# Topic 4: The Binomial Model
set.seed(2804)
library(LearnBayes)

# ---- Globe tossing: grid approximation ----
tosses <- c("W", "L", "W", "W", "W", "L", "W", "L", "W")
W <- sum(tosses == "W"); n <- length(tosses)
for (G in c(5, 20, 1000)) {
  grid <- seq(0, 1, length = G)
  post <- dbinom(W, n, grid) * 1          # uniform prior
  post <- post / sum(post)
  cat(sprintf("%4d points: P(p < 0.5) = %.4f\n", G, sum(post[grid < 0.5])))
}
pbeta(0.5, W + 1, n - W + 1)               # exact
draws <- sample(grid, 10000, replace = TRUE, prob = post)
c(P_below_half = mean(draws < 0.5), quantile(draws, c(0.05, 0.95)))
qbeta(c(0.05, 0.95), 7, 4)

# ---- Toss-by-toss updating ----
op <- par(mfrow = c(3, 3), mar = c(2, 2, 2, 1))
a <- 1; b <- 1
for (i in seq_along(tosses)) {
  curve(dbeta(x, a, b), 0, 1, lty = 2, ylim = c(0, 3), main = paste(tosses[1:i], collapse = ""))
  if (tosses[i] == "W") a <- a + 1 else b <- b + 1
  curve(dbeta(x, a, b), add = TRUE, lwd = 2)
}
par(op)

# ---- HPD interval from draws (unimodal) ----
hpd_draws <- function(x, prob = 0.9) {
  x <- sort(x); k <- floor(prob * length(x))
  w <- x[(k + 1):length(x)] - x[1:(length(x) - k)]
  i <- which.min(w); c(x[i], x[i + k])
}
hpd_draws(rbeta(100000, 7, 4))

# ---- Point estimates and one-sided intervals: premature births, Beta(32, 9) ----
# Fall 2020 S280_QuickReview
c(mean = 32 / 41, median = qbeta(0.5, 32, 9), mode = 31 / 39)
c(lower_bound_90 = qbeta(0.1, 32, 9), upper_bound_90 = qbeta(0.9, 32, 9))

# ---- Sleep: discrete prior ----
p <- seq(0.05, 0.95, by = 0.1)
prior <- c(1, 5.2, 8, 7.2, 4.6, 2.1, 0.7, 0.1, 0, 0); prior <- prior / sum(prior)
post <- pdisc(p, prior, c(11, 16))
round(cbind(p, prior, post), 3)
sum(post[p >= 0.45])

# ---- Sleep: beta prior matched to two quantiles ----
ab <- beta.select(list(p = 0.5, x = 0.3), list(p = 0.9, x = 0.5)); ab
a <- ab[1] + 11; b <- ab[2] + 16
c(mean = a / (a + b), qbeta(c(0.05, 0.95), a, b), P_above_half = 1 - pbeta(0.5, a, b))
curve(dbeta(x, a, b), 0, 1, lwd = 2, ylab = "density", main = "Sleep: prior, likelihood, posterior")
curve(dbeta(x, ab[1], ab[2]), add = TRUE, lty = 3)
curve(dbeta(x, 12, 17), add = TRUE, lty = 2)

# ---- Sleep: histogram prior on a grid ----
midpt <- seq(0.05, 0.95, by = 0.1)
grid <- seq(0.0005, 0.9995, length = 1000)
hpost <- histprior(grid, midpt, prior) * dbeta(grid, 12, 17)
hpost <- hpost / sum(hpost)
c(mean = sum(grid * hpost), grid[findInterval(c(0.05, 0.95), cumsum(hpost)) + 1])
hist(sample(grid, 10000, replace = TRUE, prob = hpost), breaks = 30, main = "Histogram prior: posterior draws", xlab = "p")

