R/standardErrorLME.est.R

In smicd: Statistical Methods for Interval-Censored Data

# Internal documentation -------------------------------------------------------
#
# Estimate standard errors

standardErrorLME.est <- function(formula, data, classes, burnin, samples, trafo,
                                 adjust, b, coef, sigmae, VaCov, nameRI, nameRS,
                                 regmodell, lambda) {

  # Fixed effect prediction to find mu
  boot_data <- data
  fixedformula <- as.formula(lme4::nobars(formula(regmodell))[-2])
  fixedeffect <- model.matrix(fixedformula, data = boot_data)
  mu <- as.matrix(fixedeffect) %*% coef

  # Get names from FE
  name.fe <- names(coef)
  # Get name from random intercept
  name.RI <- nameRI
  # Get name from random  slope
  name.RS <- nameRS
  # Get name of dependent variable from formula
  name.dep <- all.vars(formula)[1]

  coef <- NULL
  pb <- txtProgressBar(min = 1, max = b, style = 3)
  print("Bootstrapping...")
  for (i in 1:b) {
    p.ranef <- NULL
    for (j in 1:length(unique(boot_data[, tail(all.vars(formula), n = 1)]))) {
      p.ranef <- rbind(p.ranef, apply(
        rmvnorm(
          n = 1, mean = rep(0, nrow(VaCov)),
          sigma = VaCov
        ), 2,
        function(x) {
          rep(x, table(boot_data[, tail(all.vars(formula), n = 1)])[j])
        }
      ))
    }

    if (is.na(nameRS)) {
      boot_data$ynew <- mu + p.ranef[, 1] + rnorm(nrow(boot_data), 0, sigmae)
    }
    if (!is.na(nameRS)) {
      boot_data$ynew <- mu + p.ranef[, 1] + p.ranef[, 2] * boot_data[, name.RS] +
        rnorm(nrow(boot_data), 0, sigmae)
    }

    # Transform data back
    if (trafo == "log") {
      boot_data$ynew <- exp(boot_data$ynew)
    }
    if (trafo == "bc") {
      rueck <- boxcox.lme.est(dat = boot_data$ynew, lambda = lambda, inverse = T)
      boot_data$ynew <- rueck[[1]]
    }

    boot_data$ynew[boot_data$ynew < classes[1]] <- classes[1] + .Machine$double.eps
    boot_data$ynew[boot_data$ynew > classes[length(classes)]] <-
      classes[length(classes)] + .Machine$double.eps


    assign(paste0("boot_data$", name.dep), cut(boot_data$ynew, classes))

    capture.output(SEM <- semLme(
      formula = formula, data = boot_data,
      classes = classes, burnin = burnin, samples = samples,
      trafo = trafo, adjust = adjust, bootstrap.se = FALSE
    ))

    coef <- rbind(coef, SEM$coef)
    setTxtProgressBar(pb, i)
  }

  se.boot <- apply(coef, 2, function(x) {
    sqrt(mean((x - mean(x))^2))
  })

  ci <- apply(coef, 2, function(x) {
    x <- sort(x)
    ci.lower <- ifelse(round(length(x) * .025) < 1, x[1], x[round(length(x) * .025)])
    ci.upper <- ifelse(round(length(x) * .975) > length(x), x[length(x)],
      x[round(length(x) * .975)]
    )
    return(list(ci.low = ci.lower, ci.up = ci.upper))
  })

  ci.final <- NULL
  for (i in 1:length(ci)) {
    ci.final <- rbind(ci.final, c(ci[[i]]$ci.low, ci[[i]]$ci.up))
  }

  est <- list(se = se.boot, ci = ci.final)
  return(est)
}