13 Grouping data
This page covers how to group and aggregate data for descriptive analysis. It makes use of the tidyverse family of packages for common and easy-to-use functions.
Grouping data is a core component of data management and analysis. Grouped data statistically summarised by group, and can be plotted by group. Functions from the dplyr package (part of the tidyverse) make grouping and subsequent operations quite easy.
This page will address the following topics:
- Group data with the
group_by()function
- Un-group data
-
summarise()grouped data with statistics
- The difference between
count()andtally()
-
arrange()applied to grouped data
-
filter()applied to grouped data
-
mutate()applied to grouped data
-
select()applied to grouped data
- The base R
aggregate()command as an alternative
13.1 Preparation
Load packages
This code chunk shows the loading of packages required for the analyses. In this handbook we emphasize p_load() from pacman, which installs the package if necessary and loads it for use. You can also load installed packages with library() from base R. See the page on R basics for more information on R packages.
pacman::p_load(
rio, # to import data
here, # to locate files
tidyverse, # to clean, handle, and plot the data (includes dplyr)
janitor) # adding total rows and columnsImport data
We import the dataset of cases from a simulated Ebola epidemic. If you want to follow along, click to download the “clean” linelist (as .rds file). The dataset is imported using the import() function from the rio package. See the page on Import and export for various ways to import data.
linelist <- import("linelist_cleaned.rds")The first 50 rows of linelist:
13.2 Grouping
The function group_by() from dplyr groups the rows by the unique values in the column specified to it. If multiple columns are specified, rows are grouped by the unique combinations of values across the columns. Each unique value (or combination of values) constitutes a group. Subsequent changes to the dataset or calculations can then be performed within the context of each group.
For example, the command below takes the linelist and groups the rows by unique values in the column outcome, saving the output as a new data frame ll_by_outcome. The grouping column(s) are placed inside the parentheses of the function group_by().
Note that there is no perceptible change to the dataset after running group_by(), until another dplyr verb such as mutate(), summarise(), or arrange() is applied on the “grouped” data frame.
You can however “see” the groupings by printing the data frame. When you print a grouped data frame, you will see it has been transformed into a tibble class object which, when printed, displays which groupings have been applied and how many groups there are - written just above the header row.
# print to see which groups are active
ll_by_outcome## # A tibble: 5,888 × 30
## # Groups: outcome [3]
## case_id generation date_infection date_onset date_hospitalisation date_outcome outcome gender age age_unit age_years
## <chr> <dbl> <date> <date> <date> <date> <chr> <chr> <dbl> <chr> <dbl>
## 1 5fe599 4 2014-05-08 2014-05-13 2014-05-15 NA <NA> m 2 years 2
## 2 8689b7 4 NA 2014-05-13 2014-05-14 2014-05-18 Recover f 3 years 3
## 3 11f8ea 2 NA 2014-05-16 2014-05-18 2014-05-30 Recover m 56 years 56
## 4 b8812a 3 2014-05-04 2014-05-18 2014-05-20 NA <NA> f 18 years 18
## 5 893f25 3 2014-05-18 2014-05-21 2014-05-22 2014-05-29 Recover m 3 years 3
## 6 be99c8 3 2014-05-03 2014-05-22 2014-05-23 2014-05-24 Recover f 16 years 16
## 7 07e3e8 4 2014-05-22 2014-05-27 2014-05-29 2014-06-01 Recover f 16 years 16
## 8 369449 4 2014-05-28 2014-06-02 2014-06-03 2014-06-07 Death f 0 years 0
## 9 f393b4 4 NA 2014-06-05 2014-06-06 2014-06-18 Recover m 61 years 61
## 10 1389ca 4 NA 2014-06-05 2014-06-07 2014-06-09 Death f 27 years 27
## # ℹ 5,878 more rows
## # ℹ 19 more variables: age_cat <fct>, age_cat5 <fct>, hospital <chr>, lon <dbl>, lat <dbl>, infector <chr>, source <chr>,
## # wt_kg <dbl>, ht_cm <dbl>, ct_blood <dbl>, fever <chr>, chills <chr>, cough <chr>, aches <chr>, vomit <chr>, temp <dbl>,
## # time_admission <chr>, bmi <dbl>, days_onset_hosp <dbl>Unique groups
The groups created reflect each unique combination of values across the grouping columns.
To see the groups and the number of rows in each group, pass the grouped data to tally(). To see just the unique groups without counts you can pass to group_keys().
See below that there are three unique values in the grouping column outcome: “Death”, “Recover”, and NA. See that there were nrow(linelist %>% filter(outcome == "Death")) deaths, nrow(linelist %>% filter(outcome == "Recover")) recoveries, and nrow(linelist %>% filter(is.na(outcome))) with no outcome recorded.
## # A tibble: 3 × 2
## outcome n
## <chr> <int>
## 1 Death 2582
## 2 Recover 1983
## 3 <NA> 1323You can group by more than one column. Below, the data frame is grouped by outcome and gender, and then tallied. Note how each unique combination of outcome and gender is registered as its own group - including missing values for either column.
## # A tibble: 9 × 3
## # Groups: outcome [3]
## outcome gender n
## <chr> <chr> <int>
## 1 Death f 1227
## 2 Death m 1228
## 3 Death <NA> 127
## 4 Recover f 953
## 5 Recover m 950
## 6 Recover <NA> 80
## 7 <NA> f 627
## 8 <NA> m 625
## 9 <NA> <NA> 71New columns
You can also create a new grouping column within the group_by() statement. This is equivalent to calling mutate() before the group_by(). For a quick tabulation this style can be handy, but for more clarity in your code consider creating this column in its own mutate() step and then piping to group_by().
# group dat based on a binary column created *within* the group_by() command
linelist %>%
group_by(
age_class = ifelse(age >= 18, "adult", "child")) %>%
tally(sort = T)## # A tibble: 3 × 2
## age_class n
## <chr> <int>
## 1 child 3618
## 2 adult 2184
## 3 <NA> 86Add/drop grouping columns
By default, if you run group_by() on data that are already grouped, the old groups will be removed and the new one(s) will apply. If you want to add new groups to the existing ones, include the argument .add = TRUE.
# Grouped by outcome
by_outcome <- linelist %>%
group_by(outcome)
# Add grouping by gender in addition
by_outcome_gender <- by_outcome %>%
group_by(gender, .add = TRUE)** Keep all groups**
If you group on a column of class factor there may be levels of the factor that are not currently present in the data. If you group on this column, by default those non-present levels are dropped and not included as groups. To change this so that all levels appear as groups (even if not present in the data), set .drop = FALSE in your group_by() command.
13.3 Un-group
Data that have been grouped will remain grouped until specifically ungrouped via ungroup(). If you forget to ungroup, it can lead to incorrect calculations! Below is an example of removing all groupings:
You can also remove grouping for only specific columns, by placing the column name inside ungroup().
linelist %>%
group_by(outcome, gender) %>%
tally() %>%
ungroup(gender) # remove the grouping by gender, leave grouping by outcomeNOTE: The verb count() automatically ungroups the data after counting.
13.4 Summarise
See the dplyr section of the Descriptive tables page for a detailed description of how to produce summary tables with summarise(). Here we briefly address how its behavior changes when applied to grouped data.
The dplyr function summarise() (or summarize()) takes a data frame and converts it into a new summary data frame, with columns containing summary statistics that you define. On an ungrouped data frame, the summary statistics will be calculated from all rows. Applying summarise() to grouped data produces those summary statistics for each group.
The syntax of summarise() is such that you provide the name(s) of the new summary column(s), an equals sign, and then a statistical function to apply to the data, as shown below. For example, min(), max(), median(), or sd(). Within the statistical function, list the column to be operated on and any relevant argument (e.g. na.rm = TRUE). You can use sum() to count the number of rows that meet a logical criteria (with double equals ==).
Below is an example of summarise() applied without grouped data. The statistics returned are produced from the entire dataset.
# summary statistics on ungrouped linelist
linelist %>%
summarise(
n_cases = n(),
mean_age = mean(age_years, na.rm=T),
max_age = max(age_years, na.rm=T),
min_age = min(age_years, na.rm=T),
n_males = sum(gender == "m", na.rm=T))## n_cases mean_age max_age min_age n_males
## 1 5888 16.01831 84 0 2803In contrast, below is the same summarise() statement applied to grouped data. The statistics are calculated for each outcome group. Note how grouping columns will carry over into the new data frame.
