I would second Terry's advice. 39 predictors seems like an awful lot of predictors. What is the theoretical model behind the 39 predictors? Most models aren't that complex. Perhaps a composite score that reflects some/all of the 39 predictors is in order. Lastly, with the dependent variable split the way it is, would you be answering a worthwhile question in predicting only 10 cases? If the answer is yes, then exact logistic regression may be the best way to approach the problem.

------------ Karen Scheltema, M.A., M.S. Statistician Quality Resources North Memorial Health Care Robbinsdale, MN 55422 (612) 520-2744 fax (612) 520-4686 karen.scheltema@northmail.nmmc.com

____________________Reply Separator____________________ Subject: Re: Logistic Regression and Sample Size Author: Terry Taerum <Terry.Taerum@UALBERTA.CA> Date: 12/19/97 9:23 AM

White, Robin HSURC wrote in message <4C681945EB2DD11189AF00A024D6360412F027@master.sdh.sk.ca>... >Hi Everyone, > >I'm new to logistic regression and would welcome an applied perspective >on the following: > >We are predicting Y after controlling for (entering) 39 control >variables. Our X variable of interest (the 40th variable to enter the >model) is dichotomous. The 0 value has an n of 648; and the 1 value has >an n of 10. > >Can we still get meaningful results given the total number of variables >and the sample sizes (n) for X?

<snip> Along that line of reasoning then, you need to come to a better understanding of the 39 variables you are entering into the logistic regression. From the strictly statistical point of view, you might want to at least consider factor analysis of the predictor variables.

You also need to come to a better understanding of the dependent variable. You need to ask yourself, why are there only 10 1's on the dependent side of the equation.

You need to come to a better understanding of what the relationship might look like between the 39 predictors and the 1 dichotomous variable - if it was meaningful. Would, for instance, only extreme values of the predictor variables be associated with the 10 1's on the dependent side of the equation.

Finally, you need to ask yourself whether there is a better way to test whatever hypothesis it is you are examining.

<snip>

Terry.Taerum@ualberta.ca