Wednesday, January 11, 2012

Calculate umpire strikezone sizes

Head on over to THT, and you should see a post by yours truly about calculating the size of umpire strikezones. Here I will open source the code for how to do that. The hard part though is that in order to make use out of the code, you need to have every single pitch from 2011 available at your disposal. If you have that, then you can use the following code to generate umpire strikezone sizes. The script makes very heavy use of the "plyr" package (packages load libraries of extra functions and/or data into R), which may be my favorite R package. It was designed by Hadley Wickham, who is a celebrity of sorts in the R community. The plyr package is great for data manipulation and group-wise summaries. You can install it with:

install.packages("plyr")

You also need the "mgcv" package for creating the model that is used in the calculations.

You can find the code below in blue, and comments in green. Hope it's readable:

#load library
library("plyr")

#initalize -- add necessary variables. Best way is with vectorized ifelse function. 
pitcher$babipn = ifelse(pitcher$type=="X" & pitcher$event %in% c("Single", "Double", "Triple"), 1, 0)
pitcher$inplay = ifelse(pitcher$type=="X" & pitcher$event != "Home Run", 1, 0)
pitcher$swing = ifelse(pitcher$type=='X' | pitcher$des %in% c('Foul', 'Foul Tip', 'Foul (Runner Going)', 
            'Swinging Strike', 'Swinging Strike (Blocked)'), 1, 0)
pitcher$walk = ifelse(pitcher$des %in% c("Ball", "Ball In Dirt", "Intent Ball") 
            & pitcher$event %in% c("Intent Walk", "Walk") & pitcher$ball==3, 1, 0)
pitcher$strikeout = ifelse(pitcher$des %in% c("Called Strike", "Foul Tip", "Strike",
            "Swinging Strike", "Swinging Strike (Blocked)")
            & pitcher$event %in% c("Strikeout","Strikeout - DP") & pitcher$strike==2, 1, 0)
pitcher$homerun = ifelse(pitcher$des == "In play, run(s)" & pitcher$event == 'Home Run', 1, 0)
pitcher$full_name = paste(pitcher$first, pitcher$last)

#find number of plate appearances for each umpire, don't know why I named it atbats
#calculate kwERA, FIP, babip, and swing rate
pitcher = ddply(pitcher, .(ump_id), transform, atbats = length(unique(ab_id)))
pitcher = ddply(pitcher, .(ump_id), transform, fip = ((13 * sum(homerun)
              + 3 * sum(walk) - 1.93 * sum(strikeout)) / (.23 * atbats)) + 3.10,
              kwERA = 5.24 - (12 * (sum(strikeout) - sum(walk)) / atbats),
              babip = sum(babipn) / sum(inplay), swing = mean(swing))

#subset data to only look at pitches where umpire makes a call
#this won't affect the stats we just calculated
pitcher = subset(pitcher, des %in% c("Automatic Ball", "Ball","Ball In Dirt", "Called Strike",
                                   "Intent Ball", "Pitchout", "Strike"))
#add in variable to describe called strike
pitcher$cs = ifelse(pitcher$des == 'Called Strike', 1, 0)

#find number of pitches seen by each umpire
pitcher = ddply(pitcher, .(ump_id), transform, n = length(px))

#only look at umpires with greater than 3000 pitches
pitcher = subset(pitcher, n > 3000)

#define function to calculate strikezone area
csarea = function(x) {require('mgcv')
  #create model with mgcv package, use smoothing                  
  model <- gam(cs ~ s(px) + s(pz), data = x, family = binomial(link='logit'))
  mygrid <- expand.grid(x=seq(-4, 4, length=50), y=seq(-2, 6, length=50))
  #find predicted called strike rates for a bunch of tiny bins                  
mygrid$z <- as.numeric(predict(model, data.frame(px=mygrid$x, pz=mygrid$y), type='response'))
thresh <- 0.5
  #find number of tiny bins where predicted rate is at least .5                  
nbins=length(mygrid$z[mygrid$z >= thresh])
  #multiply number of tiny bins by size of each bin                  
areas=nbins * 0.02665556
  #add in fields to data.frame                  
x$area <-rep(areas, length(x[,1]))
x$nbins <- rep(nbins, length(x[,1]))
return(x)
}

#call + evaluate csarea() on each subset of umpire data
pitcher=ddply(pitcher, .(ump_id), "csarea")

#create a summary dataset for each umpire, then sort by area
sum.pitch=ddply(pitcher, .(ump_id), summarize, area=mean(area), full_name=full_name[1],
                fip=mean(fip), kwERA=mean(kwERA), babip=mean(babip), swing=mean(swing))
sum.pitch=arrange(sum.pitch, area)

3 comments:

  1. Hmm. Interesting. Here's a few suggestions:

    1. I have found that a joint smooth is more appropriate than additive smooths for px and pz. UBRE score is lower for the former model than the latter (and the zone ends up being more interesting as well).

    2. If you're low on RAM, try using the "bam" version of "gam" in the mgcv package. This saves a lot of memory.

    3. I actually created a "for loop" for my plots. You can do it for lots of functions...prolly inefficient but it helps with making lots of similar pictures saved as files. Just do:

    myx <- matrix(data=seq(from=-2, to=2,
    length=100), nrow=100, ncol=100)

    myz <- t(matrix(data=seq(from=0,to=5,
    length=100), nrow=100, ncol=100))

    fitdata <- data.frame(px=as.vector(myx),
    pz=as.vector(myz))

    for(i in unique(umpire_id)) {

    d <- subset(data, data$umpire_id==i)

    umpname <- paste(i, data$ump_FN,
    data$umpLN, sep="")

    filename <- paste(i, "Heat.png", sep="")

    png(file=filename, height=850, width=750)

    fit <- gam(strike_call ~s(px,pz,k=51),
    method="GCV.Cp",
    family=binomial(link="logit"), data=d)

    mypredict <- predict(fit,
    fitdata, type="response")

    mypredict <- matrix(mypredict,
    nrow=c(100,100))

    contour(x=seq(from=-1.5, to=1.5,
    length=100), y=seq(from=1, to=4,
    length=100), z=mypredict, lwd=2,
    lty="dashed", axes=T, levels=.5,
    labels="", labcex=1.3, col="darkred",
    xlab="Horizontal Location (ft.,
    Umpire's View)", ylab="Vertical
    Location (ft.)", main=umpname)

    dev.off()
    }

    4. Once I have my model for an umpire, I could also add the simple code to calculate area in inches, or just do it manually:

    library(splancs)

    ###Area Contour
    clines <- contourLines(x=seq(from=-1.5, to=1.5,
    length=100), y=seq(from=1, to=4,
    length=100), mypredict)

    ump_area <- round(with(clines[[5]],
    areapl(cbind(x,y)))*12, 3)

    ***
    Anyway, good stuff. I'll have to play with "ddply". My code when doing crosstabs and stuff like that is really awful and inefficient. I have reached the limits of "tapply".

    :-)

    ReplyDelete
  2. Also, sorry for the crappy code format through the comments. Ick!

    ReplyDelete
  3. Thanks for the info/comments, Millsy. And about the ddply function and plyr package, this pdf shows some ways to make use out of it. http://www.jstatsoft.org/v40/i01/paper

    After getting comfortable with the package the amount of for loops and lapply/split that I use has declined a lot.

    Another way to generate the plots would be
    pdf("filename.pdf")

    d_ply(pitcher, .(ump_id), function (x) {code to make plot})

    dev.off()

    And the way you are finding area with the splancs library seems better than what I'm doing, so I will have to check that out.

    ReplyDelete