1

I am at the univariable stage of my analysis, and I am looking at significance of individual variables/predictors. I've tested each predictor for significance by Wilcoxon Rank-Sum for continuous variables (not using t-test because data are not necessarily normal) and chi-square for discrete variables. However, in addition to p-values, I also need to calculate odds ratios (and 95%CI for the OR) for each predictor.

I was able to calculate ORs for the discrete variables using the (ad)/(cb) shortcut with a simple 2x2 table. But I cannot find instructions for how to calculate univariable ORs for the continuous variables.

Should I run a separate logistic regression for each of the variables, where the independent variable is the continuous predictor and the dependent variable is my 0/1 outcome?

Any help is appreciated!

mountainave
  • 137
  • 1
  • 2
  • 6
  • This sounds like some kind of variable selection for a regression model. Is that what you're doing? – Dave Jun 07 '21 at 20:53
  • 2
    What is a "univariable stage of my analysis" ? Please do NOT use the "significance" of variables in univariable regression models to inform which variables to use in a multivariable regression model. This is a terrible mistake. – Robert Long Jun 07 '21 at 20:56
  • @RobertLong Yes, understood. No, I'm not hunting for p-values. Working on a manuscript and need to report significance of predictors in both univariable and multivariable analysis. – mountainave Jun 07 '21 at 21:01
  • Why in the univariate case? You don't have to do every statistically dubious suggestion a reviewer makes. – Dave Jun 07 '21 at 21:02
  • I'm glad you understand :) but in what sense is the significance of variables in univariate models relevant to your research question(s) ? How will these be used ? – Robert Long Jun 07 '21 at 21:03

1 Answers1

2

There is no straightforward way to compute odds ratios manually for continuous predictors. You need to run a one-predictor logistic regression, exponentiate the coefficient on the predictor and its confidence bounds, and then report that as the odds ratio and its 95% confidence interval. You can (and maybe should) do this with your binary predictors as well.

Noah
  • 33,180
  • 3
  • 47
  • 105
  • I realized this was a possibility just after posting. Thanks, Noah. I'm so used to working with multivariable regressions that I forgot univariable regressions are a thing, too :-) – mountainave Jun 07 '21 at 20:56