metainf.R
#' Influence analysis in meta-analysis using leave-one-out method
#'
#' @description
#' Performs an influence analysis. Pooled estimates are calculated
#' omitting one study at a time.
#'
#' @param x An object of class \code{meta}.
#' @param pooled A character string indicating whether a fixed effect
#' or random effects model is used for pooling. Either missing (see
#' Details), \code{"fixed"} or \code{"random"}, can be abbreviated.
#' @param sortvar An optional vector used to sort the individual
#' studies (must be of same length as \code{x$TE}).
#'
#' @details
#' Performs a influence analysis; pooled estimates are calculated
#' omitting one study at a time. Studies are sorted according to
#' \code{sortvar}.
#'
#' Information from object \code{x} is utilised if argument
#' \code{pooled} is missing. A fixed effect model is assumed
#' (\code{pooled="fixed"}) if argument \code{x$fixed} is
#' \code{TRUE}; a random effects model is assumed
#' (\code{pooled="random"}) if argument \code{x$random} is
#' \code{TRUE} and \code{x$fixed} is \code{FALSE}.
#'
#' @return
#' An object of class \code{c("metainf", "meta")} with corresponding
#' \code{print}, and \code{forest} functions. The object is a list
#' containing the following components:
#' \item{TE, seTE}{Estimated treatment effect and standard error of
#' pooled estimate in influence analysis.}
#' \item{lower, upper}{Lower and upper confidence interval limits.}
#' \item{statistic}{Statistic for test of overall effect.}
#' \item{pval}{P-value for test of overall effect.}
#' \item{studlab}{Study label describing omission of studies.}
#' \item{w}{Sum of weights from fixed effect or random effects model.}
#' \item{I2}{Heterogeneity statistic I\eqn{^2}.}
#' \item{Rb}{Heterogeneity statistic R\eqn{_b}.}
#' \item{tau}{Square-root of between-study variance.}
#' \item{df.hakn}{Degrees of freedom for test of treatment effect for
#' Hartung-Knapp method (only if \code{hakn = TRUE}).}
#' \item{sm}{Summary measure.}
#' \item{method}{Method used for pooling.}
#' \item{k}{Number of studies combined in meta-analysis.}
#' \item{pooled}{As defined above.}
#' \item{fixed}{A logical indicating whether analysis is based on
#' fixed effect model.}
#' \item{random}{A logical indicating whether analysis is based
#' on random effects model.}
#' \item{TE.fixed, seTE.fixed}{Value is \code{NA}.}
#' \item{TE.random, seTE.random}{Value is \code{NA}.}
#' \item{Q}{Value is \code{NA}.}
#' \item{level.ma}{The level used to calculate confidence intervals
#' for pooled estimates.}
#' \item{hakn}{A logical indicating whether the method by Hartung and
#' Knapp is used to adjust test statistics and confidence
#' intervals.}
#' \item{adhoc.hakn}{A character string indicating whether \emph{ad
#' hoc} variance correction should be used for Hartung-Knapp
#' method.}
#' \item{method.tau}{A character string indicating which method is
#' used to estimate the between-study variance \eqn{\tau^2}.}
#' \item{tau.preset}{Prespecified value for the square root of the
#' between-study variance \eqn{\tau^2}.}
#' \item{TE.tau}{Overall treatment effect used to estimate the
#' between-study variance \eqn{\tau^2}.}
#' \item{n.harmonic.mean}{Harmonic mean of number of observations (for
#' back transformation of Freeman-Tukey Double arcsine
#' transformation).}
#' \item{version}{Version of R package \bold{meta} used to create
#' object.}
#'
#' @author Guido Schwarzer \email{sc@@imbi.uni-freiburg.de}
#'
#' @seealso \code{\link{metabin}}, \code{\link{metacont}},
#' \code{\link{print.meta}}
#'
#' @references
#' Cooper H & Hedges LV (1994):
#' \emph{The Handbook of Research Synthesis}.
