R/csem_resample.R
In M-E-Steiner/cSEM: Composite-Based Structural Equation Modeling
Defines functions
resamplecSEMResultsCore
resamplecSEMResults
resampleData
Documented in
resamplecSEMResults
resampleData
#' Resample data
#'
#' Resample data from a data set using common resampling methods.
#' For bootstrap or jackknife resampling, package users usually do not need to
#' call this function but directly use [resamplecSEMResults()] instead.
#'
#' The function `resampleData()` is general purpose. It simply resamples data
#' from a data set according to the resampling method provided
#' via the `.resample_method` argument and returns a list of resamples.
#' Currently, `bootstrap`, `jackknife`, `permutation`, and `cross-validation`
#' (both leave-one-out (LOOCV) and k-fold cross-validation) are implemented.
#'
#' The user may provide the data set to resample either explicitly via the `.data`
#' argument or implicitly by providing a [cSEMResults] objects to `.object`
#' in which case the original data used in the call that created the
#' [cSEMResults] object is used for resampling.
#' If both, a [cSEMResults] object and a data set via `.data` are provided
#' the former is ignored.
#'
#' As [csem()] accepts a single data set, a list of data sets as well as data sets
#' that contain a column name used to split the data into groups,
#' the [cSEMResults] object may contain multiple data sets.
#' In this case, resampling is done by data set or group. Note that depending
#' on the number of data sets/groups provided this computation may be slower
#' as resampling will be repeated for each data set/group.
#'
#' To split data provided via the `.data` argument into groups, the column name or
#' the column index of the column containing the group levels to split the data
#' must be given to `.id`. If data that contains grouping is taken from
#' a [cSEMResults] object, `.id` is taken from the object information. Hence,
#' providing `.id` is redundant in this case and therefore ignored.
#'
#' The number of bootstrap or permutation runs as well as the number of
#' cross-validation repetitions is given by `.R`. The default is
#' `499` but should be increased in real applications. See e.g.,
#' \insertCite{Hesterberg2015;textual}{cSEM}, p.380 for recommendations concerning
#' the bootstrap. For jackknife `.R` is ignored as it is based on the N leave-one-out data sets.
#'
#' Choosing `resample_method = "permutation"` for ungrouped data causes an error
#' as permutation will simply reorder the observations which is usually not
#' meaningful. If a list of data is provided
#' each list element is assumed to represent the observations belonging to one
#' group. In this case, data is pooled and group adherence permuted.
#'
#' For cross-validation the number of folds (`k`) defaults to `10`. It may be
#' changed via the `.cv_folds` argument. Setting `k = 2` (not 1!) splits
#' the data into a single training and test data set. Setting `k = N` (where `N` is the
#' number of observations) produces leave-one-out cross-validation samples.
#' Note: 1.) At least 2 folds required (`k > 1`); 2.) `k` can not be larger than `N`;
#' 3.) If `N/k` is not not an integer the last fold will have less observations.
#'
#' Random number generation (RNG) uses the L'Ecuyer-CRMR RGN stream as implemented
#' in the \href{https://github.com/HenrikBengtsson/future.apply/}{future.apply package}
#' \insertCite{Bengtsson2018a}{cSEM}.
#' See [?future_lapply][future.apply::future_lapply] for details. By default
#' a random seed is chosen.
#'
#' @usage resampleData(
#' .object = NULL,
#' .data = NULL,
#' .resample_method = c("bootstrap", "jackknife", "permutation",
#' "cross-validation"),
#' .cv_folds = 10,
#' .id = NULL,
#' .R = 499,
#' .seed = NULL
#' )
#'
#' @param .data A `data.frame`, a `matrix` or a `list` of data of either type.
#' Possible column types or classes of the data provided are:
#' "`logical`", "`numeric`" ("`double`" or "`integer`"), "`factor`" (ordered and unordered)
#' or a mix of several types. The data may also include
#' **one** character column whose column name must be given to `.id`.
#' This column is assumed to contain group identifiers used to split the data into groups.
#' If `.data` is provided, `.object` is ignored. Defaults to `NULL`.
#' @param .resample_method Character string. The resampling method to use. One of:
#' "*bootstrap*", "*jackknife*", "*permutation*", or "*cross-validation*".
#' Defaults to "*bootstrap*".
#' @param .R Integer. The number of bootstrap runs, permutation runs
#' or cross-validation repetitions to use. Defaults to `499`.
#' @inheritParams csem_arguments
#'
#' @return The structure of the output depends on the type of input and the
#' resampling method:
#' \describe{
#' \item{Bootstrap}{If a `matrix` or `data.frame` without grouping variable
#' is provided (i.e., `.id = NULL`), the result is a list of length `.R`
#' (default `499`). Each element of that list is a bootstrap (re)sample.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data for one group),
#' resampling is done by group. Hence,
#' the result is a list of length equal to the number of groups
#' with each list element containing `.R` bootstrap samples based on the
#' `N_g` observations of group `g`.}
#' \item{Jackknife}{If a `matrix` or `data.frame` without grouping variable
#' is provided (`.id = NULL`), the result is a list of length equal to the number
#' of observations/rows (`N`) of the data set provided.
#' Each element of that list is a jackknife (re)sample.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data for one group),
#' resampling is done by group. Hence,
#' the result is a list of length equal to the number of group levels
#' with each list element containing `N` jackknife samples based on the
#' `N_g` observations of group `g`.}
#' \item{Permutation}{If a `matrix` or `data.frame` without grouping variable
#' is provided an error is returned as permutation will simply reorder the observations.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data of one group),
#' group membership is permuted. Hence, the result is a list of length `.R`
#' where each element of that list is a permutation (re)sample.}
#' \item{Cross-validation}{If a `matrix` or `data.frame` without grouping variable
#' is provided a list of length `.R` is returned. Each list element
#' contains a list containing the `k` splits/folds subsequently
#' used as test and training data sets.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data for one group),
#' cross-validation is repeated `.R` times for each group. Hence,
#' the result is a list of length equal to the number of groups,
#' each containing `.R` list elements (the repetitions) which in turn contain
#' the `k` splits/folds.
#' }
#' }
#' @references
#' \insertAllCited{}
#'
#' @seealso [csem()], [cSEMResults], [resamplecSEMResults()]
#'
#' @example inst/examples/example_resampleData.R
#'
#' @export
#'
resampleData <- function(
.object = NULL,
.data = NULL,
.resample_method = c("bootstrap", "jackknife", "permutation",
"cross-validation"),
.cv_folds = 10,
.id = NULL,
.R = 499,
.seed = NULL
) {
.resample_method <- match.arg(.resample_method,
c("bootstrap", "jackknife", "permutation", "cross-validation"))
## Set plan on how to resolve futures to "sequential" as it is virtually always
## slower to resample data using "multiprocess".; reset at the end
oplan <- future::plan()
on.exit(future::plan(oplan), add = TRUE)
future::plan("sequential")
## Get data set
if(is.null(.data)) {
## Get information according to class of object
if(any(class(.object) %in% "cSEMResults_default")) {
data <- as.data.frame(.object$Information$Arguments$.data)
id <- NULL
} else if(any(class(.object) %in% "cSEMResults_multi")) {
data <- .object[[1]]$Information$Data_pooled
data_split <- lapply(.object, function(x) x$Information$Arguments$.data)
id <- ifelse(is.null(.object[[1]]$Information$Arguments$.id),
"id", .object[[1]]$Information$Arguments$.id)
} else if(any(class(.object) %in% "cSEMResults_2ndorder")) {
data <- .object$First_stage$Information$Arguments$.data
id <- NULL
} else {
stop2("The following error occured in the `resampleData()` function:\n",
"`object` must be of class cSEMResults."
)
}
} else {
## Check data
if(!any(class(.data) %in% c("data.frame", "matrix", "list"))) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"Data must be provided as a `matrix`, a `data.frame` or a `list`. ",
".data has class: ", class(.data)
)
}
## Select cases
if(!is.null(.id) && !inherits(.data, "list")) {
if(length(.id) != 1) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"`.id` must be a character string or an integer identifying one single column."
)
}
if(is.matrix(.data)) {
.data <- as.data.frame(.data)
}
data <- .data
data_split <- split(.data, f = .data[, .id])
id <- .id
} else if(any(class(.data) == "list")) {
data <- do.call(rbind, .data)
if(is.null(names(.data))) {
data[, "id"] <- rep(paste0("Data_", 1:length(.data)),
times = sapply(.data, nrow))
} else {
data[, "id"] <- rep(names(.data), times = sapply(.data, nrow))
}
data_split <- .data
id <- "id"
##add names here
} else {
data <- .data
id <- NULL
}
}
# Save old seed and restore on exit! This is important since users may have
# set a seed before, in which case the global seed would be
# overwritten if not explicitly restored
old_seed <- .Random.seed
on.exit({.Random.seed <<- old_seed})
## Create seed if not already set
if(is.null(.seed)) {
set.seed(seed = NULL)
# Note (08.12.2019): Its crucial to call set.seed(seed = NULL) before
# drawing a random seed out of .Random.seed. If set.seed(seed = NULL) is not
# called sample(.Random.seed, 1) would result in the same random seed as
# long as .Random.seed remains unchanged. By resetting the seed we make
# sure that sample draws a different element everytime it is called.
.seed <- sample(.Random.seed, 1)
}
## Set seed
set.seed(.seed)
## Choose resampling method
out <- switch (.resample_method,
"jackknife" = {
if(exists("data_split")) {
future.apply::future_lapply(data_split, function(y) future.apply::future_lapply(1:nrow(y), function(x) y[-x, ]))
} else {
future.apply::future_lapply(1:nrow(data), function(x) data[-x, ])
}
},
"bootstrap" = {
if(exists("data_split")) {
future.apply::future_lapply(data_split, function(y) {
future.apply::future_lapply(1:.R, function(x) {
y[sample(1:nrow(y), size = nrow(y), replace = TRUE), ]
}, future.seed = .seed)
})
} else {
future.apply::future_lapply(1:.R, function(x) {
data[sample(1:nrow(data), size = nrow(data), replace = TRUE), ]},
future.seed = .seed
)
}
},
"permutation" = {
if(is.null(id)) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"No id column specified to permutate the data with."
)
} else {
future.apply::future_lapply(1:.R, function(x) {
cbind(data[,-which(colnames(data) == id)], "id" = sample(data[, id]))},
future.seed = .seed)
}
},
"cross-validation" = {
if(.cv_folds < 2) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"A minimum of 2 cross-validation folds required.")
