Nothing
R/glm.sdf.R
In EdSurvey: Analysis of NCES Education Survey and Assessment Data
Defines functions
calc.glm.sdf
probit.sdf
logit.sdf
glm.sdf
Documented in
glm.sdf
logit.sdf
probit.sdf
#' @title EdSurvey Generalized Linear Models
#'
#' @description Fits a logit or probit that
#' uses weights and variance estimates
#' appropriate for the \code{edsurvey.data.frame},
#' the \code{light.edsurvey.data.frame}, or the \code{edsurvey.data.frame.list}.
#'
#' @param formula a \ifelse{latex}{\code{formula}}{\code{\link[stats]{formula}}} for the
#' linear model. See \ifelse{latex}{\code{glm}}{\code{\link[stats]{glm}}}.
#' For logit and probit, we recommend using the \code{I()} function
#' to define the level used for success. (See Examples.)
#' @param data an \code{edsurvey.data.frame}
#' @param family the \code{glm.sdf} function currently fits only the binomial
#' outcome models, such as logit and probit, although other link
#' functions are available for binomial models. See the \code{link}
#' argument in the help for
#' \ifelse{latex}{\code{family}}{\code{\link[stats]{family}}}.
#' @param jrrIMax the \code{Vjrr} sampling variance term (see
#' \href{https://www.air.org/sites/default/files/EdSurvey-Statistics.pdf}{\emph{Statistical Methods Used in EdSurvey}})
#' can be estimated with
#' any positive number of plausible values and is estimated on
#' the lower
#' of the number of available plausible values and \code{jrrIMax}. When
#' \code{jrrIMax} is set to \code{Inf}, all plausible values will be used.
#' Higher values of \code{jrrIMax} lead to longer computing times and more
#' accurate variance estimates.
#' @param weightVar character indicating the weight variable to use (see Details).
#' The \code{weightVar} must be one of the weights for the
#' \code{edsurvey.data.frame}. If \code{NULL}, uses the default
#' for the \code{edsurvey.data.frame}.
#' @param relevels a list; used to change the contrasts from the
#' default treatment contrasts to the treatment contrasts with a chosen omitted
#' group. The name of each element should be the variable name, and the value
#' should be the group to be omitted.
#' @param varMethod a character set to \dQuote{jackknife} or \dQuote{Taylor} that indicates the variance
#' estimation method to be used. See Details.
#' @param dropOmittedLevels a logical value. When set to the default value of \code{TRUE}, drops
#' those levels of all factor variables that are specified
#' in \code{edsurvey.data.frame}. Use \code{print} on an
#' \code{edsurvey.data.frame} to see the omitted levels.
#' @param defaultConditions a logical value. When set to the default value of \code{TRUE}, uses
#' the default conditions stored in an \code{edsurvey.data.frame}
#' to subset the data. Use \code{print} on an
#' \code{edsurvey.data.frame} to see the default conditions.
#' @param recode a list of lists to recode variables. Defaults to \code{NULL}. Can be set as
#' \code{recode=} \code{list(}\code{var1=} \code{list(from=} \code{c("a",} \code{"b",} \code{"c"),} \code{to=}\code{"d"))}.
#' @param returnNumberOfPSU a logical value set to \code{TRUE} to return the number of
#' primary sampling units (PSUs)
#' @param returnVarEstInputs a logical value set to \code{TRUE} to return the
#' inputs to the jackknife and imputation variance
#' estimates, which allow for
#' the computation
#' of covariances between estimates.
#'
#' @param omittedLevels this argument is deprecated. Use \code{dropOmittedLevels}
#'
#' @details
#' This function implements an estimator that correctly handles left-hand side
#' variables that are logical, allows for survey sampling weights, and estimates
#' variances using the jackknife replication or Taylor series.
#' The vignette titled
#' \href{https://www.air.org/sites/default/files/EdSurvey-Statistics.pdf}{\emph{Statistical Methods Used in EdSurvey}}
#' describes estimation of the reported statistics and how it depends on \code{varMethod}.
#'
#' The coefficients are estimated
#' using the sample weights according to the section
#' \dQuote{Estimation of Weighted Means When Plausible Values Are Not Present}
#' or the section
#' \dQuote{Estimation of Weighted Means When Plausible Values Are Present,}
#' depending on if there are assessment variables or variables with plausible values
#' in them.
#'
#' How the standard errors of the coefficients are estimated depends on the
#' presence of plausible values (assessment variables),
#' But once it is obtained, the \emph{t} statistic
#' is given by \deqn{t=\frac{\hat{\beta}}{\sqrt{\mathrm{var}(\hat{\beta})}}} where
#' \eqn{ \hat{\beta} } is the estimated coefficient and \eqn{\mathrm{var}(\hat{\beta})} is
#' its variance of that estimate.
#'
#' \code{logit.sdf} and \code{probit.sdf} are included for convenience only;
#' they give the same results as a call to \code{glm.sdf} with the binomial family
#' and the link function named in the function call (logit or probit).
#' By default, \code{glm} fits a logistic regression when \code{family} is not set,
#' so the two are expected to give the same results in that case.
#' Other types of generalized linear models are not supported.
#'
#' \subsection{Variance estimation of coefficients}{
#' All variance estimation methods are shown in the vignette titled
#' \href{https://www.air.org/sites/default/files/EdSurvey-Statistics.pdf}{\emph{Statistical Methods Used in EdSurvey}}.
#' When the predicted
#' value does not have plausible values and \code{varMethod} is set to
#' \code{jackknife}, the variance of the coefficients
#' is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Not Present, Using the Jackknife Method.}
#'
#' When plausible values are present and \code{varMethod} is set to
#' \code{jackknife}, the
#' variance of the coefficients is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Present, Using the Jackknife Method.}
#'
#' When the predicted
#' value does not have plausible values and \code{varMethod} is set to
#' \code{Taylor}, the variance of the coefficients
#' is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Not Present, Using the Taylor Series Method.}
#'
#' When plausible values are present and \code{varMethod} is set to
#' \code{Taylor}, the
#' variance of the coefficients is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Present, Using the Taylor Series Method.}
#' }
#'
#' @section Testing:
#' Of the common hypothesis tests for joint parameter testing, only the Wald
#' test is widely used with plausible values and sample weights. As such, it
#' replaces, if imperfectly, the Akaike Information Criteria (AIC), the
#' likelihood ratio test, chi-squared, and analysis of variance (ANOVA, including \emph{F}-tests).
#' See \code{\link{waldTest}} or
#' the vignette titled
#' \href{https://www.air.org/sites/default/files/EdSurvey-WaldTest.pdf}{\emph{Methods and Overview of Using EdSurvey for Running Wald Tests}}.
#'
#' @aliases logit.sdf probit.sdf
#'
#' @return
#' An \code{edsurveyGlm} with the following elements:
#' \item{call}{the function call}
#' \item{formula}{the formula used to fit the model}
#' \item{coef}{the estimates of the coefficients}
#' \item{se}{the standard error estimates of the coefficients}
#' \item{Vimp}{the estimated variance caused by uncertainty in the scores (plausible value variables)}
#' \item{Vjrr}{the estimated variance from sampling}
#' \item{M}{the number of plausible values}
#' \item{nPSU}{the number of PSUs used in the calculation}
#' \item{varm}{the variance estimates under the various plausible values}
#' \item{coefm}{the values of the coefficients under the various plausible values}
#' \item{coefmat}{the coefficient matrix (typically produced by the summary of a model)}
#' \item{weight}{the name of the weight variable}
#' \item{npv}{the number of plausible values}
#' \item{njk}{the number of the jackknife replicates used}
#' \item{varMethod}{always \code{jackknife}}
#' \item{varEstInputs}{when \code{returnVarEstInputs} is \code{TRUE},
#' this element is returned. These are
#' used for calculating covariances with
#' \code{\link{varEstToCov}}.}
#'
#' @seealso \ifelse{latex}{\code{glm}}{\code{\link[stats]{glm}}}
#' @author Paul Bailey
#'
#' @example man/examples/glm.sdf.R
#' @importFrom stats binomial predict
#' @importFrom glm2 glm2 glm.fit2
#' @importFrom Matrix rankMatrix
#' @export
#' @usage
#' glm.sdf(formula, family = binomial(link = "logit"), data,
#' weightVar = NULL, relevels = list(),
#' varMethod=c("jackknife", "Taylor"), jrrIMax = 1,
#' dropOmittedLevels = TRUE, defaultConditions = TRUE, recode = NULL,
#' returnNumberOfPSU=FALSE, returnVarEstInputs = FALSE,
#' omittedLevels = deprecated())
#'
glm.sdf <- function(formula,
family = binomial(link = "logit"),
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
dropOmittedLevels = TRUE,
defaultConditions = TRUE,
recode = NULL,
returnNumberOfPSU = FALSE,
returnVarEstInputs = FALSE,
omittedLevels = deprecated()) {
call <- match.call()
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
if (lifecycle::is_present(omittedLevels)) {
lifecycle::deprecate_soft("4.0.0", "glm.sdf(omittedLevels)", "glm.sdf(dropOmittedLevels)")
dropOmittedLevels <- omittedLevels
}
if (!is.logical(dropOmittedLevels)) stop("The ", sQuote("dropOmittedLevels"), " argument must be logical.")
