estimate_contrast(): accept any "predict" + backends divergence

DESCRIPTION

@@ -1,7 +1,7 @@
 Type: Package
 Package: modelbased
 Title: Estimation of Model-Based Predictions, Contrasts and Means
-Version: 0.10.0.31
+Version: 0.10.0.32
 Authors@R:
     c(person(given = "Dominique",
              family = "Makowski",

R/clean_names.R

@@ -28,13 +28,7 @@
 
   if (length(vars) == 1) {
     if (type == "contrast") {
-      if (minfo$is_logit && predict == "response") {
-        names(means)[names(means) == vars] <- "Odds_ratio"
-      } else if (minfo$is_poisson && predict == "response") {
-        names(means)[names(means) == vars] <- "Ratio"
-      } else {
-        names(means)[names(means) == vars] <- "Difference"
-      }
+      names(means)[names(means) == vars] <- "Difference"
     } else if (type == "mean") {
       if (minfo$is_logit && predict == "response") {
         names(means)[names(means) == vars] <- "Probability"

R/estimate_contrast_methods.R

@@ -31,7 +31,11 @@ estimate_contrasts.estimate_predicted <- function(model,
   minfo <- insight::model_info(model, response = 1)
 
   # model df
-  dof <- insight::get_df(model, type = "wald", verbose = FALSE)
+  if (minfo$is_bayesian) {
+    dof <- Inf
+  } else {
+    dof <- insight::get_df(model, type = "wald", verbose = FALSE)
+  }
   crit_factor <- (1 + ci) / 2
 
   ## TODO: For Bayesian models, we always use the returned standard errors

R/get_emcontrasts.R

@@ -41,6 +41,22 @@
   # extract first focal term
   first_focal <- my_args$contrast[1]
 
+  # setup arguments
+  fun_args <- list(model)
+
+  # handle distributional parameters
+  if (predict %in% .brms_aux_elements(model) && inherits(model, "brmsfit")) {
+    dpars <- TRUE
+    fun_args$dpar <- predict
+  } else {
+    dpars <- FALSE
+    fun_args$type <- predict
+  }
+
+  # add dots
+  dots <- list(...)
+  fun_args <- insight::compact_list(c(fun_args, dots))
+
   # if first focal term is numeric, we contrast slopes
   if (is.numeric(model_data[[first_focal]]) &&
     !first_focal %in% on_the_fly_factors &&
@@ -51,23 +67,13 @@
       insight::format_error("Please specify the `by` argument to calculate contrasts of slopes.") # nolint
     }
     # Run emmeans
-    estimated <- suppressMessages(emmeans::emtrends(
-      model,
-      specs = my_args$by,
-      var = my_args$contrast,
-      type = predict,
-      ...
-    ))
+    fun_args <- c(fun_args, list(specs = my_args$by, var = my_args$contrast))
+    estimated <- suppressMessages(do.call(emmeans::emtrends, fun_args))
     emm_by <- NULL
   } else {
     # Run emmeans
-    estimated <- suppressMessages(emmeans::emmeans(
-      model,
-      specs = my_args$emmeans_specs,
-      at = my_args$emmeans_at,
-      type = predict,
-      ...
-    ))
+    fun_args <- c(fun_args, list(specs = my_args$emmeans_specs, at = my_args$emmeans_at))
+    estimated <- suppressMessages(do.call(emmeans::emmeans, fun_args))
     # Find by variables
     emm_by <- my_args$emmeans_specs[!my_args$emmeans_specs %in% my_args$contrast]
     if (length(emm_by) == 0) {
@@ -76,7 +82,7 @@
   }
 
   # If means are on the response scale (e.g., probabilities), need to regrid
-  if (predict == "response") {
+  if (predict == "response" || dpars) {
     estimated <- emmeans::regrid(estimated)
   }
 
@@ -142,7 +148,7 @@
 
 
 .format_emmeans_contrasts <- function(model, estimated, ci, p_adjust, ...) {
   predict <- attributes(estimated)$predict
   m_info <- insight::model_info(model, response = 1)
 
   # Summarize and clean

R/get_emmeans.R

@@ -58,8 +58,10 @@
 
   # handle distributional parameters
   if (predict %in% .brms_aux_elements(model) && inherits(model, "brmsfit")) {
+    dpars <- TRUE
     fun_args$dpar <- predict
   } else {
+    dpars <- FALSE
     fun_args$type <- predict
   }
 
@@ -70,6 +72,11 @@
   # Run emmeans
   estimated <- suppressMessages(suppressWarnings(do.call(emmeans::emmeans, fun_args)))
 
+  # backtransform to response scale for dpars
+  if (dpars) {
+    estimated <- emmeans::regrid(estimated)
+  }
+
   # Special behaviour for transformations #138 (see below)
   if ("retransform" %in% names(my_args) && length(my_args$retransform) > 0) {
     for (var in names(my_args$retransform)) {
@@ -146,7 +153,7 @@
 # Table formatting emmeans ----------------------------------------------------
 
 .format_emmeans_means <- function(x, model, ci = 0.95, verbose = TRUE, ...) {
   predict <- attributes(x)$predict
   m_info <- insight::model_info(model, response = 1)
 
   # Summarize and clean

R/get_marginalmeans.R

@@ -131,7 +131,7 @@
   dots[c("by", "conf_level", "type", "digits", "bias_correction", "sigma", "offset")] <- NULL
 
   # model df - can be passed via `...`
-  if (is.null(dots$df)) {
+  if (is.null(dots$df) && !model_info$is_bayesian) {
     dots$df <- insight::get_df(model, type = "wald", verbose = FALSE)
   }
 
@@ -225,7 +225,7 @@
   # just need to add "hypothesis" argument
   means <- .call_marginaleffects(fun_args)
 
-  # Fifth step: post-processin marginal means----------------------------------
+  # Fifth step: post-processing marginal means----------------------------------
   # ---------------------------------------------------------------------------
 
   # filter "by" rows when we have "average" marginalization, because we don't
@@ -299,13 +299,13 @@
 
 .marginaleffects_errors <- function(out, fun_args) {
   # what was requested?
   if (!is.null(fun_args$hypothesis)) {
     fun <- "marginal contrasts"
   } else {
     fun <- "marginal means"
   }
   # clean original error message
   out$message <- gsub("\\s+", " ", gsub("\n", "", out$message))
   # setup clear error message
   msg <- c(
     paste0("Sorry, calculating ", fun, " failed with following error:"),
@@ -317,7 +317,7 @@
     msg <- c(msg, "\nIt seems that not all required levels of the focal terms are available in the provided data. If you want predictions extrapolated to a hypothetical target population, try setting `estimate=\"population\".") # nolint
   }
   # we get this error for models with complex random effects structures in glmmTMB
-  if (grepl("map factor length must equal", out$message, fixed = TRUE)) {
+  if (grepl("map factor length must equal", out$message, fixed = TRUE) || grepl("cannot allocate", out$message, fixed = TRUE)) { # nolint
     msg <- c(
       msg,
       paste0(
@@ -335,7 +335,7 @@
 }
 
 
-# filter datagrid foe `estimate = "average"`---------------------------------
+# filter datagrid for `estimate = "average"`---------------------------------
 
 .filter_datagrid_average <- function(means, estimate, datagrid, datagrid_info) {
   # filter "by" rows when we have "average" marginalization, because we don't
@@ -472,22 +472,22 @@
       # if no offset argument was specified, tell user what this means
       msg <- switch(estimate,
         specific = ,
         typical = "Model contains an offset-term, which is set to its mean value. If you want to average predictions over the distribution of the offset (if appropriate), use `estimate = \"average\"`. If you want to fix the offset to a specific value, for instance `1`, use `offset = 1`.",
         "Model contains an offset-term and you average predictions over the distribution of that offset. If you want to fix the offset to a specific value, for instance `1`, use `offset = 1` and set `estimate = \"typical\"`."
       )
       # if offset term is log-transformed, tell user. offset should be fixed then
       log_offset <- insight::find_transformation(insight::find_offset(model, as_term = TRUE))
       if (!is.null(log_offset) && startsWith(log_offset, "log")) {
         msg <- c(
           msg,
           "We also found that the model has a log-transformed offset term. If you use the `offset` argument, the log-transformation will automatically be applied to the provided offset-value. I.e., consider using, for instance, `offset = 10` and not `offset = log(10)`."
         )
       }
     } else {
       # if offset was specified, and estimate averages over predictions, tell this
       msg <- switch(estimate,
         average = ,
         population = paste0("For `estimate = \"", estimate, "\"`, predictions are averaged over the distribution of the offset and the `offset` argument is ignored. If you want to fix the offset to a specific value, for instance `1`, use `offset = 1` and set `estimate = \"typical\"`.")
       )
     }
     if (!is.null(msg)) {

R/get_marginaltrends.R

@@ -85,7 +85,7 @@
   }
 
   # model df - can be passed via `...`
-  if (is.null(dots$df)) {
+  if (is.null(dots$df) && !model_info$is_bayesian) {
     dots$df <- insight::get_df(model, type = "wald", verbose = FALSE)
   }
 
@@ -204,10 +204,10 @@
   # check if user provided values in `trend`, e.g. `trend=1:10`. We then pass
   # this argument to also create a data grid, but we also need to "clean" trend
   if (grepl("=", trend, fixed = TRUE)) {
     range <- trend
     trend <- gsub("=.*", "\\1", trend)
   } else {
     range <- NULL
   }
 
   # make sure range in `trend` is not also in `by`