}
if((ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) >= ceiling(nrow(data)/.cv_folds))){
warning2(
"The following error occured in the `resampleData()` function:\n",
"The number of .cv_folds is not plausible for the given sample size.\n",
"Change .cv_folds such that \n",
"ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) <= ceiling(nrow(data)/.cv_folds)"
)
}
# k-fold cross-validation (=draw k samples of equal size.).
# Note the last sample may contain less observations if equal sized
# samples are not possible
if(exists("data_split")) {
## The number of folds cannot be larger than the minimum number of
## rows/observations for all data sets in data_split
if(max(sapply(data_split, nrow)) < .cv_folds) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"The number of folds is larger than the number of observations",
" in at least one of the groups."
)
}
future.apply::future_lapply(data_split, function(y) {
future.apply::future_lapply(1:.R, function(x) {
# shuffle data set
y <- y[sample(1:nrow(y)), ]
if(ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) < ceiling(nrow(data)/.cv_folds)){
#In case that warnings occur, the splitting of the data set might be changed
#suppressWarnings(
split(as.data.frame(y), rep(1:.cv_folds,
each = ceiling(nrow(y)/.cv_folds)))
#)
}else{
#suppressWarnings(
split(as.data.frame(y), rep(1:.cv_folds,
each = floor(nrow(y)/.cv_folds)))
#)
}
}, future.seed = .seed)
})
} else {
## The number of folds cannot be larger than the minimum number of
## rows/observations for all data sets in data_split
if(nrow(data) < .cv_folds) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"The number of folds is larger than the number of observations."
)
}
# shuffle data
future.apply::future_lapply(1:.R, function(x) {
data <- data[sample(1:nrow(data)), ]
if(ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) < ceiling(nrow(data)/.cv_folds)){
suppressWarnings(
split(as.data.frame(data), rep(1:.cv_folds,
each = ceiling(nrow(data)/.cv_folds)))
)
}else{
suppressWarnings(
split(as.data.frame(data), rep(1:.cv_folds,
each = floor(nrow(data)/.cv_folds)))
)
}
}, future.seed = .seed)
}
} # END cross-validation
) # END switch
## Return samples
out
}
#' Resample cSEMResults
#'
#' Resample a [cSEMResults] object using bootstrap or jackknife resampling.
#' The function is called by [csem()] if the user sets
#' `csem(..., .resample_method = "bootstrap")` or
#' `csem(..., .resample_method = "jackknife")` but may also be called directly.
#'
#' Given `M` resamples (for bootstrap `M = .R` and for jackknife `M = N`, where
#' `N` is the number of observations) based on the data used to compute the
#' [cSEMResults] object provided via `.object`, `resamplecSEMResults()` essentially calls
#' [csem()] on each resample using the arguments of the original call (ignoring any arguments
#' related to resampling) and returns estimates for each of a subset of
#' practically useful resampled parameters/statistics computed by [csem()].
#' Currently, the following estimates are computed and returned by default based
#' on each resample: Path estimates, Loading estimates, Weight estimates.
#'
#' In practical application users may need to resample a specific statistic (e.g,
#' the heterotrait-monotrait ratio of correlations (HTMT) or differences between path
#' coefficients such as beta_1 - beta_2).
#' Such statistics may be provided by a function `fun(.object, ...)` or a list of
#' such functions via the `.user_funs` argument. The first argument of
#' these functions must always be `.object`.
#' Internally, the function will be applied on each
#' resample to produce the desired statistic. Hence, arbitrary complicated statistics
#' may be resampled as long as the body of the function draws on elements contained
#' in the [cSEMResults] object only. Output of `fun(.object, ...)` should preferably
#' be a (named) vector but matrices are also accepted.
#' However, the output will be vectorized (columnwise) in this case.
#' See the examples section for details.
#'
#' Both resampling the original [cSEMResults] object (call it "first resample")
#' and resampling based on a resampled [cSEMResults] object (call it "second resample")
#' are supported. Choices for the former
#' are "*bootstrap*" and "*jackknife*". Resampling based on a resample is turned off
#' by default (`.resample_method2 = "none"`) as this significantly
#' increases computation time (there are now `M * M2` resamples to compute, where
#' `M2` is `.R2` or `N`).
#' Resamples of a resample are required, e.g., for the studentized confidence
#' interval computed by the [infer()] function. Typically, bootstrap resamples
#' are used in this case \insertCite{Davison1997}{cSEM}.
#'
#' As [csem()] accepts a single data set, a list of data sets as well as data sets
#' that contain a column name used to split the data into groups,
#' the [cSEMResults] object may contain multiple data sets.
#' In this case, resampling is done by data set or group. Note that depending
#' on the number of data sets/groups, the computation may be considerably
#' slower as resampling will be repeated for each data set/group. However, apart
#' from speed considerations users don not need to worry about the type of
#' input used to compute the [cSEMResults] object as `resamplecSEMResults()`
#' is able to deal with each case.
#'
#' The number of bootstrap runs for the first and second run are given by `.R` and `.R2`.
#' The default is `499` for the first and `199` for the second run
#' but should be increased in real applications. See e.g.,
#' \insertCite{Hesterberg2015;textual}{cSEM}, p.380,
#' \insertCite{Davison1997;textual}{cSEM}, and
#' \insertCite{Efron2016;textual}{cSEM} for recommendations.
#' For jackknife `.R` are `.R2` are ignored.
#'
#' Resampling may produce inadmissible results (as checked by [verify()]).
#' By default these results are dropped however users may choose to `"ignore"`
#' or `"replace"` inadmissible results in which resampling continuous until
#' the necessary number of admissible results is reached.
#'
#' The \pkg{cSEM} package supports (multi)processing via the \href{https://github.com/HenrikBengtsson/future/}{future}
#' framework \insertCite{Bengtsson2018}{cSEM}. Users may simply choose an evaluation plan
#' via `.eval_plan` and the package takes care of all the complicated backend
#' issues. Currently, users may choose between standard single-core/single-session
#' evaluation (`"sequential"`) and multiprocessing (`"multisession"` or `"multicore"`). The future package
#' provides other options (e.g., `"cluster"` or `"remote"`), however, they probably
#' will not be needed in the context of the \pkg{cSEM} package as simulations usually
#' do not require high-performance clusters. Depending on the operating system, the future
#' package will manage to distribute tasks to multiple R sessions (Windows)
#' or multiple cores. Note that multiprocessing is not necessary always faster
#' when only a "small" number of replications is required as the overhead of
#' initializing new sessions or distributing tasks to different cores
#' will not immediately be compensated by the availability of multiple sessions/cores.
#'
#' Random number generation (RNG) uses the L'Ecuyer-CRMR RGN stream as implemented in the
#' \href{https://github.com/HenrikBengtsson/future.apply/}{future.apply package} \insertCite{Bengtsson2018a}{cSEM}.
#' It is independent of the evaluation plan. Hence, setting e.g., `.seed = 123` will
#' generate the same random number and replicates
#' for both `.eval_plan = "sequential"`, `.eval_plan = "multisession"`, and `.eval_plan = "multicore"`.
#' See [?future_lapply][future.apply::future_lapply] for details.
#'
#' @usage resamplecSEMResults(
#' .object = NULL,
#' .resample_method = c("bootstrap", "jackknife"),
#' .resample_method2 = c("none", "bootstrap", "jackknife"),
#' .R = 499,
#' .R2 = 199,
#' .handle_inadmissibles = c("drop", "ignore", "replace"),
#' .user_funs = NULL,
#' .eval_plan = c("sequential", "multicore", "multisession"),
#' .force = FALSE,
#' .seed = NULL,
#' .sign_change_option = c("none","individual","individual_reestimate",
#' "construct_reestimate"),
#' ...
#' )
#'
#' @inheritParams csem_arguments
#' @param .resample_method Character string. The resampling method to use. One of:
#' "*bootstrap*" or "*jackknife*". Defaults to "*bootstrap*".
#' @param ... Further arguments passed to functions supplied to `.user_funs`.
#'
#' @return The core structure is the same structure as that of `.object` with
#' the following elements added:
#' \itemize{
#' \item{ `$Estimates_resamples`: A list containing the `.R` resamples and
#' the original estimates for each of the resampled quantities (Path_estimates,
#' Loading_estimates, Weight_estimates, user defined functions).
#' Each list element is a list containing elements
#' `$Resamples` and `$Original`. `$Resamples` is a `(.R x K)` matrix with each
#' row representing one resample for each of the `K` parameters/statistics.
#' `$Original` contains the original estimates (vectorized by column if the output of
#' the user provided function is a matrix.}
#' \item {`$Information_resamples`: A list containing additional information.}
#' }
#' Use `str(<.object>, list.len = 3)` on the resulting object for an overview.
#'
#' @references
#' \insertAllCited{}
#'
#' @seealso [csem], [summarize()], [infer()], [cSEMResults]
#'
#' @example inst/examples/example_resamplecSEMResults.R
#'
#' @export
#'
resamplecSEMResults <- function(
.object = NULL,
.resample_method = c("bootstrap", "jackknife"),
.resample_method2 = c("none", "bootstrap", "jackknife"),
.R = 499,
.R2 = 199,
.handle_inadmissibles = c("drop", "ignore", "replace"),
.user_funs = NULL,
.eval_plan = c("sequential", "multicore", "multisession"),
.force = FALSE,
.seed = NULL,
.sign_change_option = c("none","individual","individual_reestimate",
"construct_reestimate"),
...
) {
## Match arguments
.resample_method <- match.arg(.resample_method)
.resample_method2 <- match.arg(.resample_method2)
.handle_inadmissibles <- match.arg(.handle_inadmissibles)
.eval_plan <- match.arg(.eval_plan)
.sign_change_option <- match.arg(.sign_change_option)
if(inherits(.object, "cSEMResults_multi")) {
out <- lapply(.object, function(x) {
resamplecSEMResults(
.object = x,
.resample_method = .resample_method,
.resample_method2 = .resample_method2,
.R = .R,
.R2 = .R2,
.handle_inadmissibles = .handle_inadmissibles,
.user_funs = .user_funs,
.eval_plan = .eval_plan,
.force = FALSE,
.seed = .seed,
.sign_change_option = .sign_change_option,
...
