collapse is a C/C++ based package for data transformation and statistical computing in R. It’s aims are:
This vignette focuses on the integration of collapse and the popular dplyr package by Hadley Wickham. In particular it will demonstrate how using collapse’s fast functions and some fast alternatives for dplyr verbs can substantially facilitate and speed up basic data manipulation, grouped and weighted aggregations and transformations, and panel data computations (i.e. between- and within-transformations, panel-lags, differences and growth rates) in a dplyr (piped) workflow.
Notes:
This vignette is targeted at dplyr / tidyverse users. collapse is a standalone package and can be programmed efficiently without pipes or dplyr verbs.
The ‘Introduction to collapse’ vignette provides a thorough introduction to the package and a built-in structured documentation is available under help("collapse-documentation") after installing the package. In addition help("collapse-package") provides a compact set of examples for quick-start.
Documentation and vignettes can also be viewed online.
A key feature of collapse is it’s broad set of Fast Statistical Functions (fsum, fprod, fmean, fmedian, fmode, fvar, fsd, fmin, fmax, fnth, ffirst, flast, fNobs, fNdistinct) which are able to substantially speed-up column-wise, grouped and weighted computations on vectors, matrices or data frames. The functions are S3 generic, with a default (vector), matrix and data frame method, as well as a grouped_df method for grouped tibbles used by dplyr. The grouped tibble method has the following arguments:
FUN.grouped_df(x, [w = NULL,] TRA = NULL, [na.rm = TRUE,]
use.g.names = FALSE, keep.group_vars = TRUE, [keep.w = TRUE,] ...)where w is a weight variable, and TRA and can be used to transform x using the computed statistics and one of 10 available transformations ("replace_fill", "replace", "-", "-+", "/", "%", "+", "*", "%%", "-%%", discussed in section 2). na.rm efficiently removes missing values and is TRUE by default. use.g.names generates new row-names from the unique combinations of groups (default: disabled), whereas keep.group_vars (default: enabled) will keep the grouping columns as is custom in the native data %>% group_by(...) %>% summarize(...) workflow in dplyr. Finally, keep.w regulates whether a weighting variable used is also aggregated and saved in a column. For fsum, fmean, fmedian, fnth, fvar, fsd and fmode this will compute the sum of the weights in each group, whereas fprod returns the product of the weights.
With that in mind, let’s consider some straightforward applications.
Consider the Groningen Growth and Development Center 10-Sector Database included in collapse and introduced in the main vignette:
library(collapse)
head(GGDC10S)
# Country Regioncode Region Variable Year AGR MIN MAN PU
# 1 BWA SSA Sub-saharan Africa VA 1960 NA NA NA NA
# 2 BWA SSA Sub-saharan Africa VA 1961 NA NA NA NA
# 3 BWA SSA Sub-saharan Africa VA 1962 NA NA NA NA
# 4 BWA SSA Sub-saharan Africa VA 1963 NA NA NA NA
# 5 BWA SSA Sub-saharan Africa VA 1964 16.30154 3.494075 0.7365696 0.1043936
# 6 BWA SSA Sub-saharan Africa VA 1965 15.72700 2.495768 1.0181992 0.1350976
# CON WRT TRA FIRE GOV OTH SUM
# 1 NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA
# 3 NA NA NA NA NA NA NA
# 4 NA NA NA NA NA NA NA
# 5 0.6600454 6.243732 1.658928 1.119194 4.822485 2.341328 37.48229
# 6 1.3462312 7.064825 1.939007 1.246789 5.695848 2.678338 39.34710
# Summarize the Data:
# descr(GGDC10S, cols = is.categorical)
# aperm(qsu(GGDC10S, ~Variable, cols = is.numeric))Simple column-wise computations using the fast functions and pipe operators are performed as follows:
library(dplyr)
GGDC10S %>% fNobs # Number of Observations
# Country Regioncode Region Variable Year AGR MIN MAN PU
# 5027 5027 5027 5027 5027 4364 4355 4355 4354
# CON WRT TRA FIRE GOV OTH SUM
# 4355 4355 4355 4355 3482 4248 4364
GGDC10S %>% fNdistinct # Number of distinct values
# Country Regioncode Region Variable Year AGR MIN MAN PU
# 43 6 6 2 67 4353 4224 4353 4237
# CON WRT TRA FIRE GOV OTH SUM
# 4339 4344 4334 4349 3470 4238 4364
GGDC10S %>% select_at(6:16) %>% fmedian # Median
# AGR MIN MAN PU CON WRT TRA FIRE GOV
# 4394.5194 173.2234 3718.0981 167.9500 1473.4470 3773.6430 1174.8000 960.1251 3928.5127
# OTH SUM
# 1433.1722 23186.1936
GGDC10S %>% select_at(6:16) %>% fmean # Mean
# AGR MIN MAN PU CON WRT TRA FIRE GOV
# 2526696.5 1867908.9 5538491.4 335679.5 1801597.6 3392909.5 1473269.7 1657114.8 1712300.3
# OTH SUM
# 1684527.3 21566436.8
GGDC10S %>% fmode # Mode
# Country Regioncode Region Variable Year
# "USA" "ASI" "Asia" "EMP" "2010"
# AGR MIN MAN PU CON
# "171.315882316326" "0" "4645.12507642586" "0" "1.34623115930777"
# WRT TRA FIRE GOV OTH
# "21.8380052682527" "8.97743416914571" "40.0701608636442" "0" "3626.84423577048"
# SUM
# "37.4822945751317"
GGDC10S %>% fmode(drop = FALSE) # Keep data structure intact
# Country Regioncode Region Variable Year AGR MIN MAN PU CON WRT TRA
# 1 USA ASI Asia EMP 2010 171.3159 0 4645.125 0 1.346231 21.83801 8.977434
# FIRE GOV OTH SUM
# 1 40.07016 0 3626.844 37.48229Moving on to grouped statistics, we can compute the average value added and employment by sector and country using:
GGDC10S %>%
group_by(Variable, Country) %>%
select_at(6:16) %>% fmean
# # A tibble: 85 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1420. 52.1 1932. 1.02e2 7.42e2 1.98e3 6.49e2 628. 2043. 9.92e2 1.05e4
# 2 EMP BOL 964. 56.0 235. 5.35e0 1.23e2 2.82e2 1.15e2 44.6 NA 3.96e2 2.22e3
# 3 EMP BRA 17191. 206. 6991. 3.65e2 3.52e3 8.51e3 2.05e3 4414. 5307. 5.71e3 5.43e4
# 4 EMP BWA 188. 10.5 18.1 3.09e0 2.53e1 3.63e1 8.36e0 15.3 61.1 2.76e1 3.94e2
# 5 EMP CHL 702. 101. 625. 2.94e1 2.96e2 6.95e2 2.58e2 272. NA 1.00e3 3.98e3
# 6 EMP CHN 287744. 7050. 67144. 1.61e3 2.09e4 2.89e4 1.39e4 4929. 22669. 3.10e4 4.86e5
# 7 EMP COL 3091. 145. 1175. 3.39e1 5.24e2 2.07e3 4.70e2 649. NA 1.73e3 9.89e3
# 8 EMP CRI 231. 1.70 136. 1.43e1 5.76e1 1.57e2 4.24e1 54.9 128. 6.51e1 8.87e2
# 9 EMP DEW 2490. 407. 8473. 2.26e2 2.09e3 4.44e3 1.48e3 1689. 3945. 9.99e2 2.62e4
# 10 EMP DNK 236. 8.03 507. 1.38e1 1.71e2 4.55e2 1.61e2 181. 549. 1.11e2 2.39e3
# # ... with 75 more rowsSimilarly we can aggregate using any other of the above functions.
It is important to not use dplyr’s summarize together with these functions since that would eliminate their speed gain. These functions are fast because they are executed only once and carry out the grouped computations in C++, whereas summarize will apply the function to each group in the grouped tibble.
To better explain this point it is perhaps good to shed some light on what is happening behind the scenes of dplyr and collapse. Fundamentally both packages follow different computing paradigms:
dplyr is an efficient implementation of the Split-Apply-Combine computing paradigm. Data is split into groups, these data-chunks are then passed to a function carrying out the computation, and finally recombined to produce the aggregated data.frame. This modus operandi is evident in the grouping mechanism of dplyr. When a data.frame is passed through group_by, a ‘groups’ attribute is attached:
GGDC10S %>% group_by(Variable, Country) %>% attr("groups")
# # A tibble: 85 x 3
# Variable Country .rows
# * <chr> <chr> <list<int>>
# 1 EMP ARG [62]
# 2 EMP BOL [61]
# 3 EMP BRA [62]
# 4 EMP BWA [52]
# 5 EMP CHL [63]
# 6 EMP CHN [62]
# 7 EMP COL [61]
# 8 EMP CRI [62]
# 9 EMP DEW [61]
# 10 EMP DNK [64]
# # ... with 75 more rowsThis object is a data.frame giving the unique groups and in the third (last) column vectors containing the indices of the rows belonging to that group. A command like summarize uses this information to split the data.frame into groups which are then passed sequentially to the function used and later recombined. These steps are also done in C++ which makes dplyr quite efficient.
Now collapse is based around one-pass grouped computations at the C++ level using its own grouped statistical functions. In other words the data is not split and recombined at all but the entire computation is performed in a single C++ loop running through that data and completing the computations for each group simultaneously. This modus operandi is also evident in collapse grouping objects. The method GRP.grouped_df takes a dplyr grouping object from a grouped tibble and efficiently converts it to a collapse grouping object:
GGDC10S %>% group_by(Variable, Country) %>% GRP %>% str
# List of 8
# $ N.groups : int 85
# $ group.id : int [1:5027] 46 46 46 46 46 46 46 46 46 46 ...
# $ group.sizes: int [1:85] 62 61 62 52 63 62 61 62 61 64 ...
# $ groups :List of 2
# ..$ Variable: chr [1:85] "EMP" "EMP" "EMP" "EMP" ...
# .. ..- attr(*, "label")= chr "Variable"
# .. ..- attr(*, "format.stata")= chr "%9s"
# ..$ Country : chr [1:85] "ARG" "BOL" "BRA" "BWA" ...
# .. ..- attr(*, "label")= chr "Country"
# .. ..- attr(*, "format.stata")= chr "%9s"
# $ group.vars : chr [1:2] "Variable" "Country"
# $ ordered : logi [1:2] TRUE TRUE
# $ order : NULL
# $ call : language GRP.grouped_df(X = .)
# - attr(*, "class")= chr "GRP"This object is a list where the first three elements give the number of groups, the group-id to which each row belongs and a vector of group-sizes. A function like fsum uses this information to (for each column) create a result vector of size ‘N.groups’ and the run through the column using the ‘group.id’ vector to add the i’th data point to the ’group.id[i]’th element of the result vector. When the loop is finished, the grouped computation is also finished.
It is obvious that collapse is faster than dplyr since it’s method of computing involves less steps, and it does not need to call statistical functions multiple times. See the benchmark section.
collapse fast functions do not develop their maximal performance on a grouped tibble created with group_by because of the additional conversion cost of the grouping object incurred by GRP.grouped_df. This cost is already minimized through the use of C++, but we can do even better replacing group_by with collapse::fgroup_by. fgroup_by works like group_by but does the grouping with collapse::GRP (up to 10x faster than group_by) and simply attaches a collapse grouping object to the grouped_df. Thus the speed gain is 2-fold: Faster grouping and no conversion cost when calling collapse functions.
Another improvement comes from replacing the dplyr verb select with collapse::fselect, and, for selection using column names, indices or functions use collapse::get_vars instead of select_at or select_if. Next to get_vars, collapse also introduces the predicates num_vars, cat_vars, char_vars, fact_vars, logi_vars and Date_vars to efficiently select columns by type.
