Shame on these people:
“One limitation of this modeling approach is that for some genes, there is a low degree of similarity between the observed expression profile and the one predicted by the most appropriate model. …It is possible that other curve-fitting methods…might also be applicable to this kind of data. However these methods do not provide the statistical machinery that comes with the regression modeling approach that we have taken. The main advantage of our method is being able to easily apply valid statistical tests to determine which one of two models is more likely in light of the data. We are also able to explicitly define statistical significance in a meaningful way that protects against a specified false discovery rate.”
It’s from the paper Decomposition of Gene Expression State Space Trajectories. Summarized, it says, “We don’t know how to use any appropriate tools, so we’ll use inappropriate ones we found in our introductory statistics book. Everyone else calculates p-values for lots of things with it, so it must be okay here, too.”
Let us review what you must have if you are not to drive a statistically inclined mad scientist to sic his ants upon you:
- Your procedure must measure something which is demonstrably relevant to what you are investigating.
- You must know the assumptions under which a procedure is valid.
- You must have tested your data and apparatus against the assumptions and found that they hold.
That’s it. Really. But if my giant mechanical ants were to truly rend those who fail to accomplish these three simple goals, the majority of inhabitants of the groves of academe would be ant food.