posterior predictive check for binomial glm with matrix response

DESCRIPTION

@@ -1,7 +1,7 @@
 Type: Package
 Package: performance
 Title: Assessment of Regression Models Performance
-Version: 0.10.7.1
+Version: 0.10.7.2
 Authors@R: 
     c(person(given = "Daniel",
              family = "Lüdecke",
@@ -153,3 +153,4 @@ Config/Needs/website:
     r-lib/pkgdown,
     easystats/easystatstemplate
 Config/rcmdcheck/ignore-inconsequential-notes: true
+Remotes: easystats/insight

NEWS.md

@@ -1,3 +1,16 @@
+# performance 0.10.8
+
+## Changes
+
+* Changed behaviour of `check_predictions()` for models from binomial family,
+  to get comparable plots for different ways of outcome specification. Now,
+  if the outcome is a proportion, or defined as matrix of trials and successes,
+  the produced plots are the same (because the models should be the same, too).
+
+## Bug fixes
+
+* Fixed CRAN check errors.
+
 # performance 0.10.7
 
 ## Breaking changes

R/binned_residuals.R

@@ -29,6 +29,7 @@
 #'   time-consuming. By default, `show_dots = NULL`. In this case `binned_residuals()`
 #'   tries to guess whether performance will be poor due to a very large model
 #'   and thus automatically shows or hides dots.
+#' @param verbose Toggle warnings and messages.
 #' @param ... Currently not used.
 #'
 #' @return A data frame representing the data that is mapped in the accompanying
@@ -83,11 +84,20 @@ binned_residuals <- function(model,
                              ci_type = c("exact", "gaussian", "boot"),
                              residuals = c("deviance", "pearson", "response"),
                              iterations = 1000,
+                             verbose = TRUE,
                              ...) {
   # match arguments
   ci_type <- match.arg(ci_type)
   residuals <- match.arg(residuals)
 
+  # for non-bernoulli models, `"exact"` doesn't work
+  if (isFALSE(insight::model_info(model)$is_bernoulli)) {
+    ci_type <- "gaussian"
+    if (verbose) {
+      insight::format_alert("Using `ci_type = \"gaussian\"` because model is not bernoulli.")
+    }
+  }
+
   fitted_values <- stats::fitted(model)
   mf <- insight::get_data(model, verbose = FALSE)
 
@@ -186,7 +196,7 @@ binned_residuals <- function(model,
   }
   out <- out / n
 
-  quant <- stats::quantile(out, c((1 - ci) / 2, (1 + ci) / 2))
+  quant <- stats::quantile(out, c((1 - ci) / 2, (1 + ci) / 2), na.rm = TRUE)
   c(CI_low = quant[1L], CI_high = quant[2L])
 }
 

R/check_model.R

@@ -49,7 +49,7 @@
 #'
 #' @details For Bayesian models from packages **rstanarm** or **brms**,
 #' models will be "converted" to their frequentist counterpart, using
 #' [`bayestestR::bayesian_as_frequentist`](https://easystats.github.io/bayestestR/reference/convert_bayesian_as_frequentist.html).
 #' A more advanced model-check for Bayesian models will be implemented at a
 #' later stage.
 #'
@@ -77,7 +77,7 @@
 #' plots are helpful to check model assumptions, they do not necessarily indicate
 #' so-called "lack of fit", e.g. missed non-linear relationships or interactions.
 #' Thus, it is always recommended to also look at
 #' [effect plots, including partial residuals](https://strengejacke.github.io/ggeffects/articles/introduction_partial_residuals.html).
 #'
 #' @section Homogeneity of Variance:
 #' This plot checks the assumption of equal variance (homoscedasticity). The
@@ -374,7 +374,7 @@
   dat$INFLUENTIAL <- .influential_obs(model, threshold = threshold)
   dat$PP_CHECK <- .safe(check_predictions(model, ...))
   if (isTRUE(model_info$is_binomial)) {
-    dat$BINNED_RESID <- binned_residuals(model, ...)
+    dat$BINNED_RESID <- binned_residuals(model, verbose = verbose, ...)
   }
   if (isTRUE(model_info$is_count)) {
     dat$OVERDISPERSION <- .diag_overdispersion(model)

R/check_predictions.R

@@ -104,7 +104,7 @@
   minfo <- insight::model_info(object, verbose = FALSE)
 
   # try to find sensible default for "type" argument
   suggest_dots <- (minfo$is_bernoulli || minfo$is_count || minfo$is_ordinal || minfo$is_categorical || minfo$is_multinomial)
   if (missing(type) && suggest_dots) {
     type <- "discrete_interval"
   }
@@ -197,10 +197,10 @@
   out <- .check_re_formula(out, object, iterations, re_formula, verbose, ...)
 
