R/summary.mitml.R

Defines functions .reducedACF summary.mitml

Documented in summary.mitml

summary.mitml <- function(object, n.Rhat = 3, goodness.of.appr = FALSE,
                          autocorrelation = FALSE, ...){

# summary method for objects of class "mitml"

  inc <- object$data
  ngr <- length(unique(attr(object$data, "group")))
  prm <- object$par.imputation
  iter <- dim(prm[[1]])[3]
  k <- object$iter$iter
  isML <- attr(object$model, "is.ML")
  isL2 <- attr(object$model, "is.L2")

  # parameter chains (for backwards compatibility)
  if(is.null(object$keep.chains)) object$keep.chains <- "full"

  # percent missing
  mdr <- sapply(inc, FUN = function(x){mean(is.na(x))})
  mdr[] <- sprintf(mdr*100, fmt = "%.1f")
  mdr <- gsub("^0.0$", "0", mdr)

  # convergence for imputation phase
  conv <- NULL
  Rhat <- ifelse(is.null(n.Rhat), FALSE, n.Rhat >= 2)

  SDprop <- goodness.of.appr
  ACF <- autocorrelation
  if(Rhat|SDprop|ACF){

    conv <- c(list(beta = NULL), if(isL2) list(beta2 = NULL), if(isML) list(psi = NULL),
              list(sigma = NULL))
    for(pp in names(conv)){

      ni <- dim(prm[[pp]])[1]
      nj <- dim(prm[[pp]])[2]
      nl <- dim(prm[[pp]])[4]
      cmat <- matrix(NA_real_, ni*nj*nl, 3+Rhat+SDprop+3*ACF)
      cmat[,1] <- rep(1:ni, nj*nl)
      cmat[,2] <- rep(1:nj, each = ni, times = nl)
      cmat[,3] <- rep(1:nl, each = ni*nj)
      colnames(cmat) <- c("i1", "i2", "grp", if(Rhat) "Rhat", if(SDprop) "SDprop",
                          if(ACF) c("lag-1", "lag-k", "lag-2k"))

      for(ll in 1:nl){ # by group

        for(jj in 1:nj){
          for(ii in 1:ni){

            # check for redundant entries
            if(pp == "psi"){
              if(jj > ii) next
            }
            if(pp == "sigma"){
              if(jj > ((ii-1)%%nj)+1) next
            }
            ind <- ( cmat[,1] == ii & cmat[,2] == jj & cmat[,3] == ll )
            chn <- matrix(prm[[pp]][ii,jj,,ll], 1, iter)
            # potential scale reduction (Rhat)
            if(Rhat) cmat[ind, "Rhat"] <- .GelmanRubin(chn, n.Rhat)
            # goodness of approximation
            if(SDprop) cmat[ind, "SDprop"] <- .SDprop(chn)
            # autocorrelation
            if(ACF){
              cmat[ind, "lag-1"] <- .reducedACF(chn, lag = 1, smooth = 0)
              cmat[ind, "lag-k"] <- .reducedACF(chn, lag = k, smooth = 2, sd=.5)
              cmat[ind, "lag-2k"] <- .reducedACF(chn, lag = 2*k, smooth = 2, sd=.5)
            }
          }
        }
      }
      conv[[pp]] <- cmat[!apply(cmat, 1, function(x) any(is.na(x))),,drop = F]
    }

  attr(conv, "stats") <- c("Rhat", "SDprop", "ACF")[c(Rhat, SDprop, ACF)]
  }

  smr <- list(
    call = object$call,
    model = object$model,
    prior = object$prior,
    iter = object$iter,
    keep.chains = object$keep.chains,
    ngr = ngr,
    missing.rates = mdr,
    conv = conv
  )

  class(smr) <- "mitml.summary"
  return(smr)

}

.reducedACF <- function(x, lag, smooth = 0, sd=.5){

  # check NA
  if(all(is.na(x))) return(NA)

  n <- length(x)
  lag0 <- lag
  lag <- lag + (-smooth:smooth)

  ac <- numeric(length(lag))
  y <- x - mean(x)
  ss.y <- sum(y^2)

  for(li in 1:length(lag)){
    ll <- lag[li]
    # leave at 0 for constant value
    ac[li] <- if(ss.y>0) sum( y[1:(n-ll)] * y[1:(n-ll)+ll] ) / ss.y else 0
  }

  if(smooth>0){
    # weights based on normal density
    w <- dnorm(-smooth:smooth, 0, sd)
    ac <- sum( ac * (w/sum(w)) )
  }

  ac

}

Try the mitml package in your browser

Any scripts or data that you put into this service are public.

mitml documentation built on March 31, 2023, 7:01 p.m.