anonymous-submission-acl2025's picture
add 36
b1a2e61
####### post #######
setGeneric("post",
function(model,x1name=NULL,x1vals=NULL,x2name=NULL,x2vals=NULL,holds=NULL,
n.sims=1000,cut=NULL,quantiles=c(.025,.975),did=NULL,weights=NULL, digits=2){
standardGeneric("post")
}
)
setClassUnion("arrayORNULL", c("array","NULL"))
setClassUnion("listORcharacter", c("list","character"))
setClass("post",
slots = c(est = "array",
did = "arrayORNULL",
sims = "array",
model = "character",
link = "listORcharacter",
quantiles = "numeric",
call = "call")
)
post.glm <- function(model,x1name=NULL,x1vals=NULL,x2name=NULL,x2vals=NULL,holds=NULL,
n.sims=1000,cut=NULL,quantiles=c(.025,.975),did=NULL,weights=NULL, digits=2){
call <- match.call()
sims <- postSim(model, n.sims=n.sims)
if (family(model)[2]=="identity"){link <- identity}
else if (family(model)[2]=="probit"){link <- pnorm}
else if (family(model)[2]=="logit"){link <- plogis}
else if (family(model)[2]=="log"){link <- exp}
else if (family(model)[2]=="cloglog"){link <- function(x){1-exp(-exp(x))}}
else {stop("Link function is not supported")}
if (is.null(weights)){wi <- c(rep(1, length(model$model[,1])))} else{wi <- weights}
n.obs <- length(model.matrix(model)[,1])
k <- length(model.matrix(model)[1,])
n.q <- length(quantiles)
if (is.null(x1name)){
X <- array(NA, c(n.obs,k))
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (i in 1:length(holds)){
newdata[ ,names(holds)[i]] <- as.numeric(holds[i])
}
}
X <- aperm(model.matrix(lm(formula(model), data=newdata)))
l1 <- array(NA, c(nrow(sims@coef),1))
l1[,1] <- apply(link(sims@coef %*% X), 1, function(x) weighted.mean(x, wi))
l2 <- array(NA, c(1,n.q+1))
l2[1,1] <- mean(l1)
l2[1,2:(n.q+1)] <- quantile(l1, probs=quantiles)
colnames(l2) <- c("mean",quantiles)
ans <- new("post",
est=round(l2, digits=digits),
did=NULL,
sims=l1,
model=class(model),
link=family(model)[2],
quantiles=quantiles,
call=call)
return(ans)
}
else if (is.null(x2name)){
n.x1 <- length(x1vals)
X <- array(NA, c(n.obs,k,n.x1))
for (i in 1:(n.x1)){
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (j in 1:length(holds)){
newdata[ ,names(holds)[j]] <- as.numeric(holds[j])
}
}
newdata[ ,x1name] <- x1vals[i]
X[ , ,i] <- model.matrix(lm(formula(model), data=newdata))
}
X <- aperm(X, c(2,1,3))
l1 <- apply(apply(X, c(2,3), function(x) link(sims@coef %*% x)), c(1,3), function(x) weighted.mean(x, wi))
l2 <- array(NA, c(n.x1+1,n.q+1))
l2[1:n.x1,1] <- apply(l1, 2, mean)
l2[1:n.x1,2:(n.q+1)] <- aperm(apply(l1, 2, function(x) quantile(x, probs=quantiles)))
l2[nrow(l2),1] <- mean(l1[ ,n.x1] - l1[ ,1])
l2[nrow(l2),2:(n.q+1)] <- quantile(l1[ ,n.x1] - l1[ ,1], probs=quantiles)
rownames(l2) <- c(paste(c(rep(paste(x1name,"="),n.x1),
paste("\u0394","(",x1vals[1],",",x1vals[length(x1vals)],")")),
