I'm trying to calculate a scaled odds ratio for an analysis that I'm working on. I have two types of independent variables in the model: 1) percentage data (0-100%) and 2) Euclidean distance to nearest roads. The distance to road data was collected in 1-meter increments; therefore, the odds ratio below for roads_eucl would read for every 1 unit change in distance to nearest road, I would expect a 0.3% decline in use. This doesn't seem very informative given the scale of 1-meter. How would you scale the distance to represent every 10-meters or every 100-meters? In addition, how would you interpret the odds ratio for the percentage data? For example, would you state: for every 1-unit increase in MPHW_Perc, I would expect a 0.8% increase in use? This seems low; therefore, would you scale the percentage somehow? Thanks for the help!

                  OR     2.5 %    97.5 %

(Intercept) 0.7143840 0.6550213 0.7789487
MPHW_Perc   1.0089051 1.0077426 1.0100719
HW_Perc     1.0016561 0.9998965 1.0034183
AG_Perc     1.0084294 1.0071395 1.0097238
Shrub_Perc  1.0129566 1.0095038 1.0164309
roads_eucl  0.9976233 0.9973553 0.9978902
MP_Perc     1.0050919 1.0039986 1.0061888

If you want to look at the decline in odds for a 100 meter change instead of 1 meter change in distance you just raise the odds ratio to the power 100, so 100 meter increase in distance to a road decreases the odds by 21%. If you want to look at a 10 percentage point increase instead of a 1 percentage point increase in MPHW (whatever that may be) you raise that odds ratio to the power 10. So 10 percentage point increase in MPHW results in 9% increase in the odds of whatever you are studying.

Alternatively, you can just create new variables that are roads_eucl / 100 and MPHW_Perc / 10 and add those new variables instead of the old variables.This is my prefered solution, as it is generally a good idea to make sure your variables have the scale you want. It makes it easier to quickly spot errors while doing your analysis and to re-read your log-files when you have received your Revice and Resubmit after 6 months.

| cite | improve this answer | |
  • $\begingroup$ ...is there a way to calculate scaled odds ratios in R rather than by hand? I would like to also include a confidence interval for the scaled odds ratio. Thanks again for the help! Current R code: #odds ratios and 95% CI exp(cbind(OR = coef(glm.catfinal), confint(glm.catfinal))) $\endgroup$ – Buck2079 Feb 10 '15 at 15:47
  • $\begingroup$ I updated the answer to answer that question too. $\endgroup$ – Maarten Buis Feb 11 '15 at 8:31
  • $\begingroup$ I assume you meant roads_eucl / 100 and not roads_eucl / 1000? Would this affect the model fit at all if a number of your road_eucl distances range from 0-200-meters? Thanks! $\endgroup$ – Buck2079 Feb 11 '15 at 16:09
  • $\begingroup$ thanks, I corrected the answer. It won't affect the model fit at all. The two models are mathematically equivalent, as it just involves a linear transformation of variables. $\endgroup$ – Maarten Buis Feb 12 '15 at 8:27
  • $\begingroup$ Great! Thank you for your input! Due to model convergence issues on another modelling procedure, I need to re-scale my variables (or standardize by subtracting the mean and dividing by the standard deviation. How would you interpret the odds ratios on scaled data like this? $\endgroup$ – Buck2079 Feb 12 '15 at 11:06

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.