Binary Logistic SNAFU

#1
Hello,

I'm curious if I have coded my variables incorrectly for an analysis. Here's a brief explanation:

Independent variables:
X1: temperature (ordinal: 10 or 4 degrees C)
X2: treatment group (categorical: exposed to pathogen or not)
X3: sex (categorical...)
X4: index of body condition (scale: ranging from 0.15-0.24)

The Wald statistic was significant for all variables except X2 (P = 0.11), which confuses me, because looking at the results, the results (0 or 1, of course) are so different between the two categories of X2. Looking at the data, it seems 'wrong' that this variable is not significant. Could it be that there is an issue with my designation of body condition as a scale variable? While the distribution of these values are normal, the range is small, but the Exp(B) result is enormous 1.626e18, and the B = 48.8. I'm wondering if that could be falsely inflated somehow as a result of my variable coding.

Thoughts?

THANK YOU.
 

CB

Super Moderator
#2
The Wald statistic was significant for all variables except X2 (P = 0.11), which confuses me, because looking at the results, the results (0 or 1, of course) are so different between the two categories of X2.
There may well be a bivariate relationship between treatment/not and your DV that disappears once you control for the other variables. This'd be plausible if your treatment and control groups differed quite a lot with respect to the other IV's. (Do they?)


Could it be that there is an issue with my designation of body condition as a scale variable? While the distribution of these values are normal, the range is small, but the Exp(B) result is enormous 1.626e18, and the B = 48.8.
Remember that the Exp(B) is telling you about the change in odds as a result of a 1-unit increase in the body scale variable. But your body scale measurement has a range of 0.15 to 0.24, so the scale of this estimate is confusing. If you did something like calculate exp(48.8/100) I believe you'd get the odds ratio for a 0.01 increase in the body scale variable, which would probably be a lot more informative given the scale you have at hand.
 

rogojel

TS Contributor
#3
Hi,
I usually get such extreme odds ratios only when something is wrong with the balance of my data - for instance the vast majority of the DV values is from one type or something like that. Could it be that this is the problem?

regards
rogojel
 
#4
The Wald statistic was significant for all variables except X2 (P = 0.11), which confuses me, because looking at the results, the results (0 or 1, of course) are so different between the two categories of X2.
How are they different? Do you mean you ran something like a contigency table and there were different ratios of the DV at different levels of the IV? Or that the levels of the IV seem very different to you?