c(x1vals,"")))
colnames(l2) <- c("mean",quantiles)
ans <- new("post",
est=round(l2, digits=digits),
did=NULL,
sims=l1,
model=class(model),
link=family(model)[2],
quantiles=quantiles,
call=call)
return(ans)
}
else{
n.x1 <- length(x1vals)
n.x2 <- length(x2vals)
X <- array(NA, c(n.obs,k,n.x1,n.x2))
for (j in 1:n.x2){
for (i in 1:n.x1){
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (k in 1:length(holds)){
newdata[ ,names(holds)[k]] <- as.numeric(holds[k])
}
}
newdata[ ,x1name] <- x1vals[i]
newdata[ ,x2name] <- x2vals[j]
X[ , ,i,j] <- model.matrix(lm(formula(model), data=newdata))
}
}
X <- aperm(X, c(2,1,3,4))
l1 <- apply(apply(X, c(2,3,4), function(x) link(sims@coef %*% x)), c(1,3,4), function(x) weighted.mean(x, wi))
l2 <- array(NA, c(n.x1+1,n.q+1,n.x2))
l2[1:n.x1,1,1:n.x2] <- apply(l1,c(2,3),mean)
l2[1:n.x1,2:(n.q+1),1:n.x2] <- aperm(apply(l1, c(2,3), function(x) quantile(x, probs=quantiles)), c(2,1,3))
l2[nrow(l2),1,1:n.x2] <- apply(l1[ ,n.x1,1:n.x2] - l1[ ,1,1:n.x2], 2, mean)
l2[nrow(l2),2:(n.q+1),1:n.x2] <- apply(l1[ ,n.x1,1:n.x2] - l1[ ,1,1:n.x2], 2, function(x) quantile(x, probs=quantiles))
dimnames(l2) <- list(paste(c(rep(paste(x1name,"="),n.x1),paste("\u0394","(",x1vals[1],",",x1vals[length(x1vals)],")")),c(x1vals,"")),
c("mean",quantiles),
paste(c(rep(paste(x2name,"="),n.x2)),
c(x2vals)))
if (is.null(did)){did <- c(x2vals[1],x2vals[n.x2])} else{did <- did}
l3 <- array(NA, c(1,n.q+1))
l3[1,1] <- mean((l1[ ,n.x1,match(did[2],x2vals)] - l1[ ,1,match(did[2],x2vals)]) - (l1[ ,n.x1,match(did[1],x2vals)] - l1[ ,1,match(did[1],x2vals)]))
l3[1,2:(n.q+1)] <- quantile((l1[ ,n.x1,match(did[2],x2vals)] - l1[ ,1,match(did[2],x2vals)]) - (l1[ ,n.x1,match(did[1],x2vals)] - l1[ ,1,match(did[1],x2vals)]), probs=quantiles)
dimnames(l3) <- list("did",c("mean",quantiles))
ans <- new("post",
est=round(l2, digits=digits),
did=round(l3, digits=digits),
sims=l1,
model=class(model),
link=family(model)[2],
quantiles=quantiles,
call=call)
return(ans)
}
}
post.polr <- function(model,x1name=NULL,x1vals=NULL,x2name=NULL,x2vals=NULL,holds=NULL,
n.sims=1000,cut=NULL,quantiles=c(.025,.975),did=NULL,weights=NULL, digits=2){
call <- match.call()
sims <- suppressMessages(postSim(model, n.sims=n.sims))
if (is.null(weights)){wi <- c(rep(1, length(model$model[,1])))} else{wi <- weights}
if (model$method=="probit"){link <- pnorm}
else if (model$method=="logistic"){link <- plogis}
else if (model$method=="cloglog"){link <- function(x){1-exp(-exp(x))}}
else {stop("Link function is not supported")}
n.obs <- length(model$model[,1])
k <- length(model.matrix(polr(getCall(model)$formula, model$model))[1,])
n.q <- length(quantiles)
n.y <- length(levels(model$model[,1]))
n.z <- length(model$zeta)
tau <- array(NA, c(n.sims,n.z+2))
tau[,1] <- -Inf