   # save information about model
-  if (!is.null(model_info)) {
-    minfo <- model_info
-  } else {
+  if (is.null(model_info)) {
     minfo <- insight::model_info(object)
+  } else {
+    minfo <- model_info
   }
 
   # glmmTMB returns column matrix for bernoulli
@@ -215,9 +215,10 @@
   }
 
   if (is.null(out)) {
-    insight::format_error(
-      sprintf("Could not simulate responses. Maybe there is no `simulate()` for objects of class `%s`?", class(object)[1])
-    )
+    insight::format_error(sprintf(
+      "Could not simulate responses. Maybe there is no `simulate()` for objects of class `%s`?",
+      class(object)[1]
+    ))
   }
 
   # get response data, and response term, to check for transformations
@@ -263,7 +264,7 @@
   out <- tryCatch(
     {
       matrix_sim <- stats::simulate(object, nsim = iterations, re.form = re_formula, ...)
-      as.data.frame(sapply(matrix_sim, function(i) i[, 1] / i[, 2], simplify = TRUE))
+      as.data.frame(sapply(matrix_sim, function(i) i[, 1] / rowSums(i, na.rm = TRUE), simplify = TRUE))
     },
     error = function(e) {
       NULL
@@ -274,9 +275,10 @@
   out <- .check_re_formula(out, object, iterations, re_formula, verbose, ...)
 
   if (is.null(out)) {
-    insight::format_error(
-      sprintf("Could not simulate responses. Maybe there is no `simulate()` for objects of class `%s`?", class(object)[1])
-    )
+    insight::format_error(sprintf(
+      "Could not simulate responses. Maybe there is no `simulate()` for objects of class `%s`?",
+      class(object)[1]
+    ))
   }
 
   # get response data, and response term
@@ -285,13 +287,13 @@
   )
   resp_string <- insight::find_terms(object)$response
 
-  out$y <- response[, 1] / response[, 2]
+  out$y <- response[, 1] / rowSums(response, na.rm = TRUE)
 
   # safe information about model
-  if (!is.null(model_info)) {
-    minfo <- model_info
-  } else {
+  if (is.null(model_info)) {
     minfo <- insight::model_info(object)
+  } else {
+    minfo <- model_info
   }
 
   attr(out, "check_range") <- check_range
@@ -363,14 +365,20 @@
     if (is.numeric(original)) {
       if (min(replicated) > min(original)) {
         insight::print_color(
-          insight::format_message("Warning: Minimum value of original data is not included in the replicated data.", "Model may not capture the variation of the data."),
+          insight::format_message(
+            "Warning: Minimum value of original data is not included in the replicated data.",
+            "Model may not capture the variation of the data."
+          ),
           "red"
         )
       }
 
       if (max(replicated) < max(original)) {
         insight::print_color(
-          insight::format_message("Warning: Maximum value of original data is not included in the replicated data.", "Model may not capture the variation of the data."),
+          insight::format_message(
+            "Warning: Maximum value of original data is not included in the replicated data.",
+            "Model may not capture the variation of the data."
+          ),
           "red"
         )
       }
@@ -423,7 +431,7 @@
     if (is.null(plus_minus)) {
       sims[] <- lapply(sims, exp)
     } else {
       sims[] <- lapply(sims, function(i) exp(i) - plus_minus)
     }
   } else if (grepl("log1p(", resp_string, fixed = TRUE)) {
     sims[] <- lapply(sims, expm1)

R/looic.R

@@ -12,6 +12,7 @@
 #' @return A list with four elements, the ELPD, LOOIC and their standard errors.
 #'
 #' @examplesIf require("rstanarm")
+#' \donttest{
 #' model <- suppressWarnings(rstanarm::stan_glm(
 #'   mpg ~ wt + cyl,
 #'   data = mtcars,
@@ -20,6 +21,7 @@
 #'   refresh = 0
 #' ))
 #' looic(model)
+#' }
 #' @export
 looic <- function(model, verbose = TRUE) {
   insight::check_if_installed("loo")