r - How to Nest Data for Pairwise Group Comparisons in dplyr? - Stack Overflow

admin管理员组

文章数量:1123433

I often use the nest() function in dplyr to perform models on nested tibbles. For example, testing within each group if the value changes over time:

library(dplyr)
library(lme4)
set.seed(123)

data <- tibble(
  ID = rep(1:50, each = 2),            
  time = rep(c(0, 1), times = 50),
  group = rep(sample(c("A", "B", "C"), 50, replace = TRUE), each = 2),
  value = runif(100)
)

data %>% 
  group_by(group) %>% 
  nest() %>% 
  mutate(lmm = map(data, \(x) lme4::lmer(value ~ time + (1|ID),
                                         data = x)))

Now, I want to compare how groups differ over time by excluding one group at a time. Specifically, I want to fit the model value ~ time * group + (1 | ID) for all combinations of two groups.

Is it possible to nest the data such that each row in the nested tibble represents combinations of two groups? How can this be achieved efficiently in dplyr?

I often use the nest() function in dplyr to perform models on nested tibbles. For example, testing within each group if the value changes over time:

library(dplyr)
library(lme4)
set.seed(123)

data <- tibble(
  ID = rep(1:50, each = 2),            
  time = rep(c(0, 1), times = 50),
  group = rep(sample(c("A", "B", "C"), 50, replace = TRUE), each = 2),
  value = runif(100)
)

data %>% 
  group_by(group) %>% 
  nest() %>% 
  mutate(lmm = map(data, \(x) lme4::lmer(value ~ time + (1|ID),
                                         data = x)))

Now, I want to compare how groups differ over time by excluding one group at a time. Specifically, I want to fit the model value ~ time * group + (1 | ID) for all combinations of two groups.

Is it possible to nest the data such that each row in the nested tibble represents combinations of two groups? How can this be achieved efficiently in dplyr?

• r
• dplyr

Share Improve this question Follow edited 15 hours ago ThomasIsCoding 99.9k99 gold badges3434 silver badges9999 bronze badges asked 15 hours ago user23485480user23485480 4744 bronze badges 3

• 4 I have no doubt this is programmatically possible, but I am not sure it is statistically sensible. I suspect that constructing contrasts to test each pair of treatments based on the full model is likely to be more powerful than fitting "pairwise models". Crossvalidated would be the place to confirm this. – Limey Commented 15 hours ago
• 1 FWIW, nest() is a function in tidyr and not in dplyr. (Perhaps you loaded both with tidyverse?) – Jon Spring Commented 10 hours ago
• 1 Use m <- lmer(value ~ time * group + (1 | ID), data = data), then emmeans::emtrends(m, pairwise ~ group, 'time'). That should give you the pairwise difference in slopes between groups. – Axeman Commented 8 hours ago

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 1

If I understand correctly, you could use combn to create a tibble with pairwise combinations, then map it:

mdl_results <- tibble(groups = combn(unique(data$group), 2, simplify = FALSE)) %>%
  mutate(
    lmm = purrr::map(groups, ~ lme4::lmer(value ~ time * group + (1 | ID), 
                             data = filter(data, group %in% .x))),
    summary = purrr::map(lmm, summary)
  )

#   groups    lmm       summary   
#   <list>    <list>    <list>    
#   1 <chr [2]> <lmerMod> <smmry.mM>
#   2 <chr [2]> <lmerMod> <smmry.mM>
#   3 <chr [2]> <lmerMod> <smmry.mM>

Though this is just programmatic - I agree with the comment that you may want to ask the statistical validity on Cross Validated. Good luck!

本文标签： rHow to Nest Data for Pairwise Group Comparisons in dplyrStack Overflow