R/RLRTSim.R

In RLRsim: Exact (Restricted) Likelihood Ratio Tests for Mixed and Additive Models

#' @export RLRTSim
#' @import Rcpp
#' @importFrom stats rchisq
RLRTSim <- function(X, Z, qrX=qr(X), sqrt.Sigma, lambda0 = NA,
  seed = NA, nsim = 10000, use.approx = 0, log.grid.hi = 8,
  log.grid.lo = -10, gridlength = 200,
  parallel = c("no", "multicore", "snow"),
  ncpus = 1L, cl = NULL) {
  if (is.na(lambda0)) {
    lambda0 <- 0
  }
  #checking args:
  if (!is.numeric(lambda0) | (lambda0 < 0) | length(lambda0) != 1) {
    stop("Invalid lambda0 specified. \n")
  }
  if (lambda0 > exp(log.grid.hi)) {
    log.grid.hi <- log(10 * lambda0)
    warning(paste0("lambda0 smaller than upper end of grid: \n",
      "Setting log.grid.hi to ln(10*lambda0).\n"),
      immediate. = TRUE)
  }
  if ((lambda0 != 0) && (lambda0 < exp(log.grid.lo))) {
    log.grid.lo <- log(-10 * lambda0)
    warning(paste0("lambda0 > 0 and larger than lower end of grid: \n",
      "Setting log.grid.lo to ln(-10*lambda0).\n"),
      immediate. = TRUE)
  }

  parallel <- match.arg(parallel)
  have_mc <- have_snow <- FALSE
  if (parallel != "no" && ncpus > 1L) {
    if (parallel == "multicore")
      have_mc <- .Platform$OS.type != "windows"
    else if (parallel == "snow")
      have_snow <- TRUE
    if (!have_mc && !have_snow)
      ncpus <- 1L
  }

  n <- NROW(X)
  p <- NCOL(X)
  K <- min(n, NCOL(Z))
  if (any(is.na(sqrt.Sigma)))
    sqrt.Sigma <- diag(NCOL(Z))
  mu <- (svd(sqrt.Sigma %*% t(qr.resid(qrX, Z)), nu = 0,
    nv = 0)$d)^2
  #normalize
  mu <- mu/max(mu)
  if (!is.na(seed))
    set.seed(seed)
  if (use.approx) {
    #eigenvalue pattern of balanced ANOVA: mu_s=const for s=1,..,K-1, mu_K approx. 0
    if ((length(unique(round(mu, 6))) == 2) & (1000 * mu[K] < mu[1])) {
      message("using simplified distribution for balanced ANOVA \n")
      approx.constantmu <- function(nsim, n, p, K, mu)
        #simplified distribution for balanced ANOVA:
        #mu_s=const for s=1,..,K-1 and mu_K=0
      {
        w.K <- rchisq(nsim, (K - 1))
        w.n <- rchisq(nsim, (n - p - K + 1))
        lambda <- pmax(rep(0, nsim),
          ((((n - p - K +1) / (K - 1)) * w.K/w.n - 1)/mu[1]))
        rlrt <- rep(0, nsim)
        rlrt[lambda != 0] = ((n - p) * log((w.K + w.n)/(n -p)) -
            (n - p - K + 1) *
            log(w.n/(n - p - K + 1)) - (K - 1) *
            log(w.K/(K - 1)))[lambda != 0]
        return(cbind(lambda, rlrt))
      }
      res <- approx.constantmu(nsim, n, p, K, mu)
      return(res)
    }
    #eigenvalue pattern for P-splines: exponential decrease
    if (mu[1]/sum(mu) > use.approx) {
      message("using simplified distribution for 1 single dominating eigenvalue \n")
      approx.scalarmu <- function(nsim, n, p, K, mu)
        #simplified distribution for B-splines:
        #mu_1 >>> mu_s for s=2,..,K
      {
        mu <- mu[1]
        w.1 <- rchisq(nsim, 1)
        w.n <- rchisq(nsim, (n - p - 1))
        lambda <- pmax(rep(0, nsim), ((((n - p - 1) * w.1)/w.n) - 1)/mu)
        rlrt <- rep(0, nsim)
        rlrt[lambda != 0] <- log(((w.1 + w.n)/(n - p))^(n -p) /
            (w.1 *(w.n/(n - p - 1)) ^
                (n - p - 1)))[lambda != 0]
        return(cbind(lambda, rlrt))
      }
      res <- approx.scalarmu(nsim, n, p, K, mu)
      return(res)
    }
    #use only first k elements of mu, adapt K<-k accordingly
    #how many eigenvalues are needed to represent at least approx.ratio of
    #the sum of all eigenvalues  (at least 1, of course)
    new.K <- max(sum((cumsum(mu)/sum(mu)) < use.approx),
      1)
    if (new.K < K)
      message(paste("Approximation used:", new.K,
        "biggest eigenvalues instead of", K, "\n"))
    mu <- mu[1:new.K]
    K <- new.K
  }
  #generate symmetric grid around lambda0 that is log-equidistant to the right,
  make.lambdagrid <- function(lambda0, gridlength, log.grid.lo, log.grid.hi) {
    
