Respect column names when "splicing" matrices in summarize()?

[MichaelChirico]

Compare:

penguins %>%
  summarize(.by = species, rbind(c(a = mean(bill_len), b = sd(bill_len))))
#     species rbind(c(a = mean(bill_len), b = sd(bill_len))).1 rbind(c(a = mean(bill_len), b = sd(bill_len))).2
# 1    Adelie                                               NA                                               NA
# 2    Gentoo                                               NA                                               NA
# 3 Chinstrap                                        48.833824                                         3.339256

penguins %>%
  summarize(.by = species, data.frame(as.list(c(a = mean(bill_len), b = sd(bill_len)))))
#     species        a        b
# 1    Adelie       NA       NA
# 2    Gentoo       NA       NA
# 3 Chinstrap 48.83382 3.339256

That is, when summarize() auto-converts a matrix into columns, we get ugly deparse()-inferred names, including totally losing a & b (now 1 and 2), but when it does so for a data.frame, we nicely get the nested data.frame's own column names.

When we have control over the function, we can prefer returning a data.frame/tibble, but sometimes we don't want to touch the function and the rbind() approach will look nicer (even if the same operation results under the hood).

The context here is related to #7689 -- trying to convert some plyr::ddply() into "standard" {dplyr} code as well & simply as I can, without needing to drastically refactor downstream package code if I can avoid it.

[DavisVaughan]

Just as an FYI, the matrix is not actually spliced in

library(dplyr, warn.conflicts = FALSE)

xx <- penguins %>%
  summarize(.by = species, rbind(c(a = mean(bill_len), b = sd(bill_len))))

xx$`rbind(c(a = mean(bill_len), b = sd(bill_len)))`
#>          [,1]     [,2]
#> [1,]       NA       NA
#> [2,]       NA       NA
#> [3,] 48.83382 3.339256

We call this a "matrix column". The whole matrix is right there in the column. It's just getting the same "auto naming" treatment as almost any other arbitrary expression, where we use the deparsed expression as the name:

library(dplyr, warn.conflicts = FALSE)

penguins %>%
  transmute(.by = species, bill_len + bill_dep) %>%
  as_tibble()
#> # A tibble: 344 × 2
#>    .by    `bill_len + bill_dep`
#>    <fct>                  <dbl>
#>  1 Adelie                  57.8
#>  2 Adelie                  56.9
#>  3 Adelie                  58.3
#>  4 Adelie                  NA  
#>  5 Adelie                  56  
#>  6 Adelie                  59.9
#>  7 Adelie                  56.7
#>  8 Adelie                  58.8
#>  9 Adelie                  52.2
#> 10 Adelie                  62.2
#> # ℹ 334 more rows

The exception to this is if you return a data frame (without naming it!), where we do "auto unpack" the data frame into its individual columns, as you saw. Compare what happens if you name it:

library(dplyr, warn.conflicts = FALSE)

penguins <- as_tibble(penguins)

penguins %>%
  summarize(
    .by = species,
    tibble(mean = mean(bill_len), sd = mean(bill_dep))
  )
#> # A tibble: 3 × 3
#>   species    mean    sd
#>   <fct>     <dbl> <dbl>
#> 1 Adelie     NA    NA  
#> 2 Gentoo     NA    NA  
#> 3 Chinstrap  48.8  18.4

penguins %>%
  summarize(
    .by = species,
    df_col = tibble(mean = mean(bill_len), sd = mean(bill_dep))
  )
#> # A tibble: 3 × 2
#>   species   df_col$mean   $sd
#>   <fct>           <dbl> <dbl>
#> 1 Adelie           NA    NA  
#> 2 Gentoo           NA    NA  
#> 3 Chinstrap        48.8  18.4

# the column itself is another data frame!
penguins %>%
  summarize(
    .by = species,
    df_col = tibble(mean = mean(bill_len), sd = mean(bill_dep))
  ) %>%
  pull(df_col)
#> # A tibble: 3 × 2
#>    mean    sd
#>   <dbl> <dbl>
#> 1  NA    NA  
#> 2  NA    NA  
#> 3  48.8  18.4

I suppose you could argue that we should do the same thing for matrices (very specifically, 2D arrays, no other arity allowed), but I imagine we'd rather keep this feature tightly scoped to matrices instead. I don't imagine we are going to relax this to include matrices, so I'm going to go ahead and close because of that!

[MichaelChirico]

Thanks for the added details!

do the same thing for matrices... tightly scoped to matrices

I guess the second one should be 'data.frame's? instead?

[DavisVaughan]

oh yes sorry