Multicollinearity. Confusing concept, but common issue to watch for w/ regression. Disclaimer, I'm just a clinician, but here's how I explain it --> Redundancy. If you have 2 covariates essentially describing the same thing (e.g. weight and BMI which both represent body size) 1/ https://twitter.com/AJWPharm/status/1388165785793667074
...the computer software doesn't know which covariate to assign the effect to & you get a weird arbitrary split across the covariates which result in confusing/meaningless coefficient estimates. This imprecision is reflected in large CIs for both weight & BMI coefficients. 2/
Good news: Clinicians can usually guess which covariates will be collinear (e.g. ICU admission & APACHE II score since both describe severity of illness; SCr & eCrCl since both describe renal fxn). Using both predictors is redundant, so just PICK one when building a model. 3/
As a good practice, your statistician can also check for collinearity across your model via a correlation matrix, variance inflation factors (VIF), etc. Then they'll ask you which one you prefer. If you don't know, try one & run it. Then replace it & run the other. Compare. 4/
Then pick the model w/ better overall fit of your data. Note, collinearity matters most when I want to know the magnitude of the predictor's effect (i.e. coefficient estimate). Here's an article discussing when multicollinearity can be safely ignored: https://statisticalhorizons.com/multicollinearity 5/
You can follow @julie_justo.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: