collapse for tidyverse Users

collapse is a C/C++ based package for data transformation and statistical computing in R that aims to enable greater performance and statistical complexity in data manipulation tasks and offers a stable, class-agnostic, and lightweight API. It is part of the core fastverse, a suite of lightweight packages with similar objectives.

The tidyverse set of packages provides a rich, expressive, and consistent syntax for data manipulation in R centering on the tibble object and tidy data principles (each observation is a row, each variable is a column).

collapse fully supports the tibble object and provides many tidyverse-like functions for data manipulation. It can thus be used to write tidyverse-like data manipulation code that, thanks to low-level vectorization of many statistical operations and optimized R code, typically runs much faster than native tidyverse code (in addition to being much more lightweight in dependencies).

Its aim is not to create a faster tidyverse, i.e., it does not implements all aspects of the rich tidyverse grammar or changes to it¹, and also takes inspiration from other leading data manipulation libraries to serve broad aims of performance, parsimony, complexity, and robustness in data manipulation for R.

Namespace and Global Options

collapse data manipulation functions familiar to tidyverse users include fselect, fgroup_by, fsummarise, fmutate, across, frename, and fcount. Other functions like fsubset, ftransform, and get_vars are inspired by base R, while again other functions like join, pivot, roworder, colorder, rowbind, etc. are inspired by other data manipulation libraries such as data.table and polars.

By virtue of the f- prefixes, the collapse namespace has no conflicts with the tidyverse, and these functions can easily be substituted in a tidyverse workflow.

R users willing to replace the tidyverse have the additional option to mask functions and eliminate the prefixes with set_collapse. For example

library(collapse)
set_collapse(mask = "manip") # version >= 2.0.0

makes available functions select, group_by, summarise, mutate, rename, count, subset, and transform in the collapse namespace and detaches and re-attaches the package, such that the following code is executed by collapse:

mtcars |>
  subset(mpg > 11) |>
  group_by(cyl, vs, am) |>
  summarise(across(c(mpg, carb, hp), mean), 
            qsec_wt = weighted.mean(qsec, wt))
#   cyl vs am      mpg     carb        hp  qsec_wt
# 1   4  0  1 26.00000 2.000000  91.00000 16.70000
# 2   4  1  0 22.90000 1.666667  84.66667 21.04028
# 3   4  1  1 28.37143 1.428571  80.57143 18.75509
# 4   6  0  1 20.56667 4.666667 131.66667 16.33306
# 5   6  1  0 19.12500 2.500000 115.25000 19.21275
# 6   8  0  0 15.98000 2.900000 191.00000 17.01239
# 7   8  0  1 15.40000 6.000000 299.50000 14.55297

Note that the correct documentation still needs to be called with prefixes, i.e., ?fsubset. See ?set_collapse for further options to the package, which also includes optimization options such as nthreads, na.rm, sort, and stable.algo. Note also that if you use collapse’s namespace masking, you can use fastverse::fastverse_conflicts() to check for namespace conflicts with other packages.

Using the Fast Statistical Functions

A key feature of collapse is that it not only provides functions for data manipulation, but also a full set of statistical functions and algorithms to speed up statistical calculations and perform more complex statistical operations (e.g. involving weights or time series data).

Notably among these, the Fast Statistical Functions is a consistent set of S3-generic statistical functions providing fully vectorized statistical operations in R.

Specifically, operations such as calculating the mean via the S3 generic fmean() function are vectorized across columns and groups and may also involve weights or transformations of the original data:

fmean(mtcars$mpg)     # Vector
# [1] 20.09062
fmean(EuStockMarkets) # Matrix
#      DAX      SMI      CAC     FTSE 
# 2530.657 3376.224 2227.828 3565.643
fmean(mtcars)         # Data Frame
#        mpg        cyl       disp         hp       drat         wt       qsec         vs         am 
#  20.090625   6.187500 230.721875 146.687500   3.596563   3.217250  17.848750   0.437500   0.406250 
#       gear       carb 
#   3.687500   2.812500

fmean(mtcars$mpg, w = mtcars$wt)  # Weighted mean
# [1] 18.54993
fmean(mtcars$mpg, g = mtcars$cyl) # Grouped mean
#        4        6        8 
# 26.66364 19.74286 15.10000
fmean(mtcars$mpg, g = mtcars$cyl, w = mtcars$wt)   # Weighted group mean
#        4        6        8 
# 25.93504 19.64578 14.80643
fmean(mtcars[5:10], g = mtcars$cyl, w = mtcars$wt) # Of data frame
#       drat       wt     qsec        vs        am     gear
# 4 4.031264 2.414750 19.38044 0.9148868 0.6498031 4.047250
# 6 3.569170 3.152060 18.12198 0.6212191 0.3787809 3.821036
# 8 3.205658 4.133116 16.88529 0.0000000 0.1203808 3.240762
fmean(mtcars$mpg, g = mtcars$cyl, w = mtcars$wt, TRA = "fill") # Replace data by weighted group mean
#  [1] 19.64578 19.64578 25.93504 19.64578 14.80643 19.64578 14.80643 25.93504 25.93504 19.64578
# [11] 19.64578 14.80643 14.80643 14.80643 14.80643 14.80643 14.80643 25.93504 25.93504 25.93504
# [21] 25.93504 14.80643 14.80643 14.80643 14.80643 25.93504 25.93504 25.93504 14.80643 19.64578
# [31] 14.80643 25.93504
# etc...

