Evaluate the predictions using a paired analysis
Closed this issue · 5 comments
Each of the patient's in cohort 1 are paired with another who is seemingly identical besides having the opposite outcome. These pairs are indicated by the pair column.
A paired analysis can help extract much better models out of data of the same size. It looks like what we would want to do is use a conditional logistic regression model instead of typical logistic regression. pylogit seems to implement this:
This seems to be the relevant example:
http://www.metafor-project.org/doku.php/tips:clogit_paired_binary_data This was the example I mentioned using R. It also only seems to do fitting, but I assume there's a predict method somewhere?
This comment also mentions trying prediction on a different dataset (and potential problems):
https://stat.ethz.ch/pipermail/r-help/2014-June/375765.html
Great, thanks!
I'm looking into getting pylogit to work, but while I was searching something for it I found this: https://www.statsmodels.org/stable/generated/statsmodels.discrete.conditional_models.ConditionalLogit.html#statsmodels.discrete.conditional_models.ConditionalLogit. It looks like it would fit our purposes and seems much simpler to use. Is there any reason not to try/use this? I can also try it alongside pylogit and see if they product similar answers.
This looks great.