https://github.com/cran/MuMIn
Tip revision: f8469f452d8a1be30d399d978de9550b4632bb51 authored by Kamil BartoĊ on 31 January 2012, 16:44:51 UTC
version 1.7.2
version 1.7.2
Tip revision: f8469f4
dredge.R
`dredge` <-
function(global.model, beta = FALSE, evaluate = TRUE, rank = "AICc",
fixed = NULL, m.max = NA, m.min = 0, subset, marg.ex = NULL,
trace = FALSE, varying, extra, ...) {
gmEnv <- parent.frame()
gmCall <- .getCall(global.model)
gmNobs <- nobs(global.model)
if (is.null(gmCall)) {
gmCall <- substitute(global.model)
if(!is.call(gmCall)) {
if(inherits(global.model, c("gamm", "gamm4")))
message("for 'gamm' models use 'MuMIn::gamm' wrapper")
stop("could not retrieve the call to 'global.model'")
}
#"For objects without a 'call' component the call to the fitting function \n",
#" must be used directly as an argument to 'dredge'.")
# NB: this is unlikely to happen:
if(!exists(as.character(gmCall[[1L]]), parent.frame(), mode="function"))
stop("could not find function '", gmCall[[1L]], "'")
} else {
# if 'update' method does not expand dots, we have a problem
# with expressions like ..1, ..2 in the call.
# So, try to replace them with respective arguments in the original call
is.dotted <- grep("^\\.\\.", sapply(as.list(gmCall), deparse))
if(length(is.dotted) > 0L) {
substGmCall <- substitute(global.model)
if(is.name(substGmCall)) {
stop("call to 'global.model' contains '...' arguments and ",
"cannot be updated: ", deparse(gmCall, control = NULL))
} else gmCall[is.dotted] <-
substitute(global.model)[names(gmCall[is.dotted])]
}
}
# *** Rank ***
rank.custom <- !missing(rank)
rankArgs <- list(...)
IC <- .getRank(rank, rankArgs)
ICName <- as.character(attr(IC, "call")[[1L]])
allTerms <- allTerms0 <- getAllTerms(global.model, intercept = TRUE,
data = eval(gmCall$data, envir = gmEnv))
# Intercept(s)
interceptLabel <- attr(allTerms, "interceptLabel")
if(is.null(interceptLabel)) interceptLabel <- "(Intercept)"
nInts <- sum(attr(allTerms, "intercept"))
#XXX: use.ranef <- FALSE
#if(use.ranef && inherits(global.model, "mer")) {
#allTerms <- c(allTerms, paste("(", attr(allTerms0, "random.terms"), ")",
#sep = ""))
#}
if(length(grep(":", all.vars(reformulate(allTerms))) > 0L))
stop("variable names in the formula cannot contain \":\"")
logLik <- .getLogLik()
# Check for na.omit
if (!is.null(gmCall$na.action) &&
as.character(gmCall$na.action) %in% c("na.omit", "na.exclude")) {
stop("'global.model' should not use 'na.action' = ", gmCall$na.action)
}
if(names(gmCall)[2L] == "") names(gmCall)[2L] <-
names(formals(deparse(gmCall[[1L]]))[1L])
# TODO: other classes: model, fixed, etc...
gmCoefNames <- fixCoefNames(names(coeffs(global.model)))
n.vars <- length(allTerms)
if(isTRUE(rankArgs$REML) || (isTRUE(.isREMLFit(global.model)) && is.null(rankArgs$REML)))
warning("comparing models fitted by REML")
if (beta && is.null(tryCatch(beta.weights(global.model), error = function(e) NULL,
warning = function(e) NULL))) {
warning("do not know how to calculate beta weights for ",
class(global.model)[1L], ", argument 'beta' ignored")
beta <- FALSE
}
m.max <- if (missing(m.max)) (n.vars - nInts) else min(n.vars - nInts, m.max)
# fixed variables:
if (!is.null(fixed)) {
if (inherits(fixed, "formula")) {
if (fixed[[1L]] != "~" || length(fixed) != 2L)
warning("'fixed' should be a one-sided formula")
fixed <- c(getAllTerms(fixed))
} else if (!is.character(fixed)) {
stop ("'fixed' should be either a character vector with"
+ " names of variables or a one-sided formula")
}
if (!all(fixed %in% allTerms)) {
warning("not all terms in 'fixed' exist in 'global.model'")
fixed <- fixed[fixed %in% allTerms]
}
}
fixed <- c(fixed, allTerms[allTerms %in% interceptLabel])
n.fixed <- length(fixed)
termsOrder <- order(allTerms %in% fixed)
allTerms <- allTerms[termsOrder]
gmFormulaEnv <- environment(as.formula(formula(global.model), env = gmEnv))
# TODO: gmEnv <- gmFormulaEnv ???
### BEGIN:
## varying BEGIN
if(!missing(varying) && !is.null(varying)) {
nvarying <- length(varying)
varying.names <- names(varying)
fvarying <- unlist(varying, recursive = FALSE)
vlen <- vapply(varying, length, 1L)
nvariants <- prod(vlen)
variants <- as.matrix(expand.grid(split(seq_len(sum(vlen)),
rep(seq_along(varying), vlen))))
} else {
variants <- varying.names <- NULL
nvariants <- 1L
nvarying <- 0L
}
## varying END
## extra BEGIN
if(!missing(extra) && length(extra) != 0L) {
extraNames <- sapply(extra, function(x) switch(mode(x),
call = deparse(x[[1]]), name = deparse(x), character = , x))
if(!is.null(names(extra)))
extraNames <- ifelse(names(extra) != "", names(extra), extraNames)
extra <- structure(as.list(unique(extra)), names = extraNames)
if(any(c("adjR^2", "R^2") %in% extra)) {
null.fit <- null.fit(global.model, TRUE, gmFormulaEnv)
extra[extra == "R^2"][[1L]] <- function(x) r.squaredLR(x, null.fit)
extra[extra == "adjR^2"][[1L]] <-
function(x) attr(r.squaredLR(x, null.fit), "adj.r.squared")
}
extra <- sapply(extra, match.fun, simplify = FALSE)
applyExtras <- function(x) unlist(lapply(extra, function(f) f(x)))
extraResult <- applyExtras(global.model)
if(!is.numeric(extraResult))
stop("function in 'extra' returned non-numeric result")
nextra <- length(extraResult)
extraNames <- names(extraResult)
} else {
nextra <- 0L
extraNames <- character(0L)
}
## extra END
nov <- as.integer(n.vars - n.fixed)
ncomb <- (2L ^ nov) * nvariants
if(nov > 31L) stop(gettextf("maximum number of predictors is 31, but %d is given", nov))
#if(nov > 10L) warning(gettextf("%d predictors will generate up to %.0f combinations", nov, ncomb))
nmax <- ncomb * nvariants
if(evaluate) {
ret.nchunk <- 25L
ret.ncol <- n.vars + nvarying + 3L + nextra
ret <- matrix(NA_real_, ncol = ret.ncol, nrow = ret.nchunk)
coefTables <- vector(ret.nchunk, mode = "list")
} else {
ret.nchunk <- nmax
}
calls <- vector(mode = "list", length = ret.nchunk)
if(hasSubset <- !missing(subset)) {
if(!tryCatch(is.language(subset), error = function(e) FALSE))
subset <- substitute(subset)
if(inherits(subset, "formula")) {
if (subset[[1L]] != "~" || length(subset) != 2L)
stop("'subset' should be a one-sided formula")
subset <- subset[[2L]]
}
if(!all(all.vars(subset) %in% allTerms))
warning("not all terms in 'subset' exist in 'global.model'")
}
comb.sfx <- rep(TRUE, n.fixed)
comb.seq <- if(nov != 0L) seq_len(nov) else 0L
k <- 0L
extraResult1 <- integer(0L)
ord <- integer(ret.nchunk)
argsOptions <- list(
response = attr(allTerms0, "response"),
intercept = nInts,
interceptLabel = interceptLabel,
random = attr(allTerms0, "random"),
gmCall = gmCall,
gmEnv = gmEnv,
allTerms = allTerms0,
gmCoefNames = gmCoefNames,
gmDataHead = if(!is.null(gmCall$data)) {
if(eval(call("is.data.frame", gmCall$data), gmEnv))
eval(call("head", gmCall$data, 1L), gmEnv) else gmCall$data
} else NULL,
gmFormulaEnv = gmFormulaEnv
)
# XXX XXX: adjusting 'marg.ex'
# newArgs <- makeArgs(global.model, allTerms, rep(TRUE, length(allTerms), argsOptions)
# formulaList <- if(is.null(attr(newArgs, "formulaList"))) newArgs else
# attr(newArgs, "formulaList")
# marg.ex <- unlist(lapply(lapply(formulaList, formulaAllowed), "attr", "marg.ex"))
retColIdx <- if(nvarying) -n.vars - seq_len(nvarying) else TRUE
prevJComb <- 0L
for(iComb in seq.int(ncomb)) {
jComb <- ceiling(iComb / nvariants)
if(jComb != prevJComb) {
isok <- TRUE
prevJComb <- jComb
comb <- c(as.logical(intToBits(jComb - 1L)[comb.seq]), comb.sfx)
nvar <- sum(comb) - nInts
if(nvar > m.max || nvar < m.min || (hasSubset &&
!eval(subset, structure(as.list(comb), names = allTerms)))) {
isok <- FALSE
next;
}
newArgs <- makeArgs(global.model, allTerms[comb], comb, argsOptions)
