4
$\begingroup$

The coefficients estimated for logistic regression are in log odds, and I understand it is common -- at least when interpreting the output -- to convert the log odds to odds so they're more easily understandable.

When reporting results for logistic regression, should log odds and their standard errors be reported, say, in a table, and then odds used to interpret the coefficients?

Or, can odds (and, uh, standard error odds?) be reported in a table?

$\endgroup$
5
$\begingroup$

It's fine to report odds ratios rather than log-odds differences if that will be easier for your audience to understand. If you report odds ratios it probably makes sense to report the confidence intervals. That is, if $\beta_i$ are the log-odds differences, $\sigma_i$ is the standard error of $\beta_i$ (also on the log-odds scale), and $q$ is 1.96 for $\alpha=0.05$ (or whatever Normal quantile you want use), then you should report the back-transformed confidence intervals:

$$ \exp(\beta_i \pm q \sigma_i) $$

rather than back-transformed values $\pm$ standard errors, as the confidence intervals are often asymmetric on the odds scale (they're not necessarily symmetric on the log-odds scale either, but generally much closer).

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks, should $\sigma_i$ be in terms of odds ratios - or its original units (which I think are in terms of log odds)? $\endgroup$ – Joshua Rosenberg Nov 30 '16 at 13:10
  • $\begingroup$ original units. $\endgroup$ – Ben Bolker Nov 30 '16 at 13:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.