Pending version 1.10.4, awaiting CRAN opening

DESCRIPTION

@@ -1,8 +1,8 @@
 Package: emmeans
 Type: Package
 Title: Estimated Margin2l Means, aka Least-Squares Means
-Version: 1.10.3-090006
-Date: 2024-07-22
+Version: 1.10.3099
+Date: 2024-08-10
 Authors@R: c(person("Russell V.", "Lenth", role = c("aut", "cre", "cph"), 
     email = "[email protected]"),
     person("Ben", "Bolker", role = "ctb"),
@@ -18,8 +18,8 @@ Authors@R: c(person("Russell V.", "Lenth", role = c("aut", "cre", "cph"),
 Depends: R (>= 4.1.0)
 Imports: estimability (>= 1.4.1), graphics, methods, numDeriv, stats, utils,
          mvtnorm
-Suggests: bayesplot, bayestestR, biglm, brms, car, coda (>= 0.17), ggplot2, lattice, 
-          logspline, mediation, mgcv, multcomp, multcompView, nlme,
+Suggests: bayesplot, bayestestR, biglm, brms, car, coda (>= 0.17), compositions,
+          ggplot2, lattice, logspline, mediation, mgcv, multcomp, multcompView, nlme,
           ordinal (>= 2014.11-12), pbkrtest (>= 0.4-1), lme4, lmerTest (>= 2.0.32), 
           MASS, MuMIn, rsm, knitr, rmarkdown, sandwich, scales, splines, testthat, 
           tibble, xtable (>= 1.8-2)

NEWS.md

@@ -2,7 +2,7 @@
 title: "NEWS for the emmeans package"
 ---
 
-## emmeans 1.10.3-090xxxx
+## emmeans 1.10.4
   * Refinements in tracking static offsets
   * Made d.f. consistent for `geeglm` and `glmgee` (#496)
   * Fixed suggestion for installing from GitHub (#497)

R/multinom-support.R

@@ -53,7 +53,7 @@ emm_basis.multinom = function(object, trms, xlev, grid,
     X = kronecker(pat, X)
     nbasis = estimability::all.estble
     nbasis = kronecker(rep(1,k), nbasis)
-    misc = list(tran = "log", inv.lbl = "e^y")
+    misc = list(tran = "clr")   ### misc = list(tran = "log", inv.lbl = "e^y")
     dfargs = list(df = object$edf)
     dffun = function(k, dfargs) dfargs$df
     if(is.null(ylevs <- object$lev))
@@ -70,6 +70,8 @@ emm_basis.multinom = function(object, trms, xlev, grid,
 
 # post-processing of ref_grid for "prob" mode
 # also allows simulated outcomes
+## Note - now that we have mvregrid, we could just pass the latent grid
+## through with the clrInv transform. But this works so I'm just leaving it here
 .multinom.postGrid = function(object, N.sim, ...) {
     linfct = object@linfct
     misc = object@misc

R/multiv.R

@@ -190,6 +190,9 @@ mvregrid = function(object, newname = "component", newlevs = seq_len(ncol(newy))
                     fcn = paste0(tran, "Inv"), ...) {
     levels = levels(object)
     tran = object@misc$tran
+    if(is.character(tran))
+        tran = rev(strsplit(tran, "\\.")[[1]])[1]
+    by = object@misc$by.vars
     if (mult.name != names(levels)[length(levels)]) {
         idx = which(names(levels) == mult.name)
         if(length(idx) == 0) 
@@ -198,32 +201,41 @@ mvregrid = function(object, newname = "component", newlevs = seq_len(ncol(newy))
     }
     ord = .std.order(object@grid, levels)
     if(any(ord != seq_along(ord)))
-        object[ord]
+        object = object[ord] |> update(by.vars = by)
 
-    yhat = matrix(predict(object), ncol = length(levels[[mult.name]]))
+    yhat = matrix(predict(object), ncol = (p <- length(levels[[mult.name]])))
     if(is.character(fcn)) {
-        if(!requireNamespace("compositions"))
+        if(requireNamespace("compositions"))
+            fcn = get(fcn, envir = asNamespace("compositions"))
+        else
             stop("The 'compositions' package must be installed to proceed")
-        fcn = get(fcn, envir = asNamespace("compositions"))
     }
     newy = fcn(yhat, ...)
     if(length(newlevs) != (k <- ncol(newy)))
         stop("Nonconforming 'newlevs': Must be of length ", k)
+    if(!is.na(object@post.beta)) {
+        pb = apply(object@linfct %*% t(object@post.beta), 2, function(y) {
+            as.numeric(fcn(matrix(y, ncol = p), ...))
+        })
+        object@post.beta = t(pb)
+    }
     Jac = numDeriv::jacobian(\(x) as.numeric(fcn(x)), yhat)
     object@V = Jac %*% vcov(object) %*% t(Jac)
     object@bhat = as.numeric(newy)
     object@linfct = diag(length(object@bhat))
     names(levels)[length(levels)] = object@roles$multresp = newname
+    object@roles$predictors = setdiff(object@roles$predictors, mult.name)
     levels[[newname]] = newlevs
     object@levels = levels
     wgt = object@grid[[".wgt."]]
-    offs = object@grid[[".offset."]]
-    object@grid = do.call(expand.grid, levels)
+    g = object@grid = object@grid[seq_len(n <- nrow(newy)), setdiff(names(object@grid), c(mult.name, ".wgt.", ".offset.")), drop = FALSE]
+    for (i in seq_len(k - 1)) object@grid = rbind(object@grid, g)
+    object@grid[["newname"]] = rep(newlevs, each = n)
     if(!is.null(wgt))
-        object@grid[[".wgt."]] = rep(wgt[seq_len(nrow(newy))], k)
-    if(!is.null(offs))
-        object@grid[[".offset."]] = rep(offs[seq_len(nrow(newy))], k)
-    object@misc$tran = object@misc$by = NULL
+        object@grid[[".wgt."]] = rep(wgt[seq_len(n)], k)
+    object@misc$tran = object@misc$inv.lbl = object@misc$sigma = NULL
+    object@misc$pri.vars = setdiff(names(object@grid), c(object@misc$by.vars, ".wgt'"))
+    object@misc$methDesc = "mvregrid"
 
     object
 }

R/qdrg.R

@@ -170,7 +170,7 @@ qdrg = function(formula, data, coef, vcov, df, mcmc, object,
         if (!missing(vcov)) result$vcov = vcov
         if(!missing(df)) result$df = df
         if(missing(contrasts))
-            contrasts = attr(model.matrix(result$formula, data = data[1:2, ]), "contrasts")
+            contrasts = attr(model.matrix(result$formula, data = data), "contrasts")
     }
     if(!missing(df)) result$df = df
     if(is.null(result$df))

R/summary.R

@@ -1,5 +1,5 @@
 ##############################################################################
-#    Copyright (c) 2012-2018 Russell V. Lenth                                #
+#    Copyright (c) 2012-2024 Russell V. Lenth                                #
 #                                                                            #
 #    This file is part of the emmeans package for R (*emmeans*)              #
 #                                                                            #
@@ -154,7 +154,11 @@ as.data.frame.summary_emm = function(x, ...) {
 #'   \samp{object@misc$inv.lbl}. Any \eqn{t} or \eqn{z} tests are still performed
 #'   on the scale of the linear predictor, not the inverse-transformed one.
 #'   Similarly, confidence intervals are computed on the linear-predictor scale,
-#'   then inverse-transformed.
+#'   then inverse-transformed. 
+#'   
+#'   Be aware that only univariate transformations and links are
+#'   supported in this way. Some multivariate transformations are supported by 
+#'   \code{\link{mvregrid}}.
 #'   
 #' @section Bias adjustment when back-transforming:
 #'   When \code{bias.adjust} is \code{TRUE}, then back-transformed estimates
@@ -466,6 +470,10 @@ summary.emmGrid <- function(object, infer, level, adjust, by,
     }
     else
         two.trans = FALSE
+    # Check if we have a log-ratio transformation. At least then we can guess the problem.
+    if ((type == "response") && is.character(object@misc$tran) && endsWith(object@misc$tran, "lr"))
+        warning("It looks like you have a multivariate transformation -- not handled by type = 'response'.\n",
+                "Perhaps instead you wanted to use 'mvregrid()'?")
 
 
 

inst/extexamples/mvregrid.R

@@ -1,11 +1,13 @@
-data(AnimalVegetation)
+data(AnimalVegetation, package = "compositions")
 AV <- as.data.frame(AnimalVegetation)
 
-AVmod <- lm(ilr(cbind(disc,spick,din,spin)) ~ regA, data = AV)
+AVmod <- lm(compositions::ilr(cbind(disc,spick,din,spin)) ~ regA, data = AV)
 
 AVRG <- ref_grid(AVmod)   |> suppressWarnings()
-AVrg
+AVRG
 
-mvregrid(AVRG, newname = "comp", newlevs = c("disc","spick","din","spin")) |>
+mvregrid(AVRG, newname = "comp", 
+         newlevs = c("disc","spick","din","spin"),
+         fcn = "ilrInv") |>
     confint(by = "regA")
 

vignettes/models.Rmd

@@ -202,6 +202,13 @@ typically the tests and confidence intervals are asymptotic. Thus the
 Some objects in this group may require that the original or reference dataset 
 be provided when calling `ref_grid()` or `emmeans()`.
 
+For `coxph` objects, the estimates we obtain are comparable to running 
+`predict.coxph()` with `reference = "zero"`; that is, no covariate centering is done.
+The user may use `at` to specify adjusted covariate values. For example, the 
+default reference grid sets each covariate to its mean, so estimates comparable
+to `predict.coxph(..., reference = "sample")` would be obtained by specifying
+`at = list(x = 0)`.
+
 In the case of `mgcv::gam` objects, there are optional `freq` and
 `unconditional` arguments as is detailed in the documentation for
 `mgcv::vcov.gam()`. Both default to `FALSE`. The value of `unconditional`