# summary statistics on grouped linelist
linelist %>%
group_by(outcome) %>%
summarise(
n_cases = n(),
mean_age = mean(age_years, na.rm=T),
max_age = max(age_years, na.rm=T),
min_age = min(age_years, na.rm=T),
n_males = sum(gender == "m", na.rm=T))## # A tibble: 3 × 6
## outcome n_cases mean_age max_age min_age n_males
## <chr> <int> <dbl> <dbl> <dbl> <int>
## 1 Death 2582 15.9 76 0 1228
## 2 Recover 1983 16.1 84 0 950
## 3 <NA> 1323 16.2 69 0 625TIP: The summarise function works with both UK and US spelling - summarise() and summarize() call the same function.
13.5 Counts and tallies
count() and tally() provide similar functionality but are different. Read more about the distinction between tally() and count() here
tally()
tally() is shorthand for summarise(n = n()), and does not group data. Thus, to achieve grouped tallys it must follow a group_by() command. You can add sort = TRUE to see the largest groups first.
## n
## 1 5888## # A tibble: 3 × 2
## outcome n
## <chr> <int>
## 1 Death 2582
## 2 Recover 1983
## 3 <NA> 1323
count()
In contrast, count() does the following:
- applies
group_by()on the specified column(s)
- applies
summarise()and returns columnnwith the number of rows per group
- applies
ungroup()
## outcome n
## 1 Death 2582
## 2 Recover 1983
## 3 <NA> 1323Just like with group_by() you can create a new column within the count() command:
## age_class n
## 1 child 3618
## 2 adult 2184
## 3 <NA> 86count() can be called multiple times, with the functionality “rolling up”. For example, to summarise the number of hospitals present for each gender, run the following. Note, the name of the final column is changed from default “n” for clarity (with name =).
linelist %>%
# produce counts by unique outcome-gender groups
count(gender, hospital) %>%
# gather rows by gender (3) and count number of hospitals per gender (6)
count(gender, name = "hospitals per gender" ) ## gender hospitals per gender
## 1 f 6
## 2 m 6
## 3 <NA> 6Add counts
In contrast to count() and summarise(), you can use add_count() to add a new column n with the counts of rows per group while retaining all the other data frame columns.
This means that a group’s count number, in the new column n, will be printed in each row of the group. For demonstration purposes, we add this column and then re-arrange the columns for easier viewing. See the section below on filter on group size for another example.
linelist %>%
as_tibble() %>% # convert to tibble for nicer printing
add_count(hospital) %>% # add column n with counts by hospital
select(hospital, n, everything()) # re-arrange for demo purposes## # A tibble: 5,888 × 31
## hospital n case_id generation date_infection date_onset date_hospitalisation date_outcome outcome gender age age_unit
## <chr> <int> <chr> <dbl> <date> <date> <date> <date> <chr> <chr> <dbl> <chr>
## 1 Other 885 5fe599 4 2014-05-08 2014-05-13 2014-05-15 NA <NA> m 2 years
## 2 Missing 1469 8689b7 4 NA 2014-05-13 2014-05-14 2014-05-18 Recover f 3 years
## 3 St. Mark… 422 11f8ea 2 NA 2014-05-16 2014-05-18 2014-05-30 Recover m 56 years
## 4 Port Hos… 1762 b8812a 3 2014-05-04 2014-05-18 2014-05-20 NA <NA> f 18 years
## 5 Military… 896 893f25 3 2014-05-18 2014-05-21 2014-05-22 2014-05-29 Recover m 3 years
## 6 Port Hos… 1762 be99c8 3 2014-05-03 2014-05-22 2014-05-23 2014-05-24 Recover f 16 years
## 7 Missing 1469 07e3e8 4 2014-05-22 2014-05-27 2014-05-29 2014-06-01 Recover f 16 years
## 8 Missing 1469 369449 4 2014-05-28 2014-06-02 2014-06-03 2014-06-07 Death f 0 years
## 9 Missing 1469 f393b4 4 NA 2014-06-05 2014-06-06 2014-06-18 Recover m 61 years
## 10 Missing 1469 1389ca 4 NA 2014-06-05 2014-06-07 2014-06-09 Death f 27 years
## # ℹ 5,878 more rows
## # ℹ 19 more variables: age_years <dbl>, age_cat <fct>, age_cat5 <fct>, lon <dbl>, lat <dbl>, infector <chr>, source <chr>,
## # wt_kg <dbl>, ht_cm <dbl>, ct_blood <dbl>, fever <chr>, chills <chr>, cough <chr>, aches <chr>, vomit <chr>, temp <dbl>,
## # time_admission <chr>, bmi <dbl>, days_onset_hosp <dbl>Add totals
To easily add total sum rows or columns after using tally() or count(), see the janitor section of the Descriptive tables page. This package offers functions like adorn_totals() and adorn_percentages() to add totals and convert to show percentages. Below is a brief example:
