08-23-2011 03:17 PM
Hello, I am using a binary logistic regression to estimate for the dependent variable y (0, 1). Unfortunately there model has very predictive power. What is worse, the standardized deviance residuals does not follow the standard normal distribution.
Is there any remedy for this? I don't think I can find more predictive explanatory variables.
Should I try other GLM models with other distributions, instead of binomial?
08-23-2011 03:38 PM
Hard to say. In practice, poor performing models is usually not a methodology issue as long as you've been relatively careful in approach. Problem is far more likely with the data, or lack of good data pertaining to the issue at hand.
No amount of methodology research/experimentation/wizardry can overcome poorly gathered and/or irrelevant data. Just my 2 cents worth.
08-23-2011 05:10 PM
I tend to agree with DLing here. Usually the issue lies with as DLing said - poorly gathered and/or irrelevant data. You need to ask yourself if the data collected is pertenent to the hypothesis you are testing. If so then maybe the poor fitting model is a sign that the variables you collected are not influencial to your dependent variable which could be a finding in of itself. There have been many times where I have seen clinical protocols that try to ask questions to with which the data collected cannot even answer.