    # return(c(0, exp(seq(log.grid.lo, log.grid.hi,
    #  length = gridlength - 1))))
    
    if (lambda0 == 0)
      return(c(0, exp(seq(log.grid.lo, log.grid.hi,
        length = gridlength - 1))))
    else {
      leftratio <- min(max((log(lambda0)/((log.grid.hi) - (log.grid.lo))),
        0.2), 0.8)
      leftlength <- max(round(leftratio * gridlength) - 1, 2)
      leftdistance <- lambda0 - exp(log.grid.lo)
      #make sure leftlength doesn't split the left side into too small parts:
      if (leftdistance < (leftlength * 10 * .Machine$double.eps)) {
        leftlength <- max(round(leftdistance/(10 * .Machine$double.eps)), 2)
      }
      #leftdistance approx. 1 ==> make a regular grid, since
      # (1 +- epsilon)^((1:n)/n) makes a too concentrated grid
      if (abs(leftdistance - 1) < 0.3) {
        leftgrid <- seq(exp(log.grid.lo), lambda0,
          length = leftlength + 1)[-(leftlength + 1)]
      }
      else {
        leftdiffs <- ifelse(rep(leftdistance > 1, leftlength - 1),
          leftdistance^((2:leftlength)/leftlength) -
            leftdistance^(1/leftlength),
          leftdistance^((leftlength - 1):1) -
            leftdistance^(leftlength))
        leftgrid <- lambda0 - rev(leftdiffs)
      }
      rightlength <- gridlength - leftlength
      rightdistance <- exp(log.grid.hi) - lambda0
      rightdiffs <- rightdistance^((2:rightlength)/rightlength) -
        rightdistance^(1/rightlength)
      rightgrid <- lambda0 + rightdiffs
      return(c(0, leftgrid, lambda0, rightgrid))
    }
  }
  lambda.grid <- make.lambdagrid(lambda0, gridlength, log.grid.lo = log.grid.lo,
    log.grid.hi = log.grid.hi)

  res <- if (ncpus > 1L && (have_mc || have_snow)) {
    nsim. <- as.integer(ceiling(nsim/ncpus))
    if (have_mc) {
      tmp <- parallel::mclapply(seq_len(ncpus), function(i){
        RLRsimCpp(p = as.integer(p), k = as.integer(K),
          n = as.integer(n), nsim = as.integer(nsim.),
          g = as.integer(gridlength),
          q = as.integer(0), mu = as.double(mu),
          lambda = as.double(lambda.grid),
          lambda0 = as.double(lambda0), xi = as.double(mu),
          REML = as.logical(TRUE))
      }, mc.cores = ncpus)
      do.call(mapply, c(tmp, FUN=c))
    } else {
      if (have_snow) {
        if (is.null(cl)) {
          cl <- parallel::makePSOCKcluster(rep("localhost",
            ncpus))
          if (RNGkind()[1L] == "L'Ecuyer-CMRG") {
            parallel::clusterSetRNGStream(cl)
          }
          tmp <- parallel::parLapply(cl, seq_len(ncpus), function(i){
            RLRsimCpp(p = as.integer(p), k = as.integer(K),
              n = as.integer(n), nsim = as.integer(nsim.),
              g = as.integer(gridlength),
              q = as.integer(0), mu = as.double(mu),
              lambda = as.double(lambda.grid),
              lambda0 = as.double(lambda0), xi = as.double(mu),
              REML = as.logical(TRUE))
          })
          parallel::stopCluster(cl)
          do.call(mapply, c(tmp, FUN=c))
        } else {
          tmp <- parallel::parLapply(cl, seq_len(ncpus), function(i){
            RLRsimCpp(p = as.integer(p), k = as.integer(K),
              n = as.integer(n), nsim = as.integer(nsim.),
              g = as.integer(gridlength),
              q = as.integer(0), mu = as.double(mu),
              lambda = as.double(lambda.grid),
              lambda0 = as.double(lambda0), xi = as.double(mu),
              REML = as.logical(TRUE))
          })
          do.call(mapply, c(tmp, FUN=c))
        }
      }
    }
  } else {
    RLRsimCpp(p = as.integer(p), k = as.integer(K),
      n = as.integer(n), nsim = as.integer(nsim),
      g = as.integer(gridlength),
      q = as.integer(0), mu = as.double(mu),
      lambda = as.double(lambda.grid),
      lambda0 = as.double(lambda0), xi = as.double(mu),
      REML = as.logical(TRUE))
  }

  ret <- res$res
  attr(ret, "lambda") <- lambda.grid[res$lambdaind]
  return(ret)
}