LingPipe Blog

Just in time for the last game(s) of this year’s World Series, the final installment of my example of Bayesian stats using baseball batting ability. I took us off-topic (text analytics) with a general intro to Bayesian stats (What is “Bayesian” Statistical Inference?). As a first example, I compared Bayesian calculations of binomial posteriors with point estimates (Batting Averages: Bayesian vs. MLE Estimates). In exploring the question of “where do priors come from?”, I started with simple non-Bayesian point estimates (Moment Matching for Empirical Bayes Beta Priors for Batting Averages). Finally, I realized I’d meant “ability” where I’d said “average” (former is a latent parameter, latter is a statistic calculated from at bats), when I considered Bayesian point estimates (Bayesian Estimators for the Beta-Binomial Model of Batting Ability).

Hierarchical Model Recap

For batter , the number of hits in at bats is modeled as a binomial,

for ,

where the ability, or chance of getting a hit, for batter is . Ability is modeled as having a beta distribution

for

with prior number of hits and prior number of outs . These parameters, which act as priors for the for the binomial parameter, are themselves given priors. The mean of the beta, , is given a uniform prior,

.

The prior for the scale is modeled by a Pareto distribution, which has probability proportional to ,

.

Fitting

I used the 2006 AL position player data (given in this previous blog post). That fixes the number of players and the hits and at-bats for each player .

I then used BUGS encoding of the model above (as shown in this previous post). BUGS automaticaly implements Gibbs sampling, a form of Markov Chain Monte Carlo (MCMC) inference. I ran 3 chains of 1000 samples each, retaining the last 500 samples, for 1500 posterior samples total. All parameters had values very near 1, indicating convergence of the Markov chains. As usual, we read off statistics from the samples and used the sampled values for inference, where they allow the integrals involved in Bayesian inference (as descirbed in this blog post) to be computed.

Beta Posterior

We can inspect the posterior for the beta mean and scale parameters. We plot them as a 2D scatterplot with their means picked out as lines.

As usual, the scale is much harder to estimate than the mean.

Ability Posteriors

We can also plot the ability for particular players. Here’s histograms of sampled posteriors for the players with the best average posterior ability estimates, with their hits and at-bats provided as labels above the plot:

Notice how the variance of the posterior is determined by the number of at bats. The player with 60 at bats has a much wider posterior than the player with 695 at bats.

Multiple Comparisons: Who’s the Best Batter?

We’re going to do what Andrew, Jennifer and Masanao suggest, which is to use our hierarchical model to make a posterior comparison about player’s ability. In particular, we’ll estimate the posterior probabability that a particular player is the best player. We simply look at all 1500 posterior samples, which include ability samples as plotted above, and count how many times a player has the highest ability in a sample. Then divide by 1500, and we’re done. It’s a snap in R, and here’s the results, for the same batters as the plot above:

The .347 batter with 521 at bats not only has the highest estimated chance in our model of having the highest ability, he also won the batting crown (Joe Mauer of Minnesota). The beta-binomial hierarchical model estimate is only 12% that this batter has the highest ability. The estimate is very close for the .343 batter with 623 at bats (Derek Jeter of the Yankees). [It turns out the race to the batting crown came down to the wire.]

The larger number of at bats provides more evidence that the batter has a higher ability than average, thus pulling the posterior ability estimate further away from the prior mean. Finally, note that we’re assigning some chance that the .367 batter with only 60 at bats is best in the league. That’s because when the samples are on the high side of the distribution, this batter’s best.

This entry was posted on November 4, 2009 at 1:05 pm and is filed under Carp's Blog, Statistics. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.