One sure illustration of the practical limits of statistics is to take a look at the football statistics* compiled by the NCAA, in particular the defensive statistics. For example, take a look at Ole Miss’ season statistics through six games.
The statistic most people have been focusing on is the Rebels’ pass defense, which is giving up 345 yards per game—dead last in the NCAA (117 of 117). Yet in the past two games, the Rebels have only given up 17 points and basically shut down their opponents’ passing games; if they’re the worst pass defense in the country, shouldn’t they have been lit up by Florida and Arkansas State?
There are three forms of selection bias at work here. The first form of selection bias is that opposing teams are passing because their rushing offense is going nowhere (the Rebels are ranked 11th, conceding just 82.5 yards/game on the ground). The second form is that the Rebels have faced two of the country’s most pass-happy offenses: #1 Texas Tech and #13 Memphis, and this hurts their pass defense statistics; the #116 pass defense, North Carolina State, also faced Texas Tech. And the third form is that teams tend to pass when they are behind; the Rebels led both Memphis and Texas Tech by double-digit margins in the second half of both games, so those teams passed even more than usual.
If you just look at the numbers, you’d think the plan to beat Ole Miss would be to pass. But unless your quarterback is as good as B.J. Symons or Danny Wimprine, that may not work; heck, Florida’s Chris Leak, whose passer rating is better than Wimprine’s, threw three picks to the secondary. On the remaining schedule, LSU’s Matt Mauck and MSU’s Kevin Fant are the only QBs known as good passers; Arkansas’ Matt Jones is primarily an option quarterback, as is South Carolina’s Dondrial Pinkins, Auburn’s Jason Campbell is a mediocre passer, and Alabama will be lucky if its third-string QB can suit up with the injuries that plague that team. So exploiting this weakness—if it actually exists—is not something that these teams are likely to be able to accomplish.
* Arguably, these statistics are strictly speaking parameters. I’ll just mumble “superpopulation” and hand-wave that issue.