Hello. I need to perform a logistics regression for a small sample size (n=20) to determine predictive factors for an event. The only problem is that a few data points skew the whole sample, leading me to receive a massive Odds Ratio of >200. This is not a realistic result, and am wondering how to best negate this. How does one properly perform a logistical regression with small sample sizes with high variance in certain variables? Anyone have any tips or tricks?
Thanks in advance
See the attached screenshot paper and related work by these authors. Best wishes David Booth
Hello Michael Nordine. How many events (or successes) are there in your sample of n = 20? And how many variables are you including in your model?
Number of positive events are important. If they are less than 5, Poisson regression may be a better choice.
Large ORs may have some reasons; 1) collinearity with another covariate, 2) separation bias can be treated by penalized (Firth) logistic regression.
I am sorry to say that you need more observations.
It is worth remembering how little information you have when estimating a binary response variable. You are interested in the 'latent probability' of an event on a continuous scale but can only observe the extreme values of 1 and 0.
20 Data points might be enough if they are nicely spread (enough of all kinds of outcomes). Could you share the data, so we could have a look?
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts