Hi I was initially calculating odds ratios for the following:
Where my outcome variable sepsis - is a binary variable as 0 or 1
i.procedure is also a binary variable as 0, 1

logit sepsis i.procedure_type comorbidity  [pw=_weight], or

I now want to calculate a RR

Is it correct to use the following - just making sure:

mlogit sepsis i.procedure_type comorbidity  [pw=_weight], rrr

[Bruce Weaver]

Hello Martin Imelda Borg. I think you may be misunderstanding what the rrr option for mlogit is giving you. Please see the explanation I wrote for myself (and others) in the attached PDF.


PS- For anyone who is about to direct me to items 12.4 and 12.5 in the FAQ, I have already looked at them carefully, and I found no explicit prohibition on attaching PDFs. My apologies if I have overstepped! (In the past, I would have provided a link to that file on my website, but my university recently shut down Google sites due to concerns about accessibility, and my site is no more.)

Stata-mlogit-RRR.pdf

[Rich Goldstein]

you want to use -poisson- regression (and be sure you use robust SE's and the irr option; see

h poisson

[Martin Imelda Borg]

Bruce Weaver thanks for this, although why would stata give it ‘rrr’ rather than rename it as odds ratio as there the same

@ Rich G I don’t understand why you recommended Poisson rather than a logistic regression, the explanatory variable is binary 1 or 0 whilst explanatory variable is also 1 or 2. Why have you recommended poisson when this is to do with counts ?

[Clyde Schechter]

although why would stata give it ‘rrr’ rather than rename it as odds ratio as there the same

The relative risk ratio that -mlogit- calculates is prob(outcome i)/probability(base outcome). In the specific case where there are only two outcome categories, then this ratio of risks is the same as the odds ratio, because in that case probability(base outcome) = 1 - probability(non-base outcome). But -mlogit- is typically used in situations where there are 3 or more outcome categories (and produces the same results as -logit- when it is used with two outcome categories).

Why have you recommended poisson when this is to do with counts ?

It is true that Poisson regression is most commonly used with count outcomes. However, the Poisson distribution family is a one-parameter family, and the parameter can be any positive number. It need not be an integer, and the outcome variable in a Poisson regression need not be a count.

Its use with 0/1 outcome variables is particularly helpful when there is a desire to obtain a risk ratio rather than an odds ratio. This occurs because the Poisson regression model has a log link, not a logit link. Though it does not apply to your situation, it is worth noting that Poisson regression can also be used with continuous outcome variables--and in fact this is often preferable to the common practice of doing a linear regression of log(outcome).

Do heed Rich Goldstein's advice to use robust vce when you do this, as otherwise your standard errors and confidence intervals will almost certainly be incorrect.

[Rich Goldstein]

note in answer to your question in #4, that the FAQ says, "Before posting, search the forum for similar questions and consider other ways of finding information:" - this is very relevant here as this issue has been discussed in this forum before (and more than once)