# if data is an edsurvey.data.frame.list, simply return a list with results
# for each edsurvey.data.frame
if (inherits(data, c("edsurvey.data.frame.list"))) {
res <- itterateESDFL(match.call(), data)
class(res) <- "edsurveyGlmList"
return(res)
} else {
return(calc.glm.sdf(
formula = formula,
family = family,
data = data,
weightVar = weightVar,
relevels = relevels,
varMethod = varMethod,
jrrIMax = jrrIMax,
omittedLevels = dropOmittedLevels,
defaultConditions = defaultConditions,
missingDefaultConditions = missing(defaultConditions),
recode = recode,
returnVarEstInputs = returnVarEstInputs,
returnLm0 = FALSE,
returnNumberOfPSU = returnNumberOfPSU,
call = call
))
}
}
#' @rdname glm.sdf
#' @export
logit.sdf <- function(formula,
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
omittedLevels = TRUE,
defaultConditions = TRUE,
recode = NULL,
returnNumberOfPSU = FALSE,
returnVarEstInputs = FALSE) {
call <- match.call()
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
# if data is an edsurvey.data.frame.list, simply return a list with results
# for each edsurvey.data.frame
if (inherits(data, c("edsurvey.data.frame.list"))) {
res <- itterateESDFL(match.call(), data)
class(res) <- "edsurveyGlmList"
return(res)
} else {
return(calc.glm.sdf(
formula = formula,
family = binomial(link = "logit"),
data = data,
weightVar = weightVar,
relevels = relevels,
varMethod = varMethod,
jrrIMax = jrrIMax,
omittedLevels = omittedLevels,
defaultConditions = defaultConditions,
missingDefaultConditions = missing(defaultConditions),
recode = recode,
returnVarEstInputs = returnVarEstInputs,
returnNumberOfPSU = returnNumberOfPSU,
returnLm0 = FALSE,
call = call
))
}
}
#' @rdname glm.sdf
#' @export
probit.sdf <- function(formula,
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
omittedLevels = TRUE,
defaultConditions = TRUE,
recode = NULL,
returnVarEstInputs = FALSE) {
call <- match.call()
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
# if data is an edsurvey.data.frame.list, simply return a list with results
# for each edsurvey.data.frame
if (inherits(data, c("edsurvey.data.frame.list"))) {
res <- itterateESDFL(match.call(), data)
class(res) <- "edsurveyGlmList"
return(res)
} else {
return(calc.glm.sdf(
formula = formula,
family = binomial(link = "probit"),
data = data,
weightVar = weightVar,
relevels = relevels,
varMethod = varMethod,
jrrIMax = jrrIMax,
omittedLevels = omittedLevels,
defaultConditions = defaultConditions,
missingDefaultConditions = missing(defaultConditions),
recode = recode,
returnVarEstInputs = returnVarEstInputs,
returnLm0 = FALSE,
call = call
))
}
}
calc.glm.sdf <- function(formula,
family = binomial(link = "logit"),
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
omittedLevels = TRUE,
defaultConditions = TRUE,
missingDefaultConditions = TRUE,
recode = NULL,
returnVarEstInputs = FALSE,
returnLm0 = FALSE,
returnNumberOfPSU = FALSE,
call = NULL) {
if (is.null(call)) {
call <- match.call()
}
varMethod <- substr(tolower(match.arg(varMethod)), 0, 1)
if (!family$family %in% c("binomial", "quasibinomial")) {
stop("Only fits binomial models.")
}
#############################
######### Outline: #########
#############################
# 1) check, format inputs
# 2) get the data
# 3) relevel
# 4) yvar could have plausible values
# 5) run the main regression
# 6) form the inputs for variance estimation
# 7) form output, including final variance estimation
# 1) check, format inputs
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
# varMethod always jackknife
# if the weight var is not set, use the default
wgt <- checkWeightVar(data, weightVar)
# check if there is an outcome variable and set it to the default if it is missing
zeroLengthLHS <- attr(terms(formula), "response") == 0
if (zeroLengthLHS) {
yvar <- attributes(getAttributes(data, "pvvars"))$default
formula <- update(formula, new = substitute(yvar ~ ., list(yvar = as.name(yvar))))
} else {
yvar <- all.vars(formula[[2]])
} # End of If/Else: if (zeroLengthLHS)
# grab the variables needed for the Taylor series method, if that is the variance estimation method being used
taylorVars <- c()
psuVar <- getPSUVar(data, weightVar = wgt)
stratumVar <- getStratumVar(data, weightVar = wgt)
# give an error if these do not exist.
tv <- checkTaylorVars(psuVar, stratumVar, wgt, varMethod, returnNumberOfPSU)
psuVar <- tv$psuVar
stratumVar <- tv$stratumVar
varMethod <- tv$varMethod
returnNumberOfPSU <- tv$returnNumberOfPSU
if (varMethod == "t") {
taylorVars <- c(psuVar, stratumVar)
jrrIMax <- NA
} else {
if (all(c(psuVar, stratumVar) %in% colnames(data))) {
taylorVars <- c(psuVar, stratumVar)
}
}
getDataVarNames <- c(all.vars(formula), wgt, taylorVars)
tryCatch(getDataVarNames <- unique(c(getDataVarNames, PSUStratumNeeded(returnNumberOfPSU, data))),
error = function(e) {
warning(paste0("Stratum and PSU variables are required for this call and are not on the incoming data. Ignoring ", dQuote("returnNumberOfPSU=TRUE"), "."))
returnNumberOfPSU <<- FALSE
}
)
# 2) get the data
# This is most of the arguments
getDataArgs <- list(
data = data,
varnames = getDataVarNames,
returnJKreplicates = (varMethod == "j"),
drop = FALSE,
omittedLevels = omittedLevels,
recode = recode,
includeNaLabel = TRUE,
dropUnusedLevels = TRUE
)
# Default conditions should be included only if the user set it. This adds the argument only if needed
if (!missingDefaultConditions) {
getDataArgs <- c(getDataArgs, list(defaultConditions = defaultConditions))
}
# edf is the actual data
edf <- do.call(getData, getDataArgs)
relVars <- colnames(edf)
relVars <- relVars[!relVars %in% taylorVars]
incomplete <- !complete.cases(edf[ , relVars])
if (any(incomplete)) {
warning("Removing ", sum(incomplete), " rows with NAs from analysis.")
edf <- edf[!incomplete, ]
}
# if doing Taylor series, check Taylor vars too
if (varMethod == "t") {
incomplete <- !complete.cases(edf[ , taylorVars])
if (any(incomplete)) {
if (any(incomplete[!edf[ , wgt] %in% c(NA, 0)])) {
warning("Removing ", sum(incomplete), " rows with NA PSU or stratum variables from analysis.")
}
edf <- edf[!incomplete, ]
}
}
# remove non-positive (full sample) weights
if (any(edf[ , wgt] <= 0)) {
posWeights <- edf[ , wgt] > 0
warning("Removing ", sum(!posWeights), " rows with nonpositive weight from analysis.")
edf <- edf[posWeights, ]
}
if (varMethod == "t" && sum(is.na(edf[ , c(stratumVar, psuVar)])) != 0) {
stop("Taylor series variance estimation requested but stratum or PSU variables contian an NA value.")
}
# check that there are not factors with only one level and give error message
lapply(
names(edf)[names(edf) %in% all.names(formula)],
function(z) {
if (is.factor(edf[ , z]) & length(levels(edf[ , z])) < 2) {
stop(paste0(
"The covariate ",
dQuote(z),
" has fewer than two levels and cannot be included in the regression."
))
}
}
)
# check that there is some data to work with
if (nrow(edf) <= 0) {
stop(paste0(
sQuote("data"), " must have more than 0 rows after a call to ",
sQuote("getData"), "."
))
}
# 3) deal with relevels.
# An argument that allows the user to change the omitted level for a factor variable
if (length(relevels) > 0) {
for (i in 1:length(relevels)) {
vari <- names(relevels)[i]
if (!vari %in% names(edf)) {
stop(paste0(
"In the ", sQuote("relevels"),
" argument, cannot find the variable named ", sQuote(vari), "."
))
} # End of if statment: ! vari %in% names(edf)
if (length(relevels[[i]]) != 1) {
stop(paste0(
"In the ", sQuote("relevels"),
" argument, each relevel must have exactly one level."
))
} # End of if statment: length(relevels[[i]]) != 1
# check that the level exists
lvls <- levels(edf[ , vari])
if (inherits(edf[ , vari], "lfactor")) {
# for a factor it can be either a level or a label
lvls <- c(lvls, labels(edf[ , vari]))
} # End of if statment: inherits(edf[ ,vari], "lfactor")
if (!relevels[[i]] %in% lvls) {
stop(paste0(
"In the ", sQuote("relevels"),
" argument, for the variable ", dQuote(vari), ", the level ",
dQuote(relevels[[i]]), " not found. Found levels are ",
pasteItems(dQuote(lvls)), "."
))
} # End of if statment !relevels[[i]] %in% lvls
edf[ , vari] <- relevel(edf[ , vari], ref = relevels[[i]])
} # end for(i in 1:length(relevels))
} # end if(length(relevels) > 0)
# 4) yvar and plausible values
pvy <- hasPlausibleValue(yvar, data) # pvy is the plausible values of the y variable
yvars <- yvar
linkingError <- detectLinkingError(data, yvars)
if(linkingError){
if(varMethod == "t") {
stop("Taylor series variance estimation not supported with NAEP linking error.")
}
if (jrrIMax != 1) {
warning("The linking error variance estimator only supports ", dQuote("jrrIMax=1"), ". Resetting to 1.")
jrrIMax <- 1
}
}
lyv <- length(yvars)
if (any(pvy)) {
if (linkingError) {
pp <- getPlausibleValue(yvars[max(pvy)], data)
suffix <- ifelse(grepl("_imp_", pp, fixed = TRUE), "_imp_", "_est_")
suffix <- ifelse(grepl("_samp_", pp, fixed = TRUE), "_samp_", suffix)
yvars <- paste0("outcome", suffix, 1:length(pp))
} else {
yvars <- paste0("outcome", 1:length(getPlausibleValue(yvars[max(pvy)], data)))
}
} else {
# if not, make sure that this variable is numeric
edf[ , "yvar"] <- as.numeric(eval(formula[[2]], edf))
formula <- update(formula, new = substitute(yvar ~ ., list(yvar = as.name(yvar))))
yvars <- "yvar"
} # End of if statment: any(pvy)
# 5) run the main regression
# run a regression, starting with the first PV or maybe the only outcome variable
yvar0 <- yvars[1]
# this allows that variable to not be dynamic variable, it is explicitly defined to be yvar0
if (family$family %in% c("binomial", "quasibinomial")) { # if using binomial set y to 1 only if it has the highest value
if (any(pvy)) {
for (i in 1:length(yvars)) {
# PV, so we have not evaluated the I() yet (if any)
for (yvi in 1:length(pvy)) {
if (pvy[yvi]) {
edf[ , yvar[yvi]] <- edf[ , getPlausibleValue(yvar[yvi], data)[i]]
}
}
edf[ , yvars[i]] <- as.numeric(eval(formula[[2]], edf))
}
oneDef <- max(edf[ , yvars], na.rm = TRUE)
for (i in yvars) {
edf[ , i] <- ifelse(edf[ , i] %in% oneDef, 1, 0)
}
edf$yvar0 <- edf[ , yvar0]
} else {
# for non-PV, I() has been evaluated
oneDef <- max(edf[ , yvars], na.rm = TRUE)
edf$yvar0 <- ifelse(edf$yvar %in% oneDef, 1, 0)
}
}
# set up formula for initial regression with yvar0 predicted by all other vars
frm <- update(formula, yvar0 ~ .)