)
})
## Add/ set class
class(out) <- c(class(.object), "cSEMResults_resampled")
return(out)
}
## Does .object already contain resamples
if(inherits(.object, "cSEMResults_resampled")) {
if(.force) {
# Delete resamples
if(inherits(.object, "cSEMResults_default")) {
.object$Estimates[["Estimates_resample"]] <- NULL
.object$Information[["Information_resample"]] <- NULL
class(.object) <- setdiff(class(.object), "cSEMResults_resampled")
} else {
.object$Second_stage$Information[["Resamples"]] <- NULL
}
} else {
stop2(
"The following issue was encountered in the `resamplecSEMResults()` function:\n",
"The object provided already contains resamples. Use `.force = TRUE` to rerun.")
}
}
## Check for the minimum number of necessary resamples
if(.R < 3 | .R2 < 3) {
stop2("The following error occured in the `resamplecSEMResults()` function:\n",
"At least 3 resamples required.")
}
## Has the object to use the data to resample from produced admissible results?
if(sum(unlist(verify(.object))) != 0) {
warning2(
"The following issue was encountered in the `resamplecSEMResults()` function:\n",
"Estimation based on the original data has produced inadmissible results.\n",
"This may be a sign that something is wrong.",
" Resampling will continue but may not produce reliable results.")
}
# Set plan on how to resolve futures; reset at the end
oplan <- future::plan()
on.exit(future::plan(oplan), add = TRUE)
future::plan(.eval_plan)
### Process original data ----------------------------------------------------
## Select relevant statistics/parameters/quantities and vectorize
Est_original <- selectAndVectorize(.object)
## Apply user defined function if specified
user_funs <- if(!is.null(.user_funs)) {
applyUserFuns(.object, .user_funs = .user_funs, ...)
}
## Add output of the user functions to Est_original
if(!is.null(.user_funs)) {
Est_original <- c(Est_original, user_funs)
}
### Resample and compute -----------------------------------------------------
# Save old seed and restore on exit! This is important since users may have
# set a seed before, in which case the global seed would be
# overwritten if not explicitly restored
# Note (07.10.2020): for some tests performed by testthat .Random.seed is not
# available. This is when it is not initialized. To initialize
# the .Random.seed object a random number is generated.
runif(1)
old_seed <- .Random.seed
on.exit({.Random.seed <<- old_seed})
## Create seed if not already set
if(is.null(.seed)) {
set.seed(seed = NULL)
# Note (08.12.2019): Its crucial to call set.seed(seed = NULL) before
# drawing a random seed out of .Random.seed. If set.seed(seed = NULL) is not
# called sample(.Random.seed, 1) would result in the same random seed as
# long as .Random.seed remains unchanged. By resetting the seed we make
# sure that sample draws a different element every time it is called.
.seed <- sample(.Random.seed, 1)
}
out <- resamplecSEMResultsCore(
.object = .object,
.resample_method = .resample_method,
.resample_method2 = .resample_method2,
.R = .R,
.R2 = .R2,
.handle_inadmissibles = .handle_inadmissibles,
.handle_inadmissibles2 = .handle_inadmissibles,
.user_funs = .user_funs,
.eval_plan = .eval_plan,
.seed = .seed,
.sign_change_option = .sign_change_option,
...
)
# Check if at least 3 admissible results were obtained
n_admissibles <- length(out)
if(n_admissibles < 3) {
stop("The following error occured in the `resamplecSEMResults()` functions:\n",
"Less than 2 admissible results produced.",
" Consider setting `.handle_inadmissibles == 'replace'` instead.",
call. = FALSE)
}
# Turn list "inside out" and bind bootstrap samples to matrix
# - columns are variables
# - rows are bootstrap runs
out <- purrr::transpose(out)
if(.resample_method2 != "none") {
out_2 <- out$Estimates2
out <- purrr::transpose(out$Estimates1)
}
out <- out %>%
lapply(function(x) do.call(rbind, x))
# Add estimated quantities based on the the original sample/data
out <- mapply(function(l1, l2) list("Resampled" = l1, "Original" = l2),
l1 = out,
l2 = Est_original,
SIMPLIFY = FALSE)
## Return --------------------------------------------------------------------
# The desired output is: a list of two:
# 1. Resamples := a list of two
# 1.1 Estimates1 := the .R resamples of the "outer" resampling run.
# 1.2 Estimates2 := the .R2 resamples of the "inner" run or NA
# 2. Information := a list of three
# 2.1 Method := the method used to obtain the "outer" resamples
# 2.2 Method2 := the method used to obtain the "inner" resamples or NA
# 2.3 Number_of_observations := the number of observations.
# 2.4 Original_object := the original cSEMResults object
#
# Since resamplecSEMResults is called recursivly within its body it is
# difficult to produce the desired output. There is now straightforward way
# to determine if a call is a recursive call or not
# My workaround for now is to simply check if the argument provided for ".object"
# is "Est_temp" since this is the argument for the recurive call. This
# is of course not general but a dirty workaround. I guess it is save enough
# though.
is_recursive_call <- eval.parent(as.list(match.call()))$.object == "Est_temp"
## Return
if(is_recursive_call) {
out
} else {
if(.resample_method2 != "none") {
out <- list("Estimates1" = out, "Estimates2" = out_2)
} else {
out <- list("Estimates1" = out, "Estimates2" = NA)
}
## Add resamples and additional information
if(inherits(.object, "cSEMResults_2ndorder")) {
resample_out <- c(.object$Second_stage$Information)
resample_out[[length(resample_out) + 1]] <- list(
"Estimates" = c(out),
"Information_resample" = list(
"Method" = .resample_method,
"Method2" = .resample_method2,
"Number_of_admissibles" = n_admissibles,
"Number_of_observations" = nrow(.object$Second_stage$Information$Data),
"Number_of_runs" = .R,
"Number_of_runs2" = .R2,
"Seed" = .seed,
"Handle_inadmissibles" = .handle_inadmissibles,
"Sign_change_option" = .sign_change_option
)
)
names(resample_out)[length(resample_out)] <- "Resamples"
out <- list(
"First_stage" = .object$First_stage,
"Second_stage" = list(
"Estimates" = .object$Second_stage$Estimates,
"Information" = resample_out
)
)
# Renew class for 2nd stage
class(out$Second_stage) <- class(.object$Second_stage)
} else {
# Estimates
estim <- c(.object$Estimates)
estim[[length(estim) + 1]] <- c(out)
names(estim)[length(estim)] <- "Estimates_resample"
# Information
info <- c(.object$Information)
info[[length(info) + 1]] <- list(
"Method" = .resample_method,
"Method2" = .resample_method2,
"Number_of_admissibles" = n_admissibles,
"Number_of_observations" = nrow(.object$Information$Data), # for infer()
"Number_of_runs" = .R,
"Number_of_runs2" = .R2,
"Seed" = .seed,
"Handle_inadmissibles" = .handle_inadmissibles,
"Sign_change_option" = .sign_change_option
)
names(info)[length(info)] <- "Information_resample"
out <- list(
"Estimates" = estim,
"Information" = info
)
}
## Add/ set class
class(out) <- c(class(.object), "cSEMResults_resampled")
return(out)
}
}
#' Core tasks of the resamplecSEMResults function
#' @noRd
#'
resamplecSEMResultsCore <- function(
.object = args_default()$.object,
.resample_method = args_default()$.resample_method,
.resample_method2 = args_default()$.resample_method2,
.R = args_default()$.R,
.R2 = args_default()$.R2,
.handle_inadmissibles = args_default()$.handle_inadmissibles,
.handle_inadmissibles2 = NULL,
.user_funs = args_default()$.user_funs,
.eval_plan = args_default()$.eval_plan,
.seed = args_default()$.seed,
.sign_change_option = args_default()$.sign_change_option,
...
) {
## Get arguments
if(inherits(.object, "cSEMResults_2ndorder")) {
info1 <- .object$First_stage$Information
info2 <- .object$Second_stage$Information
est1_normal <- .object$First_stage$Estimates
est2_normal <- .object$Second_stage$Estimates
summary_original <- summarize(.object)
est1 <- summary_original$First_stage$Estimates
est2 <- summary_original$Second_stage$Estimates
args <- info2$Arguments_original
## Append the 2ndorder approach to args
args$.approach_2ndorder <- info2$Approach_2ndorder
## Replace original data (which contains either an id column or is a list)
## by the data used for the group that is being resampled.
if(!is.null(args$.id) | inherits(args$.data, "list")) {
args$.data <- info1$Data
args$.id <- NULL
}
} else {
info1 <- .object$Information
est_normal <- .object$Estimates
summary_original <- summarize(.object)
est <- summary_original$Estimates
args <- info1$Arguments
}
## Resample jackknife
if(.resample_method == "jackknife") {
resample_jack <- resampleData(.object, .resample_method = "jackknife")
.R <- length(resample_jack)
}
### Start resampling loop ====================================================
progressr::with_progress({
if(.R >= 10) {
progress_bar_csem <- progressr::progressor(steps = floor(.R / 10))
} else {
progress_bar_csem <- progressr::progressor(steps = .R)
}
Est_ls <- future.apply::future_lapply(1:.R, function(i) {
if (i %% 10 == 0) {
progress_bar_csem(message = sprintf("Resample run = %g", i))
}
# Est_ls <- lapply(1:.R, function(i) {
# Replace the old data set by a resampled data set (resampleData always returns
# a list so for just one draw we need to pick the first list element)
data_temp <- if(.resample_method == "jackknife") {
resample_jack[[i]]
} else {
# We could use resampleData here but, bootstrap resampling directly is faster
# (not surprising)
# (compared both approaches using microbenchmark)
data <- args[[".data"]]
data[sample(1:nrow(data), size = nrow(data), replace = TRUE), ]
}
args[[".data"]] <- data_temp
# Estimate model
Est_temp <- if(inherits(.object, "cSEMResults_2ndorder")) {
## NOTE: using do.call(csem, args) would be more elegant but is much
# much much! slower (especially for larger data sets).
do.call(csem, args)
