tests/tinytest/test-WAIC.R

In texmex: Statistical Modelling of Extreme Values

m <- evm(log(ALT.M / ALT.B), data = liver, qu=.7, method = "sim")
aic <- AIC(m)

expect_equal(length(aic), 3, label = "WAIC: AIC returns AIC, DIC and WAIC by default")

aic <- AIC(m, DIC = FALSE)
expect_equal(names(aic), c("AIC", "WAIC"), info="WAIC: AIC returns AIC and WAIC when DIC=FALSE")

aic <- AIC(m, WAIC = FALSE)
expect_equal(names(aic), c("AIC", "DIC"), info="WAIC: AIC returns AIC and DIC when WAIC=FALSE")

aic <- AIC(m, DIC = FALSE, WAIC = FALSE)
expect_equal(names(aic), "AIC", label = "WAIC: AIC returns only AIC when DIC=FALSE, WAIC=FALSE")


set.seed(1234)
b <- function(){
  i <- sample(1:nrow(liver), size=nrow(liver), replace=TRUE)
  m <- evm(log(ALT.M / ALT.B), data = liver[i, ], qu=.7, eta = ~ as.numeric(dose),
           method = "sim", iter = 2000, thin = 1, family = cgpd)
  # Use cgpd to avoid xi < -0.5 whilst bootstrapping

  AIC(m)
}

aic <- t(replicate(100, b()))
expect_true(min(cor(aic)) > .98, info="WAIC: AIC, DIC and WAIC are well correlated: cgpd liver")

# We've analysed the liver data to death, so we "know" there is a dependence
# on dose
m0 <- evm(log(ALT.M / ALT.B), data = liver, qu=.7, method = "sim")
m1 <- evm(log(ALT.M / ALT.B), data = liver, qu=.7, xi = ~ as.numeric(dose), method = "sim")

expect_true(AIC(m1)["WAIC"] < AIC(m0)["WAIC"],
          info = "WAIC: 'correct' model selected for liver data: gpd")

b <- function(){
  i <- sample(1:nrow(portpirie), size=nrow(portpirie), replace=TRUE)
  m <- evm(SeaLevel, data = portpirie[i, ], family = gev,
           method = "sim", iter = 2000, thin = 1)
  AIC(m)
}

aic <- t(replicate(100, b()))
expect_true(min(cor(aic)) > .98, info="WAIC: AIC, DIC and WAIC are well correlated: gev portpirie")