tau[,2:(ncol(tau)-1)] <- sims@zeta[,1:n.z]
tau[,ncol(tau)] <- Inf
beta <- sims@coef
if (is.null(cut)){
if (is.null(x1name)){
X_temp <- array(NA, c(n.obs,k))
X <- array(NA, c(n.obs,k-1))
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (j in 1:length(holds)){
newdata[ ,names(holds)[j]] <- as.numeric(holds[j])
}
}
X_temp[ , ] <- suppressWarnings(model.matrix(polr(getCall(model)$formula, data=newdata)))
X[ , ] <- X_temp[,-1]
X <- aperm(X)
l1 <- array(NA, c(n.sims, n.obs, n.y))
for (z in 1:n.y){
l1[,,z] <- link(tau[,z+1] - beta %*% X) - link(tau[,z] - beta %*% X)
}
l2 <- apply(l1, c(1,3), function(x) weighted.mean(x, wi))
l3 <- array(NA, c(n.y,n.q+1))
for (i in 1:n.y){
l3[i,1] <- mean(l2[,i])
l3[i,2:(n.q+1)] <- quantile(l2[,i], probs=quantiles)
}
rownames(l3) <- paste(c(rep("Y =",n.y)), c(1:n.y))
colnames(l3) <- c("mean",quantiles)
ans <- new("post",
est=round(l3, digits=digits),
did=NULL,
sims=l2,
model=class(model),
link=model$method,
quantiles=quantiles,
call=call)
return(ans)
}
else if (is.null(x2name)){
n.x1 <- length(x1vals)
X_temp <- array(NA, c(n.obs,k,n.x1))
X <- array(NA, c(n.obs,k-1,n.x1))
for (i in 1:(n.x1)){
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (j in 1:length(holds)){
newdata[ ,names(holds)[j]] <- as.numeric(holds[j])
}
}
newdata[ ,x1name] <- x1vals[i]
X_temp[ , ,i] <- suppressWarnings(model.matrix(polr(getCall(model)$formula, data=newdata)))
X[ , ,i] <- X_temp[,-1,i]
}
l1 <- array(NA, c(n.sims, n.obs, n.x1, n.y))
X <- aperm(X, c(2,1,3))
for (z in 1:n.y){
l1[,,,z] <- apply(X, c(2,3), function(x) (link(tau[,z+1] - beta %*% x) - link(tau[,z] - beta %*% x)))
}
l2 <- apply(l1, c(1,3,4), function(x) weighted.mean(x, wi))
l3 <- array(NA, c(n.x1+1, n.q+1, n.y))
for (j in 1:n.y){
for (i in 1:n.x1){
l3[i,1,j] <- mean(l2[,i,j])
l3[i,2:(n.q+1),j] <- quantile(l2[,i,j], probs=quantiles)
}
l3[nrow(l3),1,j] <- mean(l2[ ,n.x1,j] - l2[ ,1,j])
l3[nrow(l3),2:(n.q+1),j] <- quantile(l2[ ,n.x1,j] - l2[ ,1,j], probs=quantiles)
}
dimnames(l3) <- list(paste(c(rep(paste(x1name,"="),n.x1),paste("\u0394","(",x1vals[1],",",x1vals[length(x1vals)],")")),c(x1vals,"")),
c("mean",quantiles),
paste(c(rep("Y =",length(levels(model$model[,1])))),
c(1:length(levels(model$model[,1])))))
ans <- new("post",
est=round(l3, digits=digits),
did=NULL,
sims=l2,
model=class(model),
link=model$method,
quantiles=quantiles,
call=call)
return(ans)
}
else{
n.x1 <- length(x1vals)
n.x2 <- length(x2vals)
X_temp <- array(NA, c(n.obs,k,n.x1,n.x2))
X <- array(NA, c(n.obs,k-1,n.x1,n.x2))
for (j in 1:n.x2){
for (i in 1:(n.x1)){
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (k in 1:length(holds)){
newdata[ ,names(holds)[k]] <- as.numeric(holds[k])
}
}
newdata[ ,x1name] <- x1vals[i]
newdata[ ,x2name] <- x2vals[j]