The data manipulation functions of collapse are integrated with these Fast Statistical Functions to enable vectorized statistical operations. For example, the following code

mtcars |>
  subset(mpg > 11) |>
  group_by(cyl, vs, am) |>
  summarise(across(c(mpg, carb, hp), fmean), 
            qsec_wt = fmean(qsec, wt))
#   cyl vs am      mpg     carb        hp  qsec_wt
# 1   4  0  1 26.00000 2.000000  91.00000 16.70000
# 2   4  1  0 22.90000 1.666667  84.66667 21.04028
# 3   4  1  1 28.37143 1.428571  80.57143 18.75509
# 4   6  0  1 20.56667 4.666667 131.66667 16.33306
# 5   6  1  0 19.12500 2.500000 115.25000 19.21275
# 6   8  0  0 15.98000 2.900000 191.00000 17.01239
# 7   8  0  1 15.40000 6.000000 299.50000 14.55297

gives exactly the same result as above, but the execution is much faster (especially on larger data), because with Fast Statistical Functions, the data does not need to be split by groups, and there is no need to call lapply() inside the across() statement: fmean.data.frame() is simply applied to a subset of the data containing columns mpg, carb and hp.

The Fast Statistical Functions also have a method for grouped data, so if we did not want to calculate the weighted mean of qsec, the code would simplify as follows:

mtcars |>
  subset(mpg > 11) |>
  group_by(cyl, vs, am) |>
  select(mpg, carb, hp) |> 
  fmean()
#   cyl vs am      mpg     carb        hp
# 1   4  0  1 26.00000 2.000000  91.00000
# 2   4  1  0 22.90000 1.666667  84.66667
# 3   4  1  1 28.37143 1.428571  80.57143
# 4   6  0  1 20.56667 4.666667 131.66667
# 5   6  1  0 19.12500 2.500000 115.25000
# 6   8  0  0 15.98000 2.900000 191.00000
# 7   8  0  1 15.40000 6.000000 299.50000

Note that all functions in collapse, including the Fast Statistical Functions, have the default na.rm = TRUE, i.e., missing values are skipped in calculations. This can be changed using set_collapse(na.rm = FALSE) to give behavior more consistent with base R.

Another thing to be aware of when using Fast Statistical Functions inside data manipulation functions is that they toggle vectorized execution wherever they are used. E.g.

mtcars |> group_by(cyl) |> summarise(mpg = fmean(mpg) + min(qsec)) # Vectorized
#   cyl      mpg
# 1   4 41.16364
# 2   6 34.24286
# 3   8 29.60000

calculates a grouped mean of mpg but adds the overall minimum of qsec to the result, whereas

mtcars |> group_by(cyl) |> summarise(mpg = fmean(mpg) + fmin(qsec)) # Vectorized
#   cyl      mpg
# 1   4 43.36364
# 2   6 35.24286
# 3   8 29.60000
mtcars |> group_by(cyl) |> summarise(mpg = mean(mpg) + min(qsec))   # Not vectorized
#   cyl      mpg
# 1   4 43.36364
# 2   6 35.24286
# 3   8 29.60000

both give the mean + the minimum within each group, but calculated in different ways: the former is equivalent to fmean(mpg, g = cyl) + fmin(qsec, g = cyl), whereas the latter is equal to sapply(gsplit(mpg, cyl), function(x) mean(x) + min(x)).

See ?fsummarise and ?fmutate for more detailed examples. This eager vectorization approach is intentional as it allows users to vectorize complex expressions and fall back to base R if this is not desired. This blog post by Andrew Ghazi provides an excellent example of computing a p-value test statistic by groups.

To take full advantage of collapse, it is highly recommended to use the Fast Statistical Functions as much as possible. You can also set set_collapse(mask = "all") to replace statistical functions in base R like sum and mean with the collapse versions (toggling vectorized execution in all cases), but this may affect other parts of your code².

