Creating models and augmenting data without losing additional columns in dplyr/broom

Question

Consider the following data / example. Each dataset contains a number of samples with one observation and one estimate:

library(tidyverse)
library(broom)

data = read.table(text = '
dataset sample_id   observation estimate
A   A1  4.8 4.7
A   A2  4.3 4.5
A   A3  3.1 2.9
A   A4  2.1 2
A   A5  1.1 1
B   B1  4.5 4.3
B   B2  3.9 4.1
B   B3  2.9 3
B   B4  1.8 2
B   B5  1   1.2
', header = TRUE)

I want to calculate a linear model per dataset to remove any linear bias between observation and estimate, and get the fitted values next to the original ones:

data %>%
  group_by(dataset) %>% 
  do(lm(observation ~ estimate, data = .) %>% augment)

What this does, however, is remove the sample_id column, which I need to keep in order to continue calculations with this dataset based on that unique ID:

# A tibble: 10 x 10
# Groups:   dataset [2]
   dataset observation estimate .fitted .se.fit   .resid  .hat .sigma  .cooksd .std.resid
   <fct>         <dbl>    <dbl>   <dbl>   <dbl>    <dbl> <dbl>  <dbl>    <dbl>      <dbl>
 1 A              4.80     4.70   4.68   0.107   0.115   0.478 0.152  0.491        1.04  
 2 A              4.30     4.50   4.49   0.0996 -0.193   0.416 0.0609 0.957       -1.64  
 3 A              3.10     2.90   2.97   0.0693  0.135   0.201 0.156  0.120        0.976 
 4 A              2.10     2.00   2.11   0.0849 -0.00583 0.303 0.189  0.000444    -0.0452
 5 A              1.10     1.00   1.15   0.120  -0.0508  0.602 0.180  0.206       -0.521 
 6 B              4.50     4.30   4.31   0.109   0.191   0.468 0.0597 1.20         1.65  
 7 B              3.90     4.10   4.09   0.100  -0.193   0.396 0.0844 0.798       -1.56  
 8 B              2.90     3.00   2.91   0.0713 -0.00630 0.201 0.195  0.000247    -0.0443
 9 B              1.80     2.00   1.83   0.0898 -0.0275  0.319 0.193  0.0103      -0.210 
10 B              1.00     1.20   0.964  0.125   0.0355  0.616 0.191  0.104        0.360

How can I keep the additional column from my original dataset?

I have seen this answer which uses nest to collapse the data before, but I still only get the model parameters using this approach. I guess I could extract the parameters per dataset:

data %>%
  group_by(dataset) %>% 
  nest() %>% 
  mutate(
    mod = map(data, linear_adj_model),
    pars = map(mod, tidy)
  ) %>%
  unnest(pars) %>%
  select(dataset, term, estimate) %>%
  spread(term, estimate)

… which gives me this:

# A tibble: 2 x 3
  dataset `(Intercept)` estimate
* <fct>           <dbl>    <dbl>
1 A               0.196    0.955
2 B              -0.330    1.08

… and then left-join that with the original data, and then mutate each estimate to get the linearly-adjusted one, but that seems much too complicated.

Another ugly hack I've found consists in adding the column as a dummy variable to the model:

data %>%
  group_by(dataset) %>% 
  do(lm(observation ~ estimate + 0 * sample_id, data = .) %>% augment)

Is there an easier (tidy) solution that does not involve manually specifying the variables I want to keep?

Another hack-ish approach is to `join()` your data back onto the resulting dataframe. If you pipe the result of `do(lm(observation ~ estimate, data = .) %>% augment)` into `left_join(data %>% select(dataset, observation, sample_id), by=c('dataset', 'observation'))`, you get what you want. This depends on the observations in your data being unique (per dataset) though! — Curt F., Feb 05 '18 at 23:42
Right, I deliberated that, but alas – the observations are not unique. I guess I could count on everything being in the same order though after the `do(augment)` call? — slhck, Feb 06 '18 at 07:51
See my answer. Adding a row_number dummy column gets rid of the uniqueness issue... — Curt F., Feb 06 '18 at 15:52

markus · Accepted Answer · 2018-02-07T13:36:37.763

You could use broom::augment_columns instead of augment. The two arguments of the function we need are x -- "a model" -- and data -- "original data onto which columns should be added".

library(tidyverse)
library(broom)
split(data, data$dataset) %>% 
 map(., ~lm(formula = observation ~ estimate, data = .)) %>% 
 map2(.x = ., .y = split(data, f = data$dataset), .f = ~augment_columns(x = .x, data = .y)) %>% 
 bind_rows() %>% 
 select(-.rownames)

