我有一个GAM模型（用mgcv拟合），它在by参数中有两个因子的交互作用

gam(y ~ s(percent, by = source:treatment), data = dat)

如果我试图从那个模型中提取emmeans，它会给我一个错误

Error in `[.data.frame`(tbl, , vars, drop = FALSE) : 
  undefined columns selected

这里有一个完全可复制的例子

library(tidyverse)
library(mgcv)
library(emmeans)

# Simulate Data -----------------------------------------------------------
# similar to ?emmeans::pigs
mycurve <- function(x, h = 4, s = 2){
  30-h + (2*h)/(1+exp((9-x)/s))
}
d_source <- tibble(source = factor(c("fish", "soy", "skim")),
                   hi = 4:6)
d_treatment <- tibble(treatment = factor(c("control", "treatment")),
                      si = c(2, 0.4))
dat <- expand_grid(d_source, d_treatment,
                   percent = seq(9, 18, by = 0.5))
set.seed(1)
dat <- dat %>% 
  rowwise() %>% 
  mutate(mu = mycurve(x = percent, h = hi, s = si),
         y = mu + rnorm(n = 1, mean = 0, sd = 0.1))
dat %>% 
  ggplot(aes(x = percent, y = y, color = source, shape = treatment,
             group = interaction(treatment, source))) +
  geom_point() +
  geom_line()

# Fit model ---------------------------------------------------------------
fit <- gam(y ~ s(percent, by = source:treatment), data = dat)
plot(fit, pages = 1)

# Extract EMM -------------------------------------------------------------
emmeans(fit, specs = "percent", by = c("treatment", "source"),
        at = list(percent = 9:18))

你知道为什么这不起作用吗？或者是emmeans::recover_data.gam()函数中的一个bug？
一个简单的解决方法是定义一个新的因素，但这不便于以后进行具体的对比。

dat <- dat %>% 
  mutate(st = interaction(source, treatment))
fit2 <- gam(y ~ s(percent, by = st), data = dat)
emmeans(fit2, specs = "percent", by = "st",
        at = list(percent = 9:18))