#' Newbury Park, CA: Russell Sage Foundation
#'
#' @examples
#' data(Fleiss1993bin)
#' m1 <- metabin(d.asp, n.asp, d.plac, n.plac,
#' data = Fleiss1993bin, studlab = study,
#' sm = "RR", method = "I")
#' m1
#' metainf(m1)
#' metainf(m1, pooled = "random")
#'
#' forest(metainf(m1))
#' forest(metainf(m1), layout = "revman5")
#' forest(metainf(m1, pooled = "random"))
#'
#' metainf(m1, sortvar = study)
#' metainf(m1, sortvar = 7:1)
#'
#' m2 <- update(m1, title = "Fleiss1993bin meta-analysis",
#' backtransf = FALSE)
#' metainf(m2)
#'
#' data(Fleiss1993cont)
#' m3 <- metacont(n.psyc, mean.psyc, sd.psyc, n.cont, mean.cont, sd.cont,
#' data = Fleiss1993cont, sm = "SMD")
#' metainf(m3)
#'
#' @export metainf
metainf <- function(x, pooled, sortvar) {
##
##
## (1) Check for meta object and upgrade older meta objects
##
##
chkclass(x, "meta")
x <- updateversion(x)
##
k.all <- length(x$TE)
if (k.all < 2) {
warning("Nothing calculated (minimum number of studies: 2).")
return(invisible(NULL))
}
##
##
## (2) Check other arguments
##
##
if (!missing(pooled))
pooled <- setchar(pooled, c("fixed", "random"))
else
if (!x$fixed & x$random)
pooled <- "random"
else
pooled <- "fixed"
##
mf <- match.call()
error <- try(sortvar <- eval(mf[[match("sortvar", names(mf))]],
as.data.frame(x, stringsAsFactors = FALSE),
enclos = sys.frame(sys.parent())),
silent = TRUE)
if (class(error) == "try-error") {
xd <- x$data
sortvar <- eval(mf[[match("sortvar", names(mf))]],
xd, enclos = NULL)
if (isCol(x$data, ".subset"))
sortvar <- sortvar[x$data$.subset]
}
sort <- !is.null(sortvar)
if (sort && (length(sortvar) != k.all))
stop("Number of studies in object 'x' and argument 'sortvar' have ",
"different length.")
if (!sort)
sortvar <- 1:k.all
##
##
## (3) Sort variables
##
##
o <- order(sortvar)
##
n.e <- x$n.e[o]
n.c <- x$n.c[o]
n <- x$n[o]
##
event.e <- x$event.e[o]
event.c <- x$event.c[o]
event <- x$event[o]
##
mean.e <- x$mean.e[o]
mean.c <- x$mean.c[o]
mean <- x$mean[o]
##
sd.e <- x$sd.e[o]
sd.c <- x$sd.c[o]
sd <- x$sd[o]
##
time.e <- x$time.e[o]
time.c <- x$time.c[o]
time <- x$time[o]
##
cor <- x$cor[o]
##
TE <- x$TE[o]
seTE <- x$seTE[o]
##
if (length(x$incr) > 1)
incr <- x$incr[o]
else if (!is.null(x$incr))
incr <- rep_len(x$incr, k.all)
else
incr <- x$incr
##
studlab <- x$studlab[o]
slab <- c(paste("Omitting", studlab), "Pooled estimate")
studlab <- c(rev(rev(slab)[-1]), " ", rev(slab)[1])
##
## Exclude studies from meta-analysis
##
if (!is.null(x$exclude))
exclude <- x$exclude[o]
else
exclude <- rep_len(FALSE, k.all)
##
##
## (4) Do sensitivity analysis
##
##
res.i <- matrix(NA, ncol = 21, nrow = k.all)
add.i <- matrix(NA, ncol = 3, nrow = k.all)
##
for (i in 1:k.all) {
sel <- -i
##
if (length(incr) > 1)
incr.i <- incr[sel]
else
incr.i <- incr
##
if (inherits(x, "metabin"))
m <- metabin(event.e[sel], n.e[sel], event.c[sel], n.c[sel],
##
exclude = exclude[sel],
##
method = x$method, sm = x$sm,
incr = incr.i, allincr = x$allincr, addincr = x$addincr,
allstudies = x$allstudies, MH.exact = x$MH.exact,
RR.Cochrane = x$RR.Cochrane, Q.Cochrane = x$Q.Cochrane,
model.glmm =
if (!is.null(x$model.glmm)) x$model.glmm else "UM.FS",