#   dataset sample_id observation estimate   .fitted    .se.fit       .resid      .hat     .sigma      .cooksd  .std.resid
#1        A        A1         4.8      4.7 4.6845093 0.10675590  0.115490737 0.4781238 0.15157780 0.4911635931  1.03547990
#2        A        A2         4.3      4.5 4.4934963 0.09956065 -0.193496255 0.4158455 0.06089193 0.9570799385 -1.63978525
#3        A        A3         3.1      2.9 2.9653922 0.06929022  0.134607804 0.2014190 0.15623754 0.1200409795  0.97563873
#4        A        A4         2.1      2.0 2.1058337 0.08491818 -0.005833662 0.3025227 0.18902495 0.0004439221 -0.04524332
#5        A        A5         1.1      1.0 1.1507686 0.11979870 -0.050768624 0.6020891 0.18032220 0.2055920869 -0.52129162
#6        B        B1         4.5      4.3 4.3087226 0.10879087  0.191277434 0.4679235 0.05965705 1.1954021471  1.64881395
#7        B        B2         3.9      4.1 4.0929657 0.10006757 -0.192965672 0.3958920 0.08438937 0.7984863377 -1.56105324
#8        B        B3         2.9      3.0 2.9063028 0.07128455 -0.006302757 0.2009004 0.19471901 0.0002470587 -0.04433279
#9        B        B4         1.8      2.0 1.8275183 0.08983650 -0.027518289 0.3190771 0.19335019 0.0103015495 -0.20968503
#10       B        B5         1.0      1.2 0.9644907 0.12484420  0.035509285 0.6162071 0.19051943 0.1042741368  0.36040302

The idea is to split the data by dataset, fit a model to each component of the list and then use map2 to iterate over the models and the (complete) data used for model building, i.e. the outcome of split(data, f = data$dataset) in parallel.

augment_columns adds a .rownames column, hence the select in the last line.

edit

The same solution but hopefully easier to read.

data_split <- split(data, data$dataset)
models <- map(data_split, ~lm(formula = observation ~ estimate, data = .))

map2(.x = models, .y = data_split, .f = ~augment_columns(x = .x, data = .y)) %>%
 bind_rows() %>% 
 select(-.rownames)

The first code block as a function that has four arguments: df, split_var, dependend_var, and explanatory_var.

augment_df <- function(df, split_var, dependend_var, explanatory_var) {

  require(tidyverse)
  require(broom)

  split(df, df[split_var]) %>% 
     map(., ~lm(formula = as.formula(paste0(dependend_var, " ~ ", explanatory_var)), data = .)) %>% 
     map2(.x = ., .y = split(df, df[split_var]), .f = ~augment_columns(x = .x, data = .y)) %>% 
     bind_rows() %>% 
     select(-.rownames)

  }

augment_df(df = data, split_var = "dataset", dependend_var = "observation", explanatory_var = "estimate")

score 3 · Answer 2 · answered Feb 05 '18 at 23:51

numbered_data <-
    data %>%
        mutate(row = row_number())

numbered_data %>%
    group_by(dataset) %>% 
    do(augment(lm(observation ~ estimate + 0*row, data = .))) %>%
    left_join(numbered_data %>% select(-observation, -estimate), by=c('dataset', 'row')) %>%
    select(-row)

This approach also relies on the use of a dummy variable in the model, but it does it in a column-agnostic way. By defining a new dummy column, row, you can left_join() your original data back onto the results of augment(), restoring any arbitrary numbers of columns without manually specifying them.

I find this a bit more readable than the other solutions, but it is still a bit hackish. Getting rid of the duplicate columns from the left_join is a bit tedious. You don't presumably want columns like observation.x and estimate.y in your output, which you will have unless you dod the select(-observation, -estimate) part.

I came back to this answer 4 years later and I notice that the linear fit terms (when extracted with `broom::tidy`) no longer include an intercept when using the `0*row` hack. Any idea how to avoid that? — slhck, Feb 13 '22 at 19:18

score 1 · Answer 3 · answered Feb 07 '18 at 16:07

This is essentially the same as Markus's answer but perhaps a little cleaner.

library(tidyverse)
library(broom)

data = read.table(text = '
dataset sample_id   observation estimate
A   A1  4.8 4.7
A   A2  4.3 4.5
A   A3  3.1 2.9
A   A4  2.1 2
A   A5  1.1 1
B   B1  4.5 4.3
B   B2  3.9 4.1
B   B3  2.9 3
B   B4  1.8 2
B   B5  1   1.2
', header = TRUE)


data %>%
  group_by(dataset) %>% 
  nest() %>% 
  mutate(mod = map(data, ~lm(observation ~ estimate, data = .)),
         aug = map2(mod, data, ~augment_columns(.x, .y))) %>% 
  unnest(aug)

score 0 · Answer 4 · answered Feb 05 '18 at 17:16

0

How about this:

DF %>%
  group_by(dataset) %>% 
  do(cbind(sample_id = .$sample_id, lm(observation ~ estimate, data = .) %>% augment))

Too ugly still?

answered Feb 05 '18 at 17:16

RolandASc

3,863
1
11
30

1

This is not general solution for keeping multiple columns. – David Arenburg Feb 05 '18 at 17:26
Yes, as @DavidArenburg said, that'd be my main issue with using such an approach. (Although to be fair, a second variable was never specified in the question.) – slhck Feb 05 '18 at 17:27
Thinking again, one could `cbind` a selection of everything in the original dataframe *except* the `observation` and `estimate` columns. – slhck Feb 06 '18 at 07:51

Creating models and augmenting data without losing additional columns in dplyr/broom

4 Answers4

Linked