Log Regression: how to interpret that the model is significant and the variables in the equation are not.?

#1
Hi,
In a small sample of 20 subjects, three continuous independent variables significantly predict a dichotomous dependent variable (Method = Enter, SPSS). Step 0 (constant) is not significant, but all three variables not in the equation are. Step 1 (Block and Model) are significant Chi (3) = 12.665, p = .005, percentage correct = 90% (very high). But in the last table, no variable in the equation (neither the constant nor the three independent ones) presents a statistically significant coefficient, and Exp (B) are very small: between 0.872 and 1.620. The results are incongruous to me, how should I interpret them?
Thanks
Matt
 

noetsi

Fortran must die
#2
20 cases is really small for logistic regression. You model or variables might be significant and your test not say it. I am surprised anything was significant,
 

Dason

Ambassador to the humans
#3
You're teaching a class and writing on the chalkboard. You get hit with a spitball. The door is closed. The windows are closed. You have evidence that something is going on. A few kids are a smirking. One is even laughing. But none of that means that you can point out who spit it.

You have enough evidence to conclude that at least one of these students is a rotten bastard but you don't quite have enough evidence to point out who exactly that bastard is.
 
#4
Thanks! Actually I have not found any standard to define what is the minimum number of subjects to perform a Log Reg, but I realize that my sample is small. So it came to my mind to add a simple bootstrap analysis (sample = 1000), and I found interesting results: Two independent variables now have significant p-values, and one doesn't.

Well, in the end maybe there were two, and not one rotten bastard, and maybe I found them. They don't know what they're in for! ;-)

Best

Matt
 

noetsi

Fortran must die
#5
There are various rules of thumb for how many cases you need. I guarantee 20 is too low. Its tied to how many variables there are in your model