from Amazon
Wednesday, January 10, 2007
More Clutch
In this blog entry, Charlie Pavitt looks at hypothesis testing and clutch. I made a couple of comments in that thread, most notably that you can achieve a correlation coefficient of .999 if two things have even the slightest possible relationship.
In the BTF thread linking to the Pavitt entry, Wille Keeler asked:
This is only tangentially related, but has anyone figured out how to convert PWA into a winning percentage? (IIRC, they are not equal). The reason that I ask is that the Fangraphs site has a measure called Clutchiness, but it’s a stock stat as opposed to flow stat (counting vs rate).
Let’s take a look at Albert Pujols:
http://www.fangraphs.com/statss.aspx?playerid=1177
He had 18.24 win advancements and 9.0 loss advancements, with +2.67 wins in clutchiness.
His PWA score would be 18.24 / (18.24+9.0) = .670
However, if you remove his clutchiness (say by adding 1.335 to his loss advancements and subtracting 1.335 from his win advancements), his “clutch-free” PWA would have been .621.
Further understanding can be found here:
http://www.insidethebook.com/ee/index.php/site/comments/more_wpa_wa_minus_la_calculations/
***
Also in that BTF thread, Mike E. wrote:
If a typical .850 OPS batter posts a .775 OPS in clutch situations, an .850 OPS batter who posts an .825 OPS can be fairly said to be a clutch performer. The baseline for measuring clutch performance for a player is not that player’s performance in “normal” situations, but the performance of a typical player of similar skills in clutch situations.
While perhaps not expressly stated in conversation, this is certainly implied in analysis. The standard is to always measure a player’s performance above the mean in that context. This is true when you look at playoff performances in hockey, where the mean player will score say 20% less goals. They didn’t all succumb to the pressure. You simply have a different style of play, higher quality goalies, and whatnot. Even if they did all succumb to the pressure, we’d still want to know who succumbed the least.
So, the long and short of it is to always compare the player to the average player for that context.
Now, Mike may also be implying that each quality of player will change his performance at a different rate. That perhaps you won’t get a straight-line change, and so you need to create your study to handle that. I don’t necessarily disagree, but I would be surprised if this is the case.
In Andy’s chapter in The Book, he did control for the base/out situation. However, if you really want to get down to it, you need to create a LWTS equation for every inning/score/base/out, since a power hitter and contact hitter would adapt differently based on what’s more valuable for that game situation.
For example, when comparing Barry Bonds and Mayne, in that fateful game when Bonds was walked with the bases loaded in the bottom of the ninth, I postulated as much:
http://www.insidethebook.com/ee/index.php/site/comments/walking_bonds_with_the_bases_loaded/#2
In Bonds’ case, the power was important. In Mayne’s case, it was mostly a matter of not making an out.
***
It is rather easy to create game-based Linear Weights (take the change in WE for each event, and divide it by the LI for that game-state).
You can try it out yourself, by taking the run value by event by base/out here:
http://www.tangotiger.net/RE9902event.html
and dividing by the LI per base/out state here:
http://www.insidethebook.com/ee/index.php/site/comments/leverage_index_by_base_out_states/