Should I use null models when I do GLM? I'm using generalized linear models (with negative bin. most of the time) to examine relations between waterbird abundance (response variable) and some environmental variables such as water level (predictor variables). I've built models, performed model selection and validated my models. All is fine. I have working models. Now, I wanna know whether I have to build null models as well and compare them to my models. Is it necessary? I'm really new to statistics so I'm not really familiar with basic concepts like null models. I'd appreciate some simple explanation. Thanks in advance. 
 A: Assessing whether your model is doing more than just overfitting your training data is of course of value (this is what you would get from a test versus a null model). Knowing whether each predictor contributes something would also be of interest (not something you would get fruition such a test). 
Clearly,  you have done extensive model building and variable selection on the training data. Thus, any simple test of your model versus a null model on the training data is invalid. Such a test would need to be conducted on an independent hold out test set of data. 
Alternatively, if you can codify everything you did,  you could try to bootstrap your whole model building process and assess your model versus a null model / assess each predictor. 
A: This is good practice, to build a null model and compare it to your model, to check if your model is any good, or at least better then the null model. If it is not then you should throw your model away and accept the null model, since you do not have enough confidence in it's validity. 
If you are doing this in r the glm function will return a null model deviance as well, which you can use to test for significance and get the p-value. Or alternatively build a null model and compare it with your model with an ANOVA.