# ---- Posterior predictive for 20 new students ----
pred <- pbetap(c(a, b), 20, 0:20)
round(pred, 3)
c(mean = sum(0:20 * pred), sd = sqrt(sum((0:20)^2 * pred) - sum(0:20 * pred)^2))
discint(cbind(0:20, pred), 0.9)
ysim <- rbinom(10000, 20, rbeta(10000, a, b))
table(ysim) / 10000
# Fall 2018 R04PostPredSleepData: the prior predictive for 20 students, before the data
round(rbind(discrete = pdiscp(p, prior, 20, 0:20), beta = pbetap(ab, 20, 0:20)), 3)

# ---- Premature births: central and HPD intervals ----
qbeta(c(0.05, 0.95), 32, 9)
hpd_draws(rbeta(100000, 32, 9))
c(mle = 31 / 39, mean = 32 / 41, post_sd = sqrt(32 * 9 / (41^2 * 42)), se = sqrt(31 / 39 * 8 / 39 / 39))

# ---- Zero events: 0 of 29 ----
qbeta(0.95, 1, 30)
c(score_upper = qnorm(0.975)^2 / (29 + qnorm(0.975)^2), exact_upper = 1 - 0.025^(1 / 29))

# ---- Race and promotion: two proportions ----
tb <- rbeta(10000, 27, 23); tw <- rbeta(10000, 207, 54)
c(P_white_higher = mean(tw > tb), mean_diff = mean(tw - tb), quantile(tw - tb, c(0.025, 0.975)),
  median_ratio = median(tb / tw))
sapply(list(c(1, 1), c(0.5, 0.5), c(10, 10), c(100, 100)), function(pr) (pr[1] + 26) / (sum(pr) + 48))
rbind(black = qbeta(c(0.05, 0.95), 27, 23), white = qbeta(c(0.05, 0.95), 207, 54))   # Fall 2018 R02RacePromotion

# ---- Perchlorate and thyroid tumours ----
fisher.test(matrix(c(2, 28, 0, 30), 2))$p.value
# Fall 2018 R03Perchlorate adds Dunson's Beta(0.11, 2.6), from earlier experiments
for (pr in list(c(1, 1), c(0.5, 0.5), c(1, 9), c(0.11, 2.6))) {
  tt <- rbeta(10000, pr[1] + 2, pr[2] + 28); tc <- rbeta(10000, pr[1], pr[2] + 30)
  cat(sprintf("prior Beta(%g, %g): P(T > C) = %.3f, 95%% interval for T - C (%.3f, %.3f)\n",
              pr[1], pr[2], mean(tt > tc), quantile(tt - tc, 0.025), quantile(tt - tc, 0.975)))
}

# ---- Exercises: ESP, penny, lead, Joe and Sam ----
pe <- seq(0, 1, by = 0.125)
pdisc(pe, c(0.001, 0.001, 0.950, rep(0.008, 6)), c(6, 4))
penny <- histprior(grid, midpt, c(0, 0, 1, 2, 7, 7, 2, 1, 0, 0) / 20) * dbeta(grid, 14, 8)
penny <- penny / sum(penny)
round(tapply(penny, cut(grid, seq(0, 1, 0.1)), sum), 3)
# Fall 2018 R05Exercises (c): the same prior after s heads in 20 spins, s = 1, ..., 20
pen <- sapply(1:20, function(s) {
  ps <- histprior(grid, midpt, c(0, 0, 1, 2, 7, 7, 2, 1, 0, 0) / 20) * dbeta(grid, s + 1, 21 - s)
  ps <- ps / sum(ps); c(s = s, mean = sum(grid * ps), grid[findInterval(c(0.05, 0.95), cumsum(ps)) + 1])
})
round(t(pen), 3)
c(qbeta(c(0.05, 0.95), 23, 8), P_above_0.6 = 1 - pbeta(0.6, 23, 8), P_9_or_10 = sum(pbetap(c(23, 8), 10, 9:10)))
round(rbind(Joe = pdiscp(c(0.1, 0.2, 0.3, 0.4, 0.5), c(0.5, 0.2, 0.2, 0.05, 0.05), 12, 0:12),
            Sam = pbetap(c(3, 12), 12, 0:12)), 3)
