Form over function

After my retirement from cycling and before I got really serious about running, I raced cars for a few years (I'm not sure I'm capable of an existence where I'm not racing something). Just amateur SCCA stuff but real-live car racing nonetheless. After driving a couple years in the stock classes, which don't allow you to change much beyond tires, I decided to dump some money into the car to make it go faster. The hot class for my car was called Street Touring. It was basically the "Rice Boy" class. You could lower the car, stiffen the shocks and springs, make some modest intake and exhaust adjustments. As my car was a Honda product, that's where everybody expected me to go. When I chose a different class where I would have to spend my money a bit differently and wind up with a less competitive car, one of the other driver's asked me, "Why didn't you go to Street Touring?" My response, "I didn't want to spend four grand to make my car slower."

You see, unlike stock, Street Touring didn't let you run real race tires. You were stuck with the typical low profile, three-season traction stuff you see on hopped up street rods. It looks cool, but the grip is way worse than a soft-compound tire with a race profile and low-void tread. And, tires are by far the most important part of a race car. My stock car with race tires was faster than anything I could build for Street Touring. So, I went for a class that allowed race tires. I didn't do as well nationally, but I did win a bunch of local races. And, it was fun to be turning in times faster than all the Street Touring cars (as well as a lot of much more expensive sports cars from the stock classes).

I bring this up because I'm at a similar point with the dissertation research. There are better opportunities for "good looking" (read, publishable) results by ditching the stratification and focusing on the correlation. And, I guess that's the way we're going to go. But, I'll still glance longingly at those graphs from time to time knowing that all the theorems in the world won't make a block sampler converge as fast as CISS.

Yeah, that's what I was talking about.