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
if (inherits(x, "metacont"))
m <- metacont(n.e[sel], mean.e[sel], sd.e[sel],
n.c[sel], mean.c[sel], sd.c[sel],
##
exclude = exclude[sel],
##
sm = x$sm, pooledvar = x$pooledvar,
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
if (inherits(x, "metacor"))
m <- metacor(cor[sel], n[sel],
##
exclude = exclude[sel],
##
sm = x$sm, null.effect = x$null.effect,
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
##
control = x$control)
##
if (inherits(x, "metagen"))
m <- metagen(TE[sel], seTE[sel],
##
exclude = exclude[sel],
##
sm = x$sm, null.effect = x$null.effect,
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
if (inherits(x,"metainc"))
m <- metainc(event.e[sel], time.e[sel],
event.c[sel], time.c[sel],
##
exclude = exclude[sel],
##
method = x$method, sm = x$sm,
incr = incr.i, allincr = x$allincr, addincr = x$addincr,
model.glmm =
if (!is.null(x$model.glmm)) x$model.glmm else "UM.FS",
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
if (inherits(x, "metamean"))
m <- metamean(n[sel], mean[sel], sd[sel],
##
exclude = exclude[sel],
##
sm = x$sm, null.effect = x$null.effect,
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
if (inherits(x, "metaprop"))
m <- metaprop(event[sel], n[sel],
##
exclude = exclude[sel],
##
method = x$method, sm = x$sm, null.effect = x$null.effect,
##
incr = incr.i, allincr = x$allincr, addincr = x$addincr,
method.ci = x$method.ci,
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
if (inherits(x, "metarate"))
m <- metarate(event[sel], time[sel],
##
exclude = exclude[sel],
##
method = x$method, sm = x$sm, null.effect = x$null.effect,
##
incr = incr.i, allincr = x$allincr, addincr = x$addincr,
##
level.ma = x$level.ma,
##
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset, TE.tau = x$TE.tau,
##
keepdata = FALSE,
warn = FALSE,
##
control = x$control)
##
sel.pft <- inherits(x, "metaprop") & x$sm == "PFT"
sel.irft <- inherits(x, "metarate") & x$sm == "IRFT"
##
add.i[i, ] <- c(m$method.tau.ci, # 1
m$sign.lower.tau, # 2
m$sign.upper.tau # 3
)
##
if (pooled == "fixed") {
res.i[i, ] <- c(m$TE.fixed, # 1
m$seTE.fixed, # 2
m$lower.fixed, # 3
m$upper.fixed, # 4
m$statistic.fixed, # 5
m$pval.fixed, # 6
m$tau2, # 7
m$lower.tau2, # 8
m$upper.tau2, # 9
m$se.tau2, # 10
m$tau, # 11
m$lower.tau, # 12
m$upper.tau, # 13
m$I2, # 14
m$lower.I2, # 15
m$upper.I2, # 16
sum(m$w.fixed, na.rm = TRUE), # 17
if (sel.pft) 1 / mean(1 / n[sel]) else NA, # 18
NA, # 19
if (sel.irft) 1 / mean(1 / time[sel]) else NA, # 20
m$Rb # 21
)
}
##
else if (pooled == "random") {
res.i[i, ] <- c(m$TE.random, # 1
m$seTE.random, # 2
m$lower.random, # 3
m$upper.random, # 4
m$statistic.random, # 5
m$pval.random, # 6
m$tau2, # 7
m$lower.tau2, # 8
m$upper.tau2, # 9
m$se.tau2, # 10
m$tau, # 11
m$lower.tau, # 12
m$upper.tau, # 13
m$I2, # 14
m$lower.I2, # 15
m$upper.I2, # 16
sum(m$w.random, na.rm = TRUE), # 17
if (sel.pft) 1 / mean(1 / n[sel]) else NA, # 18
if (x$hakn) m$df.hakn else NA, # 19
if (sel.irft) 1 / mean(1 / time[sel]) else NA, # 20
m$Rb # 21
)
}
}
##
TE.i <- res.i[, 1]
seTE.i <- res.i[, 2]
lower.i <- res.i[, 3]