GGDC10S %>% fgroup_by(Variable, Country) %>% get_vars(6:16) %>% fmedian
# # A tibble: 85 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1325. 47.4 1988. 1.05e2 7.82e2 1.85e3 5.80e2 464. 1739. 866. 9.74e3
# 2 EMP BOL 943. 53.5 167. 4.46e0 6.60e1 1.32e2 9.70e1 15.3 NA 384. 1.84e3
# 3 EMP BRA 17481. 225. 7208. 3.76e2 4.05e3 6.45e3 1.58e3 4355. 4450. 4479. 5.19e4
# 4 EMP BWA 175. 12.2 13.1 3.71e0 1.90e1 2.11e1 6.75e0 10.4 53.8 31.2 3.61e2
# 5 EMP CHL 690. 93.9 607. 2.58e1 2.30e2 4.84e2 2.05e2 106. NA 900. 3.31e3
# 6 EMP CHN 293915 8150. 61761. 1.14e3 1.06e4 1.70e4 9.56e3 4328. 19468. 9954. 4.45e5
# 7 EMP COL 3006. 84.0 1033. 3.71e1 4.19e2 1.55e3 3.91e2 655. NA 1430. 8.63e3
# 8 EMP CRI 216. 1.49 114. 7.92e0 5.50e1 8.98e1 2.55e1 19.6 122. 60.6 7.19e2
# 9 EMP DEW 2178 320. 8459. 2.47e2 2.10e3 4.45e3 1.53e3 1656 3700 900 2.65e4
# 10 EMP DNK 187. 3.75 508. 1.36e1 1.65e2 4.61e2 1.61e2 169. 642. 104. 2.42e3
# # ... with 75 more rows
microbenchmark(collapse = GGDC10S %>% fgroup_by(Variable, Country) %>% get_vars(6:16) %>% fmedian,
hybrid = GGDC10S %>% group_by(Variable, Country) %>% select_at(6:16) %>% fmedian,
dplyr = GGDC10S %>% group_by(Variable, Country) %>% select_at(6:16) %>% summarise_all(median, na.rm = TRUE))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# collapse 945.152 1038.64 1087.736 1062.961 1123.65 1395.861 100 a
# hybrid 12194.148 12581.71 13352.293 12830.273 13593.58 30239.488 100 b
# dplyr 59320.071 61269.95 64786.978 63674.774 67513.17 87006.580 100 cBenchmarks on the different components of this code and with larger data are provided under ‘Benchmarks’. Note that a grouped tibble created with fgroup_by can no longer be used for grouped computations with dplyr verbs like mutate or summarize. To avoid errors with these functions and print.grouped_df, [.grouped_df etc., the classes assigned after fgroup_by are reshuffled, so that the data.frame is treated by the dplyr ecosystem like a normal tibble:
class(group_by(GGDC10S, Variable, Country))
# [1] "grouped_df" "tbl_df" "tbl" "data.frame"
class(fgroup_by(GGDC10S, Variable, Country))
# [1] "tbl_df" "tbl" "grouped_df" "data.frame"Also note that fselect and get_vars are not full drop-in replacements for select because they do not have a grouped_df method:
GGDC10S %>% group_by(Variable, Country) %>% select_at(6:16) %>% head(3)
# # A tibble: 3 x 13
# # Groups: Variable, Country [1]
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 2 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 3 VA BWA NA NA NA NA NA NA NA NA NA NA NA
GGDC10S %>% group_by(Variable, Country) %>% get_vars(6:16) %>% head(3)
# # A tibble: 3 x 11
# AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 NA NA NA NA NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA NA NA NA NA
# 3 NA NA NA NA NA NA NA NA NA NA NASince by default keep.group_vars = TRUE in the Fast Statistical Functions, the end result is nevertheless the same:
GGDC10S %>% group_by(Variable, Country) %>% select_at(6:16) %>% fmean %>% head(3)
# # A tibble: 3 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1420. 52.1 1932. 102. 742. 1982. 649. 628. 2043. 992. 10542.
# 2 EMP BOL 964. 56.0 235. 5.35 123. 282. 115. 44.6 NA 396. 2221.
# 3 EMP BRA 17191. 206. 6991. 365. 3525. 8509. 2054. 4414. 5307. 5710. 54273.
GGDC10S %>% group_by(Variable, Country) %>% get_vars(6:16) %>% fmean %>% head(3)
# # A tibble: 3 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1420. 52.1 1932. 102. 742. 1982. 649. 628. 2043. 992. 10542.
# 2 EMP BOL 964. 56.0 235. 5.35 123. 282. 115. 44.6 NA 396. 2221.
# 3 EMP BRA 17191. 206. 6991. 365. 3525. 8509. 2054. 4414. 5307. 5710. 54273.Another useful verb introduced by collapse is fgroup_vars, which can be used to efficiently obtain the grouping columns or grouping variables from a grouped tibble:
# fgroup_by fully supports grouped tibbles created with group_by or fgroup_by:
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars %>% head(3)
# # A tibble: 3 x 2
# Variable Country
# <chr> <chr>
# 1 VA BWA
# 2 VA BWA
# 3 VA BWA
GGDC10S %>% fgroup_by(Variable, Country) %>% fgroup_vars %>% head(3)
# # A tibble: 3 x 2
# Variable Country
# <chr> <chr>
# 1 VA BWA
# 2 VA BWA
# 3 VA BWA
# The other possibilities:
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars("unique") %>% head(3)
# # A tibble: 3 x 2
# Variable Country
# <chr> <chr>
# 1 EMP ARG
# 2 EMP BOL
# 3 EMP BRA
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars("names")
# [1] "Variable" "Country"
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars("indices")
# [1] 4 1
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars("named_indices")
# Variable Country
# 4 1
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars("logical")
# [1] TRUE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE
GGDC10S %>% group_by(Variable, Country) %>% fgroup_vars("named_logical")
# Country Regioncode Region Variable Year AGR MIN MAN PU
# TRUE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE
# CON WRT TRA FIRE GOV OTH SUM
# FALSE FALSE FALSE FALSE FALSE FALSE FALSEAnother collapse verb to mention here is fsubset, a faster alternative to dplyr::filter which also provides an option to flexibly subset columns after the select argument:
# Two equivalent calls, the first is substantially faster
GGDC10S %>% fsubset(Variable == "VA" & Year > 1990, Country, Year, AGR:GOV) %>% head(3)
# Country Year AGR MIN MAN PU CON WRT TRA FIRE GOV
# 1 BWA 1991 303.1157 2646.950 472.6488 160.6079 580.0876 806.7509 232.7884 432.6965 1073.263
# 2 BWA 1992 333.4364 2690.939 537.4274 178.4532 678.7320 725.2577 285.1403 517.2141 1234.012
# 3 BWA 1993 404.5488 2624.928 567.3420 219.2183 634.2797 771.8253 349.7458 673.2540 1487.193
GGDC10S %>% filter(Variable == "VA" & Year > 1990) %>% select(Country, Year, AGR:GOV) %>% head(3)
# Country Year AGR MIN MAN PU CON WRT TRA FIRE GOV
# 1 BWA 1991 303.1157 2646.950 472.6488 160.6079 580.0876 806.7509 232.7884 432.6965 1073.263
# 2 BWA 1992 333.4364 2690.939 537.4274 178.4532 678.7320 725.2577 285.1403 517.2141 1234.012
# 3 BWA 1993 404.5488 2624.928 567.3420 219.2183 634.2797 771.8253 349.7458 673.2540 1487.193collapse also offers roworder, frename, colorder, ftransform/TRA and fungroup as fast replacements for dplyr::arrange, dplyr::rename, dplyr::relocate, dplyr::mutate, and dplyr::ungroup.
One can also aggregate with multiple functions at the same time. For such operations it is often necessary to use curly braces { to prevent first argument injection so that %>% cbind(FUN1(.), FUN2(.)) does not evaluate as %>% cbind(., FUN1(.), FUN2(.)):
GGDC10S %>%
fgroup_by(Variable, Country) %>%
get_vars(6:16) %>% {
cbind(fmedian(.),
add_stub(fmean(., keep.group_vars = FALSE), "mean_"))
} %>% head(3)
# Variable Country AGR MIN MAN PU CON WRT TRA
# 1 EMP ARG 1324.5255 47.35255 1987.5912 104.738825 782.40283 1854.612 579.93982
# 2 EMP BOL 943.1612 53.53538 167.1502 4.457895 65.97904 132.225 96.96828
# 3 EMP BRA 17480.9810 225.43693 7207.7915 375.851832 4054.66103 6454.523 1580.81120
# FIRE GOV OTH SUM mean_AGR mean_MIN mean_MAN mean_PU mean_CON
# 1 464.39920 1738.836 866.1119 9743.223 1419.8013 52.08903 1931.7602 101.720936 742.4044
# 2 15.34259 NA 384.0678 1842.055 964.2103 56.03295 235.0332 5.346433 122.7827
# 3 4354.86210 4449.942 4478.6927 51881.110 17191.3529 206.02389 6991.3710 364.573404 3524.7384
# mean_WRT mean_TRA mean_FIRE mean_GOV mean_OTH mean_SUM
# 1 1982.1775 648.5119 627.79291 2043.471 992.4475 10542.177
# 2 281.5164 115.4728 44.56442 NA 395.5650 2220.524
# 3 8509.4612 2054.3731 4413.54448 5307.280 5710.2665 54272.985The function add_stub used above is a collapse function adding a prefix (default) or suffix to variables names. The collapse predicate add_vars provides a more efficient alternative to cbind.data.frame. The idea here is ‘adding’ variables to the data.frame in the first argument i.e. the attributes of the first argument are preserved, so the expression below still gives a tibble instead of a data.frame:
GGDC10S %>%
fgroup_by(Variable, Country) %>% {
add_vars(get_vars(., "Reg", regex = TRUE) %>% ffirst, # Regular expression matching column names
num_vars(.) %>% fmean(keep.group_vars = FALSE) %>% add_stub("mean_"), # num_vars selects all numeric variables
fselect(., PU:TRA) %>% fmedian(keep.group_vars = FALSE) %>% add_stub("median_"),
fselect(., PU:CON) %>% fmin(keep.group_vars = FALSE) %>% add_stub("min_"))
} %>% head(3)
# # A tibble: 3 x 22
# Variable Country Regioncode Region mean_Year mean_AGR mean_MIN mean_MAN mean_PU mean_CON mean_WRT
# <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG LAM Latin~ 1980. 1420. 52.1 1932. 102. 742. 1982.
# 2 EMP BOL LAM Latin~ 1980 964. 56.0 235. 5.35 123. 282.
# 3 EMP BRA LAM Latin~ 1980. 17191. 206. 6991. 365. 3525. 8509.
# # ... with 11 more variables: mean_TRA <dbl>, mean_FIRE <dbl>, mean_GOV <dbl>, mean_OTH <dbl>,
# # mean_SUM <dbl>, median_PU <dbl>, median_CON <dbl>, median_WRT <dbl>, median_TRA <dbl>,
# # min_PU <dbl>, min_CON <dbl>Another nice feature of add_vars is that it can also very efficiently reorder columns i.e. bind columns in a different order than they are passed. This can be done by simply specifying the positions the added columns should have in the final data frame, and then add_vars shifts the first argument columns to the right to fill in the gaps.
GGDC10S %>%
fsubset(Variable == "VA", Country, AGR, SUM) %>%
fgroup_by(Country) %>% {
add_vars(fgroup_vars(.,"unique"),
fmean(., keep.group_vars = FALSE) %>% add_stub("mean_"),
fsd(., keep.group_vars = FALSE) %>% add_stub("sd_"),
pos = c(2,4,3,5))
} %>% head(3)
# Country mean_AGR sd_AGR mean_SUM sd_SUM
# 1 ARG 14951.292 33061.413 152533.84 301316.25
# 2 BOL 3299.718 4456.331 22619.18 33172.98
# 3 BRA 76870.146 59441.696 1200562.67 976963.14A much more compact solution to multi-function and multi-type aggregation is offered by the function collapg:
# This aggregates numeric colums using the mean (fmean) and categorical columns with the mode (fmode)
GGDC10S %>% fgroup_by(Variable, Country) %>% collapg %>% head(3)
# # A tibble: 3 x 16
# Variable Country Regioncode Region Year AGR MIN MAN PU CON WRT TRA FIRE GOV
# <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG LAM Latin~ 1980. 1420. 52.1 1932. 102. 742. 1982. 649. 628. 2043.
