Hello Pranav,

I have dealt with this issue in many instances and the answer to your question depends on the goal of the use of regression. For example if you have some arbitrarily defined predictor that is normally presented as quantiles or percentiles then you might show the predictor as it might appear. However if you are evaluating a continuous variable that is not normally distributed, even after log transformation, then there is a simple procedure you can follow. Provided the data are sufficiently numerous for comparison, divide the data into increasing percentiles (increasing by units of 5 or 10 percent) and use them as dummy variables for use as predictors on the outcome. Graph the odds ratio or coefficient exponentiated in the y axis and the percentile value on the x value. At one point there will be an departure of the effect you are trying to observe, which often represents the point at which the data become clinically relevant or there is no longer an effect observed.  You can use this to guide setting a relevant cut point for your continuous predictor. 

Hope that helps!

Alex 

Why is dividing a scale into categories based on quartiles or quintiles or anything this arbitrary more clinically relevant? I know it is something that used to be done regularly in that field, but it would seem better if you must categorize a scale to have the cut points based, for example, on results from taxometric analysis (which is now used much more in clinical research thanks to Meehl and colleagues).

If the variables is only "a bit" skewed, then I would recommend using more categories so that the top one is most likely to capture just the cases that are in that tail of the distribution. After that, if you are running regression, I would use the percentiles to create a set of dummy variables, to see if the scores in the uppermost category are indeed detached from the rest of the distribution.

Hello Pranav,

I have dealt with this issue in many instances and the answer to your question depends on the goal of the use of regression. For example if you have some arbitrarily defined predictor that is normally presented as quantiles or percentiles then you might show the predictor as it might appear. However if you are evaluating a continuous variable that is not normally distributed, even after log transformation, then there is a simple procedure you can follow. Provided the data are sufficiently numerous for comparison, divide the data into increasing percentiles (increasing by units of 5 or 10 percent) and use them as dummy variables for use as predictors on the outcome. Graph the odds ratio or coefficient exponentiated in the y axis and the percentile value on the x value. At one point there will be an departure of the effect you are trying to observe, which often represents the point at which the data become clinically relevant or there is no longer an effect observed.  You can use this to guide setting a relevant cut point for your continuous predictor. 

Hope that helps!

Alex 

Thank you very much for the suggestions. That sounds very doable. I was thinking of quartiles and tertiles as that is how data has been traditionally represented, though not always clinically relevant (I agree Daniel). The idea of dividing into percentiles and looking and the outcome to determine what a clinically relevant point of deviation is, sounds like a good approach. 

An alternative and sometime useful approach is to use quantile regression where you can model different (conditional) quantiles of the response simultaneously to see if relations are different in different parts of the response distribution

see:

http://www.econ.uiuc.edu/~roger/research/rq/QRJEP.pdf

Along with Kelvyn Jones' suggestion, you might want to consider "spline" regression as a way to deal with a non-linear predictor.

I would suggest to apply several approaches and compare results (aka sensitivity analysis, in a broader definition), this gives you an idea of how much your results depend on distribution assumptions and (arbitrarily chosen) cut-offs. When applying some sort of categorization, be it tertiles, quartiles or whatever, I strongly encourage to use clinically meaningful or at least applicable cut-offs, i.e. at least rounded to the nearest reasonable magnitude.

Thanks Linus. Yes, I plan to look at data in a few different ways to come up with a more meaningful result.