0

From here: R: Calculate and interpret odds ratio in logistic regression

After running logistic regression model with glm with logit link to convert logits to odds ratio, you can exponentiate it:

exp(coef(m))

To convert logits to probabilities, you can use the function:

exp(m)/(1+exp(m))

Question: How to do the same with quasibinomial regression with log (family=quasibinomial(log)) and quasipoison regression (family=quasipoisson("log"))? How can I convert logs to odds ratio and logs to probabilities if I use log link?

I am using survey package for R (for most of regression calculations it uses glm).

vasili111
  • 6,032
  • 10
  • 50
  • 80
  • Can you give a more concrete example? Or perhaps a model specification? With a log link, your response isn't bounded to be between 0 and 1, so I don't see how probabilities or odds ratios apply... (Also, this doesn't seem to be a programming question, so if you can refine it to be clearer, it should be moved to stats.stackexchange) – Gregor Thomas Jun 28 '19 at 16:56
  • But if you look at `?family`, the allowed link functions for binomial data (where odds ratios and probabilities would make sense) are `logit`, `probit`, and `cauchit`. `log` link with a `binomial` family GLM doesn't exist AFAIK. – Gregor Thomas Jun 28 '19 at 16:59
  • @Gregor You are right. Thank you. I use survey package for R (for most of regression calculations it uses glm). I tried to simplify question and thought that glm had same calculations available as survey package. I updated my question. – vasili111 Jun 28 '19 at 17:41
  • With a quasipoisson regression, you're modeling counts, not probabilities, so I still think your off base. [This answer on stats.stackexchange talks about interpreting quasipoisson coefficients](https://stats.stackexchange.com/a/385864/7515). – Gregor Thomas Jun 28 '19 at 18:50
  • I don't see how the `survey` package is relevant. The `quasibinomial` family is documented at the same `?family` page referenced above--it doesn't look like the survey package changes anything about that. The default link for `quasibinomial` is `"logit"`. R will let you set a `log` link, but I don't think it makes any sense. In a quick google search, I couldn't find any examples of people using a quasibinomial family with a log link---I would advise you to not do it either. – Gregor Thomas Jun 28 '19 at 18:53
  • I won't be replying to more comments on this question on Stack Overflow. It's certainly a statistical question, not a programming question. If you want more help understanding this issue, I strongly suggest moving the question to stats.stackexchange. – Gregor Thomas Jun 28 '19 at 18:53

0 Answers0