For the past few days I've been fiddling with a set of data that measure the effects of single nucleotide changes on an organism's phenotype across a variety of environments and genetic backgrounds. Any good statistics manual will tell you the best way to analyze the type of data I have is through logistic regression with a generalized linear model.  So, I tried that.  Here are the results: 

> logistic <- glm(y~full_model, family=binomial(link="logit"))
> summary(logistical)
Null deviance: 20650253  on 860  degrees of freedom
Residual deviance:   118972  on 573  degrees of freedom
AIC: 128316
These are really, really bad numbers.  The statistical method that 'should'  account for interactions between variables actually makes things worse.  (An AIC should be negative, and the residual deviance should be less than the degrees of freedom, not 200 times greater).

Unfortunately, the evolutionary process does not seem to employ the same types of linear thinking that our brains do. That makes it harder to tell a story people can understand.