`marginaleffects`错误：不支持矩阵列

Question

总结

我正在尝试使用 marginaleffects::plot_predictions() 来绘制 GLMM 的预测。下面提供了一个最小的工作示例。

目标

我试图仅在以下 GLMM 中绘制四向交互的预测 scaled_exp * 协议 * 条件 * 组：

glmer(Acceptance ~  scaled_exp * Agreement * Condition * Group +
                    scaled_age * Agreement * Group +
                    (1 | Item) +
                    (1 + Condition | Subject)

但是，如果不是每个预测变量都包含在条件列表中（即，我们对绘制 scaled_age 的影响不感兴趣），那么 margineffects::plot_predictions() 似乎会遇到问题（仅适用于 GLMM，不适用于 LMM） ，它只是为了控制老化对研究结果的影响），但如果我尝试在没有 scaled_age 的情况下运行它，我会收到以下错误。

警告：矩阵列不支持作为预测变量，因此被省略。这可能会妨碍数量的计算 出于兴趣。您可以构建自己的预测数据集并将其显式提供给

newdata

参数。警告：某些列是多列类型（例如矩阵列）：[2, 6]。 setDT 将按原样保留这些列，但后续操作（如分组和联接）可能会失败。请考虑 as.data.table() ，它将为每个嵌入列创建一个新列。错误：无法使用此模型计算预测值。您可以尝试为

newdata

参数提供不同的数据集。还引发了此错误：找不到对象“scaled_age”。错误跟踪器：https://github.com/vincentarelbundock/marginaleffects/issues

我没有遇到具有类似固定和随机效应结构的相应 LMM 的这些问题，所以我怀疑这与此有关。

MWE

# Load necessary libraries
library(lme4)
library(tidyverse)
library(marginaleffects)

# Create a sample dataset
set.seed(123)
n <- 5000

sample_data <- tibble(
  Subject = as.factor(rep(1:500, each = 10)),  # Increase the number of subjects
  Group = factor(rep(c("Group1", "Group2"), each = n / 2)),
  Item = sample(1:120, n, replace = TRUE),
  Condition = factor(rep(c("Cond1", "Cond2"), times = n / 2)),
  Agreement = factor(sample(c("Agree", "Disagree"), n, replace = TRUE)),
  Acceptance = rbinom(n, 1, 0.5),
  scaled_age = scale(rnorm(n, 40, 15)),
  scaled_exp = scale(rnorm(n, 0.3, 1.4))
)

# Fit the generalized linear mixed model
glmm_mod <- glmer(Acceptance ~ 
                    scaled_exp * Agreement * Condition * Group +
                    scaled_age * Agreement * Group +
                    (1 | Item) +
                    (1 + Condition | Subject), 
                  data = sample_data, 
                  family = binomial, 
                  control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 100000)))

# Generate predictions for the four-way interaction using plot_predictions()
plot_predictions(
  glmm_mod,
  condition = list("scaled_exp", "Agreement", "Condition", "Group"),
  re.form = NA
) +
  aes(linetype = Agreement) +  # Map linetype to Agreement
  facet_wrap(~ Group + Condition, ncol = 4) +
  scale_color_manual(values = c("Agree" = "blue", "Disagree" = "red")) +
  scale_fill_manual(values = c("Agree" = "blue", "Disagree" = "red")) +
  scale_linetype_manual(values = c("Agree" = "solid", "Disagree" = "longdash")) +  
  theme_classic() +
  labs(x = "Scaled Experience", y = "Predicted Probability of Acceptance")

Answer 1

如果您再次仔细阅读错误消息，您会发现其中很大一部分与

marginaleffects

不支持“矩阵列”有关。如果您检查数据集，您会注意到某些列确实属于“矩阵”类：

> class(sample_data$scaled_age)
[1] "matrix" "array" 

事实证明

scale()

函数默认返回一个矩阵，而不是一个数值向量。要解决此问题，您可以使用

drop()

函数：

library(lme4)
library(tidyverse)
library(marginaleffects)

set.seed(123)
n <- 5000

sample_data <- tibble(
  Subject = as.factor(rep(1:500, each = 10)),  # Increase the number of subjects
  Group = factor(rep(c("Group1", "Group2"), each = n / 2)),
  Item = sample(1:120, n, replace = TRUE),
  Condition = factor(rep(c("Cond1", "Cond2"), times = n / 2)),
  Agreement = factor(sample(c("Agree", "Disagree"), n, replace = TRUE)),
  Acceptance = rbinom(n, 1, 0.5),
  scaled_age = drop(scale(rnorm(n, 40, 15))),
  scaled_exp = drop(scale(rnorm(n, 0.3, 1.4)))
)

glmm_mod <- glmer(Acceptance ~ 
                    scaled_exp * Agreement * Condition * Group +
                    scaled_age * Agreement * Group +
                    (1 | Item) +
                    (1 + Condition | Subject), 
                  data = sample_data, 
                  family = binomial, 
                  control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 100000)))

plot_predictions(
  glmm_mod,
  condition = list("scaled_exp", "Agreement", "Condition", "Group"),
  re.form = NA
)