如何可视化ordinal包有序回归中response*temp交互项的斜率？

我用ordinal包拟合了包含response*temp交互项的有序回归模型，已经提取出response:tempcold（0.14）和response:tempwarm（-0.01）的斜率估计值，但不知道怎么以有序结果变量为y轴来可视化这些斜率。相关代码和模型结果如下：

library(ordinal)
wine2 <- wine
wine2$rating[wine2$rating==5] <- 1
wine2$rating <- ordered(wine2$rating)

fm1 <- clm(rating ~ response*temp, data=wine2)
summary(fm1)

模型输出：

Coefficients:
                  Estimate Std. Error z value Pr(>|z|)    
response           0.14478    0.02965   4.883 1.04e-06 ***
tempwarm           6.89043    1.78282   3.865 0.000111 ***
response:tempwarm -0.15651    0.03874  -4.040 5.34e-05 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Threshold coefficients:
    Estimate Std. Error z value
1|2    3.389      1.022   3.317
2|3    5.610      1.171   4.789
3|4    7.830      1.294   6.053

可视化思路与代码 #

有序回归的系数对应的是**潜变量（潜在响应变量）**的斜率，直接对应你提取的数值。如果要关联到有序因变量的类别，有两种常用方式：

方式1：直接展示潜变量随response的变化（对应斜率）

这种方式能直接呈现你提取的两个斜率，潜变量的变化斜率就是response在不同temp组的系数：

library(ggplot2)

# 创建覆盖response全范围的预测网格
pred_grid <- expand.grid(
  response = seq(min(wine2$response), max(wine2$response), length.out = 100),
  temp = c("cold", "warm")
)

# 预测潜变量值（线性预测器结果）
pred_grid$latent <- predict(fm1, newdata = pred_grid, type = "linear.predictor")

# 绘图
ggplot(pred_grid, aes(x = response, y = latent, color = temp)) +
  geom_line(linewidth = 1) +
  labs(x = "Response", y = "潜在响应变量", color = "温度") +
  theme_minimal()

图中两条线的斜率就是你提取的0.14（cold组）和-0.01（warm组），能直观对比不同温度下response对潜变量的影响方向和大小。

方式2：展示各有序类别概率随response的变化

这种方式更贴合有序因变量的实际意义，能看到response变化时，每个rating等级的概率如何随温度分组变化：

library(tidyr)

# 创建包含所有rating等级的预测网格
pred_grid <- expand.grid(
  response = seq(min(wine2$response), max(wine2$response), length.out = 100),
  temp = c("cold", "warm"),
  rating = levels(wine2$rating)
)

# 预测每个类别的概率
pred_probs <- predict(fm1, newdata = pred_grid, type = "prob")
pred_grid <- cbind(pred_grid, pred_probs)

# 转换为长格式数据
pred_grid_long <- pivot_longer(pred_grid, cols = starts_with("1"), 
                               names_to = "rating_level", values_to = "probability")
pred_grid_long$rating_level <- ordered(pred_grid_long$rating_level, levels = levels(wine2$rating))

# 绘图
ggplot(pred_grid_long, aes(x = response, y = probability, color = temp)) +
  geom_line(linewidth = 0.8) +
  facet_wrap(~rating_level) +
  labs(x = "Response", y = "预测概率", color = "温度") +
  theme_minimal()

通过这个图，你能清晰看到在cold和warm温度下，response变化时各个rating等级的概率升降趋势，间接体现两个斜率的实际影响。

内容的提问来源于stack exchange，提问作者locus