} else {
# It is important to use foreman() here
# instead of csem() to allow for lapply(x, resamplecSEMResults_default) when x
# is of class cSEMResults_2ndorder.
## NOTE:
# 01.03.2019: Using do.call(foreman, args) would be more elegant but is much
# much much! slower (especially for larger data sets).
#
# 15.05.2019: Apparently do.call(foreman, args) is not that bad
# after all. I did several comparisons using microbenchmark
# but there was no speed difference (anymore?!). When I compared and
# thought that do.call is slow I fixed other parts of
# foreman as well...maybe they had been the real culprit.
# So we use do.call again, as it avoids redundancy
do.call(foreman, args)
}
# Check status
status_code <- sum(unlist(verify(Est_temp)))
# Distinguish depending on how inadmissibles should be handled
if(status_code == 0 | (status_code != 0 & .handle_inadmissibles == "ignore")) {
# ## Select relevant statistics/parameters/quantities
x1 <- selectAndVectorize(Est_temp)
### Sign change correction for PLS-PM
if(.sign_change_option != "none" && info1$Arguments$.approach_weights == "PLS-PM") {
if(inherits(.object, "cSEMResults_2ndorder")) {
est1_temp_normal <- Est_temp$First_stage$Estimates
est2_temp_normal <- Est_temp$Second_stage$Estimates
### Individual_reestimate and construct_reestimate -------------------
if(.sign_change_option == "individual_reestimate" |
.sign_change_option == "construct_reestimate") {
## First stage
# Is there a sign difference between the first stage resample
# and original weight estimates. If so, which weights differ?
sign_diff1 <- sign(est1_normal$Weight_estimates) !=
sign(est1_temp_normal$Weight_estimates)
# Only if at least one sign differs a correction is needed
if(sum(sign_diff1) != 0) {
W1_new <- est1_temp_normal$Weight_estimates
# Individual_reestimate
if(.sign_change_option == "individual_reestimate"){
W1_new[sign_diff1] <- W1_new[sign_diff1] * (-1)
} # END individual_reestimate
if(.sign_change_option == "construct_reestimate"){
# In line with Tenenhaus et al. (2005, p 177) loadings are used,
# although they have not considered 2nd order models
Lambda1 <- est1_normal$Loading_estimates
Lambda1_temp <- est1_temp_normal$Loading_estimates
Lambda_diff1 <- abs(rowSums(Lambda1 - Lambda1_temp))
Lambda_sum1 <- abs(rowSums(Lambda1 + Lambda1_temp))
W1_new[Lambda_diff1 > Lambda_sum1, ] <- W1_new[Lambda_diff1 > Lambda_sum1, ] * (-1)
} # END construct_reestimate
# Put corrected weights in a list to be able to supply them as
# fixed weights to .PLS_modes
W1_list <- lapply(1:nrow(W1_new), function(x) {
temp <- W1_new[x, ]
temp[temp !=0]
})
names(W1_list) <- rownames(W1_new)
## Replace modes by fixed sign-corrected weights
args[[".PLS_modes"]] <- W1_list
## Reestimate
# Note: calling csem() directly is faster, however, i guess
# change option wont be used that often, so for now, I will
# use the more elegant, while slower, do.call construction.
Est_temp <- do.call(csem, args)
## Update
est2_temp_normal <- Est_temp$Second_stage$Estimates
} # END first stage
## Second stage
# Is there a sign difference between the second stage (sign-corrected)
# resample weight estimates and the original weight estimates.
# If so, which weights differ?
sign_diff2 <- sign(est2_normal$Weight_estimates) !=
sign(est2_temp_normal$Weight_estimates)
# Only if at least one sign differs a correction is needed
if(sum(sign_diff2) != 0) {
W2_new <- est2_temp_normal$Weight_estimates
# Individual_reestimate
if(.sign_change_option == "individual_reestimate"){
W2_new[sign_diff2] <- W2_new[sign_diff2] * (-1)
} # END individual_reestimate
if(.sign_change_option == "construct_reestimate"){
# In line with Tenenhaus et al. (2005, p 177) loadings are used,
# although they have not considered 2nd order models
Lambda2 <- est2_normal$Loading_estimates
Lambda2_temp <- est2_temp_normal$Loading_estimates
Lambda_diff2 <- abs(rowSums(Lambda2 - Lambda2_temp))
Lambda_sum2 <- abs(rowSums(Lambda2 + Lambda2_temp))
W2_new[Lambda_diff2 > Lambda_sum2, ] <- W2_new[Lambda_diff2 > Lambda_sum2, ] * (-1)
} # END construct_reestimate
# Put corrected weights in a list to be able to supply them as
# fixed weights to .PLS_modes
W2_list <- lapply(1:nrow(W2_new), function(x){
temp <- W2_new[x, ]
temp[temp != 0]
})
names(W2_list) <- rownames(W2_new)
args[[".PLS_modes"]] = W2_list
## Reestimate
# Note: calling csem() directly is faster, however, i guess
# change option wont be used that often, so for now, I will
# use the more elegant, while slower, do.call construction.
Est_temp <- do.call(csem, args)
} # END second stage
# ## Update using estimates based on the sign-corrected weights
x1 <- selectAndVectorize(Est_temp)
} # END individual_reestimate, construct_reestimate
if(.sign_change_option == "individual") {
# Reverse the signs off ALL parameter estimates in a bootstrap run if
# their sign differs from the sign of the original estimation
x1 <- reverseSign(.Est_temp = Est_temp, .summary_original = summary_original)
}
} else {
est_temp_normal <- Est_temp$Estimates
sign_diff <- sign(est_normal$Weight_estimates) !=
sign(est_temp_normal$Weight_estimates)
# Is there a difference in the signs of th weights? Otherwise no correction of the signs is done
# Not sure whether this is a problem for the construct_reestimate approach which only compares the loadings
# I think not.
if(sum(sign_diff) !=0 ) {
# Sign change correction individual_reestimate and construct_reestimate
if(.sign_change_option == "individual_reestimate" |
.sign_change_option == "construct_reestimate") {
W_new <- est_temp_normal$Weight_estimates
# Individual_reestimate
if(.sign_change_option == "individual_reestimate"){
W_new[sign_diff] <- W_new[sign_diff] * (-1)
} # END individual_reestimate
if(.sign_change_option == "construct_reestimate"){
Lambda <- est_normal$Loading_estimates
Lambda_temp <- est_temp_normal$Loading_estimates
Lambda_diff <- abs(rowSums(Lambda - Lambda_temp))
Lambda_sum <- abs(rowSums(Lambda + Lambda_temp))
W_new[Lambda_diff > Lambda_sum, ] <- W_new[Lambda_diff > Lambda_sum, ] * (-1)
}
# Put corrected weights in a list to be able to supply them as
# fixed weights to .PLS_modes
W_list <- lapply(1:nrow(W_new), function(x){
temp <- W_new[x, ]
temp[temp != 0]
})
names(W_list) <- rownames(W_new)
args[[".PLS_modes"]] = W_list
## Reestimate
# Note: calling csem() directly is faster, however, i guess
# change option wont be used that often, so for now, I will
# use the more elegant, while slower, do.call construction.
Est_temp <- do.call(csem, args)
## Update using estimates based on the sign-corrected weights
x1 <- selectAndVectorize(Est_temp)
} # end if individual_reestimate, construct_reestimate
if(.sign_change_option == 'individual') {
# Reverse the signs off ALL parameter estimates in a bootstrap run if
# their sign differs from the sign of the original estimation
x1 <- reverseSign(.Est_temp = Est_temp, .summary_original = summary_original)
} # END .sign_change_option == "individual"
} # END sum(sign(.object$Estimates$Weight_estimates)!=sign(Est_temp$Estimates$Weight_estimates)
} # END else (= not 2ndorder)
} # END .sign_change_option
## Apply user defined function if specified
user_funs <- if(!is.null(.user_funs)) {
applyUserFuns(Est_temp, .user_funs = .user_funs, ...)
}
# user_funs <- if(!is.null(.user_funs)) {
# if(is.function(.user_funs)) {
# list("User_fun" = c(.user_funs(Est_temp, ...)))
# } else {
# x <- lapply(.user_funs, function(f) c(f(Est_temp, ...)))
# if(is.null(names(x))) {
# names(x) <- paste0("User_fun", 1:length(x))
# }
# x
# }
# }
## Add output of the user functions to x1
if(!is.null(.user_funs)) {
x1 <- c(x1, user_funs)
}
## Resampling from a bootstrap sample is required for the
## bootstraped t-interval CI (studentized CI), hence the second run
# In the second run no sign change option is used. We can think about
# applying the same correction as in the first run
if(.resample_method2 != "none") {
Est_resamples2 <- resamplecSEMResults(
.object = Est_temp,
.R = .R2,
.handle_inadmissibles = .handle_inadmissibles2,
.resample_method = .resample_method2,
.resample_method2 = "none",
.user_funs = .user_funs,
.seed = .seed,
.force = FALSE,
.sign_change_option = "none",
...
)
x1 <- list("Estimates1" = x1, "Estimates2" = Est_resamples2)
}
} else if(status_code != 0 & .handle_inadmissibles != "ignore") {
# Retrun NA if status is not okay and .handle_inadmissibles == "drop" or
# "replace"
x1 <- NA
}
## Return
x1
}, future.seed = .seed)
# })
})
## Process data --------------------------------------------------------------
# Delete potential NA's
out <- Filter(Negate(anyNA), Est_ls)
## Replace inadmissibles
if(length(out) != .R && .resample_method == "bootstrap" &&
.handle_inadmissibles == "replace") {
n_attempt <- 0
while (length(out) < .R) {
n_attempt <- n_attempt + 1
R_new <- .R - length(out)
if(is.null(.seed)) {
.seed <- sample(.Random.seed, 1)
} else {
.seed <- .seed + 1
}
Est_replace <- resamplecSEMResultsCore(
.object = .object,
.R = R_new,
.handle_inadmissibles = "drop",
.handle_inadmissibles2= .handle_inadmissibles,
.resample_method = .resample_method,
.resample_method2 = .resample_method2,
.R2 = .R2,
.user_funs = .user_funs,
.eval_plan = "sequential", # multiprocessing will generally
# not be meaningful here since we typically
# replace only a small subset of the total runs
.seed = .seed,
.sign_change_option = .sign_change_option,
...
)
out <- c(out, Est_replace)
if(n_attempt == 10*.R) {
warning2(
"The following issue was encountered in the `resamplecSEMResults()` function:\n",
"Attempting to replace inadmissibles failed after ", 10*.R, " attempts.",
" Returning all admissible results up to the last attempt."
