A logistic regression model: How can you explain a high p-value for a variable in a logistic regression (say .9587) with a point estimate (odds ratio) of >999.99. The variable has a strong Pearson correlation (p = .004) and is highly specific for the outcome variable.
Thanks Jimmy for the explanation. So, would you then report it as an independent predictor of outcome with an odds ratio >999...and for the record I have 17 observations out of 171 (10%) for this variable.
I think Jimmys answer is very much to the point. Concerning the question of whether to use the variable are not: how independent of each other are your explanatory variables? If they are highly correlated, you might want to get rid of collinearity with a PCA and the following analysis may look much better (altough you have to be sure, that the PCA-axis do really mean something in your context). If your explanatory variables are uncorrelated or your "tricky" variable strongly dominates one PCA-axis, I think you cannot exclude the "tricky" variable from your analysis without the risk to lose important information. Of course you have to mention the problem, that there are few observations in the variable and I would definitely add a comment, that the 999 odds ratio cannot be take at face value. Btw: if the Pearson correlation you mention involves the binary response variable, you can hardly use it (as normal distribution is unlikely).
Thanks Jimmy and Irene...and Irene, most explanatory variables are highly correlated.
Apart from this model, the variable in question has been correlated using a different genetic method (odds ratio of 29 in this case) and by a specificity of 99% by the same method. Even in the model and method used in question, there was a 100% specificity and a p = .001. So, I was just curious about what had happened with this second model that was used. At any rate, thanks....Your explanations are highly appreciated.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series