upper.i <- res.i[, 4]
statistic.i <- res.i[, 5]
pval.i <- res.i[, 6]
##
tau2.i <- res.i[, 7]
lower.tau2.i <- res.i[, 8]
upper.tau2.i <- res.i[, 9]
se.tau2.i <- res.i[, 10]
##
tau.i <- res.i[, 11]
lower.tau.i <- res.i[, 12]
upper.tau.i <- res.i[, 13]
##
I2.i <- res.i[, 14]
lower.I2.i <- res.i[, 15]
upper.I2.i <- res.i[, 16]
##
weight.i <- res.i[, 17]
n.harmonic.mean.i <- res.i[, 18]
if (pooled == "random" & x$hakn)
df.hakn.i <- res.i[, 19]
t.harmonic.mean.i <- res.i[, 20]
Rb.i <- res.i[, 21]
##
method.tau.ci <- unique(add.i[, 1])
sign.lower.tau.i <- add.i[, 2]
sign.upper.tau.i <- add.i[, 3]
##
if (pooled == "fixed") {
TE.s <- x$TE.fixed
seTE.s <- x$seTE.fixed
lower.TE.s <- x$lower.fixed
upper.TE.s <- x$upper.fixed
statistic.s <- x$statistic.fixed
pval.s <- x$pval.fixed
w.s <- sum(x$w.fixed, na.rm = TRUE)
}
##
else if (pooled == "random") {
TE.s <- x$TE.random
seTE.s <- x$seTE.random
lower.TE.s <- x$lower.random
upper.TE.s <- x$upper.random
statistic.s <- x$statistic.random
pval.s <- x$pval.random
w.s <- sum(x$w.random, na.rm = TRUE)
}
##
##
## (5) Generate R object
##
##
res <- list(TE = c(TE.i, NA, TE.s),
seTE = c(seTE.i, NA, seTE.s),
lower = c(lower.i, NA, lower.TE.s),
upper = c(upper.i, NA, upper.TE.s),
statistic = c(statistic.i, NA, statistic.s),
pval = c(pval.i, NA, pval.s),
studlab = studlab,
##
tau2 = c(tau2.i, NA, x$tau2),
lower.tau2 = c(lower.tau2.i, NA, x$lower.tau2),
upper.tau2 = c(upper.tau2.i, NA, x$upper.tau2),
se.tau2 = c(se.tau2.i, NA, x$se.tau2),
##
tau = c(tau.i, NA, x$tau),
lower.tau = c(lower.tau.i, NA, x$lower.tau),
upper.tau = c(upper.tau.i, NA, x$upper.tau),
##
method.tau.ci = method.tau.ci,
sign.lower.tau.i = c(sign.lower.tau.i, NA, x$sign.lower.tau),
sign.upper.tau.i = c(sign.upper.tau.i, NA, x$sign.upper.tau),
##
I2 = c(I2.i, NA, x$I2),
lower.I2 = c(lower.I2.i, NA, x$lower.I2),
upper.I2 = c(upper.I2.i, NA, x$upper.I2),
##
Rb = c(Rb.i, NA, x$Rb),
##
w = c(weight.i, NA, w.s),
df.hakn =
if (pooled == "random" & x$hakn)
c(df.hakn.i, NA, x$df.hakn) else NULL,
##
sm = x$sm, method = x$method, k = x$k,
pooled = pooled,
fixed = ifelse(pooled == "fixed", TRUE, FALSE),
random = ifelse(pooled == "random", TRUE, FALSE),
TE.fixed = NA, seTE.fixed = NA,
TE.random = NA, seTE.random = NA,
null.effect = x$null.effect,
##
Q = NA,
level.ma = x$level.ma,
hakn = x$hakn, adhoc.hakn = x$adhoc.hakn,
method.tau = x$method.tau,
tau.preset = x$tau.preset,
TE.tau = x$TE.tau,
n.harmonic.mean = c(n.harmonic.mean.i, NA, 1 / mean(1 / n)),
t.harmonic.mean = c(t.harmonic.mean.i, NA, 1 / mean(1 / time)),
prediction = FALSE,
##
backtransf = x$backtransf,
pscale = x$pscale,
irscale = x$irscale, irunit = x$irunit,
##
text.fixed = x$text.fixed, text.random = x$text.random,
text.predict = x$text.predict,
text.w.fixed = x$text.w.fixed, text.w.random = x$text.w.random,
##
title = x$title, complab = x$complab,
outclab = x$outclab,
##
x = x,
##
call = match.call())
res$version <- packageDescription("meta")$Version
##
res$x$fixed <- res$fixed
res$x$random <- res$random
class(res) <- c("metainf", "summary.meta", "meta")
##
if (inherits(x, "trimfill"))
class(res) <- c(class(res), "trimfill")
res
}