X_temp[ , ,i,j] <- suppressWarnings(model.matrix(polr(getCall(model)$formula, data=newdata)))
X[ , ,i,j] <- X_temp[,-1,i,j]
}
}
X <- aperm(X, c(2,1,3,4))
l1 <- array(NA, c(n.sims, n.obs, n.x1, n.x2, n.y))
for (z in 1:n.y){
l1[,,,,z] <- apply(X, c(2,3,4), function(x) (link(tau[,z+1] - beta %*% x) - link(tau[,z] - beta %*% x)))
}
l2 <- apply(l1, c(1,3,4,5), function(x) weighted.mean(x, wi))
l3 <- array(NA, c(n.x1+1, n.q+1, n.x2, n.y))
for (k in 1:n.y){
for (j in 1:n.x2){
for (i in 1:n.x1){
l3[i,1,j,k] <- mean(l2[,i,j,k])
l3[i,2:(n.q+1),j,k] <- quantile(l2[,i,j,k], probs=quantiles)
}
l3[n.x1+1,1,j,k] <- mean(l2[,n.x1,j,k] - l2[,1,j,k])
l3[n.x1+1,2:(n.q+1),j,k] <- quantile(l2[,n.x1,j,k] - l2[,1,j,k], probs=quantiles)
}
}
dimnames(l3) <- list(paste(c(rep(paste(x1name," ="),n.x1),paste("\u0394","(",x1vals[1],",",x1vals[length(x1vals)],")")),c(x1vals,"")),
c("mean",quantiles),
paste(c(rep(paste(x2name,"="),n.x2)),x2vals),
paste(c(rep("Y =",n.y)), c(1:n.y)))
if (is.null(did)){did <- c(x2vals[1],x2vals[n.x2])} else{did <- did}
l4 <- array(NA, c(n.y,n.q+1))
for (i in 1:n.y){
l4[i,1] <- mean((l2[ ,n.x1,match(did[2],x2vals),i] - l2[ ,1,match(did[2],x2vals),i]) - (l2[ ,n.x1,match(did[1],x2vals),i] - l2[ ,1,match(did[1],x2vals),i]))
l4[i,2:(n.q+1)] <- quantile((l2[ ,n.x1,match(did[2],x2vals),i] - l2[ ,1,match(did[2],x2vals),i]) - (l2[ ,n.x1,match(did[1],x2vals),i] - l2[ ,1,match(did[1],x2vals),i]), probs=quantiles)
}
yvals <- 1:n.y
dimnames(l4) <- list(paste(c(rep(paste("Y","="),n.y)),yvals),c("mean",quantiles))
ans <- new("post",
est=round(l3, digits=digits),
did=round(l4, digits=digits),
sims=l2,
model=class(model),
link=model$method,
quantiles=quantiles,
call=call)
return(ans)
}
}
else{
if (is.null(x1name)){
X_temp <- array(NA, c(n.obs,k))
X <- array(NA, c(n.obs,k-1))
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (j in 1:length(holds)){
newdata[ ,names(holds)[j]] <- as.numeric(holds[j])
}
}
X_temp[ , ] <- suppressWarnings(model.matrix(polr(getCall(model)$formula, data=newdata)))
X[ , ] <- X_temp[,-1]
X <- aperm(X)
l1 <- apply(link(-tau[,cut+1] + beta %*% X), 1, function(x) weighted.mean(x, wi))
l2 <- array(NA, c(1,n.q+1))
l2[1,1] <- mean(l1)
l2[1,2:(n.q+1)] <- quantile(l1, probs=quantiles)
colnames(l2) <- c("mean",quantiles)
ans <- new("post",
est=round(l2, digits=digits),
did=NULL,
sims=l1,
model=class(model),
link=model$method,
quantiles=quantiles,
call=call)
return(ans)
}
else if (is.null(x2name)){
n.x1 <- length(x1vals)
X_temp <- array(NA, c(n.obs,k,n.x1))
X <- array(NA, c(n.obs,k-1,n.x1))
for (i in 1:(n.x1)){
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (j in 1:length(holds)){
newdata[ ,names(holds)[j]] <- as.numeric(holds[j])
}
}
newdata[ ,x1name] <- x1vals[i]
X_temp[ , ,i] <- suppressWarnings(model.matrix(polr(getCall(model)$formula, data=newdata)))