edf$w <- edf[ , wgt]
# get an approximate fit without weights. Starting with weights can lead to
# convergence problems.
suppressWarnings(lm00 <- glm2(frm, data = edf, family = family))
edf$c2 <- predict(lm00, type = "response")
# get a final fit, with weights. Starting at c2
suppressWarnings(lm0 <- glm2(frm, data = edf, weights = w, family = family, mustart = c2, epsilon = 1e-14))
# get final prediction
edf$c2 <- predict(lm0, type = "response")
if (returnLm0) {
return(lm0)
}
# this grabs the list of weight variable names
wgtl <- getAttributes(data, "weights")[[wgt]]
varEstInputs <- list()
# note we have not made the B matrix
madeB <- FALSE
# there is now a very long set of conditionals
# regression with each yvar
stat_reg_glm <- function(fam = FALSE) {
ref <- function(pv, wg) {
if (family$family %in% c("binomial", "quasibinomial") & (fam)) {
# do this for each yvar, but not for each jk weight (in parent function)
edf$yvar0 <- ifelse(pv %in% oneDef, 1, 0)
}
edf$w <- wg
suppressWarnings(lmi <- glm2(frm, data = edf, weights = w, family = family, mustart = c2, epsilon = 1e-14))
return(coef(lmi))
}
}
# 6) form the inputs for variance estimation
if (any(pvy)) { # the y variable is a plausible value
# this if condition estimates both Taylor series method and jackknife method
# for equations, see the statistics vignette
if (varMethod == "t") {
jrrIMax <- length(yvars)
} else {
jrrIMax <- min(jrrIMax, length(yvars))
}
# initialize: varm is the variance matrix by coefficient and PV (for V_jrr)
varM <- list()
varm <- matrix(NA, nrow = jrrIMax, ncol = length(coef(lm0)))
if (varMethod == "j") {
if (linkingError) {
repWeights <- paste0(wgtl$jkbase, wgtl$jksuffixes)
# get the mean estimate
est <- getLinkingEst(
data = edf,
pvEst = yvars[grep("_est", yvars)],
stat = stat_reg_glm(fam = TRUE),
wgt = wgt
)
# coefficients matrix, r-squared vector
coefm <- est$coef
coef <- apply(coefm, 2, mean)
coefm0 <- t(t(coefm) - coef)
# get imputation variance
impVar <- getLinkingImpVar(
data = edf,
pvImp = yvars[grep("_imp", yvars)],
ramCols = ncol(getRAM()),
stat = stat_reg_glm(fam = TRUE),
wgt = wgt,
T0 = est$est,
T0Centered = FALSE
)
# get sampling variance
sampVar <- getLinkingSampVar(edf,
pvSamp = yvars[grep("_samp", yvars)],
stat = stat_reg_glm(fam = TRUE),
rwgt = repWeights,
T0 = est$est,
T0Centered = FALSE
)
varM[[1]] <- sampVar$Bi
varEstInputs[["JK"]] <- sampVar$veiJK
varEstInputs[["PV"]] <- impVar$veiImp
# drop r.squared term
varm[1, ] <- sampVar$V
M <- length(yvars[grep("_est", yvars)])
Vimp <- impVar$V
} else {
# coefficients by PV (getEst in lm.sdf)
res <- getEst(
data = edf,
pvEst = yvars,
stat = stat_reg_glm(fam = TRUE),
wgt = wgt
)
coefm <- res$coef
varEstInputs[["JK"]] <- data.frame()
jkSumMult <- getAttributes(data, "jkSumMultiplier")
for (pvi in 1:jrrIMax) { # for each PV (up to jrrIMax)
res <- getVarEstJK(
stat = stat_reg_glm(fam = FALSE),
yvar = edf[ , yvars[pvi]],
wgtM = edf[ , paste0(wgtl$jkbase, wgtl$jksuffixes)],
co0 = coefm[pvi, ],
jkSumMult = jkSumMult,
pvName = pvi
)
varM[[pvi]] <- res$Bi
varEstInputs[["JK"]] <- rbind(varEstInputs[["JK"]], res$veiJK)
varm[pvi, ] <- res$VsampInp
} # end for(pvi in 1:jrrIMax)
# number of PVs
M <- length(yvars)
# imputaiton variance / variance due to uncertaintly about PVs
coef <- apply(coefm, 2, mean)
coefm0 <- t(t(coefm) - coef)
coefmPVByRow <- lapply(1:ncol(coefm0), function(coli) {
data.frame(
stringsAsFactors = FALSE,
PV = 1:nrow(coefm0),
variable = rep(names(coef(lm0))[coli], nrow(coefm0)),
value = coefm0[ , coli]
)
})
coefmPV <- do.call(rbind, coefmPVByRow)
varEstInputs[["PV"]] <- coefmPV
Vimp <- (M + 1) / M * apply(coefm, 2, var)
} # end else for if(linkingError)
} else { # End of if statment: varMethod == "j"
# Taylor series variance esimation, no Plausible Values
dofNum <- matrix(0, nrow = length(coef(lm0)), ncol = length(yvars))
dofDenom <- matrix(0, nrow = length(coef(lm0)), ncol = length(yvars))
# variances are calculated iteratively for all each y variable
est <- getEst(
data = edf,
pvEst = yvars,
stat = stat_reg_glm(fam = TRUE),
wgt = wgt
)
coefm <- est$coef
for (mm in 1:length(yvars)) {
edf$yvar0 <- as.numeric(edf[ , yvars[mm]])
y <- edf[ , yvars[mm]]
suppressWarnings(lmi <- glm2(frm, data = edf, weights = w, family = family, mustart = c2, epsilon = 1e-14))
coef <- b <- co0 <- coef(lmi)
D <- vcov(lmi)
eta <- predict(lmi, type = "link")
pred <- predict(lmi, type = "response")
# precision weights
# in the notation of McCullagh and Nelder, this is d(mu)/d(eta)
# evaluated at the (latent) predicted values (eta)
mu.eta <- family$mu.eta(eta)
# this is the partial of the likelihood at the unit level
X <- sparse.model.matrix(frm, edf)
uhij <- (y - pred) / (pred * (1 - pred)) * as.matrix(X) * mu.eta
# vals by stratum (singleton PSUs removed in function)
res <- getVarTaylor(uhij, edf, D, wgt, psuVar, stratumVar)
vc <- D %*% res$vv %*% D
# collect
dofNum[ , mm] <- res$nums
dofDenom[ , mm] <- res$nums2
varM[[mm]] <- as.matrix(vc)
varm[mm, ] <- as.numeric(diag(vc))
}
# number of PVs
M <- length(yvars)
# imputaiton variance / variance due to uncertaintly about PVs
coef <- apply(coefm, 2, mean)
coefm0 <- t(t(coefm) - coef)
coefmPVByRow <- lapply(1:ncol(coefm0), function(coli) {
data.frame(
stringsAsFactors = FALSE,
PV = 1:nrow(coefm0),
variable = rep(names(coef(lm0))[coli], nrow(coefm0)),
value = coefm0[ , coli]
)
})
coefmPV <- do.call(rbind, coefmPVByRow)
varEstInputs[["PV"]] <- coefmPV
Vimp <- (M + 1) / M * apply(coefm, 2, var)
} # End of if/else statment: varMethod == "j"
# calculate van Buuren B
B <- (1 / (M - 1)) * Reduce(
"+", # add up the matrix results of the sapply
sapply(1:nrow(coefm), function(q) {
# within each PV set, calculate the outer product
# (2.19 of Van Buuren)
outer(coefm0[q, ], coefm0[q, ])
}, simplify = FALSE)
)
madeB <- TRUE
# 2.18 in var Buuren
Ubar <- (1 / length(varM)) * Reduce("+", varM)
# variance due to sampling
Vjrr <- apply(varm[1:jrrIMax, , drop = FALSE], 2, mean)
} else { # end if(any(pvy))
# the y variable is not a plausible value
# this section handles jackknife and Taylor series estimation
if (varMethod == "j") {
jkSumMult <- getAttributes(data, "jkSumMultiplier")
coef <- coef(lm0)
res <- getVarEstJK(
stat = stat_reg_glm(fam = FALSE),
yvar = edf[ , yvars[1]],
wgtM = edf[ , paste0(wgtl$jkbase, wgtl$jksuffixes)],
co0 = coef,
jkSumMult = jkSumMult,
pvName = 1
)
Ubar <- res$Bi
Vjrr <- res$VsampInp
varEstInputs[["JK"]] <- res$veiJK
} else { # end if(varMethod == "j")
# Taylor series variance esimation, no PVs
X <- sparse.model.matrix(frm, edf)
y <- edf$yvar0
coef <- b <- co0 <- coef(lm0)
D <- vcov(lm0)
eta <- predict(lm0, type = "link")
pred <- predict(lm0, type = "response")
# precision weights
# in the notation of McCullagh and Nelder, this is d(mu)/d(eta)
# evaluated at the (latent) predicted values (eta)
mu.eta <- family$mu.eta(eta)
# this is the partial of the likelihood at the unit level
uhij <- (y - pred) / (pred * (1 - pred)) * as.matrix(X) * mu.eta
# vals by stratum (singleton PSUs removed in function)
res <- getVarTaylor(uhij, edf, D, wgt, psuVar, stratumVar)
vc <- D %*% res$vv %*% D
# collect
dofNum <- res$nums
dofDenom <- res$nums2
Vjrr <- as.numeric(diag(vc))
# this is the VC matrix, store it for vcov to recover
Ubar <- as.matrix(vc)
} # end else for if(varMethod == "j")
# not pvy common
M <- 1 # only on replicate when there are no PVs
Vimp <- 0 # no imputation variance when there are no PVs
varm <- NULL # no variance by PV
coefm <- NULL # no coefficients matrix by PV
B <- 0 * Ubar # no PVs
} # end else for if(any(pvy))
# 7) form output, including final variance estimation
V <- Vjrr + Vimp
names(coef) <- names(coef(lm0))
se <- sqrt(V)
names(se) <- names(coef)
# get fitted and resid based on overall coef
X <- model.matrix(frm, edf)
fittedLatent <- as.vector(X %*% coef)
fitted1 <- family$linkinv(fittedLatent)
if (linkingError) {
ye <- grep("_est_", yvars)
Y <- sapply(ye, function(yi) {
as.vector(edf[ , yvars[yi]])
}, simplify = TRUE)
resid1 <- Y - fitted1
colnames(resid1) <- yvars[ye]
} else {
Y <- sapply(1:length(yvars), function(yi) {
as.vector(edf[ , yvars[yi]])
}, simplify = TRUE)
resid1 <- Y - fitted1
colnames(resid1) <- yvars
}
# residual df calculation
nobs <- nrow(edf)
n.ok <- nobs - sum(edf$origwt == 0)
nvars <- ncol(X)
rank <- rankMatrix(X)
resdf <- n.ok - rank
df.r <- resdf
# get residual based on coef by PV
if (!is.null(coefm)) {
if (!is.matrix(coefm)) {
coefm <- t(as.matrix(coefm))
}
fitted2 <- family$linkinv(as.matrix(X %*% t(coefm)))
resid2 <- Y - fitted2
}
coefmat <- data.frame(
coef = coef,
se = se,
t = coef / se
)
if (varMethod == "t") {
m <- length(wgtl$jksuffixes)
if (inherits(dofNum, "matrix")) {
coefmat$dof <- apply((3.16 - 2.77 / sqrt(m)) * dofNum^2 / dofDenom, 1, mean)
} else {
coefmat$dof <- (3.16 - 2.77 / sqrt(m)) * dofNum^2 / dofDenom
}
} else {
coefmat$dof <- sapply(names(coef), function(cn) {
DoFCorrection(varEstA = varEstInputs, varA = cn, method = "JR")
})
}
pti <- pt(coefmat$t, df = coefmat$dof)
coefmat[ , "Pr(>|t|)"] <- 2 * pmin(pti, 1 - pti)
njk <- length(wgtl$jksuffixes)
if (varMethod == "t") {
njk <- NA
}
varmeth <- ifelse(varMethod == "t", "Taylor series", "jackknife")
res <- list(
call = call, formula = formula, family = family, coef = coef, se = se, Vimp = Vimp,
Vjrr = Vjrr, M = M, varm = varm, coefm = coefm,
coefmat = coefmat, weight = wgt,
npv = length(yvars), jrrIMax = min(jrrIMax, length(yvars)),
njk = njk, varMethod = varmeth,
residuals = resid1, fitted.values = fitted1, fittedLatent = fittedLatent, residual.df = resdf
)
if (!is.null(coefm)) {
res <- c(res, list(PV.residuals = resid2, PV.fitted.values = fitted2))
}
if (returnVarEstInputs) {
res <- c(res, list(varEstInputs = varEstInputs))
if (varMethod == "t") {
warning(paste0(
"Taylor series method not supported with the ",
sQuote("varEstInputs"), " argument set to ", dQuote("TRUE"), "."
))
}
}
if (madeB) {
# equation 2.29 in van Buuren, pp 42
tryCatch(
rbar <- (1 + 1 / M) * (1 / nrow(Ubar)) * sum(diag(B %*% solve(Ubar))),
error = function(e) {
rbar <<- (1 + 1 / M) * (1 / nrow(Ubar)) * sum(diag(B %*% MASS::ginv(Ubar)))
frm <- Formula(frm)
if (terms(frm, lhs = 0, rhs = NULL) == ~1) {
warning("A variance estimate was replaced with NA because there was no variance across strata.", call. = FALSE)
}
}
)
Ttilde <- (1 + rbar) * Ubar
res <- c(res, list(B = B, U = Ubar, rbar = rbar, Ttilde = Ttilde))
} else {
res <- c(res, list(B = B, U = Ubar))
}
if (returnNumberOfPSU) {
stratumVar <- getAttributes(data, "stratumVar")
psuVar <- getAttributes(data, "psuVar")
if ("JK1" %in% stratumVar) {
res <- c(res, list(nPSU = nrow(edf)))
} else if (all(c(stratumVar, psuVar) %in% colnames(edf))) {
if (sum(is.na(edf[ , c(stratumVar, psuVar)])) == 0) {
res <- c(res, list(nPSU = nrow(unique(edf[ , c(stratumVar, psuVar)]))))
} else {
warning("Cannot return number of PSUs because the stratum or PSU variables contain NA values.")
}
}
}
# add waldDenomBaseDof if relevant
if (!is.null(stratumVar) && !is.null(psuVar)) {
if (all(c(stratumVar, psuVar) %in% colnames(edf))) {
if (sum(is.na(edf[ , c(stratumVar, psuVar)])) == 0) {
res <- c(res, list(waldDenomBaseDof = waldDof(edf, stratumVar, psuVar)))
}
}
if ("JK1" %in% getStratumVar(data)) {
res <- c(res, list(waldDenomBaseDof = "JK1"))
}
}
res <- c(res, list(n0 = nrow2.edsurvey.data.frame(data), nUsed = nrow(edf)))
if (inherits(data, "edsurvey.data.frame")) {
res <- c(res, list(data = data))
} else {
res <- c(res, list(lm0 = lm0))
}
class(res) <- "edsurveyGlm"
return(res)
}
#' @method print edsurveyGlm
#' @export
print.edsurveyGlm <- function(x, ...) {
print(coef(x), ...)
}
#' @method print edsurveyGlmList
#' @export
print.edsurveyGlmList <- function(x, ...) {
for (i in 1:length(x)) {
cat("glm", i, "\n")
print(coef(x[[i]]), ...)
}
}
#' @method summary edsurveyGlm
#' @export
summary.edsurveyGlm <- function(object, ...) {
class(object) <- "summary.edsurveyGlm"
object
}
#' @method summary edsurveyGlmList
#' @export
summary.edsurveyGlmList <- function(object, ...) {
class(object) <- "summary.edsurveyGlmList"
for (i in 1:length(object)) {
class(object[[i]]) <- "summary.edsurveyGlm"
}
object
}
#' @method print summary.edsurveyGlm
#' @importFrom stats printCoefmat
#' @export
print.summary.edsurveyGlm <- function(x, ...) {
cat(paste0("\nFormula: ", paste(deparse(x$formula), collapse = ""), "\n"))
cat(paste0("Family: ", x$family$family, " (", x$family$link, ")\n\n"))
if (x$npv != 1) {
cat(paste0("jrrIMax: ", x$jrrIMax, "\n"))
}
cat(paste0("Weight variable: ", sQuote(x$weight), "\n"))
cat(paste0("Variance method: ", x$varMethod, "\n"))
if (!is.na(x$njk)) {
cat(paste0("JK replicates: ", x$njk, "\n"))
}
cat(paste0("full data n: ", x$n0, "\n"))
if (!is.null(x$nPSU)) {
cat(paste0("n PSU: ", x$nPSU, "\n"))
}
cat(paste0("n used: ", x$nUsed, "\n\n"))
cat(paste0("Coefficients:\n"))
printCoefmat(x$coefmat, P.values = TRUE, has.Pvalue = TRUE)
}
#' @method print summary.edsurveyGlmList
#' @export
print.summary.edsurveyGlmList <- function(x, ...) {
for (i in 1:length(x)) {
cat("glm", i, "\n")
print(x[[i]], ...)
}
}
#' @method coef edsurveyGlm
#' @export
coef.edsurveyGlm <- function(object, ...) {
object$coef
}
#' @method coef edsurveyGlmList
#' @export
coef.edsurveyGlmList <- function(object, ...) {
sapply(object, function(li) {
li$coef
})
}
#' @method vcov edsurveyGlm
#' @export
vcov.edsurveyGlm <- function(object, ...) {
vcov.edsurveyLm(object, ...)
}
#' @title Odds Ratios for edsurveyGlm Models
#' @description Converts coefficients from \code{edsurveyGlm} logit regression model to odds ratios.
#' @param model an \code{edsurveyGlm} model
#' @param alpha the alpha level for the confidence level
#'
#' @return
#' An \code{oddsRatio.edsurveyGlm} object with the following elements:
#' \item{OR}{odds ratio coefficient estimates}
#' \item{2.5\%}{lower bound 95\% confidence interval}
#' \item{97.5\%}{upper bound 95\% confidence interval}
#'
#' @export
oddsRatio <- function(model, alpha = 0.05) {
if (!inherits(model, "edsurveyGlm")) stop("x must be of class 'edsurveyGlm'.")
if (model$family$family != "binomial" & model$family$family != "quasibinomial") {
stop("Model must be binomial.")
}
if (model$family$link != "logit") {
stop("Model must be logit.")
}
OR <- exp(coef(model))
varDiag <- diag(vcov(model))
CIlo <- exp(coef(model) - (qt(1 - alpha / 2, df = model$coefmat$dof) * sqrt(varDiag)))
CIup <- exp(coef(model) + (qt(1 - alpha / 2, df = model$coefmat$dof) * sqrt(varDiag)))
r <- cbind(OR, CIlo, CIup)
pctRound <- 1
pct <- round(100 * c(alpha / 2, 1 - alpha / 2), pctRound)
while (pct[1] == 0) {
pctRound <- pctRound + 1
pct <- round(100 * c(alpha / 2, 1 - alpha / 2), pctRound)
}
colnames(r) <- c("OR", paste0(pct, "%"))
r <- as.data.frame(r)
class(r) <- c("oddsRatio.edsurveyGlm", "data.frame")
r
}
# @export
setMethod(
"coef",
c(object = "edsurveyGlm"),
coef.edsurveyGlm
)
# @export
setMethod(
"coef",
c(object = "edsurveyGlmList"),
coef.edsurveyGlmList
)
# @export
setMethod(
"vcov",
c(object = "edsurveyGlm"),
vcov.edsurveyGlm
)
Try the EdSurvey package in your browser
Any scripts or data that you put into this service are public.
EdSurvey documentation built on Nov. 2, 2023, 6:25 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 calc.glm.sdf probit.sdf logit.sdf glm.sdf
Documented in glm.sdf logit.sdf probit.sdf
#' @title EdSurvey Generalized Linear Models
#'
#' @description Fits a logit or probit that
#' uses weights and variance estimates
#' appropriate for the \code{edsurvey.data.frame},
#' the \code{light.edsurvey.data.frame}, or the \code{edsurvey.data.frame.list}.
#'
#' @param formula a \ifelse{latex}{\code{formula}}{\code{\link[stats]{formula}}} for the
#' linear model. See \ifelse{latex}{\code{glm}}{\code{\link[stats]{glm}}}.
#' For logit and probit, we recommend using the \code{I()} function
#' to define the level used for success. (See Examples.)
#' @param data an \code{edsurvey.data.frame}
#' @param family the \code{glm.sdf} function currently fits only the binomial
#' outcome models, such as logit and probit, although other link
#' functions are available for binomial models. See the \code{link}
#' argument in the help for
#' \ifelse{latex}{\code{family}}{\code{\link[stats]{family}}}.
#' @param jrrIMax the \code{Vjrr} sampling variance term (see
#' \href{https://www.air.org/sites/default/files/EdSurvey-Statistics.pdf}{\emph{Statistical Methods Used in EdSurvey}})
#' can be estimated with
#' any positive number of plausible values and is estimated on
#' the lower
#' of the number of available plausible values and \code{jrrIMax}. When
#' \code{jrrIMax} is set to \code{Inf}, all plausible values will be used.
#' Higher values of \code{jrrIMax} lead to longer computing times and more
#' accurate variance estimates.
#' @param weightVar character indicating the weight variable to use (see Details).
#' The \code{weightVar} must be one of the weights for the
#' \code{edsurvey.data.frame}. If \code{NULL}, uses the default
#' for the \code{edsurvey.data.frame}.
#' @param relevels a list; used to change the contrasts from the
#' default treatment contrasts to the treatment contrasts with a chosen omitted
#' group. The name of each element should be the variable name, and the value
#' should be the group to be omitted.
#' @param varMethod a character set to \dQuote{jackknife} or \dQuote{Taylor} that indicates the variance
#' estimation method to be used. See Details.
#' @param dropOmittedLevels a logical value. When set to the default value of \code{TRUE}, drops
#' those levels of all factor variables that are specified
#' in \code{edsurvey.data.frame}. Use \code{print} on an
#' \code{edsurvey.data.frame} to see the omitted levels.
#' @param defaultConditions a logical value. When set to the default value of \code{TRUE}, uses
#' the default conditions stored in an \code{edsurvey.data.frame}
#' to subset the data. Use \code{print} on an
#' \code{edsurvey.data.frame} to see the default conditions.
#' @param recode a list of lists to recode variables. Defaults to \code{NULL}. Can be set as
#' \code{recode=} \code{list(}\code{var1=} \code{list(from=} \code{c("a",} \code{"b",} \code{"c"),} \code{to=}\code{"d"))}.
#' @param returnNumberOfPSU a logical value set to \code{TRUE} to return the number of
#' primary sampling units (PSUs)
#' @param returnVarEstInputs a logical value set to \code{TRUE} to return the
#' inputs to the jackknife and imputation variance
#' estimates, which allow for
#' the computation
#' of covariances between estimates.
#'
#' @param omittedLevels this argument is deprecated. Use \code{dropOmittedLevels}
#'
#' @details
#' This function implements an estimator that correctly handles left-hand side
#' variables that are logical, allows for survey sampling weights, and estimates
#' variances using the jackknife replication or Taylor series.
#' The vignette titled
#' \href{https://www.air.org/sites/default/files/EdSurvey-Statistics.pdf}{\emph{Statistical Methods Used in EdSurvey}}
#' describes estimation of the reported statistics and how it depends on \code{varMethod}.
#'
#' The coefficients are estimated
#' using the sample weights according to the section
#' \dQuote{Estimation of Weighted Means When Plausible Values Are Not Present}
#' or the section
#' \dQuote{Estimation of Weighted Means When Plausible Values Are Present,}
#' depending on if there are assessment variables or variables with plausible values
#' in them.
#'
#' How the standard errors of the coefficients are estimated depends on the
#' presence of plausible values (assessment variables),
#' But once it is obtained, the \emph{t} statistic
#' is given by \deqn{t=\frac{\hat{\beta}}{\sqrt{\mathrm{var}(\hat{\beta})}}} where
#' \eqn{ \hat{\beta} } is the estimated coefficient and \eqn{\mathrm{var}(\hat{\beta})} is
#' its variance of that estimate.
#'
#' \code{logit.sdf} and \code{probit.sdf} are included for convenience only;
#' they give the same results as a call to \code{glm.sdf} with the binomial family
#' and the link function named in the function call (logit or probit).
#' By default, \code{glm} fits a logistic regression when \code{family} is not set,
#' so the two are expected to give the same results in that case.
#' Other types of generalized linear models are not supported.
#'
#' \subsection{Variance estimation of coefficients}{
#' All variance estimation methods are shown in the vignette titled
#' \href{https://www.air.org/sites/default/files/EdSurvey-Statistics.pdf}{\emph{Statistical Methods Used in EdSurvey}}.
#' When the predicted
#' value does not have plausible values and \code{varMethod} is set to
#' \code{jackknife}, the variance of the coefficients
#' is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Not Present, Using the Jackknife Method.}
#'
#' When plausible values are present and \code{varMethod} is set to
#' \code{jackknife}, the
#' variance of the coefficients is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Present, Using the Jackknife Method.}
#'
#' When the predicted
#' value does not have plausible values and \code{varMethod} is set to
#' \code{Taylor}, the variance of the coefficients
#' is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Not Present, Using the Taylor Series Method.}
#'
#' When plausible values are present and \code{varMethod} is set to
#' \code{Taylor}, the
#' variance of the coefficients is estimated according to the section
#' \dQuote{Estimation of Standard Errors of Weighted Means When
#' Plausible Values Are Present, Using the Taylor Series Method.}
#' }
#'
#' @section Testing:
#' Of the common hypothesis tests for joint parameter testing, only the Wald
#' test is widely used with plausible values and sample weights. As such, it
#' replaces, if imperfectly, the Akaike Information Criteria (AIC), the
#' likelihood ratio test, chi-squared, and analysis of variance (ANOVA, including \emph{F}-tests).
#' See \code{\link{waldTest}} or
#' the vignette titled
#' \href{https://www.air.org/sites/default/files/EdSurvey-WaldTest.pdf}{\emph{Methods and Overview of Using EdSurvey for Running Wald Tests}}.
#'
#' @aliases logit.sdf probit.sdf
#'
#' @return
#' An \code{edsurveyGlm} with the following elements:
#' \item{call}{the function call}
#' \item{formula}{the formula used to fit the model}
#' \item{coef}{the estimates of the coefficients}
#' \item{se}{the standard error estimates of the coefficients}
#' \item{Vimp}{the estimated variance caused by uncertainty in the scores (plausible value variables)}
#' \item{Vjrr}{the estimated variance from sampling}
#' \item{M}{the number of plausible values}
#' \item{nPSU}{the number of PSUs used in the calculation}
#' \item{varm}{the variance estimates under the various plausible values}
#' \item{coefm}{the values of the coefficients under the various plausible values}
#' \item{coefmat}{the coefficient matrix (typically produced by the summary of a model)}
#' \item{weight}{the name of the weight variable}
#' \item{npv}{the number of plausible values}
#' \item{njk}{the number of the jackknife replicates used}
#' \item{varMethod}{always \code{jackknife}}
#' \item{varEstInputs}{when \code{returnVarEstInputs} is \code{TRUE},
#' this element is returned. These are
#' used for calculating covariances with
#' \code{\link{varEstToCov}}.}
#'
#' @seealso \ifelse{latex}{\code{glm}}{\code{\link[stats]{glm}}}
#' @author Paul Bailey
#'
#' @example man/examples/glm.sdf.R
#' @importFrom stats binomial predict
#' @importFrom glm2 glm2 glm.fit2
#' @importFrom Matrix rankMatrix
#' @export
#' @usage
#' glm.sdf(formula, family = binomial(link = "logit"), data,
#' weightVar = NULL, relevels = list(),
#' varMethod=c("jackknife", "Taylor"), jrrIMax = 1,
#' dropOmittedLevels = TRUE, defaultConditions = TRUE, recode = NULL,
#' returnNumberOfPSU=FALSE, returnVarEstInputs = FALSE,
#' omittedLevels = deprecated())
#'
glm.sdf <- function(formula,
family = binomial(link = "logit"),
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
dropOmittedLevels = TRUE,
defaultConditions = TRUE,
recode = NULL,
returnNumberOfPSU = FALSE,
returnVarEstInputs = FALSE,
omittedLevels = deprecated()) {
call <- match.call()
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
if (lifecycle::is_present(omittedLevels)) {
lifecycle::deprecate_soft("4.0.0", "glm.sdf(omittedLevels)", "glm.sdf(dropOmittedLevels)")
dropOmittedLevels <- omittedLevels
}
if (!is.logical(dropOmittedLevels)) stop("The ", sQuote("dropOmittedLevels"), " argument must be logical.")
# if data is an edsurvey.data.frame.list, simply return a list with results
# for each edsurvey.data.frame
if (inherits(data, c("edsurvey.data.frame.list"))) {
res <- itterateESDFL(match.call(), data)
class(res) <- "edsurveyGlmList"
return(res)
} else {
return(calc.glm.sdf(
formula = formula,
family = family,
data = data,
weightVar = weightVar,
relevels = relevels,
varMethod = varMethod,
jrrIMax = jrrIMax,
omittedLevels = dropOmittedLevels,
defaultConditions = defaultConditions,
missingDefaultConditions = missing(defaultConditions),
recode = recode,
returnVarEstInputs = returnVarEstInputs,
returnLm0 = FALSE,
returnNumberOfPSU = returnNumberOfPSU,
call = call
))
}
}
#' @rdname glm.sdf
#' @export
logit.sdf <- function(formula,
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
omittedLevels = TRUE,
defaultConditions = TRUE,
recode = NULL,
returnNumberOfPSU = FALSE,
returnVarEstInputs = FALSE) {
call <- match.call()
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
# if data is an edsurvey.data.frame.list, simply return a list with results
# for each edsurvey.data.frame
if (inherits(data, c("edsurvey.data.frame.list"))) {
res <- itterateESDFL(match.call(), data)
class(res) <- "edsurveyGlmList"
return(res)
} else {
return(calc.glm.sdf(
formula = formula,
family = binomial(link = "logit"),
data = data,
weightVar = weightVar,
relevels = relevels,
varMethod = varMethod,
jrrIMax = jrrIMax,
omittedLevels = omittedLevels,
defaultConditions = defaultConditions,
missingDefaultConditions = missing(defaultConditions),
recode = recode,
returnVarEstInputs = returnVarEstInputs,
returnNumberOfPSU = returnNumberOfPSU,
returnLm0 = FALSE,
call = call
))
}
}
#' @rdname glm.sdf
#' @export
probit.sdf <- function(formula,
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
omittedLevels = TRUE,
defaultConditions = TRUE,
recode = NULL,
returnVarEstInputs = FALSE) {
call <- match.call()
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
# if data is an edsurvey.data.frame.list, simply return a list with results
# for each edsurvey.data.frame
if (inherits(data, c("edsurvey.data.frame.list"))) {
res <- itterateESDFL(match.call(), data)
class(res) <- "edsurveyGlmList"
return(res)
} else {
return(calc.glm.sdf(
formula = formula,
family = binomial(link = "probit"),
data = data,
weightVar = weightVar,
relevels = relevels,
varMethod = varMethod,
jrrIMax = jrrIMax,
omittedLevels = omittedLevels,
defaultConditions = defaultConditions,
missingDefaultConditions = missing(defaultConditions),
recode = recode,
returnVarEstInputs = returnVarEstInputs,
returnLm0 = FALSE,
call = call
))
}
}
calc.glm.sdf <- function(formula,
family = binomial(link = "logit"),
data,
weightVar = NULL,
relevels = list(),
varMethod = c("jackknife", "Taylor"),
jrrIMax = 1,
omittedLevels = TRUE,
defaultConditions = TRUE,
missingDefaultConditions = TRUE,
recode = NULL,
returnVarEstInputs = FALSE,
returnLm0 = FALSE,
returnNumberOfPSU = FALSE,
call = NULL) {
if (is.null(call)) {
call <- match.call()
}
varMethod <- substr(tolower(match.arg(varMethod)), 0, 1)
if (!family$family %in% c("binomial", "quasibinomial")) {
stop("Only fits binomial models.")
}
#############################
######### Outline: #########
#############################
# 1) check, format inputs
# 2) get the data
# 3) relevel
# 4) yvar could have plausible values
# 5) run the main regression
# 6) form the inputs for variance estimation
# 7) form output, including final variance estimation
# 1) check, format inputs
checkDataClass(data, c("edsurvey.data.frame", "light.edsurvey.data.frame", "edsurvey.data.frame.list"))
# varMethod always jackknife
# if the weight var is not set, use the default
wgt <- checkWeightVar(data, weightVar)
# check if there is an outcome variable and set it to the default if it is missing
zeroLengthLHS <- attr(terms(formula), "response") == 0
if (zeroLengthLHS) {
yvar <- attributes(getAttributes(data, "pvvars"))$default
formula <- update(formula, new = substitute(yvar ~ ., list(yvar = as.name(yvar))))
} else {
yvar <- all.vars(formula[[2]])
} # End of If/Else: if (zeroLengthLHS)
# grab the variables needed for the Taylor series method, if that is the variance estimation method being used
taylorVars <- c()
psuVar <- getPSUVar(data, weightVar = wgt)
stratumVar <- getStratumVar(data, weightVar = wgt)
# give an error if these do not exist.
tv <- checkTaylorVars(psuVar, stratumVar, wgt, varMethod, returnNumberOfPSU)
psuVar <- tv$psuVar
stratumVar <- tv$stratumVar
varMethod <- tv$varMethod
returnNumberOfPSU <- tv$returnNumberOfPSU
if (varMethod == "t") {
taylorVars <- c(psuVar, stratumVar)
jrrIMax <- NA
} else {
if (all(c(psuVar, stratumVar) %in% colnames(data))) {
taylorVars <- c(psuVar, stratumVar)
}
}
getDataVarNames <- c(all.vars(formula), wgt, taylorVars)
tryCatch(getDataVarNames <- unique(c(getDataVarNames, PSUStratumNeeded(returnNumberOfPSU, data))),
error = function(e) {
warning(paste0("Stratum and PSU variables are required for this call and are not on the incoming data. Ignoring ", dQuote("returnNumberOfPSU=TRUE"), "."))
returnNumberOfPSU <<- FALSE
}
)
# 2) get the data
# This is most of the arguments
getDataArgs <- list(
data = data,
varnames = getDataVarNames,
returnJKreplicates = (varMethod == "j"),
drop = FALSE,
omittedLevels = omittedLevels,
recode = recode,
includeNaLabel = TRUE,
dropUnusedLevels = TRUE
)
# Default conditions should be included only if the user set it. This adds the argument only if needed
if (!missingDefaultConditions) {
getDataArgs <- c(getDataArgs, list(defaultConditions = defaultConditions))
}
# edf is the actual data
edf <- do.call(getData, getDataArgs)
relVars <- colnames(edf)
relVars <- relVars[!relVars %in% taylorVars]
incomplete <- !complete.cases(edf[ , relVars])
if (any(incomplete)) {
warning("Removing ", sum(incomplete), " rows with NAs from analysis.")
edf <- edf[!incomplete, ]
}
# if doing Taylor series, check Taylor vars too
if (varMethod == "t") {
incomplete <- !complete.cases(edf[ , taylorVars])
if (any(incomplete)) {
if (any(incomplete[!edf[ , wgt] %in% c(NA, 0)])) {
warning("Removing ", sum(incomplete), " rows with NA PSU or stratum variables from analysis.")
}
edf <- edf[!incomplete, ]
}
}
# remove non-positive (full sample) weights
if (any(edf[ , wgt] <= 0)) {
posWeights <- edf[ , wgt] > 0
warning("Removing ", sum(!posWeights), " rows with nonpositive weight from analysis.")
edf <- edf[posWeights, ]
}
if (varMethod == "t" && sum(is.na(edf[ , c(stratumVar, psuVar)])) != 0) {
stop("Taylor series variance estimation requested but stratum or PSU variables contian an NA value.")
}
# check that there are not factors with only one level and give error message
lapply(
names(edf)[names(edf) %in% all.names(formula)],
function(z) {
if (is.factor(edf[ , z]) & length(levels(edf[ , z])) < 2) {
stop(paste0(
"The covariate ",
dQuote(z),
" has fewer than two levels and cannot be included in the regression."
))
}
}
)
# check that there is some data to work with
if (nrow(edf) <= 0) {
stop(paste0(
sQuote("data"), " must have more than 0 rows after a call to ",
sQuote("getData"), "."
))
}
# 3) deal with relevels.
# An argument that allows the user to change the omitted level for a factor variable
if (length(relevels) > 0) {
for (i in 1:length(relevels)) {
vari <- names(relevels)[i]
if (!vari %in% names(edf)) {
stop(paste0(
"In the ", sQuote("relevels"),
" argument, cannot find the variable named ", sQuote(vari), "."
))
} # End of if statment: ! vari %in% names(edf)
if (length(relevels[[i]]) != 1) {
stop(paste0(
"In the ", sQuote("relevels"),
" argument, each relevel must have exactly one level."
))
} # End of if statment: length(relevels[[i]]) != 1
# check that the level exists
lvls <- levels(edf[ , vari])
if (inherits(edf[ , vari], "lfactor")) {
# for a factor it can be either a level or a label
lvls <- c(lvls, labels(edf[ , vari]))
} # End of if statment: inherits(edf[ ,vari], "lfactor")
if (!relevels[[i]] %in% lvls) {
stop(paste0(
"In the ", sQuote("relevels"),
" argument, for the variable ", dQuote(vari), ", the level ",
dQuote(relevels[[i]]), " not found. Found levels are ",
pasteItems(dQuote(lvls)), "."
))
} # End of if statment !relevels[[i]] %in% lvls
edf[ , vari] <- relevel(edf[ , vari], ref = relevels[[i]])
} # end for(i in 1:length(relevels))
} # end if(length(relevels) > 0)
# 4) yvar and plausible values
pvy <- hasPlausibleValue(yvar, data) # pvy is the plausible values of the y variable
yvars <- yvar
linkingError <- detectLinkingError(data, yvars)
if(linkingError){
if(varMethod == "t") {
stop("Taylor series variance estimation not supported with NAEP linking error.")
}
if (jrrIMax != 1) {
warning("The linking error variance estimator only supports ", dQuote("jrrIMax=1"), ". Resetting to 1.")
jrrIMax <- 1
}
}
lyv <- length(yvars)
if (any(pvy)) {
if (linkingError) {
pp <- getPlausibleValue(yvars[max(pvy)], data)
suffix <- ifelse(grepl("_imp_", pp, fixed = TRUE), "_imp_", "_est_")
suffix <- ifelse(grepl("_samp_", pp, fixed = TRUE), "_samp_", suffix)
yvars <- paste0("outcome", suffix, 1:length(pp))
} else {
yvars <- paste0("outcome", 1:length(getPlausibleValue(yvars[max(pvy)], data)))
}
} else {
# if not, make sure that this variable is numeric
edf[ , "yvar"] <- as.numeric(eval(formula[[2]], edf))
formula <- update(formula, new = substitute(yvar ~ ., list(yvar = as.name(yvar))))
yvars <- "yvar"
} # End of if statment: any(pvy)
# 5) run the main regression
# run a regression, starting with the first PV or maybe the only outcome variable
yvar0 <- yvars[1]
# this allows that variable to not be dynamic variable, it is explicitly defined to be yvar0
if (family$family %in% c("binomial", "quasibinomial")) { # if using binomial set y to 1 only if it has the highest value
if (any(pvy)) {
for (i in 1:length(yvars)) {
# PV, so we have not evaluated the I() yet (if any)
for (yvi in 1:length(pvy)) {
if (pvy[yvi]) {
edf[ , yvar[yvi]] <- edf[ , getPlausibleValue(yvar[yvi], data)[i]]
}
}
edf[ , yvars[i]] <- as.numeric(eval(formula[[2]], edf))
}
oneDef <- max(edf[ , yvars], na.rm = TRUE)
for (i in yvars) {
edf[ , i] <- ifelse(edf[ , i] %in% oneDef, 1, 0)
}
edf$yvar0 <- edf[ , yvar0]
} else {
# for non-PV, I() has been evaluated
oneDef <- max(edf[ , yvars], na.rm = TRUE)
edf$yvar0 <- ifelse(edf$yvar %in% oneDef, 1, 0)
}
}
# set up formula for initial regression with yvar0 predicted by all other vars
frm <- update(formula, yvar0 ~ .)
edf$w <- edf[ , wgt]
# get an approximate fit without weights. Starting with weights can lead to
# convergence problems.
suppressWarnings(lm00 <- glm2(frm, data = edf, family = family))
edf$c2 <- predict(lm00, type = "response")
# get a final fit, with weights. Starting at c2
suppressWarnings(lm0 <- glm2(frm, data = edf, weights = w, family = family, mustart = c2, epsilon = 1e-14))
# get final prediction
edf$c2 <- predict(lm0, type = "response")
if (returnLm0) {
return(lm0)
}
# this grabs the list of weight variable names
wgtl <- getAttributes(data, "weights")[[wgt]]
varEstInputs <- list()
# note we have not made the B matrix
madeB <- FALSE
# there is now a very long set of conditionals
# regression with each yvar
stat_reg_glm <- function(fam = FALSE) {
ref <- function(pv, wg) {
if (family$family %in% c("binomial", "quasibinomial") & (fam)) {
# do this for each yvar, but not for each jk weight (in parent function)
edf$yvar0 <- ifelse(pv %in% oneDef, 1, 0)
}
edf$w <- wg
suppressWarnings(lmi <- glm2(frm, data = edf, weights = w, family = family, mustart = c2, epsilon = 1e-14))
return(coef(lmi))
}
}
# 6) form the inputs for variance estimation
if (any(pvy)) { # the y variable is a plausible value
# this if condition estimates both Taylor series method and jackknife method
# for equations, see the statistics vignette
if (varMethod == "t") {
jrrIMax <- length(yvars)
} else {
jrrIMax <- min(jrrIMax, length(yvars))
}
# initialize: varm is the variance matrix by coefficient and PV (for V_jrr)
varM <- list()
varm <- matrix(NA, nrow = jrrIMax, ncol = length(coef(lm0)))
if (varMethod == "j") {
if (linkingError) {
repWeights <- paste0(wgtl$jkbase, wgtl$jksuffixes)
# get the mean estimate
est <- getLinkingEst(
data = edf,
pvEst = yvars[grep("_est", yvars)],
stat = stat_reg_glm(fam = TRUE),
wgt = wgt
)
# coefficients matrix, r-squared vector
coefm <- est$coef
coef <- apply(coefm, 2, mean)
coefm0 <- t(t(coefm) - coef)
# get imputation variance
impVar <- getLinkingImpVar(
data = edf,
pvImp = yvars[grep("_imp", yvars)],
ramCols = ncol(getRAM()),
stat = stat_reg_glm(fam = TRUE),
wgt = wgt,
T0 = est$est,
T0Centered = FALSE
)
# get sampling variance
sampVar <- getLinkingSampVar(edf,
pvSamp = yvars[grep("_samp", yvars)],
stat = stat_reg_glm(fam = TRUE),
rwgt = repWeights,
T0 = est$est,
T0Centered = FALSE
)
varM[[1]] <- sampVar$Bi
varEstInputs[["JK"]] <- sampVar$veiJK
varEstInputs[["PV"]] <- impVar$veiImp
# drop r.squared term
varm[1, ] <- sampVar$V
M <- length(yvars[grep("_est", yvars)])
Vimp <- impVar$V
} else {
# coefficients by PV (getEst in lm.sdf)
res <- getEst(
data = edf,
pvEst = yvars,
stat = stat_reg_glm(fam = TRUE),
wgt = wgt
)
coefm <- res$coef
varEstInputs[["JK"]] <- data.frame()
jkSumMult <- getAttributes(data, "jkSumMultiplier")
for (pvi in 1:jrrIMax) { # for each PV (up to jrrIMax)
res <- getVarEstJK(
stat = stat_reg_glm(fam = FALSE),
yvar = edf[ , yvars[pvi]],
wgtM = edf[ , paste0(wgtl$jkbase, wgtl$jksuffixes)],
co0 = coefm[pvi, ],
jkSumMult = jkSumMult,
pvName = pvi
)
varM[[pvi]] <- res$Bi
varEstInputs[["JK"]] <- rbind(varEstInputs[["JK"]], res$veiJK)
varm[pvi, ] <- res$VsampInp
} # end for(pvi in 1:jrrIMax)
# number of PVs
M <- length(yvars)
# imputaiton variance / variance due to uncertaintly about PVs
coef <- apply(coefm, 2, mean)
coefm0 <- t(t(coefm) - coef)
coefmPVByRow <- lapply(1:ncol(coefm0), function(coli) {
data.frame(
stringsAsFactors = FALSE,
PV = 1:nrow(coefm0),
variable = rep(names(coef(lm0))[coli], nrow(coefm0)),
value = coefm0[ , coli]
)
})
coefmPV <- do.call(rbind, coefmPVByRow)
varEstInputs[["PV"]] <- coefmPV
Vimp <- (M + 1) / M * apply(coefm, 2, var)
} # end else for if(linkingError)
} else { # End of if statment: varMethod == "j"
# Taylor series variance esimation, no Plausible Values
dofNum <- matrix(0, nrow = length(coef(lm0)), ncol = length(yvars))
dofDenom <- matrix(0, nrow = length(coef(lm0)), ncol = length(yvars))
# variances are calculated iteratively for all each y variable
est <- getEst(
data = edf,
pvEst = yvars,
stat = stat_reg_glm(fam = TRUE),
wgt = wgt
)
coefm <- est$coef
for (mm in 1:length(yvars)) {
edf$yvar0 <- as.numeric(edf[ , yvars[mm]])
y <- edf[ , yvars[mm]]
suppressWarnings(lmi <- glm2(frm, data = edf, weights = w, family = family, mustart = c2, epsilon = 1e-14))
coef <- b <- co0 <- coef(lmi)
D <- vcov(lmi)
eta <- predict(lmi, type = "link")
pred <- predict(lmi, type = "response")
# precision weights
# in the notation of McCullagh and Nelder, this is d(mu)/d(eta)
# evaluated at the (latent) predicted values (eta)
mu.eta <- family$mu.eta(eta)
# this is the partial of the likelihood at the unit level
X <- sparse.model.matrix(frm, edf)
uhij <- (y - pred) / (pred * (1 - pred)) * as.matrix(X) * mu.eta
# vals by stratum (singleton PSUs removed in function)
res <- getVarTaylor(uhij, edf, D, wgt, psuVar, stratumVar)
vc <- D %*% res$vv %*% D
# collect
dofNum[ , mm] <- res$nums
dofDenom[ , mm] <- res$nums2
varM[[mm]] <- as.matrix(vc)
varm[mm, ] <- as.numeric(diag(vc))
}
# number of PVs
M <- length(yvars)
# imputaiton variance / variance due to uncertaintly about PVs
coef <- apply(coefm, 2, mean)
coefm0 <- t(t(coefm) - coef)
coefmPVByRow <- lapply(1:ncol(coefm0), function(coli) {
data.frame(
stringsAsFactors = FALSE,
PV = 1:nrow(coefm0),
variable = rep(names(coef(lm0))[coli], nrow(coefm0)),
value = coefm0[ , coli]
)
})
coefmPV <- do.call(rbind, coefmPVByRow)
varEstInputs[["PV"]] <- coefmPV
Vimp <- (M + 1) / M * apply(coefm, 2, var)
} # End of if/else statment: varMethod == "j"
# calculate van Buuren B
B <- (1 / (M - 1)) * Reduce(
"+", # add up the matrix results of the sapply
sapply(1:nrow(coefm), function(q) {
# within each PV set, calculate the outer product
# (2.19 of Van Buuren)
outer(coefm0[q, ], coefm0[q, ])
}, simplify = FALSE)
)
madeB <- TRUE
# 2.18 in var Buuren
Ubar <- (1 / length(varM)) * Reduce("+", varM)
# variance due to sampling
Vjrr <- apply(varm[1:jrrIMax, , drop = FALSE], 2, mean)
} else { # end if(any(pvy))
# the y variable is not a plausible value
# this section handles jackknife and Taylor series estimation
if (varMethod == "j") {
jkSumMult <- getAttributes(data, "jkSumMultiplier")
coef <- coef(lm0)
res <- getVarEstJK(
stat = stat_reg_glm(fam = FALSE),
yvar = edf[ , yvars[1]],
wgtM = edf[ , paste0(wgtl$jkbase, wgtl$jksuffixes)],
co0 = coef,
jkSumMult = jkSumMult,
pvName = 1
)
Ubar <- res$Bi
Vjrr <- res$VsampInp
varEstInputs[["JK"]] <- res$veiJK
} else { # end if(varMethod == "j")
# Taylor series variance esimation, no PVs
X <- sparse.model.matrix(frm, edf)
y <- edf$yvar0
coef <- b <- co0 <- coef(lm0)
D <- vcov(lm0)
eta <- predict(lm0, type = "link")
pred <- predict(lm0, type = "response")
# precision weights
# in the notation of McCullagh and Nelder, this is d(mu)/d(eta)
# evaluated at the (latent) predicted values (eta)
mu.eta <- family$mu.eta(eta)
# this is the partial of the likelihood at the unit level
uhij <- (y - pred) / (pred * (1 - pred)) * as.matrix(X) * mu.eta
# vals by stratum (singleton PSUs removed in function)
res <- getVarTaylor(uhij, edf, D, wgt, psuVar, stratumVar)
vc <- D %*% res$vv %*% D
# collect
dofNum <- res$nums
dofDenom <- res$nums2
Vjrr <- as.numeric(diag(vc))
# this is the VC matrix, store it for vcov to recover
Ubar <- as.matrix(vc)
} # end else for if(varMethod == "j")
# not pvy common
M <- 1 # only on replicate when there are no PVs
Vimp <- 0 # no imputation variance when there are no PVs
varm <- NULL # no variance by PV
coefm <- NULL # no coefficients matrix by PV
B <- 0 * Ubar # no PVs
} # end else for if(any(pvy))
