Tuesday, February 23, 2010

A Little Butt-Whacking Never Hurt Anyone

Via ESPN's TrueHoop, I read yesterday a New York Times article about a study by UC-Berkeley researchers showing that physical touching between NBA players is positively correlated with good on-court performance.

The article is yet unpublished, so it is difficult to thoroughly understand or critique it. A few points come to mind:

  • According to the NYT summary, the data consisted of "every bump, hug and high five in a single game played by each team in the National Basketball Association early last season." As I interpret that, they followed one game for each of the 30 NBA teams -- i.e. they followed 15 games total. That's an awfully small sample size, considering that the total number of games in one season is 82 times that number. Possibly, they followed 30 separate games (focusing on just one team in each game), but that's still not much.

  • According to the NYT write-up, the study uses assist-to-turnover ratio, rather than points scored or plus-minus point differential, as the measure of team and player performance. If they were attempting to measure the effect of positive point performance upon touching, then we could interpret the use of assist-to-turnover ratio as an "instrumental variable" -- i.e. an alternative explanatory variable that is correlated with the real explanator (points) but has no reverse causality from the dependent variable (touching). Does this make sense, though? If touching encourages better point scoring and better defense, then touching could also encourage better ball-sharing and ball-husbanding.

  • On the other hand, if the researchers' working model is that touching leads causally to better play, then I don't fully understand why they would want to replace the purported dependent variable (point differential) with another, related dependent variable (AST/TO ratio). Surely there is a reverse causality problem here: more assists could mean that players develop a mutual affinity, encouraging touching. Fewer turnovers would reduce enmity and resentment, again encouraging touching. That good play encourages touching seems like a fairly trivial finding, even if some numerical confirmation would be nice.

  • How, then, to tease out causality? I might recommend something like the "Propensity Score Matching" statistical technique (a brute-force technique in non-experimental settings to group together similar subjects of study, thus controlling for background variables, and then observing what happens when the treatment is applied vs. when it's not applied). But analyzing data this way is definitely not my day job.
  • No comments: