Level of significance of a continuous variable in logistical regression

[SJK10]

Hi everyone,
Hoping to get some help with some data analysis. I've collected some data on predictors of Coronary Artery Disease (CAD) in patients with Atrial Fibrillation. The outcome CAD is a dichotomous variable - there are multiple continuous, categorical and dichotomous predictor variables which I've run Logistical Regression analysis as well as the case control command on stata to find an Odds Ratio. One of the predictor variables is TSH (Thyroid Stimulating Hormone) which has a normal range of 0.5 to 4, however I get a statistically significant result. I suspect that this is because the actually level of significance lies within the normal range. I was hoping someone might be able to help as to how to find out at what level would of TSH is this significant. It may just be within the reference range.
Thanks for your help
Stephen

[SJK10]

[Carlo Lazzaro]

Stephen (please, as per FAQ, re-register with your full name surname, too. Just click on the Contact us button and follow the instructions):
- as per FAQ, your chance to receive helpful replies would increase if you post exactly what you typed and what Stata gave you back in turn.

Kind regards,
Carlo

[Carlo Lazzaro]

Stephen (please, as per FAQ, re-register with your full name surname, too. Just click on the Contact us button and follow the instructions):
- per each 1-unit increase in TSH, the odds for CAD increases by 37.5%.
I'm not clear with the remainig part of your question.

Kind regards,
Carlo

[ben earnhart]

*if* I understand your question correctly, you might look at the residual plots, or you might want something like peicewise regression if you suspect that your effects are taking place just within a certain range of your independent variable.

ps. Two other possible solutions are to ad a squared term (nice for its simplicity, but doesn't really capture things well if, for example, all the action is somewhere in the midrange), or some sort of quantile regression (say chop the independent variable up into five dummy variables). Piecewise is probably your best bet, especially since you seem to have an a priori idea where the action is.

[Andrew Lover]

Hi Stephen,

There are some inherent issues with 'binning' a continuous predictor that might be worth considering:

'Dichotomizing continuous predictors in multiple regression: a bad idea'
P Royston, DG Altman, W Sauerbrei - Statistics in medicine, 2006

That said, I certainly have done it; and sometimes it is a necessary evil. Would be good check if there are well-established cutoffs for TSH in the literature (likely), and then also do a sensitivity analysis on the breakpoints (tertiles, quartiles etc as above).