topic Removal of terms from the modelsbased on change in effect estimate. in Statistical Procedures

Removal of terms from the modelsbased on change in effect estimate.

Ashwini_uci — Mon, 17 Dec 2012 02:55:07 GMT

Hi,

In the following text, I am trying to understand the red colored text?

Any help/suggestions what it really means and how to actaully do that?

Multivariable logistic regression was used to adjust for

confounding. Factors in Table
I as well as AIDS, liver disease, peptic ulcer disease, obesity, alcohol or drug abuse, andpsychiatric disorders were included in an initial logistic regression model. Age, sex, race, and renal function were forced into the model, and backward stepwise logistic regression was used to generate the final model. Terms whose removal did not alter the effect estimate for dialysis status by >10% were removed in order of significance.

Thanks,

Ashwini

Re: Removal of terms from the modelsbased on change in effect estimate.

Reeza — Mon, 17 Dec 2012 03:50:14 GMT

My interpretation:

Fit a model with the dialysis and age, sex, race, renal function. The remaining variables are added all at once. Each one is removed one at a time with the highest p value first, if the dialysis status had not changed by more than 10% from the previous or original model (I'm not sure what the 10% would be in comparison to).

Re: Removal of terms from the modelsbased on change in effect estimate.

SteveDenham — Mon, 17 Dec 2012 18:40:49 GMT

And mine: They fit the full model. Then they fit reduced models, deleting one predictor at a time. They made a list of the "deleted models" where the predicted value of dialysis status was changed less than 10%, ordered by the significance of the independent variables that happened to be deleted. The model with the least significant deleted independent variable from this list was then selected. Iterate this process until the deletion resuts in a change in dialysis status of 10% or more.

I just don't see why they used significance to remove an independent variable. Why not choose the variable that had the least effect on the predicted value (i.e., smallest effect size in some context)?

Steve Denham