)
break
}
}
}
out
}
M-E-Steiner/cSEM documentation built on March 18, 2024, 12:18 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Defines functions resamplecSEMResultsCore resamplecSEMResults resampleData
Documented in resamplecSEMResults resampleData
#' Resample data
#'
#' Resample data from a data set using common resampling methods.
#' For bootstrap or jackknife resampling, package users usually do not need to
#' call this function but directly use [resamplecSEMResults()] instead.
#'
#' The function `resampleData()` is general purpose. It simply resamples data
#' from a data set according to the resampling method provided
#' via the `.resample_method` argument and returns a list of resamples.
#' Currently, `bootstrap`, `jackknife`, `permutation`, and `cross-validation`
#' (both leave-one-out (LOOCV) and k-fold cross-validation) are implemented.
#'
#' The user may provide the data set to resample either explicitly via the `.data`
#' argument or implicitly by providing a [cSEMResults] objects to `.object`
#' in which case the original data used in the call that created the
#' [cSEMResults] object is used for resampling.
#' If both, a [cSEMResults] object and a data set via `.data` are provided
#' the former is ignored.
#'
#' As [csem()] accepts a single data set, a list of data sets as well as data sets
#' that contain a column name used to split the data into groups,
#' the [cSEMResults] object may contain multiple data sets.
#' In this case, resampling is done by data set or group. Note that depending
#' on the number of data sets/groups provided this computation may be slower
#' as resampling will be repeated for each data set/group.
#'
#' To split data provided via the `.data` argument into groups, the column name or
#' the column index of the column containing the group levels to split the data
#' must be given to `.id`. If data that contains grouping is taken from
#' a [cSEMResults] object, `.id` is taken from the object information. Hence,
#' providing `.id` is redundant in this case and therefore ignored.
#'
#' The number of bootstrap or permutation runs as well as the number of
#' cross-validation repetitions is given by `.R`. The default is
#' `499` but should be increased in real applications. See e.g.,
#' \insertCite{Hesterberg2015;textual}{cSEM}, p.380 for recommendations concerning
#' the bootstrap. For jackknife `.R` is ignored as it is based on the N leave-one-out data sets.
#'
#' Choosing `resample_method = "permutation"` for ungrouped data causes an error
#' as permutation will simply reorder the observations which is usually not
#' meaningful. If a list of data is provided
#' each list element is assumed to represent the observations belonging to one
#' group. In this case, data is pooled and group adherence permuted.
#'
#' For cross-validation the number of folds (`k`) defaults to `10`. It may be
#' changed via the `.cv_folds` argument. Setting `k = 2` (not 1!) splits
#' the data into a single training and test data set. Setting `k = N` (where `N` is the
#' number of observations) produces leave-one-out cross-validation samples.
#' Note: 1.) At least 2 folds required (`k > 1`); 2.) `k` can not be larger than `N`;
#' 3.) If `N/k` is not not an integer the last fold will have less observations.
#'
#' Random number generation (RNG) uses the L'Ecuyer-CRMR RGN stream as implemented
#' in the \href{https://github.com/HenrikBengtsson/future.apply/}{future.apply package}
#' \insertCite{Bengtsson2018a}{cSEM}.
#' See [?future_lapply][future.apply::future_lapply] for details. By default
#' a random seed is chosen.
#'
#' @usage resampleData(
#' .object = NULL,
#' .data = NULL,
#' .resample_method = c("bootstrap", "jackknife", "permutation",
#' "cross-validation"),
#' .cv_folds = 10,
#' .id = NULL,
#' .R = 499,
#' .seed = NULL
#' )
#'
#' @param .data A `data.frame`, a `matrix` or a `list` of data of either type.
#' Possible column types or classes of the data provided are:
#' "`logical`", "`numeric`" ("`double`" or "`integer`"), "`factor`" (ordered and unordered)
#' or a mix of several types. The data may also include
#' **one** character column whose column name must be given to `.id`.
#' This column is assumed to contain group identifiers used to split the data into groups.
#' If `.data` is provided, `.object` is ignored. Defaults to `NULL`.
#' @param .resample_method Character string. The resampling method to use. One of:
#' "*bootstrap*", "*jackknife*", "*permutation*", or "*cross-validation*".
#' Defaults to "*bootstrap*".
#' @param .R Integer. The number of bootstrap runs, permutation runs
#' or cross-validation repetitions to use. Defaults to `499`.
#' @inheritParams csem_arguments
#'
#' @return The structure of the output depends on the type of input and the
#' resampling method:
#' \describe{
#' \item{Bootstrap}{If a `matrix` or `data.frame` without grouping variable
#' is provided (i.e., `.id = NULL`), the result is a list of length `.R`
#' (default `499`). Each element of that list is a bootstrap (re)sample.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data for one group),
#' resampling is done by group. Hence,
#' the result is a list of length equal to the number of groups
#' with each list element containing `.R` bootstrap samples based on the
#' `N_g` observations of group `g`.}
#' \item{Jackknife}{If a `matrix` or `data.frame` without grouping variable
#' is provided (`.id = NULL`), the result is a list of length equal to the number
#' of observations/rows (`N`) of the data set provided.
#' Each element of that list is a jackknife (re)sample.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data for one group),
#' resampling is done by group. Hence,
#' the result is a list of length equal to the number of group levels
#' with each list element containing `N` jackknife samples based on the
#' `N_g` observations of group `g`.}
#' \item{Permutation}{If a `matrix` or `data.frame` without grouping variable
#' is provided an error is returned as permutation will simply reorder the observations.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data of one group),
#' group membership is permuted. Hence, the result is a list of length `.R`
#' where each element of that list is a permutation (re)sample.}
#' \item{Cross-validation}{If a `matrix` or `data.frame` without grouping variable
#' is provided a list of length `.R` is returned. Each list element
#' contains a list containing the `k` splits/folds subsequently
#' used as test and training data sets.
#' If a grouping variable is specified or a list of data is provided
#' (where each list element is assumed to contain data for one group),
#' cross-validation is repeated `.R` times for each group. Hence,
#' the result is a list of length equal to the number of groups,
#' each containing `.R` list elements (the repetitions) which in turn contain
#' the `k` splits/folds.
#' }
#' }
#' @references
#' \insertAllCited{}
#'
#' @seealso [csem()], [cSEMResults], [resamplecSEMResults()]
#'
#' @example inst/examples/example_resampleData.R
#'
#' @export
#'
resampleData <- function(
.object = NULL,
.data = NULL,
.resample_method = c("bootstrap", "jackknife", "permutation",
"cross-validation"),
.cv_folds = 10,
.id = NULL,
.R = 499,
.seed = NULL
) {
.resample_method <- match.arg(.resample_method,
c("bootstrap", "jackknife", "permutation", "cross-validation"))
## Set plan on how to resolve futures to "sequential" as it is virtually always
## slower to resample data using "multiprocess".; reset at the end
oplan <- future::plan()
on.exit(future::plan(oplan), add = TRUE)
future::plan("sequential")
## Get data set
if(is.null(.data)) {
## Get information according to class of object
if(any(class(.object) %in% "cSEMResults_default")) {
data <- as.data.frame(.object$Information$Arguments$.data)
id <- NULL
} else if(any(class(.object) %in% "cSEMResults_multi")) {
data <- .object[[1]]$Information$Data_pooled
data_split <- lapply(.object, function(x) x$Information$Arguments$.data)
id <- ifelse(is.null(.object[[1]]$Information$Arguments$.id),
"id", .object[[1]]$Information$Arguments$.id)
} else if(any(class(.object) %in% "cSEMResults_2ndorder")) {
data <- .object$First_stage$Information$Arguments$.data
id <- NULL
} else {
stop2("The following error occured in the `resampleData()` function:\n",
"`object` must be of class cSEMResults."
)
}
} else {
## Check data
if(!any(class(.data) %in% c("data.frame", "matrix", "list"))) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"Data must be provided as a `matrix`, a `data.frame` or a `list`. ",
".data has class: ", class(.data)
)
}
## Select cases
if(!is.null(.id) && !inherits(.data, "list")) {
if(length(.id) != 1) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"`.id` must be a character string or an integer identifying one single column."
)
}
if(is.matrix(.data)) {
.data <- as.data.frame(.data)
}
data <- .data
data_split <- split(.data, f = .data[, .id])
id <- .id
} else if(any(class(.data) == "list")) {
data <- do.call(rbind, .data)
if(is.null(names(.data))) {
data[, "id"] <- rep(paste0("Data_", 1:length(.data)),
times = sapply(.data, nrow))
} else {
data[, "id"] <- rep(names(.data), times = sapply(.data, nrow))
}
data_split <- .data
id <- "id"
##add names here
} else {
data <- .data
id <- NULL
}
}
# Save old seed and restore on exit! This is important since users may have
# set a seed before, in which case the global seed would be
# overwritten if not explicitly restored
old_seed <- .Random.seed
on.exit({.Random.seed <<- old_seed})
## Create seed if not already set
if(is.null(.seed)) {
set.seed(seed = NULL)
# Note (08.12.2019): Its crucial to call set.seed(seed = NULL) before
# drawing a random seed out of .Random.seed. If set.seed(seed = NULL) is not
# called sample(.Random.seed, 1) would result in the same random seed as
# long as .Random.seed remains unchanged. By resetting the seed we make
# sure that sample draws a different element everytime it is called.
.seed <- sample(.Random.seed, 1)
}
## Set seed
set.seed(.seed)
## Choose resampling method
out <- switch (.resample_method,
"jackknife" = {
if(exists("data_split")) {
future.apply::future_lapply(data_split, function(y) future.apply::future_lapply(1:nrow(y), function(x) y[-x, ]))
} else {
future.apply::future_lapply(1:nrow(data), function(x) data[-x, ])
}
},
"bootstrap" = {
if(exists("data_split")) {
future.apply::future_lapply(data_split, function(y) {
future.apply::future_lapply(1:.R, function(x) {
y[sample(1:nrow(y), size = nrow(y), replace = TRUE), ]
}, future.seed = .seed)
})
} else {
future.apply::future_lapply(1:.R, function(x) {
data[sample(1:nrow(data), size = nrow(data), replace = TRUE), ]},
future.seed = .seed
)
}
},
"permutation" = {
if(is.null(id)) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"No id column specified to permutate the data with."
)
} else {
future.apply::future_lapply(1:.R, function(x) {
cbind(data[,-which(colnames(data) == id)], "id" = sample(data[, id]))},
future.seed = .seed)
}
},
"cross-validation" = {
if(.cv_folds < 2) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"A minimum of 2 cross-validation folds required.")
