Constructs heat map of probability of a hit or home run for a specific player from pitchFX data

heat_plot.R

heat_plot <- function(player, d, HR=FALSE){
 # inputs
 # player - name of player
 # d - pitchRX data frame with variables Batter, Event, and X, Z (location of pitch)
 # will output a ggplot2 object
 # need to use print function to display the pot
  require(dplyr)
  require(ggplot2)
  require(mgcv)
  # define the strike zone
  topKzone <- 3.5
  botKzone <- 1.6
  inKzone <- -0.95
  outKzone <- 0.95
  kZone <- data.frame(
    x=c(inKzone, inKzone, outKzone, outKzone, inKzone),
    y=c(botKzone, topKzone, topKzone, botKzone, botKzone)
  )
  
  # only consider events that are official at-bats
  TT <- table(d$Event)
  To_Remove_indices <- c(1, 5, 8, 16, 18, 21, 22, 23, 24, 30)
  AB <- names(TT)[-To_Remove_indices]
  d_AB <- filter(d, Event %in% AB )
  # define the 1/0 response variable
  if(HR == FALSE) d_AB <- mutate(d_AB, 
    Hit=ifelse(Event %in% c("Single", "Double", "Triple", "Home Run"),
                            1, 0)) else
    d_AB <- mutate(d_AB, Hit=ifelse(Event == "Home Run",
                            1, 0))
  
  # implement the GAM fit (logistic link)
  pdata <- filter(d_AB, Batter==player)
  fit <- gam(Hit ~ s(X, Z), family=binomial, data=pdata)
  
  # find predicted probabilities over a 50 x 50 grid
  x <- seq(-1.5, 1.5, length.out=50)
  y <- seq(0.5, 5, length.out=50)
  data.predict <- data.frame(X = c(outer(x, y * 0 + 1)),
                             Z = c(outer(x * 0 + 1, y)))
  lp <- predict(fit, data.predict)
  data.predict$Probability <- exp(lp) / (1 + exp(lp))
  
  # construct the plot
  type <- ifelse(HR==TRUE, "HR", "HIT")
  ggplot(kZone, aes(x, y)) +
    geom_tile(data=data.predict, 
              aes(x=X, y=Z, fill= Probability)) +
    scale_fill_distiller(palette = "Spectral") +
    geom_path(lwd=1.5, col="black") +
    coord_fixed() +
    ggtitle(paste(player, type))
}

Thank you for sharing your work. I have a question about using GAM. Why did you use GAM function instead of using GLM? If you could share it, that will be so great. Thanks again!!

Hi A generalized linear model assumes that a response is a linear function of some covariates. In contrast a generalized additive model (gam) allows for input that is an arbitrary function of the covariates. This works better when you have nonlinear relationships— for example, the distance a ball travels depends on the launch speed and launch angle in a nonlinear way. Jim

…

Sent from my iPhone

On Jun 10, 2019, at 2:22 AM, hansr0518 ***@***.***> wrote: Thank you for sharing your work. I have a question about using GAM. Why did you use GAM function instead of using GLM? If you could share it, that will be so great. Thanks again!! — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

Thank you for the explanation!!