2
$\begingroup$

I'm carrying out multiple logistic regression with three independent categorical variables each with more than two categories per independent variable.

I want to use the EPV (events per variable) calculation however I'm unsure when counting the number of predictors if I count dummy variables created or I count the number of independent variables?

For example:

Vehicle Type: Car, Van, Bus

Gender: Female Male

Would this be two independent variables or three dummy variables used in the EPV calculation?

I know EPV is not the only method I should use to test sample size however it forms part of the evidence I want to provide on the subject.

$\endgroup$
1
  • $\begingroup$ At least three dummy variables: if your base cases are Female and Car, then your dummy variables are Male, Bus and Van. If you want more interaction, you can also have Male×Bus and Male×Van as well, but at that stage you are losing what I would regard as the interesting part of the model. $\endgroup$
    – Henry
    Commented Jun 19 at 17:29

1 Answer 1

4
$\begingroup$

The minimum sample size just to do the analysis at all, depends on the number of parameters estimated, so for categorical variables you should count the number of dummy variables.

The actual required sample size depends on more than that. What proportion of the dummy variable is 0/1? How large a difference do you expect between 0/1?

If you don't know what effect size to expect, you can still try to simulate the required sample size by asking yourself what the smallest relevant difference is, and then pretend that that is your expected difference. If you have sufficient power to pick up the smallest relevant difference, you have sufficient power to pick up bigger differences too.

These types of simulations require a lot of work substantiating all the input though. I would be wary of any simple 'calculators' for statistical power.

$\endgroup$
1
  • 1
    $\begingroup$ +1. Yes, doing this simulation will require a lot of work. I consider this a feature, not a bug, because it forces you to actually think through your analysis before you run it, and actually before you collect any data... and that is a Good Thing. $\endgroup$ Commented Jun 19 at 9:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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