# 2 EMP BOL LAM Latin~ 1980 964. 56.0 235. 5.35 123. 282. 115. 44.6 NA
# 3 EMP BRA LAM Latin~ 1980. 17191. 206. 6991. 365. 3525. 8509. 2054. 4414. 5307.
# # ... with 2 more variables: OTH <dbl>, SUM <dbl>By default it aggregates numeric columns using the fmean and categorical columns using fmode, and preserves the order of all columns. Changing these defaults is very easy:
# This aggregates numeric colums using the median and categorical columns using the first value
GGDC10S %>% fgroup_by(Variable, Country) %>% collapg(fmedian, flast) %>% head(3)
# # A tibble: 3 x 16
# Variable Country Regioncode Region Year AGR MIN MAN PU CON WRT TRA FIRE
# <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG LAM Latin~ 1980. 1325. 47.4 1988. 105. 782. 1855. 580. 464.
# 2 EMP BOL LAM Latin~ 1980 943. 53.5 167. 4.46 66.0 132. 97.0 15.3
# 3 EMP BRA LAM Latin~ 1980. 17481. 225. 7208. 376. 4055. 6455. 1581. 4355.
# # ... with 3 more variables: GOV <dbl>, OTH <dbl>, SUM <dbl>One can apply multiple functions to both numeric and/or categorical data:
GGDC10S %>% fgroup_by(Variable, Country) %>%
collapg(list(fmean, fmedian), list(first, fmode, flast)) %>% head(3)
# # A tibble: 3 x 32
# Variable Country first.Regioncode fmode.Regioncode flast.Regioncode first.Region fmode.Region
# <chr> <chr> <chr> <chr> <chr> <chr> <chr>
# 1 EMP ARG LAM LAM LAM Latin Ameri~ Latin Ameri~
# 2 EMP BOL LAM LAM LAM Latin Ameri~ Latin Ameri~
# 3 EMP BRA LAM LAM LAM Latin Ameri~ Latin Ameri~
# # ... with 25 more variables: flast.Region <chr>, fmean.Year <dbl>, fmedian.Year <dbl>,
# # fmean.AGR <dbl>, fmedian.AGR <dbl>, fmean.MIN <dbl>, fmedian.MIN <dbl>, fmean.MAN <dbl>,
# # fmedian.MAN <dbl>, fmean.PU <dbl>, fmedian.PU <dbl>, fmean.CON <dbl>, fmedian.CON <dbl>,
# # fmean.WRT <dbl>, fmedian.WRT <dbl>, fmean.TRA <dbl>, fmedian.TRA <dbl>, fmean.FIRE <dbl>,
# # fmedian.FIRE <dbl>, fmean.GOV <dbl>, fmedian.GOV <dbl>, fmean.OTH <dbl>, fmedian.OTH <dbl>,
# # fmean.SUM <dbl>, fmedian.SUM <dbl>Applying multiple functions to only numeric (or only categorical) data allows return in a long format:
GGDC10S %>% fgroup_by(Variable, Country) %>%
collapg(list(fmean, fmedian), cols = is.numeric, return = "long") %>% head(3)
# # A tibble: 3 x 15
# Function Variable Country Year AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 fmean EMP ARG 1980. 1420. 52.1 1932. 102. 742. 1982. 649. 628. 2043. 992.
# 2 fmean EMP BOL 1980 964. 56.0 235. 5.35 123. 282. 115. 44.6 NA 396.
# 3 fmean EMP BRA 1980. 17191. 206. 6991. 365. 3525. 8509. 2054. 4414. 5307. 5710.
# # ... with 1 more variable: SUM <dbl>Finally, collapg also makes it very easy to apply aggregator functions to certain columns only:
GGDC10S %>% fgroup_by(Variable, Country) %>%
collapg(custom = list(fmean = 6:8, fmedian = 10:12)) %>% head(3)
# # A tibble: 3 x 8
# Variable Country fmean.AGR fmean.MIN fmean.MAN fmedian.CON fmedian.WRT fmedian.TRA
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1420. 52.1 1932. 782. 1855. 580.
# 2 EMP BOL 964. 56.0 235. 66.0 132. 97.0
# 3 EMP BRA 17191. 206. 6991. 4055. 6455. 1581.To understand more about collapg, look it up in the documentation (?collapg).
Weighted aggregations are possible with the functions fsum, fprod, fmean, fmedian, fnth, fmode, fvar and fsd. The implementation is such that by default (option keep.w = TRUE) these functions also aggregate the weights, so that further weighted computations can be performed on the aggregated data. fprod saves the product of the weights, whereas the other functions save the sum of the weights in a column next to the grouping variables. If na.rm = TRUE (the default), rows with missing weights are omitted from the computation.
# This computes a frequency-weighted grouped standard-deviation, taking the total EMP / VA as weight
GGDC10S %>%
fgroup_by(Variable, Country) %>%
fselect(AGR:SUM) %>% fsd(SUM) %>% head(3)
# # A tibble: 3 x 13
# Variable Country sum.SUM AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 653615. 225. 22.2 176. 20.5 285. 856. 195. 493. 1123. 506.
# 2 EMP BOL 135452. 99.7 17.1 168. 4.87 123. 324. 98.1 69.8 NA 258.
# 3 EMP BRA 3364925. 1587. 73.8 2952. 93.8 1861. 6285. 1306. 3003. 3621. 4257.
# This computes a weighted grouped mode, taking the total EMP / VA as weight
GGDC10S %>%
fgroup_by(Variable, Country) %>%
fselect(AGR:SUM) %>% fmode(SUM) %>% head(3)
# # A tibble: 3 x 13
# Variable Country sum.SUM AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 653615. 1162. 127. 2164. 152. 1415. 3768. 1060. 1748. 4336. 1999.
# 2 EMP BOL 135452. 819. 37.6 604. 10.8 433. 893. 333. 321. NA 1057.
# 3 EMP BRA 3364925. 16451. 313. 11841. 388. 8154. 21860. 5169. 12011. 12149. 14235.The weighted variance / standard deviation is currently only implemented with frequency weights.
Weighted aggregations may also be performed with collapg. By default fsum is used to compute a sum of the weights, but it is also possible here to aggregate the weights with other functions:
# This aggregates numeric colums using the weighted mean (the default) and categorical columns using the weighted mode (the default).
# Weights (column SUM) are aggregated using both the sum and the maximum.
GGDC10S %>% group_by(Variable, Country) %>%
collapg(w = SUM, wFUN = list(fsum, fmax)) %>% head(3)
# # A tibble: 3 x 17
# Variable Country fsum.SUM fmax.SUM Regioncode Region Year AGR MIN MAN PU CON WRT
# <chr> <chr> <dbl> <dbl> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 653615. 17929. LAM Latin~ 1985. 1361. 56.5 1935. 105. 811. 2217.
# 2 EMP BOL 135452. 4508. LAM Latin~ 1987. 977. 57.9 296. 7.07 167. 400.
# 3 EMP BRA 3364925. 102572. LAM Latin~ 1989. 17746. 238. 8466. 389. 4436. 11376.
# # ... with 4 more variables: TRA <dbl>, FIRE <dbl>, GOV <dbl>, OTH <dbl>collapse also provides some fast transformations that significantly extend the scope and speed of manipulations that can be performed with dplyr::mutate.
The function ftransform can be used to manipulate columns in the same ways as mutate:
GGDC10S %>% fsubset(Variable == "VA", Country, Year, AGR, SUM) %>%
ftransform(AGR_perc = AGR / SUM * 100, # Computing % of VA in Agriculture
AGR_mean = fmean(AGR), # Average Agricultural VA
AGR = NULL, SUM = NULL) %>% # Deleting columns AGR and SUM
head
# Country Year AGR_perc AGR_mean
# 1 BWA 1960 NA 5137561
# 2 BWA 1961 NA 5137561
# 3 BWA 1962 NA 5137561
# 4 BWA 1963 NA 5137561
# 5 BWA 1964 43.49132 5137561
# 6 BWA 1965 39.96990 5137561Instead of column = value type arguments, it is also possible to pass a single list of transformed variables to ftransform, which will be regarded in the same way as an evaluated list of column = value arguments:
# This replaces variables mpg, carb and wt by their log
mtcars %>% ftransform(fselect(., mpg, carb, wt) %>% lapply(log)) %>% head
# mpg cyl disp hp drat wt qsec vs am gear carb
# Mazda RX4 3.044522 6 160 110 3.90 0.9631743 16.46 0 1 4 1.3862944
# Mazda RX4 Wag 3.044522 6 160 110 3.90 1.0560527 17.02 0 1 4 1.3862944
# Datsun 710 3.126761 4 108 93 3.85 0.8415672 18.61 1 1 4 0.0000000
# Hornet 4 Drive 3.063391 6 258 110 3.08 1.1678274 19.44 1 0 3 0.0000000
# Hornet Sportabout 2.928524 8 360 175 3.15 1.2354715 17.02 0 0 3 0.6931472
# Valiant 2.895912 6 225 105 2.76 1.2412686 20.22 1 0 3 0.0000000
# This adds the log of mpg, carb and wt
mtcars %>% ftransform(fselect(., mpg, carb, wt) %>% lapply(log) %>% add_stub("log.")) %>% head
# mpg cyl disp hp drat wt qsec vs am gear carb log.mpg log.carb log.wt
# Mazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4 3.044522 1.3862944 0.9631743
# Mazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4 3.044522 1.3862944 1.0560527
# Datsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 1 3.126761 0.0000000 0.8415672
# Hornet 4 Drive 21.4 6 258 110 3.08 3.215 19.44 1 0 3 1 3.063391 0.0000000 1.1678274
# Hornet Sportabout 18.7 8 360 175 3.15 3.440 17.02 0 0 3 2 2.928524 0.6931472 1.2354715
# Valiant 18.1 6 225 105 2.76 3.460 20.22 1 0 3 1 2.895912 0.0000000 1.2412686If only the computed columns need to be returned, fcompute provides an efficient alternative:
GGDC10S %>% fsubset(Variable == "VA", Country, Year, AGR, SUM) %>%
fcompute(AGR_perc = AGR / SUM * 100,
AGR_mean = fmean(AGR)) %>% head
# AGR_perc AGR_mean
# 1 NA 5137561
# 2 NA 5137561
# 3 NA 5137561
# 4 NA 5137561
# 5 43.49132 5137561
# 6 39.96990 5137561ftransform and fcompute are an order of magnitude faster than mutate, but they do not support grouped computations using arbitrary functions. We will see that this is hardly a limitation as collapse provides very efficient and elegant alternative programming mechanisms…
All statistical (scalar-valued) functions in the collapse package (fsum, fprod, fmean, fmedian, fmode, fvar, fsd, fmin, fmax, fnth, ffirst, flast, fNobs, fNdistinct) have a TRA argument which can be used to efficiently transforms data by either (column-wise) replacing data values with computed statistics or sweeping the statistics out of the data. Operations can be specified using either an integer or quoted operator / string. The 10 operations supported by TRA are:
1 - “replace_fill” : replace and overwrite missing values (same as mutate)
2 - “replace” : replace but preserve missing values
3 - “-” : subtract (center)
4 - “-+” : subtract group-statistics but add average of group statistics
5 - “/” : divide (scale)
6 - “%” : compute percentages (divide and multiply by 100)
7 - “+” : add
8 - "*" : multiply
9 - “%%” : modulus
10 - “-%%” : subtract modulus
Simple transformations are again straightforward to specify:
# This subtracts the median value from all data points i.e. centers on the median
GGDC10S %>% num_vars %>% fmedian(TRA = "-") %>% head
# Year AGR MIN MAN PU CON WRT TRA FIRE GOV
# 1 -22 NA NA NA NA NA NA NA NA NA
# 2 -21 NA NA NA NA NA NA NA NA NA
# 3 -20 NA NA NA NA NA NA NA NA NA
# 4 -19 NA NA NA NA NA NA NA NA NA
# 5 -18 -4378.218 -169.7294 -3717.362 -167.8456 -1472.787 -3767.399 -1173.141 -959.0059 -3923.690
# 6 -17 -4378.792 -170.7277 -3717.080 -167.8149 -1472.101 -3766.578 -1172.861 -958.8783 -3922.817
# OTH SUM
# 1 NA NA
# 2 NA NA
# 3 NA NA
# 4 NA NA
# 5 -1430.831 -23148.71
# 6 -1430.494 -23146.85
# This replaces all data points with the mode
GGDC10S %>% char_vars %>% fmode(TRA = "replace") %>% head
# Country Regioncode Region Variable
# 1 USA ASI Asia EMP
# 2 USA ASI Asia EMP
# 3 USA ASI Asia EMP
# 4 USA ASI Asia EMP
# 5 USA ASI Asia EMP
# 6 USA ASI Asia EMPSimilarly for grouped transformations:
# Replacing data with the 2nd quartile (25%)
GGDC10S %>%
fselect(Variable, Country, AGR:SUM) %>%
fgroup_by(Variable, Country) %>% fnth(0.25, TRA = "replace_fill") %>% head(3)
# # A tibble: 3 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 VA BWA 61.3 21.7 23.1 6.31 23.2 26.7 8.98 11.3 27.0 10.1 220.
