A What is a large change in deviance?

[FallenApple]

So for GLMs, what is considered a large change in deviance?

If I go from 3500 in one model to 3200 in another, does that mean that the second model is better?

 

[Dale]

That depends on the application. Does it make a practical difference in the intended purpose?

A larger model (more terms) will always fit the data better. In extreme cases this is just fitting to noise. You need to penalize the less parsimonious model. Two common ways of doing so are the AIC or the BIC.

 

[FallenApple]

That depends on the application. Does it make a practical difference in the intended purpose?

A larger model (more terms) will always fit the data better. In extreme cases this is just fitting to noise. You need to penalize the less parsimonious model. Two common ways of doing so are the AIC or the BIC.

Not sure if it would make a practical difference. Its just a small percentage loss for the deviance.

Is it because a model that is overfitted would not be able to capture the next point that is added to the model? Because the overfitted model would have the curve zigzag around the data to the point where the function is hardly even is a function.