1
$\begingroup$

I was researching logistic regression's critieria, and I found in many different sources that the ideal

sample size = (10 * #of explanatory variables) / (probability of least frequent observations)

Can anyone explain the reasoning behind this formula? I want to understand why this formula is used before I blindly use it to determine if my sample size is large enough for a logistic regression model to be accurate.

Improve this question
$\endgroup$
6
  • 3
    $\begingroup$ You have to state precisely what do you mean by ‘accurate’. $\endgroup$
    – utobi
    Dec 21, 2022 at 5:54
  • 1
    $\begingroup$ I don't think I have ever seen this rule of thumb characterized as "ideal:" usually it is a minimum. One interpretation is that if you have a smaller sample than this, be even more wary of overfitting and lack of power than you might otherwise be. It is usually employed to help you decide how many explanatory variables to use for model fitting in circumstances where you cannot increase the sample size. $\endgroup$
    – whuber ♦
    Dec 21, 2022 at 14:59
  • 1
    $\begingroup$ The minimum sample size is 96 if there are no covariates as argued in my course notes at hbiostat.org/rmsc/lrm.html $\endgroup$
    – Frank Harrell
    Dec 21, 2022 at 15:57
  • $\begingroup$ @Frank Yes, that's very good to know and alerts us to be careful. But that argument is based on a worst case and an arbitrary precision target. There are realistic logistic regression applications that need far fewer data to achieve acceptable power, depending on the purpose. Well-designed dose-response studies come to mind. $\endgroup$
    – whuber ♦
    Dec 22, 2022 at 17:17
  • $\begingroup$ You can only have acceptable power in lower sample sizes if the effect you are trying not to miss is huge. This effect size is almost always overstated. And the 96 is not a worst case. Most analysts would seek a margin of error of less than 0.1 in estimating probabilities and most analyses would involve covariates. The sample size needed for a model with covariates is greater than the sample size needed when only estimating the intercept (which is a best case analysis in this sense). $\endgroup$
    – Frank Harrell
    Dec 23, 2022 at 11:02

0

Reset to default