Logistic regression model

#1
Hi all

I was wondering if you were kind enough to give me a short advice on a statistical query i have.

I have a sample of 15 children with a rare medical condition on treatment with a special diet. During the treatment course I have serial measurements (4 measurements at 0, 15, 30, and 60 days of treatment) of a disease marker. For some of these children (n=6) the disease activity improved and for the rest 9 did not at the end of treatment (60 d).

The research question I have to address is whether the percent change from baseline of the disease marker at 15 and/or 30 days can predict the clinical outcome at 60d (disease improvement or not improvement).

What I did is:

I calculated the %change from baseline at 15 and 30 days and I applied a binary logistic regression model for each time point seperately . The response is (disease improvement or not improvement) and the predictor % percent change of the inflammatory marker concentration at 15 and 30 days.

Now although I am getting "a test of all slopes are zero" which is very significant (p<0.0000001) the predictor (%change) is not (p=0.177)! Moreover when I did Mann Whitney test for difference in %change at 30day between patients who achieved clinical remission and those did not I got a p-value <0.0001.

I am a bit mixed up whether the %change at 30 days is an actual predictor of clinical outcome (disease remission) or not. Any advice on that?

I attach/uploaded a word document with the minitab analysis

I am looking forward to hearing from you


Kostas
 
#2
Just a thought - Evaluate a frequency table of your dependent variable (Y/N) by the predictor variable. Check to see how the predictor is distributed. If all the Ys had high percent change and all the Ns had low percent change (or vice versa), you will get this problem.
 
#3
Hi there and thank you very much for your advice. Indeed my Ys have a higher precent change than my Ns. Overall which is the conclusion? Can the %change predict the outcome or not? I need that for one of my papers who is under review by a medical journal. Many nany thanks once again

I am looking forward to hearing from you
 
#4
Yes. I think %change can predict the outcome well. In fact, it is too good of a predictor for logistic regression. I think you may have to use a t test (or your nonparametric method) instead. By the way - if you want to look into this problem more. It is called 'separation'.

Good luck.