formulaList <- if(is.null(attr(newArgs, "formulaList"))) newArgs else
attr(newArgs, "formulaList")
if(!all(vapply(formulaList, formulaAllowed, logical(1L), marg.ex))) {
isok <- FALSE; next;
}
if(!is.null(attr(newArgs, "problems"))) {
print.warnings(structure(vector(mode = "list",
length = length(attr(newArgs, "problems"))),
names = attr(newArgs, "problems")))
} # end if <problems>
cl <- gmCall
cl[names(newArgs)] <- newArgs
} # end if(jComb != prevJComb)
if(!isok) next;
## --- Variants ---------------------------
clVariant <- cl
if (nvarying) clVariant[varying.names] <-
fvarying[variants[(iComb - 1L) %% nvariants + 1L, ]]
if(trace) {
cat(iComb, ": "); print(clVariant)
utils::flush.console()
}
if(evaluate) {
# begin row1: (clVariant, gmEnv, modelId, IC(), applyExtras(),
# nextra, allTerms, beta,
# if(nvarying) variantsIdx[v] else NULL
fit1 <- tryCatch(eval(clVariant, gmEnv), error = function(err) {
err$message <- paste(conditionMessage(err), "(model",
iComb, "skipped)", collapse = "")
class(err) <- c("simpleError", "warning", "condition")
warning(err)
return(NULL)
})
if (is.null(fit1)) next;
if(nextra != 0L) {
extraResult1 <- applyExtras(fit1)
if(length(extraResult1) < nextra) {
tmp <- rep(NA_real_, nextra)
tmp[match(names(extraResult1), names(extraResult))] <- extraResult1
extraResult1 <- tmp
}
#row1 <- c(row1, extraResult1)
}
mcoef1 <- matchCoef(fit1, all.terms = allTerms, beta = beta,
allCoef = TRUE)
ll <- logLik(fit1)
nobs1 <- nobs(fit1)
if(nobs1 != gmNobs) warning(gettextf(
"number of observations in model #%d (%d) differs from that in the global model (%d)",
iComb, nobs1, gmNobs))
row1 <- c(mcoef1[allTerms], extraResult1,
df = attr(ll, "df"), ll = ll, ic = IC(fit1)
)
## end -> row1
k <- k + 1L # all OK, add model to table
ret.nrow <- nrow(ret)
if(k > ret.nrow) { # append if necesarry
nadd <- min(ret.nchunk, nmax - ret.nrow)
coefTables <- c(coefTables, vector(nadd, mode = "list"))
ret <- rbind(ret, matrix(NA, ncol = ret.ncol, nrow = nadd),
deparse.level = 0L)
calls <- c(calls, vector("list", nadd))
ord <- c(ord, integer(nadd))
}
ord[k] <- iComb
ret[k, retColIdx] <- row1
coefTables[[k]] <- attr(mcoef1, "coefTable")
} else { # if evaluate
k <- k + 1L # all OK, add model to table
}
calls[[k]] <- clVariant
} ### for (iComb ...)
if(k == 0L) stop("the result is empty")
names(calls) <- ord
if(!evaluate) return(calls[seq_len(k)])
if(k < nrow(ret)) {
i <- seq_len(k)
ret <- ret[i, , drop = FALSE]
ord <- ord[i]
calls <- calls[i]
coefTables <- coefTables[i]
}
if(nvarying) {
varlev <- ord %% nvariants; varlev[varlev == 0L] <- nvariants
ret[, n.vars + seq_len(nvarying)] <- variants[varlev, ]
}
ret <- as.data.frame(ret)
row.names(ret) <- ord
# Convert columns with presence/absence of terms to factors
tfac <- which(!(allTerms %in% gmCoefNames))
ret[tfac] <- lapply(ret[tfac], factor, levels = NaN, labels="+")
i <- seq_along(allTerms)
v <- order(termsOrder)
ret[, i] <- ret[, v]
allTerms <- allTerms[v]
colnames(ret) <- c(allTerms, varying.names, extraNames, "df", "logLik", ICName)
if(nvarying) {
variant.names <- lapply(varying, function(x)
make.unique(if(is.null(names(x))) as.character(x) else names(x)))
for (i in varying.names) ret[, i] <-
factor(ret[, i], labels = variant.names[[i]])
}
o <- order(ret[, ICName], decreasing = FALSE)
ret <- ret[o, ]
coefTables <- coefTables[o]
ret$delta <- ret[, ICName] - min(ret[, ICName])
ret$weight <- exp(-ret$delta / 2) / sum(exp(-ret$delta / 2))
structure(ret,
class = c("model.selection", "data.frame"),
calls = calls[o],
global = global.model,
global.call = gmCall,
terms = structure(allTerms, interceptLabel = interceptLabel),
rank = IC,
rank.call = attr(IC, "call"),
beta = beta,
call = match.call(expand.dots = TRUE),
coefTables = coefTables,
nobs = gmNobs,
vCols = varying.names
)
} ######