X[ , ,i] <- X_temp[,-1,i]
}
X <- aperm(X, c(2,1,3))
l1 <- apply(apply(X, c(2,3), function(x) link(-tau[,cut+1] + beta %*% x)),
c(1,3), function(x) weighted.mean(x, wi))
l2 <- array(NA, c(n.x1+1,n.q+1))
for (i in 1:n.x1){
l2[i,1] <- mean(l1[,i])
l2[i,2:(n.q+1)] <- quantile(l1[,i], probs=quantiles)
}
l2[nrow(l2),1] <- mean(l1[ ,ncol(l1)] - l1[ ,1])
l2[nrow(l2),2:(n.q+1)] <- quantile(l1[ ,ncol(l1)] - l1[ ,1], probs=quantiles)
rownames(l2) <- c(paste(c(rep(paste(x1name,"="),n.x1),
paste("\u0394","(",x1vals[1],",",x1vals[length(x1vals)],")")),
c(x1vals,"")))
colnames(l2) <- c("mean",quantiles)
ans <- new("post",
est=round(l2, digits=digits),
did=NULL,
sims=l1,
model=class(model),
link=model$method,
quantiles=quantiles,
call=call)
return(ans)
}
else{
n.x1 <- length(x1vals)
n.x2 <- length(x2vals)
X_temp <- array(NA, c(n.obs,k,n.x1,n.x2))
X <- array(NA, c(n.obs,k-1,n.x1,n.x2))
for (j in 1:n.x2){
for (i in 1:n.x1){
newdata <- data.frame(model$model)
if (!is.null(holds)){
for (k in 1:length(holds)){
newdata[ ,names(holds)[k]] <- as.numeric(holds[k])
}
}
newdata[ ,x1name] <- x1vals[i]
newdata[ ,x2name] <- x2vals[j]
X_temp[ , ,i,j] <- suppressWarnings(model.matrix(polr(getCall(model)$formula, data=newdata)))
X[ , ,i,j] <- X_temp[,-1,i,j]
}
}
X <- aperm(X, c(2,1,3,4))
l1 <- apply(apply(X, c(2,3,4), function(x) link(-tau[,cut+1] + beta %*% x)), c(1,3,4), function(x) weighted.mean(x, wi))
l2 <- array(NA, c(n.x1+1,n.q+1,n.x2))
for (j in 1:n.x2){
for (i in 1:n.x1){
l2[i,1,j] <- mean(l1[,i,j])
l2[i,2:(n.q+1),j] <- quantile(l1[,i,j], probs=quantiles)
}
l2[nrow(l2),1,j] <- mean(l1[ ,n.x1,j] - l1[ ,1,j])
l2[nrow(l2),2:(n.q+1),j] <- quantile(l1[ ,n.x1,j] - l1[ ,1,j], probs=quantiles)
}
dimnames(l2) <- list(paste(c(rep(paste(x1name," ="),n.x1),paste("\u0394","(",x1vals[1],",",x1vals[length(x1vals)],")")),c(x1vals,"")),
c("mean",quantiles),
paste(c(rep(paste(x2name," ="),n.x2)),
c(x2vals)))
if (is.null(did)){did <- c(x2vals[1],x2vals[n.x2])} else{did <- did}
l3 <- array(NA, c(1,n.q+1))
l3[1,1] <- mean((l1[ ,n.x1,match(did[2],x2vals)] - l1[ ,1,match(did[2],x2vals)]) - (l1[ ,n.x1,match(did[1],x2vals)] - l1[ ,1,match(did[1],x2vals)]))
l3[1,2:(n.q+1)] <- quantile((l1[ ,n.x1,match(did[2],x2vals)] - l1[ ,1,match(did[2],x2vals)]) - (l1[ ,n.x1,match(did[1],x2vals)] - l1[ ,1,match(did[1],x2vals)]), probs=quantiles)
dimnames(l3) <- list("did",c("mean",quantiles))
ans <- new("post",
est=round(l2, digits=digits),
did=round(l3, digits=digits),
sims=l1,
model=class(model),
link=model$method,
quantiles=quantiles,
call=call)
return(ans)
}
}
}
setMethod("post", signature(model = "lm"), post.glm)
setMethod("post", signature(model = "glm"), post.glm)
setMethod("post", signature(model = "svyglm"), post.glm)
setMethod("post", signature(model = "polr"), post.polr)