linelist %>% # case linelist
tabyl(age_cat, gender) %>% # cross-tabulate counts of two columns
adorn_totals(where = "row") %>% # add a total row
adorn_percentages(denominator = "col") %>% # convert to proportions with column denominator
adorn_pct_formatting() %>% # convert proportions to percents
adorn_ns(position = "front") %>% # display as: "count (percent)"
adorn_title( # adjust titles
row_name = "Age Category",
col_name = "Gender")## Gender
## Age Category f m NA_
## 0-4 640 (22.8%) 416 (14.8%) 39 (14.0%)
## 5-9 641 (22.8%) 412 (14.7%) 42 (15.1%)
## 10-14 518 (18.5%) 383 (13.7%) 40 (14.4%)
## 15-19 359 (12.8%) 364 (13.0%) 20 (7.2%)
## 20-29 468 (16.7%) 575 (20.5%) 30 (10.8%)
## 30-49 179 (6.4%) 557 (19.9%) 18 (6.5%)
## 50-69 2 (0.1%) 91 (3.2%) 2 (0.7%)
## 70+ 0 (0.0%) 5 (0.2%) 1 (0.4%)
## <NA> 0 (0.0%) 0 (0.0%) 86 (30.9%)
## Total 2,807 (100.0%) 2,803 (100.0%) 278 (100.0%)To add more complex totals rows that involve summary statistics other than sums, see this section of the Descriptive Tables page.
13.6 Grouping by date
When grouping data by date, you must have (or create) a column for the date unit of interest - for example “day”, “epiweek”, “month”, etc. You can make this column using floor_date() from lubridate, as explained in the Epidemiological weeks section of the Working with dates page. Once you have this column, you can use count() from dplyr to group the rows by those unique date values and achieve aggregate counts.
One additional step common for date situations, is to “fill-in” any dates in the sequence that are not present in the data. Use complete() from tidyr so that the aggregated date series is complete including all possible date units within the range. Without this step, a week with no cases reported might not appear in your data!
Within complete() you re-define your date column as a sequence of dates seq.Date() from the minimum to the maximum - thus the dates are expanded. By default, the case count values in any new “expanded” rows will be NA. You can set them to 0 using the fill = argument of complete(), which expects a named list (if your counts column is named n, provide fill = list(n = 0). See ?complete for details and the Working with dates page for an example.
Linelist cases into days
Here is an example of grouping cases into days without using complete(). Note the first rows skip over dates with no cases.
daily_counts <- linelist %>%
drop_na(date_onset) %>% # remove that were missing date_onset
count(date_onset) # count number of rows per unique dateBelow we add the complete() command to ensure every day in the range is represented.
daily_counts <- linelist %>%
drop_na(date_onset) %>% # remove case missing date_onset
count(date_onset) %>% # count number of rows per unique date
complete( # ensure all days appear even if no cases
date_onset = seq.Date( # re-define date colume as daily sequence of dates
from = min(date_onset, na.rm=T),
to = max(date_onset, na.rm=T),
by = "day"),
fill = list(n = 0)) # set new filled-in rows to display 0 in column n (not NA as default) Linelist cases into weeks
The same principle can be applied for weeks. First create a new column that is the week of the case using floor_date() with unit = "week". Then, use count() as above to achieve weekly case counts. Finish with complete() to ensure that all weeks are represented, even if they contain no cases.
# Make dataset of weekly case counts
weekly_counts <- linelist %>%
drop_na(date_onset) %>% # remove cases missing date_onset
mutate(week = lubridate::floor_date(date_onset, unit = "week")) %>% # new column of week of onset
count(week) %>% # group data by week and count rows per group
complete( # ensure all days appear even if no cases
week = seq.Date( # re-define date colume as daily sequence of dates
from = min(week, na.rm=T),
to = max(week, na.rm=T),
by = "week"),
fill = list(n = 0)) # set new filled-in rows to display 0 in column n (not NA as default) Here are the first 50 rows of the resulting data frame:
Linelist cases into months
To aggregate cases into months, again use floor_date() from the lubridate package, but with the argument unit = "months". This rounds each date down to the 1st of its month. The output will be class Date. Note that in the complete() step we also use by = "months".