# 2 VA BWA 61.3 21.7 23.1 6.31 23.2 26.7 8.98 11.3 27.0 10.1 220.
# 3 VA BWA 61.3 21.7 23.1 6.31 23.2 26.7 8.98 11.3 27.0 10.1 220.
# Scaling sectoral data by Variable and Country
GGDC10S %>%
fselect(Variable, Country, AGR:SUM) %>%
fgroup_by(Variable, Country) %>% fsd(TRA = "/") %>% head
# # A tibble: 6 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 VA BWA NA NA NA NA NA NA NA NA NA
# 2 VA BWA NA NA NA NA NA NA NA NA NA
# 3 VA BWA NA NA NA NA NA NA NA NA NA
# 4 VA BWA NA NA NA NA NA NA NA NA NA
# 5 VA BWA 0.0270 5.56e-4 5.23e-4 3.88e-4 5.11e-4 0.00194 0.00154 5.23e-4 0.00134
# 6 VA BWA 0.0260 3.97e-4 7.23e-4 5.03e-4 1.04e-3 0.00220 0.00180 5.83e-4 0.00158
# # ... with 2 more variables: OTH <dbl>, SUM <dbl>The benchmarks below will demonstrate that these internal sweeping and replacement operations fully performed in C++ compute significantly faster than using dplyr::mutate, especially as the number of groups grows large. The S3 generic nature of the Fast Statistical Functions further allows us to perform grouped mutations on the fly (together with ftransform or fcompute), without the need of first creating a grouped tibble:
# AGR_gmed = TRUE if AGR is greater than it's median value, grouped by Variable and Country
# Note: This calls fmedian.default
settransform(GGDC10S, AGR_gmed = AGR > fmedian(AGR, list(Variable, Country), TRA = "replace"))
tail(GGDC10S, 3)
# Country Regioncode Region Variable Year AGR MIN MAN
# 5025 EGY MENA Middle East and North Africa EMP 2010 5205.529 28.99641 2435.549
# 5026 EGY MENA Middle East and North Africa EMP 2011 5185.919 27.56394 2373.814
# 5027 EGY MENA Middle East and North Africa EMP 2012 5160.590 24.78083 2348.434
# PU CON WRT TRA FIRE GOV OTH SUM AGR_gmed
# 5025 307.2712 2732.953 2977.063 1992.274 801.2984 5538.946 NA 22019.88 TRUE
# 5026 317.9979 2795.264 3020.236 2048.335 814.7403 5635.522 NA 22219.39 TRUE
# 5027 324.9332 2931.196 3109.522 2065.004 832.4770 5735.623 NA 22532.56 TRUEWeights are easily added to any grouped transformation:
# This subtracts weighted group means from the data, using SUM column as weights..
GGDC10S %>%
fselect(Variable, Country, AGR:SUM) %>%
fgroup_by(Variable, Country) %>% fmean(SUM, "-") %>% head
# # A tibble: 6 x 13
# Variable Country SUM AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 2 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 3 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 4 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 5 VA BWA 37.5 -1301. -13317. -2965. -529. -2746. -6540. -2157. -4431. -7551. -2613.
# 6 VA BWA 39.3 -1302. -13318. -2964. -529. -2745. -6540. -2156. -4431. -7550. -2613.Sequential operations are also easily performed:
# This scales and then subtracts the median
GGDC10S %>%
fselect(Variable, Country, AGR:SUM) %>%
fgroup_by(Variable, Country) %>% fsd(TRA = "/") %>% fmedian(TRA = "-")
# # A tibble: 5,027 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 2 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 3 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 4 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 5 VA BWA -0.182 -0.235 -0.183 -0.245 -0.118 -0.0820 -0.0724 -0.0661 -0.108 -0.0848 -0.146
# 6 VA BWA -0.183 -0.235 -0.183 -0.245 -0.117 -0.0817 -0.0722 -0.0660 -0.108 -0.0846 -0.146
# 7 VA BWA -0.180 -0.235 -0.183 -0.245 -0.117 -0.0813 -0.0720 -0.0659 -0.107 -0.0843 -0.145
# 8 VA BWA -0.177 -0.235 -0.183 -0.245 -0.117 -0.0826 -0.0724 -0.0659 -0.107 -0.0841 -0.146
# 9 VA BWA -0.174 -0.235 -0.183 -0.245 -0.117 -0.0823 -0.0717 -0.0661 -0.108 -0.0848 -0.146
# 10 VA BWA -0.173 -0.234 -0.182 -0.243 -0.115 -0.0821 -0.0715 -0.0660 -0.108 -0.0846 -0.145
# # ... with 5,017 more rowsOf course it is also possible to combine multiple functions as in the aggregation section, or to add variables to existing data:
# This adds a groupwise observation count next to each column
add_vars(GGDC10S, seq(7,27,2)) <- GGDC10S %>%
fgroup_by(Variable, Country) %>% fselect(AGR:SUM) %>%
fNobs("replace_fill") %>% add_stub("N_")
head(GGDC10S)
# Country Regioncode Region Variable Year AGR N_AGR MIN N_MIN MAN N_MAN
# 1 BWA SSA Sub-saharan Africa VA 1960 NA 47 NA 47 NA 47
# 2 BWA SSA Sub-saharan Africa VA 1961 NA 47 NA 47 NA 47
# 3 BWA SSA Sub-saharan Africa VA 1962 NA 47 NA 47 NA 47
# 4 BWA SSA Sub-saharan Africa VA 1963 NA 47 NA 47 NA 47
# 5 BWA SSA Sub-saharan Africa VA 1964 16.30154 47 3.494075 47 0.7365696 47
# 6 BWA SSA Sub-saharan Africa VA 1965 15.72700 47 2.495768 47 1.0181992 47
# PU N_PU CON N_CON WRT N_WRT TRA N_TRA FIRE N_FIRE GOV N_GOV
# 1 NA 47 NA 47 NA 47 NA 47 NA 47 NA 47
# 2 NA 47 NA 47 NA 47 NA 47 NA 47 NA 47
# 3 NA 47 NA 47 NA 47 NA 47 NA 47 NA 47
# 4 NA 47 NA 47 NA 47 NA 47 NA 47 NA 47
# 5 0.1043936 47 0.6600454 47 6.243732 47 1.658928 47 1.119194 47 4.822485 47
# 6 0.1350976 47 1.3462312 47 7.064825 47 1.939007 47 1.246789 47 5.695848 47
# OTH N_OTH SUM N_SUM AGR_gmed
# 1 NA 47 NA 47 NA
# 2 NA 47 NA 47 NA
# 3 NA 47 NA 47 NA
# 4 NA 47 NA 47 NA
# 5 2.341328 47 37.48229 47 FALSE
# 6 2.678338 47 39.34710 47 FALSE
rm(GGDC10S)There are lots of other examples one could construct using the 10 operations and 14 functions listed above, the examples provided just outline the suggested programming basics. Performance considerations make it very much worthwhile to spend some time and think how complex operations can be implemented in this programming framework, before defining some function in R and applying it to data using dplyr::mutate.
TRA FunctionTowards this end, calling TRA() directly also facilitates more complex and customized operations. Behind the scenes of the TRA = ... argument, the Fast Statistical Functions first compute the grouped statistics on all columns of the data, and these statistics are then directly fed into a C++ function that uses them to replace or sweep them out of data points in one of the 10 ways described above. This function can also be called directly by the name of TRA.
Fundamentally, TRA is a generalization of base::sweep for column-wise grouped operations1. Direct calls to TRA enable more control over inputs and outputs.
The two operations below are equivalent, although the first is slightly more efficient as it only requires one method dispatch and one check of the inputs:
# This divides by the product
GGDC10S %>%
fgroup_by(Variable, Country) %>%
get_vars(6:16) %>% fprod(TRA = "/") %>% head
# # A tibble: 6 x 11
# AGR MIN MAN PU CON WRT TRA FIRE GOV
# <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 NA NA NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA NA NA
# 3 NA NA NA NA NA NA NA NA NA
# 4 NA NA NA NA NA NA NA NA NA
# 5 1.29e-105 2.81e-127 1.40e-101 4.44e-74 4.19e-102 3.97e-113 6.91e-92 1.01e-97 2.51e-117
# 6 1.24e-105 2.00e-127 1.94e-101 5.75e-74 8.55e-102 4.49e-113 8.08e-92 1.13e-97 2.96e-117
# # ... with 2 more variables: OTH <dbl>, SUM <dbl>
# Same thing
GGDC10S %>%
fgroup_by(Variable, Country) %>%
get_vars(6:16) %>%
TRA(fprod(., keep.group_vars = FALSE), "/") %>% head # [same as TRA(.,fprod(., keep.group_vars = FALSE),"/")]
# # A tibble: 6 x 11
# AGR MIN MAN PU CON WRT TRA FIRE GOV
# <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 NA NA NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA NA NA
# 3 NA NA NA NA NA NA NA NA NA
# 4 NA NA NA NA NA NA NA NA NA
# 5 1.29e-105 2.81e-127 1.40e-101 4.44e-74 4.19e-102 3.97e-113 6.91e-92 1.01e-97 2.51e-117
# 6 1.24e-105 2.00e-127 1.94e-101 5.75e-74 8.55e-102 4.49e-113 8.08e-92 1.13e-97 2.96e-117
# # ... with 2 more variables: OTH <dbl>, SUM <dbl>TRA.grouped_df was designed such that it matches the columns of the statistics (aggregated columns) to those of the original data, and only transforms matching columns while returning the whole data frame. Thus it is easily possible to only apply a transformation to the first two sectors:
# This only demeans Agriculture (AGR) and Mining (MIN)
GGDC10S %>%
fgroup_by(Variable, Country) %>%
TRA(fselect(., AGR, MIN) %>% fmean(keep.group_vars = FALSE), "-") %>% head
# # A tibble: 6 x 16
# Country Regioncode Region Variable Year AGR MIN MAN PU CON WRT TRA FIRE GOV
# <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA SSA Sub-s~ VA 1960 NA NA NA NA NA NA NA NA NA
# 2 BWA SSA Sub-s~ VA 1961 NA NA NA NA NA NA NA NA NA
# 3 BWA SSA Sub-s~ VA 1962 NA NA NA NA NA NA NA NA NA
# 4 BWA SSA Sub-s~ VA 1963 NA NA NA NA NA NA NA NA NA
# 5 BWA SSA Sub-s~ VA 1964 -446. -4505. 0.737 0.104 0.660 6.24 1.66 1.12 4.82
# 6 BWA SSA Sub-s~ VA 1965 -446. -4506. 1.02 0.135 1.35 7.06 1.94 1.25 5.70
# # ... with 2 more variables: OTH <dbl>, SUM <dbl>Since TRA is already built into all Fast Statistical Functions as an argument, it is best used in computations where grouped statistics are computed using some other function.
# Same as above, with one line of code using fmean.data.frame and ftransform...
GGDC10S %>% ftransform(fmean(list(AGR = AGR, MIN = MIN), list(Variable, Country), TRA = "-")) %>% head
# Country Regioncode Region Variable Year AGR MIN MAN PU
# 1 BWA SSA Sub-saharan Africa VA 1960 NA NA NA NA
# 2 BWA SSA Sub-saharan Africa VA 1961 NA NA NA NA
# 3 BWA SSA Sub-saharan Africa VA 1962 NA NA NA NA
# 4 BWA SSA Sub-saharan Africa VA 1963 NA NA NA NA
# 5 BWA SSA Sub-saharan Africa VA 1964 -445.8739 -4505.178 0.7365696 0.1043936
# 6 BWA SSA Sub-saharan Africa VA 1965 -446.4485 -4506.176 1.0181992 0.1350976
# CON WRT TRA FIRE GOV OTH SUM
# 1 NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA
# 3 NA NA NA NA NA NA NA
# 4 NA NA NA NA NA NA NA
# 5 0.6600454 6.243732 1.658928 1.119194 4.822485 2.341328 37.48229
# 6 1.3462312 7.064825 1.939007 1.246789 5.695848 2.678338 39.34710Another potential use of TRA is to do computations in two- or more steps, for example if both aggregated and transformed data are needed, or if computations are more complex and involve other manipulations in-between the aggregating and sweeping part:
# Get grouped tibble
gGGDC <- GGDC10S %>% fgroup_by(Variable, Country)
# Get aggregated data
gsumGGDC <- gGGDC %>% fselect(AGR:SUM) %>% fsum
head(gsumGGDC)
# # A tibble: 6 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 8.80e4 3230. 1.20e5 6307. 4.60e4 1.23e5 4.02e4 3.89e4 1.27e5 6.15e4 6.54e5
# 2 EMP BOL 5.88e4 3418. 1.43e4 326. 7.49e3 1.72e4 7.04e3 2.72e3 NA 2.41e4 1.35e5
# 3 EMP BRA 1.07e6 12773. 4.33e5 22604. 2.19e5 5.28e5 1.27e5 2.74e5 3.29e5 3.54e5 3.36e6
# 4 EMP BWA 8.84e3 493. 8.49e2 145. 1.19e3 1.71e3 3.93e2 7.21e2 2.87e3 1.30e3 1.85e4
# 5 EMP CHL 4.42e4 6389. 3.94e4 1850. 1.86e4 4.38e4 1.63e4 1.72e4 NA 6.32e4 2.51e5
# 6 EMP CHN 1.73e7 422972. 4.03e6 96364. 1.25e6 1.73e6 8.36e5 2.96e5 1.36e6 1.86e6 2.91e7
# Get transformed (scaled) data
head(TRA(gGGDC, gsumGGDC, "/"))
# # A tibble: 6 x 16
# Country Regioncode Region Variable Year AGR MIN MAN PU CON WRT
# <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA SSA Sub-s~ VA 1960 NA NA NA NA NA NA
# 2 BWA SSA Sub-s~ VA 1961 NA NA NA NA NA NA
# 3 BWA SSA Sub-s~ VA 1962 NA NA NA NA NA NA
# 4 BWA SSA Sub-s~ VA 1963 NA NA NA NA NA NA
# 5 BWA SSA Sub-s~ VA 1964 7.50e-4 1.65e-5 1.66e-5 1.03e-5 1.57e-5 6.82e-5
# 6 BWA SSA Sub-s~ VA 1965 7.24e-4 1.18e-5 2.30e-5 1.33e-5 3.20e-5 7.72e-5
# # ... with 5 more variables: TRA <dbl>, FIRE <dbl>, GOV <dbl>, OTH <dbl>, SUM <dbl>As discussed, whether using the argument to fast statistical functions or TRA directly, these data transformations are essentially a two-step process: Statistics are first computed and then used to transform the original data.
Although both steps are efficiently done in C++, it would be even more efficient to do them in a single step without materializing all the statistics before transforming the data. Such slightly more efficient functions are provided for the very commonly applied tasks of centering and averaging data by groups (widely known as ‘between’-group and ‘within’-group transformations), and scaling and centering data by groups (also known as ‘standardizing’ data).
The functions fbetween and fwithin are slightly more memory efficient implementations of fmean invoked with different TRA options:
GGDC10S %>% # Same as ... %>% fmean(TRA = "replace")
fgroup_by(Variable, Country) %>% get_vars(6:16) %>% fbetween %>% head(2)
# # A tibble: 2 x 11
# AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 NA NA NA NA NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA NA NA NA NA
GGDC10S %>% # Same as ... %>% fmean(TRA = "replace_fill")
fgroup_by(Variable, Country) %>% get_vars(6:16) %>% fbetween(fill = TRUE) %>% head(2)
# # A tibble: 2 x 11
# AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 462. 4509. 942. 216. 895. 1948. 635. 1359. 2373. 773. 14112.
# 2 462. 4509. 942. 216. 895. 1948. 635. 1359. 2373. 773. 14112.
GGDC10S %>% # Same as ... %>% fmean(TRA = "-")
fgroup_by(Variable, Country) %>% get_vars(6:16) %>% fwithin %>% head(2)
# # A tibble: 2 x 11
# AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 NA NA NA NA NA NA NA NA NA NA NA
# 2 NA NA NA NA NA NA NA NA NA NA NAApart from higher speed, fwithin has a mean argument to assign an arbitrary mean to centered data, the default being mean = 0. A very common choice for such an added mean is just the overall mean of the data, which can be added in by invoking mean = "overall.mean":
GGDC10S %>%
fgroup_by(Variable, Country) %>%
fselect(Country, Variable, AGR:SUM) %>% fwithin(mean = "overall.mean") %>% head(3)
# # A tibble: 3 x 13
# Country Variable AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 2 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 3 BWA VA NA NA NA NA NA NA NA NA NA NA NAThis can also be done using weights. The code below uses the SUM column as weights, and then for each variable and each group subtracts out the weighted mean, and then adds the overall weighted column mean back to the centered columns. The SUM column is just kept as it is and added after the grouping columns.
GGDC10S %>%
fgroup_by(Variable, Country) %>%
fselect(Country, Variable, AGR:SUM) %>% fwithin(SUM, mean = "overall.mean") %>% head(3)
# # A tibble: 3 x 13
# Country Variable SUM AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 2 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 3 BWA VA NA NA NA NA NA NA NA NA NA NA NAAnother argument to fwithin is the theta parameter, allowing partial- or quasi-demeaning operations, e.g. fwithin(gdata, theta = theta) is equal to gdata - theta * fbetween(gdata). This is particularly useful to prepare data for variance components (also known as ‘random-effects’) estimation.
Apart from fbetween and fwithin, the function fscale exists to efficiently scale and center data, to avoid sequential calls such as ... %>% fsd(TRA = "/") %>% fmean(TRA = "-").
# This efficiently scales and centers (i.e. standardizes) the data
GGDC10S %>%
fgroup_by(Variable, Country) %>%
fselect(Country, Variable, AGR:SUM) %>% fscale
# # A tibble: 5,027 x 13
# Country Variable AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 2 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 3 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 4 BWA VA NA NA NA NA NA NA NA NA NA NA NA
# 5 BWA VA -0.738 -0.717 -0.668 -0.805 -0.692 -0.603 -0.589 -0.635 -0.656 -0.596 -0.676
# 6 BWA VA -0.739 -0.717 -0.668 -0.805 -0.692 -0.603 -0.589 -0.635 -0.656 -0.596 -0.676
# 7 BWA VA -0.736 -0.717 -0.668 -0.805 -0.692 -0.603 -0.589 -0.635 -0.656 -0.595 -0.676
# 8 BWA VA -0.734 -0.717 -0.668 -0.805 -0.692 -0.604 -0.589 -0.635 -0.655 -0.595 -0.676
# 9 BWA VA -0.730 -0.717 -0.668 -0.805 -0.692 -0.604 -0.588 -0.635 -0.656 -0.596 -0.676
# 10 BWA VA -0.729 -0.716 -0.667 -0.803 -0.690 -0.603 -0.588 -0.635 -0.656 -0.596 -0.675
# # ... with 5,017 more rowsfscale also has additional mean and sd arguments allowing the user to (group-) scale data to an arbitrary mean and standard deviation. Setting mean = FALSE just scales the data but preserves the means, and is thus different from fsd(..., TRA = "/") which simply divides all values by the standard deviation:
# Saving grouped tibble
gGGDC <- GGDC10S %>%
fgroup_by(Variable, Country) %>%
fselect(Country, Variable, AGR:SUM)
# Original means
head(fmean(gGGDC))
# # A tibble: 6 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1420. 52.1 1932. 102. 742. 1.98e3 6.49e2 628. 2043. 9.92e2 1.05e4
# 2 EMP BOL 964. 56.0 235. 5.35 123. 2.82e2 1.15e2 44.6 NA 3.96e2 2.22e3
# 3 EMP BRA 17191. 206. 6991. 365. 3525. 8.51e3 2.05e3 4414. 5307. 5.71e3 5.43e4
# 4 EMP BWA 188. 10.5 18.1 3.09 25.3 3.63e1 8.36e0 15.3 61.1 2.76e1 3.94e2
# 5 EMP CHL 702. 101. 625. 29.4 296. 6.95e2 2.58e2 272. NA 1.00e3 3.98e3
# 6 EMP CHN 287744. 7050. 67144. 1606. 20852. 2.89e4 1.39e4 4929. 22669. 3.10e4 4.86e5
# Mean Preserving Scaling
head(fmean(fscale(gGGDC, mean = FALSE)))
# # A tibble: 6 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1420. 52.1 1932. 102. 742. 1.98e3 6.49e2 628. 2043. 9.92e2 1.05e4
# 2 EMP BOL 964. 56.0 235. 5.35 123. 2.82e2 1.15e2 44.6 NA 3.96e2 2.22e3
# 3 EMP BRA 17191. 206. 6991. 365. 3525. 8.51e3 2.05e3 4414. 5307. 5.71e3 5.43e4
# 4 EMP BWA 188. 10.5 18.1 3.09 25.3 3.63e1 8.36e0 15.3 61.1 2.76e1 3.94e2
# 5 EMP CHL 702. 101. 625. 29.4 296. 6.95e2 2.58e2 272. NA 1.00e3 3.98e3
# 6 EMP CHN 287744. 7050. 67144. 1606. 20852. 2.89e4 1.39e4 4929. 22669. 3.10e4 4.86e5
head(fsd(fscale(gGGDC, mean = FALSE)))
# # A tibble: 6 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 EMP ARG 1. 1. 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.
# 2 EMP BOL 1. 1.00 1. 1.00 1.00 1. 1. 1. NA 1. 1.
# 3 EMP BRA 1. 1. 1. 1.00 1. 1.00 1.00 1.00 1. 1.00 1.00
# 4 EMP BWA 1.00 1.00 1. 1. 1. 1.00 1. 1.00 1. 1.00 1.00
# 5 EMP CHL 1. 1. 1.00 1. 1. 1. 1.00 1. NA 1. 1.00
# 6 EMP CHN 1. 1. 1. 1.00 1.00 1. 1. 1. 1.00 1.00 1.One can also set mean = "overall.mean", which group-centers columns on the overall mean as illustrated with fwithin. Another interesting option is setting sd = "within.sd". This group-scales data such that every group has a standard deviation equal to the within-standard deviation of the data:
# Just using VA data for this example
gGGDC <- GGDC10S %>%
fsubset(Variable == "VA", Country, AGR:SUM) %>%
fgroup_by(Country)
# This calculates the within- standard deviation for all columns
fsd(num_vars(ungroup(fwithin(gGGDC))))
# AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# 45046972 40122220 75608708 3062688 30811572 44125207 20676901 16030868 20358973 18780869
# SUM
# 306429102
# This scales all groups to take on the within- standard deviation while preserving group means
fsd(fscale(gGGDC, mean = FALSE, sd = "within.sd"))
# # A tibble: 43 x 12
# Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 ARG 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# 2 BOL 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 NA 1.88e7 3.06e8
# 3 BRA 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# 4 BWA 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# 5 CHL 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 NA 1.88e7 3.06e8
# 6 CHN 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# 7 COL 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 NA 1.88e7 3.06e8
# 8 CRI 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# 9 DEW 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# 10 DNK 4.50e7 4.01e7 7.56e7 3.06e6 3.08e7 4.41e7 2.07e7 1.60e7 2.04e7 1.88e7 3.06e8
# # ... with 33 more rowsA grouped scaling operation with both mean = "overall.mean" and sd = "within.sd" thus efficiently achieves a harmonization of all groups in the first two moments without changing the fundamental properties (in terms of level and scale) of the data.
This section introduces 3 further powerful collapse functions: flag, fdiff and fgrowth. The first function, flag, efficiently computes sequences of fully identified lags and leads on time series and panel data. The following code computes 1 fully-identified panel-lag and 1 fully identified panel-lead of each variable in the data:
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% flag(-1:1, Year)
# # A tibble: 5,027 x 36
# Country Variable Year F1.AGR AGR L1.AGR F1.MIN MIN L1.MIN F1.MAN MAN L1.MAN F1.PU PU
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA NA NA NA NA NA
# 4 BWA VA 1963 16.3 NA NA 3.49 NA NA 0.737 NA NA 0.104 NA
# 5 BWA VA 1964 15.7 16.3 NA 2.50 3.49 NA 1.02 0.737 NA 0.135 0.104
# 6 BWA VA 1965 17.7 15.7 16.3 1.97 2.50 3.49 0.804 1.02 0.737 0.203 0.135
# 7 BWA VA 1966 19.1 17.7 15.7 2.30 1.97 2.50 0.938 0.804 1.02 0.203 0.203
# 8 BWA VA 1967 21.1 19.1 17.7 1.84 2.30 1.97 0.750 0.938 0.804 0.203 0.203
# 9 BWA VA 1968 21.9 21.1 19.1 5.24 1.84 2.30 2.14 0.750 0.938 0.578 0.203
# 10 BWA VA 1969 23.1 21.9 21.1 10.2 5.24 1.84 4.15 2.14 0.750 1.12 0.578
# # ... with 5,017 more rows, and 22 more variables: L1.PU <dbl>, F1.CON <dbl>, CON <dbl>,
# # L1.CON <dbl>, F1.WRT <dbl>, WRT <dbl>, L1.WRT <dbl>, F1.TRA <dbl>, TRA <dbl>, L1.TRA <dbl>,
# # F1.FIRE <dbl>, FIRE <dbl>, L1.FIRE <dbl>, F1.GOV <dbl>, GOV <dbl>, L1.GOV <dbl>, F1.OTH <dbl>,
# # OTH <dbl>, L1.OTH <dbl>, F1.SUM <dbl>, SUM <dbl>, L1.SUM <dbl>If the time-variable passed does not exactly identify the data (i.e. because of gaps or repeated values in each group), all 3 functions will issue appropriate error messages. flag, fdiff and fgrowth support unbalanced panels with different start and end periods and duration of coverage for each individual, but not irregular panels. A workaround for such panels exists with the function seqid which generates a new panel-id identifying consecutive time-sequences at the sub-individual level, see ?seqid.
It is also possible to omit the time-variable if one is certain that the data is sorted:
GGDC10S %>%
fselect(Variable, Country,AGR:SUM) %>%
fgroup_by(Variable, Country) %>% flag
# # A tibble: 5,027 x 13
# Variable Country AGR MIN MAN PU CON WRT TRA FIRE GOV OTH SUM
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 2 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 3 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 4 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 5 VA BWA NA NA NA NA NA NA NA NA NA NA NA
# 6 VA BWA 16.3 3.49 0.737 0.104 0.660 6.24 1.66 1.12 4.82 2.34 37.5
# 7 VA BWA 15.7 2.50 1.02 0.135 1.35 7.06 1.94 1.25 5.70 2.68 39.3
# 8 VA BWA 17.7 1.97 0.804 0.203 1.35 8.27 2.15 1.36 6.37 2.99 43.1
# 9 VA BWA 19.1 2.30 0.938 0.203 0.897 4.31 1.72 1.54 7.04 3.31 41.4
# 10 VA BWA 21.1 1.84 0.750 0.203 1.22 5.17 2.44 1.03 5.03 2.36 41.1
# # ... with 5,017 more rowsfdiff computes sequences of lagged-leaded and iterated differences as well as quasi-differences and log-differences on time series and panel data. The code below computes the 1 and 10 year first and second differences of each variable in the data:
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% fdiff(c(1, 10), 1:2, Year)
# # A tibble: 5,027 x 47
# Country Variable Year D1.AGR D2.AGR L10D1.AGR L10D2.AGR D1.MIN D2.MIN L10D1.MIN L10D2.MIN D1.MAN
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA NA NA NA
# 4 BWA VA 1963 NA NA NA NA NA NA NA NA NA
# 5 BWA VA 1964 NA NA NA NA NA NA NA NA NA
# 6 BWA VA 1965 -0.575 NA NA NA -0.998 NA NA NA 0.282
# 7 BWA VA 1966 1.95 2.53 NA NA -0.525 0.473 NA NA -0.214
# 8 BWA VA 1967 1.47 -0.488 NA NA 0.328 0.854 NA NA 0.134
# 9 BWA VA 1968 1.95 0.488 NA NA -0.460 -0.788 NA NA -0.188
# 10 BWA VA 1969 0.763 -1.19 NA NA 3.41 3.87 NA NA 1.39
# # ... with 5,017 more rows, and 35 more variables: D2.MAN <dbl>, L10D1.MAN <dbl>, L10D2.MAN <dbl>,
# # D1.PU <dbl>, D2.PU <dbl>, L10D1.PU <dbl>, L10D2.PU <dbl>, D1.CON <dbl>, D2.CON <dbl>,
# # L10D1.CON <dbl>, L10D2.CON <dbl>, D1.WRT <dbl>, D2.WRT <dbl>, L10D1.WRT <dbl>, L10D2.WRT <dbl>,
# # D1.TRA <dbl>, D2.TRA <dbl>, L10D1.TRA <dbl>, L10D2.TRA <dbl>, D1.FIRE <dbl>, D2.FIRE <dbl>,
# # L10D1.FIRE <dbl>, L10D2.FIRE <dbl>, D1.GOV <dbl>, D2.GOV <dbl>, L10D1.GOV <dbl>,
# # L10D2.GOV <dbl>, D1.OTH <dbl>, D2.OTH <dbl>, L10D1.OTH <dbl>, L10D2.OTH <dbl>, D1.SUM <dbl>,
# # D2.SUM <dbl>, L10D1.SUM <dbl>, L10D2.SUM <dbl>Log-differences of the form \(log(x_t) - log(x_{t-s})\) are also easily computed.
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% fdiff(c(1, 10), 1, Year, log = TRUE)
# # A tibble: 5,027 x 25
# Country Variable Year Dlog1.AGR L10Dlog1.AGR Dlog1.MIN L10Dlog1.MIN Dlog1.MAN L10Dlog1.MAN
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA
# 4 BWA VA 1963 NA NA NA NA NA NA
# 5 BWA VA 1964 NA NA NA NA NA NA
# 6 BWA VA 1965 -0.0359 NA -0.336 NA 0.324 NA
# 7 BWA VA 1966 0.117 NA -0.236 NA -0.236 NA
# 8 BWA VA 1967 0.0796 NA 0.154 NA 0.154 NA
# 9 BWA VA 1968 0.0972 NA -0.223 NA -0.223 NA
# 10 BWA VA 1969 0.0355 NA 1.05 NA 1.05 NA
# # ... with 5,017 more rows, and 16 more variables: Dlog1.PU <dbl>, L10Dlog1.PU <dbl>,
# # Dlog1.CON <dbl>, L10Dlog1.CON <dbl>, Dlog1.WRT <dbl>, L10Dlog1.WRT <dbl>, Dlog1.TRA <dbl>,
# # L10Dlog1.TRA <dbl>, Dlog1.FIRE <dbl>, L10Dlog1.FIRE <dbl>, Dlog1.GOV <dbl>, L10Dlog1.GOV <dbl>,
# # Dlog1.OTH <dbl>, L10Dlog1.OTH <dbl>, Dlog1.SUM <dbl>, L10Dlog1.SUM <dbl>Finally, it is also possible to compute quasi-differences and quasi-log-differences of the form \(x_t - \rho x_{t-s}\) or \(log(x_t) - \rho log(x_{t-s})\):
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% fdiff(t = Year, rho = 0.95)
# # A tibble: 5,027 x 14
# Country Variable Year AGR MIN MAN PU CON WRT TRA FIRE GOV OTH
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA NA NA NA NA
# 4 BWA VA 1963 NA NA NA NA NA NA NA NA NA NA
# 5 BWA VA 1964 NA NA NA NA NA NA NA NA NA NA
# 6 BWA VA 1965 0.241 -0.824 0.318 0.0359 0.719 1.13 0.363 0.184 1.11 0.454
# 7 BWA VA 1966 2.74 -0.401 -0.163 0.0743 0.0673 1.56 0.312 0.174 0.955 0.449
# 8 BWA VA 1967 2.35 0.427 0.174 0.0101 -0.381 -3.55 -0.323 0.246 0.988 0.465
# 9 BWA VA 1968 2.91 -0.345 -0.141 0.0101 0.365 1.08 0.804 -0.427 -1.66 -0.780
# 10 BWA VA 1969 1.82 3.50 1.43 0.385 2.32 0.841 0.397 0.252 0.818 0.385
# # ... with 5,017 more rows, and 1 more variable: SUM <dbl>The quasi-differencing feature was added to fdiff to facilitate the preparation of time series and panel data for least-squares estimations suffering from serial correlation following Cochrane & Orcutt (1949).
Finally, fgrowth computes growth rates in the same way. By default exact growth rates are computed in percentage terms using \((x_t-x_{t-s}) / x_{t-s} \times 100\) (the default argument is scale = 100). The user can also request growth rates obtained by log-differencing using \(log(x_t/ x_{t-s}) \times 100\).
# Exact growth rates, computed as: (x/lag(x) - 1) * 100
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% fgrowth(c(1, 10), 1, Year)
# # A tibble: 5,027 x 25
# Country Variable Year G1.AGR L10G1.AGR G1.MIN L10G1.MIN G1.MAN L10G1.MAN G1.PU L10G1.PU G1.CON
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA NA NA NA
# 4 BWA VA 1963 NA NA NA NA NA NA NA NA NA
# 5 BWA VA 1964 NA NA NA NA NA NA NA NA NA
# 6 BWA VA 1965 -3.52 NA -28.6 NA 38.2 NA 29.4 NA 104.
# 7 BWA VA 1966 12.4 NA -21.1 NA -21.1 NA 50.0 NA 0
# 8 BWA VA 1967 8.29 NA 16.7 NA 16.7 NA 0 NA -33.3
# 9 BWA VA 1968 10.2 NA -20 NA -20 NA 0 NA 35.7
# 10 BWA VA 1969 3.61 NA 185. NA 185. NA 185. NA 185.
# # ... with 5,017 more rows, and 13 more variables: L10G1.CON <dbl>, G1.WRT <dbl>, L10G1.WRT <dbl>,
# # G1.TRA <dbl>, L10G1.TRA <dbl>, G1.FIRE <dbl>, L10G1.FIRE <dbl>, G1.GOV <dbl>, L10G1.GOV <dbl>,
# # G1.OTH <dbl>, L10G1.OTH <dbl>, G1.SUM <dbl>, L10G1.SUM <dbl>
# Log-difference growth rates, computed as: log(x / lag(x)) * 100
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% fgrowth(c(1, 10), 1, Year, logdiff = TRUE)
# # A tibble: 5,027 x 25
# Country Variable Year Dlog1.AGR L10Dlog1.AGR Dlog1.MIN L10Dlog1.MIN Dlog1.MAN L10Dlog1.MAN
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA
# 4 BWA VA 1963 NA NA NA NA NA NA
# 5 BWA VA 1964 NA NA NA NA NA NA
# 6 BWA VA 1965 -3.59 NA -33.6 NA 32.4 NA
# 7 BWA VA 1966 11.7 NA -23.6 NA -23.6 NA
# 8 BWA VA 1967 7.96 NA 15.4 NA 15.4 NA
# 9 BWA VA 1968 9.72 NA -22.3 NA -22.3 NA
# 10 BWA VA 1969 3.55 NA 105. NA 105. NA
# # ... with 5,017 more rows, and 16 more variables: Dlog1.PU <dbl>, L10Dlog1.PU <dbl>,
# # Dlog1.CON <dbl>, L10Dlog1.CON <dbl>, Dlog1.WRT <dbl>, L10Dlog1.WRT <dbl>, Dlog1.TRA <dbl>,
# # L10Dlog1.TRA <dbl>, Dlog1.FIRE <dbl>, L10Dlog1.FIRE <dbl>, Dlog1.GOV <dbl>, L10Dlog1.GOV <dbl>,
# # Dlog1.OTH <dbl>, L10Dlog1.OTH <dbl>, Dlog1.SUM <dbl>, L10Dlog1.SUM <dbl>fdiff and fgrowth can also perform leaded (forward) differences and growth rates (i.e. ... %>% fgrowth(-c(1, 10), 1:2, Year) would compute one and 10-year leaded first and second differences). Again it is possible to perform sequential operations:
# This computes the 1 and 10-year growth rates, for the current period and lagged by one period
GGDC10S %>%
fselect(-Region, -Regioncode) %>%
fgroup_by(Variable, Country) %>% fgrowth(c(1, 10), 1, Year) %>% flag(0:1, Year)
# # A tibble: 5,027 x 47
# Country Variable Year G1.AGR L1.G1.AGR L10G1.AGR L1.L10G1.AGR G1.MIN L1.G1.MIN L10G1.MIN
# * <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
# 1 BWA VA 1960 NA NA NA NA NA NA NA
# 2 BWA VA 1961 NA NA NA NA NA NA NA
# 3 BWA VA 1962 NA NA NA NA NA NA NA
# 4 BWA VA 1963 NA NA NA NA NA NA NA
# 5 BWA VA 1964 NA NA NA NA NA NA NA
# 6 BWA VA 1965 -3.52 NA NA NA -28.6 NA NA
# 7 BWA VA 1966 12.4 -3.52 NA NA -21.1 -28.6 NA
# 8 BWA VA 1967 8.29 12.4 NA NA 16.7 -21.1 NA
# 9 BWA VA 1968 10.2 8.29 NA NA -20 16.7 NA
# 10 BWA VA 1969 3.61 10.2 NA NA 185. -20 NA
# # ... with 5,017 more rows, and 37 more variables: L1.L10G1.MIN <dbl>, G1.MAN <dbl>,
# # L1.G1.MAN <dbl>, L10G1.MAN <dbl>, L1.L10G1.MAN <dbl>, G1.PU <dbl>, L1.G1.PU <dbl>,
# # L10G1.PU <dbl>, L1.L10G1.PU <dbl>, G1.CON <dbl>, L1.G1.CON <dbl>, L10G1.CON <dbl>,
# # L1.L10G1.CON <dbl>, G1.WRT <dbl>, L1.G1.WRT <dbl>, L10G1.WRT <dbl>, L1.L10G1.WRT <dbl>,
# # G1.TRA <dbl>, L1.G1.TRA <dbl>, L10G1.TRA <dbl>, L1.L10G1.TRA <dbl>, G1.FIRE <dbl>,
# # L1.G1.FIRE <dbl>, L10G1.FIRE <dbl>, L1.L10G1.FIRE <dbl>, G1.GOV <dbl>, L1.G1.GOV <dbl>,
# # L10G1.GOV <dbl>, L1.L10G1.GOV <dbl>, G1.OTH <dbl>, L1.G1.OTH <dbl>, L10G1.OTH <dbl>,
# # L1.L10G1.OTH <dbl>, G1.SUM <dbl>, L1.G1.SUM <dbl>, L10G1.SUM <dbl>, L1.L10G1.SUM <dbl>This section seeks to demonstrate that the functionality introduced in the preceding 2 sections indeed produces code that evaluates substantially faster than native dplyr.
