Broadly consider do() and rowwise()

[hadley]

With the idea of (finally!) coming up with a formal replacement.

Imaginary code:

gapminder %>%
  nest_by(country) %>% 
  mutate(fit = lm(lifeExp ~ year, data = data)) %>% 
  summarise(broom::tidy(fit))

Equivalent to:

library(gapminder)
library(tidyverse)

gapminder %>%
  group_by(country) %>%
  nest() %>% 
  mutate(
    fit = map(data, ~ lm(lifeExp ~ year, data = .x)),
    tidy = map(fit, broom::tidy)
  ) %>% 
  select(tidy) %>% 
  unnest(tidy)
#> Adding missing grouping variables: `country`
#> # A tibble: 284 x 6
#> # Groups:   country [142]
#>    country     term         estimate std.error statistic  p.value
#>    <fct>       <chr>           <dbl>     <dbl>     <dbl>    <dbl>
#>  1 Afghanistan (Intercept)  -508.     40.5        -12.5  1.93e- 7
#>  2 Afghanistan year            0.275   0.0205      13.5  9.84e- 8
#>  3 Albania     (Intercept)  -594.     65.7         -9.05 3.94e- 6
#>  4 Albania     year            0.335   0.0332      10.1  1.46e- 6
#>  5 Algeria     (Intercept) -1068.     43.8        -24.4  3.07e-10
#>  6 Algeria     year            0.569   0.0221      25.7  1.81e-10
#>  7 Angola      (Intercept)  -377.     46.6         -8.08 1.08e- 5
#>  8 Angola      year            0.209   0.0235       8.90 4.59e- 6
#>  9 Argentina   (Intercept)  -390.      9.68       -40.3  2.14e-12
#> 10 Argentina   year            0.232   0.00489     47.4  4.22e-13
#> # … with 274 more rows

Or

gapminder %>%
  group_by(country) %>%
  do(fit = lm(lifeExp ~ year, data = .)) %>% 
  do(data.frame(country = .$country, broom::tidy(.$fit)))

This requires:

• mutate.rowwise() needs to automatically wrap in outputs in list where needed.
• rowwise() needs to be able to capture grouping variables; or grouped_df needs some way to activate row-wise magic?
• summarise.rowwise() would return grouped_df since might no longer have 1 row per group?
• New nest_by() that works like group_nest() + rowwise(). Needs a lot of thinking about name.

[romainfrancois]

if we don't need fit to stick around, then we have :

library(gapminder)
library(dplyr, warn.conflicts = FALSE)

gapminder %>%
  filter(continent == "Asia") %>%
  group_by(country) %>% 
  summarise(broom::tidy(lm(lifeExp ~ year)))
#> # A tibble: 66 x 6
#>    country     term        estimate std.error statistic  p.value
#>    <fct>       <chr>          <dbl>     <dbl>     <dbl>    <dbl>
#>  1 Afghanistan (Intercept) -508.      40.5       -12.5  1.93e- 7
#>  2 Afghanistan year           0.275    0.0205     13.5  9.84e- 8
#>  3 Bahrain     (Intercept) -860.      54.3       -15.8  2.07e- 8
#>  4 Bahrain     year           0.468    0.0274     17.0  1.02e- 8
#>  5 Bangladesh  (Intercept) -936.      32.3       -29.0  5.63e-11
#>  6 Bangladesh  year           0.498    0.0163     30.5  3.37e-11
#>  7 Cambodia    (Intercept) -736.     186.         -3.95 2.74e- 3
#>  8 Cambodia    year           0.396    0.0942      4.20 1.82e- 3
#>  9 China       (Intercept) -989.     128.         -7.74 1.57e- 5
#> 10 China       year           0.531    0.0645      8.23 9.21e- 6
#> # … with 56 more rows

^{Created on 2020-01-09 by the reprex package (v0.3.0.9000)}

[romainfrancois]

If we were to :

• roll back summarise() ability to return size > 1
• enlist the results that are not size 1 vectors

Then we can have this:

library(gapminder)
library(dplyr, warn.conflicts = FALSE)

gapminder %>%
  filter(continent == "Asia") %>%
  group_by(country) %>% 
  summarise(
     fit = lm(lifeExp ~ year),
     results = broom::tidy(fit)
  ) 

Perhaps with a warning here because we magic enlist a data frame (so a vector) of size > 1.

And we could make it explicit with results = list(broom::tidy(fit)) instead, but we would not have to make it explicit for fit = because it does not return a vector, nor we would need to add the [[ as long as we are in the summarise() call, as we know we've just made it.

This loses some functionality that summarise() was supposed to gain, i.e. size > 1 results, but that can be the job of do() which would allow it, and potentially not do group peeling.