# Make dataset of monthly case counts
monthly_counts <- linelist %>%
drop_na(date_onset) %>%
mutate(month = lubridate::floor_date(date_onset, unit = "months")) %>% # new column, 1st of month of onset
count(month) %>% # count cases by month
complete(
month = seq.Date(
min(month, na.rm=T), # include all months with no cases reported
max(month, na.rm=T),
by="month"),
fill = list(n = 0))Daily counts into weeks
To aggregate daily counts into weekly counts, use floor_date() as above. However, use group_by() and summarize() instead of count() because you need to sum() daily case counts instead of just counting the number of rows per week.
Daily counts into months
To aggregate daily counts into months counts, use floor_date() with unit = "month" as above. However, use group_by() and summarize() instead of count() because you need to sum() daily case counts instead of just counting the number of rows per month.
13.7 Arranging grouped data
Using the dplyr verb arrange() to order the rows in a data frame behaves the same when the data are grouped, unless you set the argument .by_group =TRUE. In this case the rows are ordered first by the grouping columns and then by any other columns you specify to arrange().
13.8 Filter on grouped data
filter()
When applied in conjunction with functions that evaluate the data frame (like max(), min(), mean()), these functions will now be applied to the groups. For example, if you want to filter and keep rows where patients are above the median age, this will now apply per group - filtering to keep rows above the group’s median age.
Slice rows per group
The dplyr function slice(), which filters rows based on their position in the data, can also be applied per group. Remember to account for sorting the data within each group to get the desired “slice”.
For example, to retrieve only the latest 5 admissions from each hospital:
- Group the linelist by column
hospital
- Arrange the records from latest to earliest
date_hospitalisationwithin each hospital group
- Slice to retrieve the first 5 rows from each hospital
linelist %>%
group_by(hospital) %>%
arrange(hospital, date_hospitalisation) %>%
slice_head(n = 5) %>%
arrange(hospital) %>% # for display
select(case_id, hospital, date_hospitalisation) # for display## # A tibble: 30 × 3
## # Groups: hospital [6]
## case_id hospital date_hospitalisation
## <chr> <chr> <date>
## 1 20b688 Central Hospital 2014-05-06
## 2 d58402 Central Hospital 2014-05-10
## 3 b8f2fd Central Hospital 2014-05-13
## 4 acf422 Central Hospital 2014-05-28
## 5 275cc7 Central Hospital 2014-05-28
## 6 d1fafd Military Hospital 2014-04-17
## 7 974bc1 Military Hospital 2014-05-13
## 8 6a9004 Military Hospital 2014-05-13
## 9 09e386 Military Hospital 2014-05-14
## 10 865581 Military Hospital 2014-05-15
## # ℹ 20 more rowsslice_head() - selects n rows from the topslice_tail() - selects n rows from the endslice_sample() - randomly selects n rowsslice_min() - selects n rows with highest values in order_by = column, use with_ties = TRUE to keep tiesslice_max() - selects n rows with lowest values in order_by = column, use with_ties = TRUE to keep ties
See the De-duplication page for more examples and detail on slice().
Filter on group size
The function add_count() adds a column n to the original data giving the number of rows in that row’s group.
Shown below, add_count() is applied to the column hospital, so the values in the new column n reflect the number of rows in that row’s hospital group. Note how values in column n are repeated. In the example below, the column name n could be changed using name = within add_count(). For demonstration purposes we re-arrange the columns with select().