Writing Efficient Code

It is also performance-critical to correctly sequence operations and limit excess computations. tidyverse code is often inefficient simply because the tidyverse allows you to do everything. For example, mtcars |> group_by(cyl) |> filter(mpg > 13) |> arrange(mpg) is permissible but inefficient code as it filters and reorders grouped data, requiring modifications to both the data frame and the attached grouping object. collapse does not allow calls to fsubset() on grouped data, and messages about it in roworder(), encouraging you to write more efficient code.

The above example can also be optimized because we are subsetting the whole frame and then doing computations on a subset of columns. It would be more efficient to select all required columns during the subset operation:

mtcars |>
  subset(mpg > 11, cyl, vs, am, mpg, carb, hp, qsec, wt) |>
  group_by(cyl, vs, am) |>
  summarise(across(c(mpg, carb, hp), fmean), 
            qsec_wt = fmean(qsec, wt))
#   cyl vs am      mpg     carb        hp  qsec_wt
# 1   4  0  1 26.00000 2.000000  91.00000 16.70000
# 2   4  1  0 22.90000 1.666667  84.66667 21.04028
# 3   4  1  1 28.37143 1.428571  80.57143 18.75509
# 4   6  0  1 20.56667 4.666667 131.66667 16.33306
# 5   6  1  0 19.12500 2.500000 115.25000 19.21275
# 6   8  0  0 15.98000 2.900000 191.00000 17.01239
# 7   8  0  1 15.40000 6.000000 299.50000 14.55297

Without the weighted mean of qsec, this would simplify to

mtcars |>
  subset(mpg > 11, cyl, vs, am, mpg, carb, hp) |>
  group_by(cyl, vs, am) |> 
  fmean()
#   cyl vs am      mpg     carb        hp
# 1   4  0  1 26.00000 2.000000  91.00000
# 2   4  1  0 22.90000 1.666667  84.66667
# 3   4  1  1 28.37143 1.428571  80.57143
# 4   6  0  1 20.56667 4.666667 131.66667
# 5   6  1  0 19.12500 2.500000 115.25000
# 6   8  0  0 15.98000 2.900000 191.00000
# 7   8  0  1 15.40000 6.000000 299.50000

Finally, we could set the following options to toggle unsorted grouping, no missing value skipping, and multithreading across the three columns for more efficient execution.

mtcars |>
  subset(mpg > 11, cyl, vs, am, mpg, carb, hp) |>
  group_by(cyl, vs, am, sort = FALSE) |> 
  fmean(nthreads = 3, na.rm = FALSE)
#   cyl vs am      mpg     carb        hp
# 1   6  0  1 20.56667 4.666667 131.66667
# 2   4  1  1 28.37143 1.428571  80.57143
# 3   6  1  0 19.12500 2.500000 115.25000
# 4   8  0  0 15.98000 2.900000 191.00000
# 5   4  1  0 22.90000 1.666667  84.66667
# 6   4  0  1 26.00000 2.000000  91.00000
# 7   8  0  1 15.40000 6.000000 299.50000

Setting these options globally using set_collapse(sort = FALSE, nthreads = 3, na.rm = FALSE) avoids the need to set them repeatedly.

Using Internal Grouping

Another key to writing efficient code with collapse is to avoid fgroup_by() where possible, especially for mutate operations. collapse does not implement .by arguments to manipulation functions like dplyr, but instead allows ad-hoc grouped transformations through its statistical functions. For example, the easiest and fastest way to computed the median of mpg by cyl, vs, and am is

mtcars |>
  mutate(mpg_median = fmedian(mpg, list(cyl, vs, am), TRA = "fill")) |> 
  head(3)
#                mpg cyl disp  hp drat    wt  qsec vs am gear carb mpg_median
# Mazda RX4     21.0   6  160 110 3.90 2.620 16.46  0  1    4    4       21.0
# Mazda RX4 Wag 21.0   6  160 110 3.90 2.875 17.02  0  1    4    4       21.0
# Datsun 710    22.8   4  108  93 3.85 2.320 18.61  1  1    4    1       30.4

For the common case of averaging and centering data, collapse also provides functions fbetween() for averaging and fwithin() for centering, i.e., fbetween(mpg, list(cyl, vs, am)) is the same as fmean(mpg, list(cyl, vs, am), TRA = "fill"). There is also fscale() for (grouped) scaling and centering.