}
if((ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) >= ceiling(nrow(data)/.cv_folds))){
warning2(
"The following error occured in the `resampleData()` function:\n",
"The number of .cv_folds is not plausible for the given sample size.\n",
"Change .cv_folds such that \n",
"ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) <= ceiling(nrow(data)/.cv_folds)"
)
}
# k-fold cross-validation (=draw k samples of equal size.).
# Note the last sample may contain less observations if equal sized
# samples are not possible
if(exists("data_split")) {
## The number of folds cannot be larger than the minimum number of
## rows/observations for all data sets in data_split
if(max(sapply(data_split, nrow)) < .cv_folds) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"The number of folds is larger than the number of observations",
" in at least one of the groups."
)
}
future.apply::future_lapply(data_split, function(y) {
future.apply::future_lapply(1:.R, function(x) {
# shuffle data set
y <- y[sample(1:nrow(y)), ]
if(ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) < ceiling(nrow(data)/.cv_folds)){
#In case that warnings occur, the splitting of the data set might be changed
#suppressWarnings(
split(as.data.frame(y), rep(1:.cv_folds,
each = ceiling(nrow(y)/.cv_folds)))
#)
}else{
#suppressWarnings(
split(as.data.frame(y), rep(1:.cv_folds,
each = floor(nrow(y)/.cv_folds)))
#)
}
}, future.seed = .seed)
})
} else {
## The number of folds cannot be larger than the minimum number of
## rows/observations for all data sets in data_split
if(nrow(data) < .cv_folds) {
stop2(
"The following error occured in the `resampleData()` function:\n",
"The number of folds is larger than the number of observations."
)
}
# shuffle data
future.apply::future_lapply(1:.R, function(x) {
data <- data[sample(1:nrow(data)), ]
if(ceiling(nrow(data)/.cv_folds)*.cv_folds - nrow(data) < ceiling(nrow(data)/.cv_folds)){
suppressWarnings(
split(as.data.frame(data), rep(1:.cv_folds,
each = ceiling(nrow(data)/.cv_folds)))
)
}else{
suppressWarnings(
split(as.data.frame(data), rep(1:.cv_folds,
each = floor(nrow(data)/.cv_folds)))
)
}
}, future.seed = .seed)
}
} # END cross-validation
) # END switch
## Return samples
out
}
#' Resample cSEMResults
#'
#' Resample a [cSEMResults] object using bootstrap or jackknife resampling.
#' The function is called by [csem()] if the user sets
#' `csem(..., .resample_method = "bootstrap")` or
#' `csem(..., .resample_method = "jackknife")` but may also be called directly.
#'
#' Given `M` resamples (for bootstrap `M = .R` and for jackknife `M = N`, where
#' `N` is the number of observations) based on the data used to compute the
#' [cSEMResults] object provided via `.object`, `resamplecSEMResults()` essentially calls
#' [csem()] on each resample using the arguments of the original call (ignoring any arguments
#' related to resampling) and returns estimates for each of a subset of
#' practically useful resampled parameters/statistics computed by [csem()].
#' Currently, the following estimates are computed and returned by default based
#' on each resample: Path estimates, Loading estimates, Weight estimates.
#'
#' In practical application users may need to resample a specific statistic (e.g,
#' the heterotrait-monotrait ratio of correlations (HTMT) or differences between path
#' coefficients such as beta_1 - beta_2).
#' Such statistics may be provided by a function `fun(.object, ...)` or a list of
#' such functions via the `.user_funs` argument. The first argument of
#' these functions must always be `.object`.
#' Internally, the function will be applied on each
#' resample to produce the desired statistic. Hence, arbitrary complicated statistics
#' may be resampled as long as the body of the function draws on elements contained
#' in the [cSEMResults] object only. Output of `fun(.object, ...)` should preferably
#' be a (named) vector but matrices are also accepted.
#' However, the output will be vectorized (columnwise) in this case.
#' See the examples section for details.
#'
#' Both resampling the original [cSEMResults] object (call it "first resample")
#' and resampling based on a resampled [cSEMResults] object (call it "second resample")
#' are supported. Choices for the former
#' are "*bootstrap*" and "*jackknife*". Resampling based on a resample is turned off
#' by default (`.resample_method2 = "none"`) as this significantly
#' increases computation time (there are now `M * M2` resamples to compute, where
#' `M2` is `.R2` or `N`).
#' Resamples of a resample are required, e.g., for the studentized confidence
#' interval computed by the [infer()] function. Typically, bootstrap resamples
#' are used in this case \insertCite{Davison1997}{cSEM}.
#'
#' As [csem()] accepts a single data set, a list of data sets as well as data sets
#' that contain a column name used to split the data into groups,
#' the [cSEMResults] object may contain multiple data sets.
#' In this case, resampling is done by data set or group. Note that depending
#' on the number of data sets/groups, the computation may be considerably
#' slower as resampling will be repeated for each data set/group. However, apart
#' from speed considerations users don not need to worry about the type of
#' input used to compute the [cSEMResults] object as `resamplecSEMResults()`
#' is able to deal with each case.
#'
#' The number of bootstrap runs for the first and second run are given by `.R` and `.R2`.
#' The default is `499` for the first and `199` for the second run
#' but should be increased in real applications. See e.g.,
#' \insertCite{Hesterberg2015;textual}{cSEM}, p.380,
#' \insertCite{Davison1997;textual}{cSEM}, and
#' \insertCite{Efron2016;textual}{cSEM} for recommendations.
#' For jackknife `.R` are `.R2` are ignored.
#'
#' Resampling may produce inadmissible results (as checked by [verify()]).
#' By default these results are dropped however users may choose to `"ignore"`
#' or `"replace"` inadmissible results in which resampling continuous until
#' the necessary number of admissible results is reached.
#'
#' The \pkg{cSEM} package supports (multi)processing via the \href{https://github.com/HenrikBengtsson/future/}{future}
#' framework \insertCite{Bengtsson2018}{cSEM}. Users may simply choose an evaluation plan
#' via `.eval_plan` and the package takes care of all the complicated backend
#' issues. Currently, users may choose between standard single-core/single-session
#' evaluation (`"sequential"`) and multiprocessing (`"multisession"` or `"multicore"`). The future package
#' provides other options (e.g., `"cluster"` or `"remote"`), however, they probably
#' will not be needed in the context of the \pkg{cSEM} package as simulations usually
#' do not require high-performance clusters. Depending on the operating system, the future
#' package will manage to distribute tasks to multiple R sessions (Windows)
#' or multiple cores. Note that multiprocessing is not necessary always faster
#' when only a "small" number of replications is required as the overhead of
#' initializing new sessions or distributing tasks to different cores
#' will not immediately be compensated by the availability of multiple sessions/cores.
#'
#' Random number generation (RNG) uses the L'Ecuyer-CRMR RGN stream as implemented in the
#' \href{https://github.com/HenrikBengtsson/future.apply/}{future.apply package} \insertCite{Bengtsson2018a}{cSEM}.
#' It is independent of the evaluation plan. Hence, setting e.g., `.seed = 123` will
#' generate the same random number and replicates
#' for both `.eval_plan = "sequential"`, `.eval_plan = "multisession"`, and `.eval_plan = "multicore"`.
#' See [?future_lapply][future.apply::future_lapply] for details.
#'
#' @usage resamplecSEMResults(
#' .object = NULL,
#' .resample_method = c("bootstrap", "jackknife"),
#' .resample_method2 = c("none", "bootstrap", "jackknife"),
#' .R = 499,
#' .R2 = 199,
#' .handle_inadmissibles = c("drop", "ignore", "replace"),
#' .user_funs = NULL,
#' .eval_plan = c("sequential", "multicore", "multisession"),
#' .force = FALSE,
#' .seed = NULL,
#' .sign_change_option = c("none","individual","individual_reestimate",
#' "construct_reestimate"),
#' ...
#' )
#'
#' @inheritParams csem_arguments
#' @param .resample_method Character string. The resampling method to use. One of:
#' "*bootstrap*" or "*jackknife*". Defaults to "*bootstrap*".
#' @param ... Further arguments passed to functions supplied to `.user_funs`.
#'
#' @return The core structure is the same structure as that of `.object` with
#' the following elements added:
#' \itemize{
#' \item{ `$Estimates_resamples`: A list containing the `.R` resamples and
#' the original estimates for each of the resampled quantities (Path_estimates,
#' Loading_estimates, Weight_estimates, user defined functions).
#' Each list element is a list containing elements
#' `$Resamples` and `$Original`. `$Resamples` is a `(.R x K)` matrix with each
#' row representing one resample for each of the `K` parameters/statistics.
#' `$Original` contains the original estimates (vectorized by column if the output of
#' the user provided function is a matrix.}
#' \item {`$Information_resamples`: A list containing additional information.}
#' }
#' Use `str(<.object>, list.len = 3)` on the resulting object for an overview.
#'
#' @references
#' \insertAllCited{}
#'
#' @seealso [csem], [summarize()], [infer()], [cSEMResults]
#'
#' @example inst/examples/example_resamplecSEMResults.R
#'
#' @export
#'
resamplecSEMResults <- function(
.object = NULL,
.resample_method = c("bootstrap", "jackknife"),
.resample_method2 = c("none", "bootstrap", "jackknife"),
.R = 499,
.R2 = 199,
.handle_inadmissibles = c("drop", "ignore", "replace"),
.user_funs = NULL,
.eval_plan = c("sequential", "multicore", "multisession"),
.force = FALSE,
.seed = NULL,
.sign_change_option = c("none","individual","individual_reestimate",
"construct_reestimate"),
...
) {
## Match arguments
.resample_method <- match.arg(.resample_method)
.resample_method2 <- match.arg(.resample_method2)
.handle_inadmissibles <- match.arg(.handle_inadmissibles)
.eval_plan <- match.arg(.eval_plan)
.sign_change_option <- match.arg(.sign_change_option)
if(inherits(.object, "cSEMResults_multi")) {
out <- lapply(.object, function(x) {
resamplecSEMResults(
.object = x,
.resample_method = .resample_method,
.resample_method2 = .resample_method2,
.R = .R,
.R2 = .R2,
.handle_inadmissibles = .handle_inadmissibles,
.user_funs = .user_funs,
.eval_plan = .eval_plan,
.force = FALSE,
.seed = .seed,
.sign_change_option = .sign_change_option,
...