To do this properly, the different components of a typical piped call (selecting / subsetting, ordering, grouping, and performing some computation) are bechmarked separately on 2 different data sizes.
All benchmarks are run on a Windows 8.1 laptop with a 2x 2.2 GHZ Intel i5 processor, 8GB DDR3 RAM and a Samsung 850 EVO SSD hard drive.
Bechmarks are run on the original GGDC10S data used throughout this vignette and a larger dataset with approx. 1 million observations, obtained by replicating and row-binding GGDC10S 200 times while maintaining unique groups.
# This shows the groups in GGDC10S
GRP(GGDC10S, ~ Variable + Country)
# collapse grouping object of length 5027 with 85 ordered groups
#
# Call: GRP.default(X = GGDC10S, by = ~Variable + Country), X is unordered
#
# Distribution of group sizes:
# Min. 1st Qu. Median Mean 3rd Qu. Max.
# 4.00 53.00 62.00 59.14 63.00 65.00
#
# Groups with sizes:
# EMP.ARG EMP.BOL EMP.BRA EMP.BWA EMP.CHL EMP.CHN
# 62 61 62 52 63 62
# ---
# VA.TWN VA.TZA VA.USA VA.VEN VA.ZAF VA.ZMB
# 63 52 65 63 52 52
# This replicates the data 200 times
data <- replicate(200, GGDC10S, simplify = FALSE)
# This function adds a number i to the country and variable columns of each dataset
uniquify <- function(x, i) ftransform(x, lapply(unclass(x)[c(1,4)], paste0, i))
# Making datasets unique and row-binding them
data <- unlist2d(Map(uniquify, data, as.list(1:200)), idcols = FALSE)
fdim(data)
# [1] 1005400 16
# This shows the groups in the replicated data
GRP(data, ~ Variable + Country)
# collapse grouping object of length 1005400 with 17000 ordered groups
#
# Call: GRP.default(X = data, by = ~Variable + Country), X is unordered
#
# Distribution of group sizes:
# Min. 1st Qu. Median Mean 3rd Qu. Max.
# 4.00 53.00 62.00 59.14 63.00 65.00
#
# Groups with sizes:
# EMP1.ARG1 EMP1.BOL1 EMP1.BRA1 EMP1.BWA1 EMP1.CHL1 EMP1.CHN1
# 62 61 62 52 63 62
# ---
# VA99.TWN99 VA99.TZA99 VA99.USA99 VA99.VEN99 VA99.ZAF99 VA99.ZMB99
# 63 52 65 63 52 52
gc()
# used (Mb) gc trigger (Mb) max used (Mb)
# Ncells 1924737 102.8 3670734 196.1 3670734 196.1
# Vcells 19870476 151.6 28315229 216.1 23044570 175.9## Selecting columns
# Small
microbenchmark(dplyr = select(GGDC10S, Country, Variable, AGR:SUM),
collapse = fselect(GGDC10S, Country, Variable, AGR:SUM))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 2798.415 2980.931 3159.70634 3127.3000 3306.245 3923.851 100 b
# collapse 10.710 17.404 26.15927 28.7835 34.808 56.673 100 a
# Large
microbenchmark(dplyr = select(data, Country, Variable, AGR:SUM),
collapse = fselect(data, Country, Variable, AGR:SUM))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 2771.641 2826.752 3038.0593 2936.976 3174.825 3767.664 100 b
# collapse 11.156 15.619 25.7219 29.006 34.361 56.228 100 a
## Subsetting columns
# Small
microbenchmark(dplyr = filter(GGDC10S, Variable == "VA"),
collapse = fsubset(GGDC10S, Variable == "VA"))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 1373.102 1637.505 1797.1978 1716.0440 1888.2945 2913.547 100 b
# collapse 157.972 209.513 266.3651 229.8175 317.0585 497.566 100 a
# Large
microbenchmark(dplyr = filter(data, Variable == "VA"),
collapse = fsubset(data, Variable == "VA"))
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 16.846720 17.146822 22.628601 17.511629 19.158280 171.76675 100 b
# collapse 7.695085 7.825166 9.181815 7.993847 8.666342 23.32132 100 a
## Ordering rows
# Small
microbenchmark(dplyr = arrange(GGDC10S, desc(Country), Variable, Year),
collapse = roworder(GGDC10S, -Country, Variable, Year))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 7784.781 8136.646 8736.6524 8491.4130 8959.9725 14771.671 100 b
# collapse 559.147 624.523 786.6669 697.4845 876.4295 2740.403 100 a
# Large
microbenchmark(dplyr = arrange(data, desc(Country), Variable, Year),
collapse = roworder(data, -Country, Variable, Year), times = 2)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 2601.7312 2601.7312 2632.6191 2632.6191 2663.5069 2663.5069 2 b
# collapse 344.6215 344.6215 346.6969 346.6969 348.7724 348.7724 2 a
## Grouping
# Small
microbenchmark(dplyr = group_by(GGDC10S, Country, Variable),
collapse = fgroup_by(GGDC10S, Country, Variable))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 2763.162 2898.8215 3088.2130 2994.988 3264.074 3832.817 100 b
# collapse 348.519 367.4845 402.0911 396.044 414.564 785.395 100 a
# Large
microbenchmark(dplyr = group_by(data, Country, Variable),
collapse = fgroup_by(data, Country, Variable), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 76.63803 77.36809 78.87136 78.39401 80.88586 81.68731 10 b
# collapse 68.31864 68.74615 72.99447 71.08538 71.57669 96.47059 10 a
## Computing a new column
# Small
microbenchmark(dplyr = mutate(GGDC10S, NEW = AGR+1),
collapse = ftransform(GGDC10S, NEW = AGR+1))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 2479.349 2586.226 2717.8860 2659.4095 2825.860 3310.261 100 b
# collapse 21.866 26.775 36.8424 33.2455 45.964 70.061 100 a
# Large
microbenchmark(dplyr = mutate(data, NEW = AGR+1),
collapse = ftransform(data, NEW = AGR+1))
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 6.479508 6.751273 8.913357 7.048919 7.416626 37.09250 100 b
# collapse 1.751520 3.723040 4.420390 4.000828 4.069104 30.74999 100 a
## All combined with pipes
# Small
microbenchmark(dplyr = filter(GGDC10S, Variable == "VA") %>%
select(Country, Year, AGR:SUM) %>%
arrange(desc(Country), Year) %>%
mutate(NEW = AGR+1) %>%
group_by(Country),
collapse = fsubset(GGDC10S, Variable == "VA", Country, Year, AGR:SUM) %>%
roworder(-Country, Year) %>%
ftransform(NEW = AGR+1) %>%
fgroup_by(Country))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 14700.718 15442.380 16238.182 16080.2905 16786.254 21672.882 100 b
# collapse 698.377 779.148 841.774 813.2855 898.742 1129.898 100 a
# Large
microbenchmark(dplyr = filter(data, Variable == "VA") %>%
select(Country, Year, AGR:SUM) %>%
arrange(desc(Country), Year) %>%
mutate(NEW = AGR+1) %>%
group_by(Country),
collapse = fsubset(data, Variable == "VA", Country, Year, AGR:SUM) %>%
roworder(-Country, Year) %>%
ftransform(NEW = AGR+1) %>%
fgroup_by(Country), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 29.952104 30.812915 35.031735 32.106808 32.820580 53.31448 10 b
# collapse 7.094883 8.230582 8.733189 8.909993 9.348877 10.03476 10 a
gc()
# used (Mb) gc trigger (Mb) max used (Mb)
# Ncells 1930366 103.1 3670734 196.1 3670734 196.1
# Vcells 21386676 163.2 44612936 340.4 44612936 340.4## Grouping the data
cgGGDC10S <- fgroup_by(GGDC10S, Variable, Country) %>% fselect(-Region, -Regioncode)
gGGDC10S <- group_by(GGDC10S, Variable, Country) %>% fselect(-Region, -Regioncode)
cgdata <- fgroup_by(data, Variable, Country) %>% fselect(-Region, -Regioncode)
gdata <- group_by(data, Variable, Country) %>% fselect(-Region, -Regioncode)
rm(data, GGDC10S)
gc()
# used (Mb) gc trigger (Mb) max used (Mb)
# Ncells 1947353 104.0 3670734 196.1 3670734 196.1
# Vcells 20487216 156.4 44612936 340.4 44612936 340.4
## Conversion of Grouping object: This time would be required extra in all hybrid calls
## i.e. when calling collapse functions on data grouped with dplyr::group_by
# Small
microbenchmark(GRP(gGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval
# GRP(gGGDC10S) 167.342 190.1015 204.8187 195.68 214.868 553.793 100
# Large
microbenchmark(GRP(gdata))
# Unit: milliseconds
# expr min lq mean median uq max neval
# GRP(gdata) 30.90529 33.20525 34.67392 34.65956 35.94654 46.42175 100
## Sum
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, sum, na.rm = TRUE),
collapse = fsum(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 8585.348 9057.2545 10219.6699 9580.255 10114.4135 19238.159 100 b
# collapse 238.296 276.6735 296.1522 294.746 316.6125 490.426 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, sum, na.rm = TRUE),
collapse = fsum(cgdata), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 579.45496 584.42393 613.59644 598.77769 624.97484 738.12035 10 b
# collapse 40.33181 40.54601 42.07606 42.07909 42.54966 47.16876 10 a
## Mean
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, mean.default, na.rm = TRUE),
collapse = fmean(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 11492.648 11748.793 13004.0464 12048.226 12505.1825 31512.631 100 b
# collapse 253.469 264.179 296.0227 307.688 312.5965 420.812 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, mean.default, na.rm = TRUE),
collapse = fmean(cgdata), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 1312.46756 1506.54399 1606.47536 1664.64309 1705.4015 1779.43256 10 b
# collapse 43.48589 43.93347 45.26195 45.17917 45.9871 49.31209 10 a
## Median
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, median, na.rm = TRUE),
collapse = fmedian(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 50473.67 52100.6860 57787.6456 54558.1685 59640.700 212904.930 100 b
# collapse 489.98 515.6385 568.5235 560.7095 598.641 849.655 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, median, na.rm = TRUE),
collapse = fmedian(cgdata), times = 2)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 9880.28083 9880.28083 10086.79484 10086.79484 10293.30884 10293.30884 2 b
# collapse 89.30252 89.30252 91.89656 91.89656 94.49059 94.49059 2 a
## Standard Deviation
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, sd, na.rm = TRUE),
collapse = fsd(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 23725.62 24611.419 26140.1147 25158.518 26656.124 35508.773 100 b
# collapse 422.15 453.164 492.8132 485.517 515.862 883.123 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, sd, na.rm = TRUE),
collapse = fsd(cgdata), times = 2)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 4304.44404 4304.44404 4380.66706 4380.66706 4456.89008 4456.89008 2 b
# collapse 81.43251 81.43251 82.12664 82.12664 82.82078 82.82078 2 a
## Maximum
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, max, na.rm = TRUE),
collapse = fmax(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 11240.964 11646.603 12413.9514 12253.7220 12704.2085 20804.039 100 b
# collapse 178.945 207.059 236.9262 241.1965 251.6835 548.438 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, max, na.rm = TRUE),
collapse = fmax(cgdata), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 1044.13785 1062.44157 1100.07827 1083.97366 1103.79975 1223.72999 10 b
# collapse 24.18391 24.85061 25.75836 25.68442 26.44282 28.02878 10 a
## First Value
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, first),
collapse = ffirst(cgGGDC10S, na.rm = FALSE))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 27466.06 28773.7895 30433.8506 29823.3625 31313.381 37054.127 100 b
# collapse 60.69 74.9695 104.9933 111.1155 126.065 213.306 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, first),
collapse = ffirst(cgdata, na.rm = FALSE), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 4878.13483 5045.077204 5289.121344 5196.811698 5375.949782 6093.325884 10 b
# collapse 4.44953 4.526284 4.917643 4.608171 4.914073 6.536181 10 a
## Number of Distinct Values
# Small
microbenchmark(dplyr = summarise_all(gGGDC10S, n_distinct, na.rm = TRUE),
collapse = fNdistinct(cgGGDC10S))
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 370.935750 378.919332 399.013004 385.050768 399.750147 542.351303 100 b
# collapse 1.276267 1.327139 1.500403 1.387606 1.523264 5.796303 100 a
# Large
microbenchmark(dplyr = summarise_all(gdata, n_distinct, na.rm = TRUE),
collapse = fNdistinct(cgdata), times = 5)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 74045.7033 74201.0000 74950.3911 74891.2018 75278.3091 76335.7411 5 b
# collapse 317.5298 319.3764 326.7462 325.0972 329.0956 342.6321 5 a
gc()
# used (Mb) gc trigger (Mb) max used (Mb)
# Ncells 1950123 104.2 3670734 196.1 3670734 196.1
# Vcells 20493548 156.4 44612936 340.4 44612936 340.4Below are some additional benchmarks for weighted aggregations and aggregations using the statistical mode, which cannot easily or efficiently be performed with dplyr.
