I have 18 explanatory variables for modeling Nightjar (bird) occurrence and some are highly correlated. But which value should I use to say two are highly correlated? I read that in pairs correlated >0.5 one should be excluded. But maybe it should be lower - 0.4?
Personally, I would not exclude a predictor solely because it correlated with another predictor. The issue is whether each one makes a unique contribution to the prediction. Of course, the more they overlap (correlate with each other), the harder it is for that to happen. But unless they're virtually interchangeable, it can!
(It might be worth asking whether there's a logical link between the two, as opposed to a correlation that "just happened to turn up" in your data set. In the former case, I'd be more inclined to drop one.)
Maybe you should try running a regression equation with both variables, just to see. If they don't both contribute to a statistically significant degree, then drop the weaker one.
I would generally avoid dropping variables. A better approach is to reparameterize or combine predictors. Data reduction methods can help formalise this but you can also do it from first principles in some cases.
Dear Ivan
The importance of the relationship can be decided from the significant level of the test, in the case of your study the suitable choice is partial correlation or repeated multiple regression analysis.
Good Luck
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis