http://www.andrewheiss.com/blog/2016/04/25/convert-logistic-regression-standard-errors-to-odds-ratios-with-r/

[andrewheiss]

Convert logistic regression standard errors to odds ratios with R | Andrew Heiss
http://www.andrewheiss.com/blog/2016/04/25/convert-logistic-regression-standard-errors-to-odds-ratios-with-r/

[andrewheiss]

Originally posted by David Randahl on 2017-04-06 09:46:24

Hi Andrew,

I just have one question: WHY would you want to calculate this? While
odds-ratio adjusted standard errors apparently is a thing, I do not see
why in the world you would want to calculate this. It cannot be used for
obtaining confidence intervals, and it cannot be used for calculating
significance. It is simply a useless concept. Confidence intervals in an
odds-ratio setting are not symmetric, so you cannot use this measure to
calculate them. Similarly, the Z-value is a function of the distance
between a target value and the coefficient based on the standard error,
but since the regression is conducted in log-form, the transformation to
odds-ratio makes these values non-linear too.

It's easy for people to get confused by posts like this, so please
consider removing it or preface it with a warning.

Best,
David Randahl
PhD Candidate
MSc Statistics

--
P.S.

Your calculations are overly complicated.
Sqrt(exp(coef)^2*var(coef))=exp(coef)*std.e(coef) which is much
easier to conceptualize, and extremely trivial to implement in R

[andrewheiss]

Originally posted on 2017-04-06 14:06:45

Sure, it's messed up and not entirely useful, but it's what Stata does
and, correspondingly, it's what appears in lots of published work in
economics and political science. When replicating articles that do
report the transformed SEs, it's helpful to be able to transform them
to ensure the replicated results match the published results.

[andrewheiss]

Originally posted by Kevin R Carriere on 2018-05-16 12:10:02

Hi Andrew,

I just wanted to comment and thank you for this code. The above
commenter completely misses the point you outline - that regardless of
its clarity or usefulness, it is the expected standard and it is
necessary to be able to convert for publication!

[andrewheiss]

Originally posted by JoAnn A on 2018-08-17 22:00:54

This was very useful to me. I needed to calculate a confidence interval
for c*(RR - 1), where RR is the relative risk obtained by log binomial
regression.

[andrewheiss]

Originally posted by RoseB on 2020-02-12 18:46:54

Thank you for this tutorial, it's very helpful. How would I convert the
logistic regression SE to odds ratio SE if the only information I had
was the log odds, standard error, and N and not the variance covariance
matrix? Many thanks.

[andrewheiss]

Originally posted by Fahim Ahmed on 2020-11-12 15:28:57

Hello Dr. Heiss,

Just to note here that, a easier way to get the odd-ratio from the
estimates using broom::tidy would be:

model.df %>% tidy(exponentiate = T,
[conf.int](http://conf.int "conf.int") = T)

Thanks,
Fahim Ahmed
CEE, UofSC

[andrewheiss]

Originally posted on 2020-11-16 02:53:13

broom::tidy() exponentiates the coefficients and confidence intervals,
but not the standard errors

https://uploads.disquscdn.c...

[andrewheiss]

Originally posted by Jonathan Gellar on 2021-06-03 15:35:01

Doing a little math, var.diag is just std.error^2, so
[or.se](http://or.se "or.se") is just
or*std.error. So there are some easier ways to calculate this.