NBA: does taking more shots lead to lower accuracy?
One of the hypotheses about why you can't take a player's FG% at face value, is that good players will be asked to take a lot of worse shots, like desperation FG attempts with the shot clock running out. That will, perhaps misleadingly, lower their percentage.
While I had my 2008-09 data typed in, I'd thought I'd give that a quick test. I'm sure this has been done before, but I figured I'd post it anyway.
I tried to predict a team/position's eFG% using (a) his position; (b) the rest of the team's FG% (averaged by position); and (c) the percentage of his team's FGA he took. There were 30 rows in the regression, one for each 2008-09 NBA team.
The resulting equation:
+ .022 if he's a SG
+ .016 if he's a SF
+ .025 if he's a PF
+ .043 if he's a C
+ .643 * the eFG% of the rest of the team
- .114 * the percentage of the team's FGA he takes.
The hypothesis seems to check out. The more shot attempts, the lower the overall percentage. For instance, suppose a player takes 20% of this team's attempts, and shoots .500. If he took only 19% of his team's attempts, he'd shoot .5114. If he took 21% of his team's attempts, he'd shot .4886.
That's bigger than it looks. A .500 percentage over 20 shots is 10 FG. A .5114 percentage over 19 shots is 9.727 FG. So, the extra shot nets only 0.283 FG. That means that, at the margin, the player shoots only .283 on those extra shots.
Second, note the high correlation between a position and his teammates. For every percentage point the teammates shoot better than average, the individual player will shoot 0.643 points better than average.
Everything was statistically significant at the 1% level, except the -.114, which was significant at only 7.2%. Its standard error was 63 points, so we definitely need at least another year's data before we can say we have a true understanding of the size of the "shoots more, therefore shoots less accurately" effect.
Also, if certain positions are asked to take desperation shots more than others, the regression might benefit from interaction terms. I'll leave that to you guys. My dataset is available on request if you want to play with it a bit.
Looking forward to any comments you basketball guys may have.