I found it amusing that, when forced to try modeling my weight data (see previous blog), my DOE software recommended a fifth order polynomial* model! That’s a bit more ‘tayloring’ (Ha ha – inside joke) than I really needed. In fact, just to show how silly this is (5th order!) I offer the following scenario as a cautionary tale. Perhaps it may help to dissuade others who make similarly nonsensical models from what is really just (naturally) randomly generated data.
Looking forward to a work/vacation trip to Tampa in late March (I really will be going there, I am happy to say!), let’s pretend that I use this fifth-order model to help me decide whether to bring a swimming suit. Hmmm, extrapolating out to day 75, when I finish my conference and head for the Gulf shore, the over-fitted model (really should just use the mean!) predicts that by then I will balloon to nearly 100 pounds over my norm. In this case I may easily be mistaken for a beached whale!
It’s just not right to apply model-fitting tools to what is not a DOE, but rather simply a process run-out at steady-state conditions. Extrapolation makes this even more dangerous by far. See the graph for a case in point.
*(A math-phobic person I am acquainted with, whom I will not identify, mockingly refers to these equations as “poly moles” — hence my title for this blog.)