how to judge significant changes in odds ratio estimates

[rebekah shao]

Hi all,

I have run a multivariable logistic model and for my main exposure of interest I have got an OR 1.7 (CI 0.78 - 3.83, p = 0.17). I ran goodness of fit tests (Hosmer-Lemeshow) and found that my model was not a good fit, so I examined residuals etc, removed observations with a large effect and re-ran the model again. This time it was ok in the goodness of fit tests, and I got a new estimate of the exposure: OR 1.4 (CI 0.61-3.40, p = 0.40).

My question is: is there a statistical rule of thumb (e.g. 10% change in estimate) that would help me to decide whether this change in OR (1.7 to 1.4) is an important size? I would like to know how to interpret this, at the moment I have just said that removing a few observations doesn't change the estimate much. Clinically, a 40% increase in odds of depression (my outcome) doesn't seem that different to a 70% increase in odds.... And both times the effect of this exposure is non-significant on the outcome anyway..

Many thanks!!

[Carlo Lazzaro]

Rebekah:
some comments about your post:
- as you may still be aware of, posting what you typed and what Stata gave you back increases dramatically your chances of getting helpful replies (please read the FAQ about this and other posting-related topics. Thanks);
- you performed a multivariable logistic regression: hence the OR you mention is adjusted for the effect of other predictors;
- you removed observations with large effect: without investigating the origin of that large effect, this may sound as making-up the data and you may end up with a sample far off the original one;
- being not a p-value fan, I'm more interested in CI width. In both cases they well include 1. It may well be that your sample is too small to give back an evidence of a difference;
- as far as I know, there's no such a statistical rule of thumb. However, the clinical relevance of a given finding may be quite different from its statistical significance.