)
})
## Add/ set class
class(out) <- c(class(.object), "cSEMResults_resampled")
return(out)
}
## Does .object already contain resamples
if(inherits(.object, "cSEMResults_resampled")) {
if(.force) {
# Delete resamples
if(inherits(.object, "cSEMResults_default")) {
.object$Estimates[["Estimates_resample"]] <- NULL
.object$Information[["Information_resample"]] <- NULL
class(.object) <- setdiff(class(.object), "cSEMResults_resampled")
} else {
.object$Second_stage$Information[["Resamples"]] <- NULL
}
} else {
stop2(
"The following issue was encountered in the `resamplecSEMResults()` function:\n",
"The object provided already contains resamples. Use `.force = TRUE` to rerun.")
}
}
## Check for the minimum number of necessary resamples
if(.R < 3 | .R2 < 3) {
stop2("The following error occured in the `resamplecSEMResults()` function:\n",
"At least 3 resamples required.")
}
## Has the object to use the data to resample from produced admissible results?
if(sum(unlist(verify(.object))) != 0) {
warning2(
"The following issue was encountered in the `resamplecSEMResults()` function:\n",
"Estimation based on the original data has produced inadmissible results.\n",
"This may be a sign that something is wrong.",
" Resampling will continue but may not produce reliable results.")
}
# Set plan on how to resolve futures; reset at the end
oplan <- future::plan()
on.exit(future::plan(oplan), add = TRUE)
future::plan(.eval_plan)
### Process original data ----------------------------------------------------
## Select relevant statistics/parameters/quantities and vectorize
Est_original <- selectAndVectorize(.object)
## Apply user defined function if specified
user_funs <- if(!is.null(.user_funs)) {
applyUserFuns(.object, .user_funs = .user_funs, ...)
}
## Add output of the user functions to Est_original
if(!is.null(.user_funs)) {
Est_original <- c(Est_original, user_funs)
}
### Resample and compute -----------------------------------------------------
# Save old seed and restore on exit! This is important since users may have
# set a seed before, in which case the global seed would be
# overwritten if not explicitly restored
# Note (07.10.2020): for some tests performed by testthat .Random.seed is not
# available. This is when it is not initialized. To initialize
# the .Random.seed object a random number is generated.
runif(1)
old_seed <- .Random.seed
on.exit({.Random.seed <<- old_seed})
## Create seed if not already set
if(is.null(.seed)) {
set.seed(seed = NULL)
# Note (08.12.2019): Its crucial to call set.seed(seed = NULL) before
# drawing a random seed out of .Random.seed. If set.seed(seed = NULL) is not
# called sample(.Random.seed, 1) would result in the same random seed as
# long as .Random.seed remains unchanged. By resetting the seed we make
# sure that sample draws a different element every time it is called.
.seed <- sample(.Random.seed, 1)
}
out <- resamplecSEMResultsCore(
.object = .object,
.resample_method = .resample_method,
.resample_method2 = .resample_method2,
.R = .R,
.R2 = .R2,
.handle_inadmissibles = .handle_inadmissibles,
.handle_inadmissibles2 = .handle_inadmissibles,
.user_funs = .user_funs,
.eval_plan = .eval_plan,
.seed = .seed,
.sign_change_option = .sign_change_option,
...
)
# Check if at least 3 admissible results were obtained
n_admissibles <- length(out)
if(n_admissibles < 3) {
stop("The following error occured in the `resamplecSEMResults()` functions:\n",
"Less than 2 admissible results produced.",
" Consider setting `.handle_inadmissibles == 'replace'` instead.",
call. = FALSE)
}
# Turn list "inside out" and bind bootstrap samples to matrix
# - columns are variables
# - rows are bootstrap runs
out <- purrr::transpose(out)
if(.resample_method2 != "none") {
out_2 <- out$Estimates2
out <- purrr::transpose(out$Estimates1)
}
out <- out %>%
lapply(function(x) do.call(rbind, x))
# Add estimated quantities based on the the original sample/data
out <- mapply(function(l1, l2) list("Resampled" = l1, "Original" = l2),
l1 = out,
l2 = Est_original,
SIMPLIFY = FALSE)
## Return --------------------------------------------------------------------
# The desired output is: a list of two:
# 1. Resamples := a list of two
# 1.1 Estimates1 := the .R resamples of the "outer" resampling run.
# 1.2 Estimates2 := the .R2 resamples of the "inner" run or NA
# 2. Information := a list of three
# 2.1 Method := the method used to obtain the "outer" resamples
# 2.2 Method2 := the method used to obtain the "inner" resamples or NA
# 2.3 Number_of_observations := the number of observations.
# 2.4 Original_object := the original cSEMResults object
#
# Since resamplecSEMResults is called recursivly within its body it is
# difficult to produce the desired output. There is now straightforward way
# to determine if a call is a recursive call or not
# My workaround for now is to simply check if the argument provided for ".object"
# is "Est_temp" since this is the argument for the recurive call. This
# is of course not general but a dirty workaround. I guess it is save enough
# though.
is_recursive_call <- eval.parent(as.list(match.call()))$.object == "Est_temp"
## Return
if(is_recursive_call) {
out
} else {
if(.resample_method2 != "none") {
out <- list("Estimates1" = out, "Estimates2" = out_2)
} else {
out <- list("Estimates1" = out, "Estimates2" = NA)
}
## Add resamples and additional information
if(inherits(.object, "cSEMResults_2ndorder")) {
resample_out <- c(.object$Second_stage$Information)
resample_out[[length(resample_out) + 1]] <- list(
"Estimates" = c(out),
"Information_resample" = list(
"Method" = .resample_method,
"Method2" = .resample_method2,
"Number_of_admissibles" = n_admissibles,
"Number_of_observations" = nrow(.object$Second_stage$Information$Data),
"Number_of_runs" = .R,
"Number_of_runs2" = .R2,
"Seed" = .seed,
"Handle_inadmissibles" = .handle_inadmissibles,
"Sign_change_option" = .sign_change_option
)
)
names(resample_out)[length(resample_out)] <- "Resamples"
out <- list(
"First_stage" = .object$First_stage,
"Second_stage" = list(
"Estimates" = .object$Second_stage$Estimates,
"Information" = resample_out
)
)
# Renew class for 2nd stage
class(out$Second_stage) <- class(.object$Second_stage)
} else {
# Estimates
estim <- c(.object$Estimates)
estim[[length(estim) + 1]] <- c(out)
names(estim)[length(estim)] <- "Estimates_resample"
# Information
info <- c(.object$Information)
info[[length(info) + 1]] <- list(
"Method" = .resample_method,
"Method2" = .resample_method2,
"Number_of_admissibles" = n_admissibles,
"Number_of_observations" = nrow(.object$Information$Data), # for infer()
"Number_of_runs" = .R,
"Number_of_runs2" = .R2,
"Seed" = .seed,
"Handle_inadmissibles" = .handle_inadmissibles,
"Sign_change_option" = .sign_change_option
)
names(info)[length(info)] <- "Information_resample"
out <- list(
"Estimates" = estim,
"Information" = info
)
}
## Add/ set class
class(out) <- c(class(.object), "cSEMResults_resampled")
return(out)
}
}
#' Core tasks of the resamplecSEMResults function
#' @noRd
#'
resamplecSEMResultsCore <- function(
.object = args_default()$.object,
.resample_method = args_default()$.resample_method,
.resample_method2 = args_default()$.resample_method2,
.R = args_default()$.R,
.R2 = args_default()$.R2,
.handle_inadmissibles = args_default()$.handle_inadmissibles,
.handle_inadmissibles2 = NULL,
.user_funs = args_default()$.user_funs,
.eval_plan = args_default()$.eval_plan,
.seed = args_default()$.seed,
.sign_change_option = args_default()$.sign_change_option,
...
) {
## Get arguments
if(inherits(.object, "cSEMResults_2ndorder")) {
info1 <- .object$First_stage$Information
info2 <- .object$Second_stage$Information
est1_normal <- .object$First_stage$Estimates
est2_normal <- .object$Second_stage$Estimates
summary_original <- summarize(.object)
est1 <- summary_original$First_stage$Estimates
est2 <- summary_original$Second_stage$Estimates
args <- info2$Arguments_original
## Append the 2ndorder approach to args
args$.approach_2ndorder <- info2$Approach_2ndorder
## Replace original data (which contains either an id column or is a list)
## by the data used for the group that is being resampled.
if(!is.null(args$.id) | inherits(args$.data, "list")) {
args$.data <- info1$Data
args$.id <- NULL
}
} else {
info1 <- .object$Information
est_normal <- .object$Estimates
summary_original <- summarize(.object)
est <- summary_original$Estimates
args <- info1$Arguments
}
## Resample jackknife
if(.resample_method == "jackknife") {
resample_jack <- resampleData(.object, .resample_method = "jackknife")
.R <- length(resample_jack)
}
### Start resampling loop ====================================================
progressr::with_progress({
if(.R >= 10) {
progress_bar_csem <- progressr::progressor(steps = floor(.R / 10))
} else {
progress_bar_csem <- progressr::progressor(steps = .R)
}
Est_ls <- future.apply::future_lapply(1:.R, function(i) {
if (i %% 10 == 0) {
progress_bar_csem(message = sprintf("Resample run = %g", i))
}
# Est_ls <- lapply(1:.R, function(i) {
# Replace the old data set by a resampled data set (resampleData always returns
# a list so for just one draw we need to pick the first list element)
data_temp <- if(.resample_method == "jackknife") {
resample_jack[[i]]
} else {
# We could use resampleData here but, bootstrap resampling directly is faster
# (not surprising)
# (compared both approaches using microbenchmark)
data <- args[[".data"]]
data[sample(1:nrow(data), size = nrow(data), replace = TRUE), ]
}
args[[".data"]] <- data_temp
# Estimate model
Est_temp <- if(inherits(.object, "cSEMResults_2ndorder")) {
## NOTE: using do.call(csem, args) would be more elegant but is much
# much much! slower (especially for larger data sets).
do.call(csem, args)