# 7) form output, including final variance estimation
V <- Vjrr + Vimp
names(coef) <- names(coef(lm0))
se <- sqrt(V)
names(se) <- names(coef)
# get fitted and resid based on overall coef
X <- model.matrix(frm, edf)
fittedLatent <- as.vector(X %*% coef)
fitted1 <- family$linkinv(fittedLatent)
if (linkingError) {
ye <- grep("_est_", yvars)
Y <- sapply(ye, function(yi) {
as.vector(edf[ , yvars[yi]])
}, simplify = TRUE)
resid1 <- Y - fitted1
colnames(resid1) <- yvars[ye]
} else {
Y <- sapply(1:length(yvars), function(yi) {
as.vector(edf[ , yvars[yi]])
}, simplify = TRUE)
resid1 <- Y - fitted1
colnames(resid1) <- yvars
}
# residual df calculation
nobs <- nrow(edf)
n.ok <- nobs - sum(edf$origwt == 0)
nvars <- ncol(X)
rank <- rankMatrix(X)
resdf <- n.ok - rank
df.r <- resdf
# get residual based on coef by PV
if (!is.null(coefm)) {
if (!is.matrix(coefm)) {
coefm <- t(as.matrix(coefm))
}
fitted2 <- family$linkinv(as.matrix(X %*% t(coefm)))
resid2 <- Y - fitted2
}
coefmat <- data.frame(
coef = coef,
se = se,
t = coef / se
)
if (varMethod == "t") {
m <- length(wgtl$jksuffixes)
if (inherits(dofNum, "matrix")) {
coefmat$dof <- apply((3.16 - 2.77 / sqrt(m)) * dofNum^2 / dofDenom, 1, mean)
} else {
coefmat$dof <- (3.16 - 2.77 / sqrt(m)) * dofNum^2 / dofDenom
}
} else {
coefmat$dof <- sapply(names(coef), function(cn) {
DoFCorrection(varEstA = varEstInputs, varA = cn, method = "JR")
})
}
pti <- pt(coefmat$t, df = coefmat$dof)
coefmat[ , "Pr(>|t|)"] <- 2 * pmin(pti, 1 - pti)
njk <- length(wgtl$jksuffixes)
if (varMethod == "t") {
njk <- NA
}
varmeth <- ifelse(varMethod == "t", "Taylor series", "jackknife")
res <- list(
call = call, formula = formula, family = family, coef = coef, se = se, Vimp = Vimp,
Vjrr = Vjrr, M = M, varm = varm, coefm = coefm,
coefmat = coefmat, weight = wgt,
npv = length(yvars), jrrIMax = min(jrrIMax, length(yvars)),
njk = njk, varMethod = varmeth,
residuals = resid1, fitted.values = fitted1, fittedLatent = fittedLatent, residual.df = resdf
)
if (!is.null(coefm)) {
res <- c(res, list(PV.residuals = resid2, PV.fitted.values = fitted2))
}
if (returnVarEstInputs) {
res <- c(res, list(varEstInputs = varEstInputs))
if (varMethod == "t") {
warning(paste0(
"Taylor series method not supported with the ",
sQuote("varEstInputs"), " argument set to ", dQuote("TRUE"), "."
))
}
}
if (madeB) {
# equation 2.29 in van Buuren, pp 42
tryCatch(
rbar <- (1 + 1 / M) * (1 / nrow(Ubar)) * sum(diag(B %*% solve(Ubar))),
error = function(e) {
rbar <<- (1 + 1 / M) * (1 / nrow(Ubar)) * sum(diag(B %*% MASS::ginv(Ubar)))
frm <- Formula(frm)
if (terms(frm, lhs = 0, rhs = NULL) == ~1) {
warning("A variance estimate was replaced with NA because there was no variance across strata.", call. = FALSE)
}
}
)
Ttilde <- (1 + rbar) * Ubar
res <- c(res, list(B = B, U = Ubar, rbar = rbar, Ttilde = Ttilde))
} else {
res <- c(res, list(B = B, U = Ubar))
}
if (returnNumberOfPSU) {
stratumVar <- getAttributes(data, "stratumVar")
psuVar <- getAttributes(data, "psuVar")
if ("JK1" %in% stratumVar) {
res <- c(res, list(nPSU = nrow(edf)))
} else if (all(c(stratumVar, psuVar) %in% colnames(edf))) {
if (sum(is.na(edf[ , c(stratumVar, psuVar)])) == 0) {
res <- c(res, list(nPSU = nrow(unique(edf[ , c(stratumVar, psuVar)]))))
} else {
warning("Cannot return number of PSUs because the stratum or PSU variables contain NA values.")
}
}
}
# add waldDenomBaseDof if relevant
if (!is.null(stratumVar) && !is.null(psuVar)) {
if (all(c(stratumVar, psuVar) %in% colnames(edf))) {
if (sum(is.na(edf[ , c(stratumVar, psuVar)])) == 0) {
res <- c(res, list(waldDenomBaseDof = waldDof(edf, stratumVar, psuVar)))
}
}
if ("JK1" %in% getStratumVar(data)) {
res <- c(res, list(waldDenomBaseDof = "JK1"))
}
}
res <- c(res, list(n0 = nrow2.edsurvey.data.frame(data), nUsed = nrow(edf)))
if (inherits(data, "edsurvey.data.frame")) {
res <- c(res, list(data = data))
} else {
res <- c(res, list(lm0 = lm0))
}
class(res) <- "edsurveyGlm"
return(res)
}
#' @method print edsurveyGlm
#' @export
print.edsurveyGlm <- function(x, ...) {
print(coef(x), ...)
}
#' @method print edsurveyGlmList
#' @export
print.edsurveyGlmList <- function(x, ...) {
for (i in 1:length(x)) {
cat("glm", i, "\n")
print(coef(x[[i]]), ...)
}
}
#' @method summary edsurveyGlm
#' @export
summary.edsurveyGlm <- function(object, ...) {
class(object) <- "summary.edsurveyGlm"
object
}
#' @method summary edsurveyGlmList
#' @export
summary.edsurveyGlmList <- function(object, ...) {
class(object) <- "summary.edsurveyGlmList"
for (i in 1:length(object)) {
class(object[[i]]) <- "summary.edsurveyGlm"
}
object
}
#' @method print summary.edsurveyGlm
#' @importFrom stats printCoefmat
#' @export
print.summary.edsurveyGlm <- function(x, ...) {
cat(paste0("\nFormula: ", paste(deparse(x$formula), collapse = ""), "\n"))
cat(paste0("Family: ", x$family$family, " (", x$family$link, ")\n\n"))
if (x$npv != 1) {
cat(paste0("jrrIMax: ", x$jrrIMax, "\n"))
}
cat(paste0("Weight variable: ", sQuote(x$weight), "\n"))
cat(paste0("Variance method: ", x$varMethod, "\n"))
if (!is.na(x$njk)) {
cat(paste0("JK replicates: ", x$njk, "\n"))
}
cat(paste0("full data n: ", x$n0, "\n"))
if (!is.null(x$nPSU)) {
cat(paste0("n PSU: ", x$nPSU, "\n"))
}
cat(paste0("n used: ", x$nUsed, "\n\n"))
cat(paste0("Coefficients:\n"))
printCoefmat(x$coefmat, P.values = TRUE, has.Pvalue = TRUE)
}
#' @method print summary.edsurveyGlmList
#' @export
print.summary.edsurveyGlmList <- function(x, ...) {
for (i in 1:length(x)) {
cat("glm", i, "\n")
print(x[[i]], ...)
}
}
#' @method coef edsurveyGlm
#' @export
coef.edsurveyGlm <- function(object, ...) {
object$coef
}
#' @method coef edsurveyGlmList
#' @export
coef.edsurveyGlmList <- function(object, ...) {
sapply(object, function(li) {
li$coef
})
}
#' @method vcov edsurveyGlm
#' @export
vcov.edsurveyGlm <- function(object, ...) {
vcov.edsurveyLm(object, ...)
}
#' @title Odds Ratios for edsurveyGlm Models
#' @description Converts coefficients from \code{edsurveyGlm} logit regression model to odds ratios.
#' @param model an \code{edsurveyGlm} model
#' @param alpha the alpha level for the confidence level
#'
#' @return
#' An \code{oddsRatio.edsurveyGlm} object with the following elements:
#' \item{OR}{odds ratio coefficient estimates}
#' \item{2.5\%}{lower bound 95\% confidence interval}
#' \item{97.5\%}{upper bound 95\% confidence interval}
#'
#' @export
oddsRatio <- function(model, alpha = 0.05) {
if (!inherits(model, "edsurveyGlm")) stop("x must be of class 'edsurveyGlm'.")
if (model$family$family != "binomial" & model$family$family != "quasibinomial") {
stop("Model must be binomial.")
}
if (model$family$link != "logit") {
stop("Model must be logit.")
}
OR <- exp(coef(model))
varDiag <- diag(vcov(model))
CIlo <- exp(coef(model) - (qt(1 - alpha / 2, df = model$coefmat$dof) * sqrt(varDiag)))
CIup <- exp(coef(model) + (qt(1 - alpha / 2, df = model$coefmat$dof) * sqrt(varDiag)))
r <- cbind(OR, CIlo, CIup)
pctRound <- 1
pct <- round(100 * c(alpha / 2, 1 - alpha / 2), pctRound)
while (pct[1] == 0) {
pctRound <- pctRound + 1
pct <- round(100 * c(alpha / 2, 1 - alpha / 2), pctRound)
}
colnames(r) <- c("OR", paste0(pct, "%"))
r <- as.data.frame(r)
class(r) <- c("oddsRatio.edsurveyGlm", "data.frame")
r
}
# @export
setMethod(
"coef",
c(object = "edsurveyGlm"),
coef.edsurveyGlm
)
# @export
setMethod(
"coef",
c(object = "edsurveyGlmList"),
coef.edsurveyGlmList
)
# @export
setMethod(
"vcov",
c(object = "edsurveyGlm"),
vcov.edsurveyGlm
)
Try the EdSurvey package in your browser
Any scripts or data that you put into this service are public.
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.