## Weighted Mean
# Small
microbenchmark(fmean(cgGGDC10S, SUM))
# Unit: microseconds
# expr min lq mean median uq max neval
# fmean(cgGGDC10S, SUM) 285.152 301.4405 326.9209 325.315 334.686 453.387 100
# Large
microbenchmark(fmean(cgdata, SUM), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fmean(cgdata, SUM) 51.64685 52.21269 53.04709 52.45746 53.25825 55.89691 10
## Weighted Standard-Deviation
# Small
microbenchmark(fsd(cgGGDC10S, SUM))
# Unit: microseconds
# expr min lq mean median uq max neval
# fsd(cgGGDC10S, SUM) 433.752 441.115 468.5864 462.0895 476.592 673.387 100
# Large
microbenchmark(fsd(cgdata, SUM), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fsd(cgdata, SUM) 81.2134 82.15186 84.33088 83.91788 85.15689 89.66487 10
## Statistical Mode
# Small
microbenchmark(fmode(cgGGDC10S))
# Unit: milliseconds
# expr min lq mean median uq max neval
# fmode(cgGGDC10S) 1.587748 1.637504 1.679478 1.663387 1.697747 1.95367 100
# Large
microbenchmark(fmode(cgdata), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fmode(cgdata) 381.3137 391.2128 425.0835 404.3425 424.7772 549.6336 10
## Weighted Statistical Mode
# Small
microbenchmark(fmode(cgGGDC10S, SUM))
# Unit: milliseconds
# expr min lq mean median uq max neval
# fmode(cgGGDC10S, SUM) 1.793468 1.84791 1.935726 1.890302 2.007889 2.626164 100
# Large
microbenchmark(fmode(cgdata, SUM), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fmode(cgdata, SUM) 471.0955 478.6665 491.8299 481.7983 515.7568 517.8626 10
gc()
# used (Mb) gc trigger (Mb) max used (Mb)
# Ncells 1949570 104.2 3670734 196.1 3670734 196.1
# Vcells 20490142 156.4 44612936 340.4 44612936 340.4
## Replacing with group sum
# Small
microbenchmark(dplyr = mutate_all(gGGDC10S, sum, na.rm = TRUE),
collapse = fsum(cgGGDC10S, TRA = "replace_fill"))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 9422.508 9950.641 10949.9220 10214.373 10769.504 27304.075 100 b
# collapse 291.399 340.710 366.2888 363.692 391.359 553.793 100 a
# Large
microbenchmark(dplyr = mutate_all(gdata, sum, na.rm = TRUE),
collapse = fsum(cgdata, TRA = "replace_fill"), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 1152.99314 1200.45642 1273.494 1249.90417 1298.4063 1512.5563 10 b
# collapse 58.45658 79.04419 116.200 94.37434 113.0353 329.2473 10 a
## Dividing by group sum
# Small
microbenchmark(dplyr = mutate_all(gGGDC10S, function(x) x/sum(x, na.rm = TRUE)),
collapse = fsum(cgGGDC10S, TRA = "/"))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 9622.872 10204.11 11141.4289 10644.109 11165.1020 23135.235 100 b
# collapse 548.884 581.46 631.4444 619.615 652.6365 1326.693 100 a
# Large
microbenchmark(dplyr = mutate_all(gdata, function(x) x/sum(x, na.rm = TRUE)),
collapse = fsum(cgdata, TRA = "/"), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 1241.6062 1398.6664 1530.9812 1556.4277 1632.4495 1861.7946 10 b
# collapse 113.1424 121.9201 136.9692 133.0716 143.1489 196.7169 10 a
## Centering
# Small
microbenchmark(dplyr = mutate_all(gGGDC10S, function(x) x-mean.default(x, na.rm = TRUE)),
collapse = fwithin(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 12829.158 13595.3640 14916.2687 13954.146 14407.980 35788.124 100 b
# collapse 310.589 348.7425 379.9439 374.625 400.507 700.608 100 a
# Large
microbenchmark(dplyr = mutate_all(gdata, function(x) x-mean.default(x, na.rm = TRUE)),
collapse = fwithin(cgdata), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 2205.50124 2646.09750 2706.8577 2763.9730 2831.4105 2870.2822 10 b
# collapse 64.86068 90.56183 140.7654 112.0254 131.0632 374.7248 10 a
## Centering and Scaling (Standardizing)
# Small
microbenchmark(dplyr = mutate_all(gGGDC10S, function(x) (x-mean.default(x, na.rm = TRUE))/sd(x, na.rm = TRUE)),
collapse = fscale(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr 31800.906 32996.848 36621.5260 34130.0930 37528.0415 57398.085 100 b
# collapse 516.754 565.172 634.9027 596.1865 639.2495 1398.092 100 a
# Large
microbenchmark(dplyr = mutate_all(gdata, function(x) (x-mean.default(x, na.rm = TRUE))/sd(x, na.rm = TRUE)),
collapse = fscale(cgdata), times = 2)
# Unit: milliseconds
# expr min lq mean median uq max neval cld
# dplyr 6435.4363 6435.4363 6704.2027 6704.2027 6972.969 6972.969 2 b
# collapse 106.2894 106.2894 122.0602 122.0602 137.831 137.831 2 a
## Lag
# Small
microbenchmark(dplyr_unordered = mutate_all(gGGDC10S, dplyr::lag),
collapse_unordered = flag(cgGGDC10S),
dplyr_ordered = mutate_all(gGGDC10S, dplyr::lag, order_by = "Year"),
collapse_ordered = flag(cgGGDC10S, t = Year))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr_unordered 47350.386 48740.669 51035.0391 49742.4930 52168.069 68811.747 100 b
# collapse_unordered 375.294 442.231 468.3410 468.1135 490.426 582.799 100 a
# dplyr_ordered 128288.897 131983.154 136240.5114 134911.4270 139166.616 161151.871 100 c
# collapse_ordered 322.637 375.964 408.4056 394.4825 421.257 1137.038 100 a
# Large
microbenchmark(dplyr_unordered = mutate_all(gdata, dplyr::lag),
collapse_unordered = flag(cgdata),
dplyr_ordered = mutate_all(gdata, dplyr::lag, order_by = "Year"),
collapse_ordered = flag(cgdata, t = Year), times = 2)
# Unit: milliseconds
# expr min lq mean median uq max neval
# dplyr_unordered 8975.91694 8975.91694 9083.08228 9083.08228 9190.24763 9190.24763 2
# collapse_unordered 53.55188 53.55188 62.14437 62.14437 70.73686 70.73686 2
# dplyr_ordered 25813.19249 25813.19249 26739.83750 26739.83750 27666.48252 27666.48252 2
# collapse_ordered 96.89318 96.89318 106.13697 106.13697 115.38075 115.38075 2
# cld
# b
# a
# c
# a
## First-Difference (unordered)
# Small
microbenchmark(dplyr_unordered = mutate_all(gGGDC10S, function(x) x - dplyr::lag(x)),
collapse_unordered = fdiff(cgGGDC10S))
# Unit: microseconds
# expr min lq mean median uq max neval cld
# dplyr_unordered 62157.31 63476.4170 66662.2014 64777.896 68252.823 95926.613 100 b
# collapse_unordered 386.45 460.5275 513.5147 497.789 528.803 1019.229 100 a
# Large
microbenchmark(dplyr_unordered = mutate_all(gdata, function(x) x - dplyr::lag(x)),
collapse_unordered = fdiff(cgdata), times = 2)
# Unit: milliseconds
# expr min lq mean median uq max neval
# dplyr_unordered 12545.43710 12545.43710 13110.48386 13110.48386 13675.53062 13675.53062 2
# collapse_unordered 64.10741 64.10741 73.32665 73.32665 82.54589 82.54589 2
# cld
# b
# a
gc()
# used (Mb) gc trigger (Mb) max used (Mb)
# Ncells 1951949 104.3 3670735 196.1 3670735 196.1
# Vcells 21539057 164.4 65092104 496.7 65092101 496.7Below again some benchmarks for transformations not easily of efficiently performed with dplyr, such as centering on the overall mean, mean-preserving scaling, weighted scaling and centering, sequences of lags / leads, (iterated) panel-differences and growth rates.
# Centering on overall mean
microbenchmark(fwithin(cgdata, mean = "overall.mean"), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fwithin(cgdata, mean = "overall.mean") 65.08112 83.53835 92.52791 95.56872 103.4503 115.175 10
# Weighted Centering
microbenchmark(fwithin(cgdata, SUM), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fwithin(cgdata, SUM) 64.75179 65.81118 84.23177 81.5579 101.527 106.3509 10
microbenchmark(fwithin(cgdata, SUM, mean = "overall.mean"), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max
# fwithin(cgdata, SUM, mean = "overall.mean") 67.02944 67.83268 99.77674 87.0155 103.1862 233.4667
# neval
# 10
# Weighted Scaling and Standardizing
microbenchmark(fsd(cgdata, SUM, TRA = "/"), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fsd(cgdata, SUM, TRA = "/") 143.8683 156.2623 164.8902 167.915 170.7149 185.994 10
microbenchmark(fscale(cgdata, SUM), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fscale(cgdata, SUM) 97.76158 101.3057 117.8158 112.4665 135.0901 144.7045 10
# Sequence of lags and leads
microbenchmark(flag(cgdata, -1:1), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# flag(cgdata, -1:1) 72.85832 110.3707 200.5462 231.0337 250.2536 280.4177 10
# Iterated difference
microbenchmark(fdiff(cgdata, 1, 2), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fdiff(cgdata, 1, 2) 61.83958 65.19715 85.75602 86.90238 99.94373 110.68 10
# Growth Rate
microbenchmark(fgrowth(cgdata,1), times = 10)
# Unit: milliseconds
# expr min lq mean median uq max neval
# fgrowth(cgdata, 1) 64.35151 79.62878 90.42032 96.65288 101.8938 110.4622 10Timmer, M. P., de Vries, G. J., & de Vries, K. (2015). “Patterns of Structural Change in Developing Countries.” . In J. Weiss, & M. Tribe (Eds.), Routledge Handbook of Industry and Development. (pp. 65-83). Routledge.
Cochrane, D. & Orcutt, G. H. (1949). “Application of Least Squares Regression to Relationships Containing Auto-Correlated Error Terms”. Journal of the American Statistical Association. 44 (245): 32–61.
Prais, S. J. & Winsten, C. B. (1954). “Trend Estimators and Serial Correlation”. Cowles Commission Discussion Paper No. 383. Chicago.
Row-wise operations are not supported by TRA.↩︎