} else {
# It is important to use foreman() here
# instead of csem() to allow for lapply(x, resamplecSEMResults_default) when x
# is of class cSEMResults_2ndorder.
## NOTE:
# 01.03.2019: Using do.call(foreman, args) would be more elegant but is much
# much much! slower (especially for larger data sets).
#
# 15.05.2019: Apparently do.call(foreman, args) is not that bad
# after all. I did several comparisons using microbenchmark
# but there was no speed difference (anymore?!). When I compared and
# thought that do.call is slow I fixed other parts of
# foreman as well...maybe they had been the real culprit.
# So we use do.call again, as it avoids redundancy
do.call(foreman, args)
}
# Check status
status_code <- sum(unlist(verify(Est_temp)))
# Distinguish depending on how inadmissibles should be handled
if(status_code == 0 | (status_code != 0 & .handle_inadmissibles == "ignore")) {
# ## Select relevant statistics/parameters/quantities
x1 <- selectAndVectorize(Est_temp)
### Sign change correction for PLS-PM
if(.sign_change_option != "none" && info1$Arguments$.approach_weights == "PLS-PM") {
if(inherits(.object, "cSEMResults_2ndorder")) {
est1_temp_normal <- Est_temp$First_stage$Estimates
est2_temp_normal <- Est_temp$Second_stage$Estimates
### Individual_reestimate and construct_reestimate -------------------
if(.sign_change_option == "individual_reestimate" |
.sign_change_option == "construct_reestimate") {
## First stage
# Is there a sign difference between the first stage resample
# and original weight estimates. If so, which weights differ?
sign_diff1 <- sign(est1_normal$Weight_estimates) !=
sign(est1_temp_normal$Weight_estimates)
# Only if at least one sign differs a correction is needed
if(sum(sign_diff1) != 0) {
W1_new <- est1_temp_normal$Weight_estimates
# Individual_reestimate
if(.sign_change_option == "individual_reestimate"){
W1_new[sign_diff1] <- W1_new[sign_diff1] * (-1)
} # END individual_reestimate
if(.sign_change_option == "construct_reestimate"){
# In line with Tenenhaus et al. (2005, p 177) loadings are used,
# although they have not considered 2nd order models
Lambda1 <- est1_normal$Loading_estimates
Lambda1_temp <- est1_temp_normal$Loading_estimates
Lambda_diff1 <- abs(rowSums(Lambda1 - Lambda1_temp))
Lambda_sum1 <- abs(rowSums(Lambda1 + Lambda1_temp))
W1_new[Lambda_diff1 > Lambda_sum1, ] <- W1_new[Lambda_diff1 > Lambda_sum1, ] * (-1)
} # END construct_reestimate
# Put corrected weights in a list to be able to supply them as
# fixed weights to .PLS_modes
W1_list <- lapply(1:nrow(W1_new), function(x) {
temp <- W1_new[x, ]
temp[temp !=0]
})
names(W1_list) <- rownames(W1_new)
## Replace modes by fixed sign-corrected weights
args[[".PLS_modes"]] <- W1_list
## Reestimate
# Note: calling csem() directly is faster, however, i guess
# change option wont be used that often, so for now, I will
# use the more elegant, while slower, do.call construction.
Est_temp <- do.call(csem, args)
## Update
est2_temp_normal <- Est_temp$Second_stage$Estimates
} # END first stage
## Second stage
# Is there a sign difference between the second stage (sign-corrected)
# resample weight estimates and the original weight estimates.
# If so, which weights differ?
sign_diff2 <- sign(est2_normal$Weight_estimates) !=
sign(est2_temp_normal$Weight_estimates)
# Only if at least one sign differs a correction is needed
if(sum(sign_diff2) != 0) {
W2_new <- est2_temp_normal$Weight_estimates
# Individual_reestimate
if(.sign_change_option == "individual_reestimate"){
W2_new[sign_diff2] <- W2_new[sign_diff2] * (-1)
} # END individual_reestimate
if(.sign_change_option == "construct_reestimate"){
# In line with Tenenhaus et al. (2005, p 177) loadings are used,
# although they have not considered 2nd order models
Lambda2 <- est2_normal$Loading_estimates
Lambda2_temp <- est2_temp_normal$Loading_estimates
Lambda_diff2 <- abs(rowSums(Lambda2 - Lambda2_temp))
Lambda_sum2 <- abs(rowSums(Lambda2 + Lambda2_temp))
W2_new[Lambda_diff2 > Lambda_sum2, ] <- W2_new[Lambda_diff2 > Lambda_sum2, ] * (-1)
} # END construct_reestimate
# Put corrected weights in a list to be able to supply them as
# fixed weights to .PLS_modes
W2_list <- lapply(1:nrow(W2_new), function(x){
temp <- W2_new[x, ]
temp[temp != 0]
})
names(W2_list) <- rownames(W2_new)
args[[".PLS_modes"]] = W2_list
## Reestimate
# Note: calling csem() directly is faster, however, i guess
# change option wont be used that often, so for now, I will
# use the more elegant, while slower, do.call construction.
Est_temp <- do.call(csem, args)
} # END second stage
# ## Update using estimates based on the sign-corrected weights
x1 <- selectAndVectorize(Est_temp)
} # END individual_reestimate, construct_reestimate
if(.sign_change_option == "individual") {
# Reverse the signs off ALL parameter estimates in a bootstrap run if
# their sign differs from the sign of the original estimation
x1 <- reverseSign(.Est_temp = Est_temp, .summary_original = summary_original)
}
} else {
est_temp_normal <- Est_temp$Estimates
sign_diff <- sign(est_normal$Weight_estimates) !=
sign(est_temp_normal$Weight_estimates)
# Is there a difference in the signs of th weights? Otherwise no correction of the signs is done
# Not sure whether this is a problem for the construct_reestimate approach which only compares the loadings
# I think not.
if(sum(sign_diff) !=0 ) {
# Sign change correction individual_reestimate and construct_reestimate
if(.sign_change_option == "individual_reestimate" |
.sign_change_option == "construct_reestimate") {
W_new <- est_temp_normal$Weight_estimates
# Individual_reestimate
if(.sign_change_option == "individual_reestimate"){
W_new[sign_diff] <- W_new[sign_diff] * (-1)
} # END individual_reestimate
if(.sign_change_option == "construct_reestimate"){
Lambda <- est_normal$Loading_estimates
Lambda_temp <- est_temp_normal$Loading_estimates
Lambda_diff <- abs(rowSums(Lambda - Lambda_temp))
Lambda_sum <- abs(rowSums(Lambda + Lambda_temp))
W_new[Lambda_diff > Lambda_sum, ] <- W_new[Lambda_diff > Lambda_sum, ] * (-1)
}
# Put corrected weights in a list to be able to supply them as
# fixed weights to .PLS_modes
W_list <- lapply(1:nrow(W_new), function(x){
temp <- W_new[x, ]
temp[temp != 0]
})
names(W_list) <- rownames(W_new)
args[[".PLS_modes"]] = W_list
## Reestimate
# Note: calling csem() directly is faster, however, i guess
# change option wont be used that often, so for now, I will
# use the more elegant, while slower, do.call construction.
Est_temp <- do.call(csem, args)
## Update using estimates based on the sign-corrected weights
x1 <- selectAndVectorize(Est_temp)
} # end if individual_reestimate, construct_reestimate
if(.sign_change_option == 'individual') {
# Reverse the signs off ALL parameter estimates in a bootstrap run if
# their sign differs from the sign of the original estimation
x1 <- reverseSign(.Est_temp = Est_temp, .summary_original = summary_original)
} # END .sign_change_option == "individual"
} # END sum(sign(.object$Estimates$Weight_estimates)!=sign(Est_temp$Estimates$Weight_estimates)
} # END else (= not 2ndorder)
} # END .sign_change_option
## Apply user defined function if specified
user_funs <- if(!is.null(.user_funs)) {
applyUserFuns(Est_temp, .user_funs = .user_funs, ...)
}
# user_funs <- if(!is.null(.user_funs)) {
# if(is.function(.user_funs)) {
# list("User_fun" = c(.user_funs(Est_temp, ...)))
# } else {
# x <- lapply(.user_funs, function(f) c(f(Est_temp, ...)))
# if(is.null(names(x))) {
# names(x) <- paste0("User_fun", 1:length(x))
# }
# x
# }
# }
## Add output of the user functions to x1
if(!is.null(.user_funs)) {
x1 <- c(x1, user_funs)
}
## Resampling from a bootstrap sample is required for the
## bootstraped t-interval CI (studentized CI), hence the second run
# In the second run no sign change option is used. We can think about
# applying the same correction as in the first run
if(.resample_method2 != "none") {
Est_resamples2 <- resamplecSEMResults(
.object = Est_temp,
.R = .R2,
.handle_inadmissibles = .handle_inadmissibles2,
.resample_method = .resample_method2,
.resample_method2 = "none",
.user_funs = .user_funs,
.seed = .seed,
.force = FALSE,
.sign_change_option = "none",
...
)
x1 <- list("Estimates1" = x1, "Estimates2" = Est_resamples2)
}
} else if(status_code != 0 & .handle_inadmissibles != "ignore") {
# Retrun NA if status is not okay and .handle_inadmissibles == "drop" or
# "replace"
x1 <- NA
}
## Return
x1
}, future.seed = .seed)
# })
})
## Process data --------------------------------------------------------------
# Delete potential NA's
out <- Filter(Negate(anyNA), Est_ls)
## Replace inadmissibles
if(length(out) != .R && .resample_method == "bootstrap" &&
.handle_inadmissibles == "replace") {
n_attempt <- 0
while (length(out) < .R) {
n_attempt <- n_attempt + 1
R_new <- .R - length(out)
if(is.null(.seed)) {
.seed <- sample(.Random.seed, 1)
} else {
.seed <- .seed + 1
}
Est_replace <- resamplecSEMResultsCore(
.object = .object,
.R = R_new,
.handle_inadmissibles = "drop",
.handle_inadmissibles2= .handle_inadmissibles,
.resample_method = .resample_method,
.resample_method2 = .resample_method2,
.R2 = .R2,
.user_funs = .user_funs,
.eval_plan = "sequential", # multiprocessing will generally
# not be meaningful here since we typically
# replace only a small subset of the total runs
.seed = .seed,
.sign_change_option = .sign_change_option,
...
)
out <- c(out, Est_replace)
if(n_attempt == 10*.R) {
warning2(
"The following issue was encountered in the `resamplecSEMResults()` function:\n",
"Attempting to replace inadmissibles failed after ", 10*.R, " attempts.",
" Returning all admissible results up to the last attempt."
)
break
}
}
}
out
}
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.