linelist %>%
as_tibble() %>%
add_count(hospital) %>% # add "number of rows admitted to same hospital as this row"
select(hospital, n, everything())## # A tibble: 5,888 × 31
## hospital n case_id generation date_infection date_onset date_hospitalisation date_outcome outcome gender age age_unit
## <chr> <int> <chr> <dbl> <date> <date> <date> <date> <chr> <chr> <dbl> <chr>
## 1 Other 885 5fe599 4 2014-05-08 2014-05-13 2014-05-15 NA <NA> m 2 years
## 2 Missing 1469 8689b7 4 NA 2014-05-13 2014-05-14 2014-05-18 Recover f 3 years
## 3 St. Mark… 422 11f8ea 2 NA 2014-05-16 2014-05-18 2014-05-30 Recover m 56 years
## 4 Port Hos… 1762 b8812a 3 2014-05-04 2014-05-18 2014-05-20 NA <NA> f 18 years
## 5 Military… 896 893f25 3 2014-05-18 2014-05-21 2014-05-22 2014-05-29 Recover m 3 years
## 6 Port Hos… 1762 be99c8 3 2014-05-03 2014-05-22 2014-05-23 2014-05-24 Recover f 16 years
## 7 Missing 1469 07e3e8 4 2014-05-22 2014-05-27 2014-05-29 2014-06-01 Recover f 16 years
## 8 Missing 1469 369449 4 2014-05-28 2014-06-02 2014-06-03 2014-06-07 Death f 0 years
## 9 Missing 1469 f393b4 4 NA 2014-06-05 2014-06-06 2014-06-18 Recover m 61 years
## 10 Missing 1469 1389ca 4 NA 2014-06-05 2014-06-07 2014-06-09 Death f 27 years
## # ℹ 5,878 more rows
## # ℹ 19 more variables: age_years <dbl>, age_cat <fct>, age_cat5 <fct>, lon <dbl>, lat <dbl>, infector <chr>, source <chr>,
## # wt_kg <dbl>, ht_cm <dbl>, ct_blood <dbl>, fever <chr>, chills <chr>, cough <chr>, aches <chr>, vomit <chr>, temp <dbl>,
## # time_admission <chr>, bmi <dbl>, days_onset_hosp <dbl>It then becomes easy to filter for case rows who were hospitalized at a “small” hospital, say, a hospital that admitted fewer than 500 patients:
13.9 Mutate on grouped data
To retain all columns and rows (not summarise) and add a new column containing group statistics, use mutate() after group_by() instead of summarise().
This is useful if you want group statistics in the original dataset with all other columns present - e.g. for calculations that compare one row to its group.
For example, this code below calculates the difference between a row’s delay-to-admission and the median delay for their hospital. The steps are:
- Group the data by hospital
- Use the column
days_onset_hosp(delay to hospitalisation) to create a new column containing the mean delay at the hospital of that row
- Calculate the difference between the two columns
We select() only certain columns to display, for demonstration purposes.
linelist %>%
# group data by hospital (no change to linelist yet)
group_by(hospital) %>%
# new columns
mutate(
# mean days to admission per hospital (rounded to 1 decimal)
group_delay_admit = round(mean(days_onset_hosp, na.rm=T), 1),
# difference between row's delay and mean delay at their hospital (rounded to 1 decimal)
diff_to_group = round(days_onset_hosp - group_delay_admit, 1)) %>%
# select certain rows only - for demonstration/viewing purposes
select(case_id, hospital, days_onset_hosp, group_delay_admit, diff_to_group)## # A tibble: 5,888 × 5
## # Groups: hospital [6]
## case_id hospital days_onset_hosp group_delay_admit diff_to_group
## <chr> <chr> <dbl> <dbl> <dbl>
## 1 5fe599 Other 2 2 0
## 2 8689b7 Missing 1 2.1 -1.1
## 3 11f8ea St. Mark's Maternity Hospital (SMMH) 2 2.1 -0.1
## 4 b8812a Port Hospital 2 2.1 -0.1
## 5 893f25 Military Hospital 1 2.1 -1.1
## 6 be99c8 Port Hospital 1 2.1 -1.1
## 7 07e3e8 Missing 2 2.1 -0.1
## 8 369449 Missing 1 2.1 -1.1
## 9 f393b4 Missing 1 2.1 -1.1
## 10 1389ca Missing 2 2.1 -0.1
## # ℹ 5,878 more rows13.10 Select on grouped data
The verb select() works on grouped data, but the grouping columns are always included (even if not mentioned in select()). If you do not want these grouping columns, use ungroup() first.
13.11 Resources
Here are some useful resources for more information:
You can perform any summary function on grouped data; see the RStudio data transformation cheat sheet
The Data Carpentry page on dplyr
The tidyverse reference pages on group_by() and grouping
This page on Data manipulation