This also applies to multiple columns, where we can use fmutate(across(...)) or ftransformv(), i.e.

mtcars |>
  mutate(across(c(mpg, disp, qsec), fmedian, list(cyl, vs, am), TRA = "fill")) |> 
  head(2)
#               mpg cyl disp  hp drat    wt  qsec vs am gear carb
# Mazda RX4      21   6  160 110  3.9 2.620 16.46  0  1    4    4
# Mazda RX4 Wag  21   6  160 110  3.9 2.875 16.46  0  1    4    4

# Or 
mtcars |>
  transformv(c(mpg, disp, qsec), fmedian, list(cyl, vs, am), TRA = "fill") |> 
  head(2)
#               mpg cyl disp  hp drat    wt  qsec vs am gear carb
# Mazda RX4      21   6  160 110  3.9 2.620 16.46  0  1    4    4
# Mazda RX4 Wag  21   6  160 110  3.9 2.875 16.46  0  1    4    4

Of course, if we want to apply different functions using the same grouping, fgroup_by() is sensible, but for mutate operations it also has the argument return.groups = FALSE, which avoids materializing the unique grouping columns, saving some memory.

mtcars |>
  group_by(cyl, vs, am, return.groups = FALSE) |> 
  mutate(mpg_median = fmedian(mpg), 
         mpg_mean = fmean(mpg), # Or fbetween(mpg)
         mpg_demean = fwithin(mpg), # Or fmean(mpg, TRA = "-")
         mpg_scale = fscale(mpg), 
         .keep = "used") |>
  ungroup() |>
  head(3)
#                mpg cyl vs am mpg_median mpg_mean mpg_demean  mpg_scale
# Mazda RX4     21.0   6  0  1       21.0 20.56667  0.4333333  0.5773503
# Mazda RX4 Wag 21.0   6  0  1       21.0 20.56667  0.4333333  0.5773503
# Datsun 710    22.8   4  1  1       30.4 28.37143 -5.5714286 -1.1710339

The TRA argument supports a whole array of operations, see ?TRA. For example fsum(mtcars, TRA = "/") turns the column vectors into proportions. As an application of this, consider a generated dataset of sector-level exports.

# c = country, s = sector, y = year, v = value
exports <- expand.grid(c = paste0("c", 1:8), s = paste0("s", 1:8), y = 1:15) |>
           mutate(v = round(abs(rnorm(length(c), mean = 5)), 2)) |>
           subset(-sample.int(length(v), 360)) # Making it unbalanced and irregular
head(exports)
#    c  s y    v
# 1 c2 s1 1 5.55
# 2 c3 s1 1 4.33
# 3 c4 s1 1 5.21
# 4 c5 s1 1 5.31
# 5 c6 s1 1 6.17
# 6 c7 s1 1 5.62
nrow(exports)
# [1] 600

It is very easy then to compute Balassa’s (1965) Revealed Comparative Advantage (RCA) index, which is the share of a sector in country exports divided by the share of the sector in world exports. An index above 1 indicates that a RCA of country c in sector s.

# Computing Balassa's (1965) RCA index: fast and memory efficient
# settfm() modifies exports and assigns it back to the global environment
settfm(exports, RCA = fsum(v, list(c, y), TRA = "/") %/=% fsum(v, list(s, y), TRA = "/"))

Note that this involved a single expression with two different grouped operations, which is only possible by incorporating grouping into statistical functions themselves. Let’s summarise this dataset using pivot() to aggregate the RCA index across years. Here "mean" calls a highly efficient internal mean function.

pivot(exports, ids = "c", values = "RCA", names = "s", 
      how = "wider", FUN = "mean", sort = TRUE)
#    c        s1        s2        s3       s4       s5        s6        s7       s8
# 1 c1 0.9327521 0.9087815 0.9434970 1.105864 1.158613 0.9579166 1.1094150 1.218718
# 2 c2 1.4989832 1.0502050 0.8113781 1.024990 1.103707 1.1494829 1.0681358 1.021685
# 3 c3 1.0403483 0.9580809 0.8358023 1.024633 1.192487 0.9333733 1.0719161 1.010648
# 4 c4 0.9771630 1.0265800 0.9293951 1.007469 1.052942 0.9285248 1.4031524 1.027218
# 5 c5 0.9807908 1.1023470 0.8480027 1.080013 1.072168 0.9704144 1.1817784 1.099050
# 6 c6 0.9819940 1.1434701 0.9122508 1.164649 1.193275 0.9322847 0.9929571 1.177062
# 7 c7 1.1542193 1.1939893 0.7462051 1.109936 1.438044 1.0482547 1.5907867 1.055214
# 8 c8 1.4220817 1.2235288 0.7090515 1.189408 1.119605 1.3108897 1.3264848 1.279526

We may also wish to investigate the growth rate of RCA. This can be done using fgrowth(). Since the panel is irregular, i.e., not every sector is observed in every year, it is critical to also supply the time variable.

exports |> 
  mutate(RCA_growth = fgrowth(RCA, g = list(c, s), t = y)) |> 
  pivot(ids = "c", values = "RCA_growth", names = "s", 
        how = "wider", FUN = fmedian, sort = TRUE)
#    c         s1        s2         s3          s4          s5         s6         s7          s8
# 1 c1         NA  29.87093  56.837880   0.3513705  11.9750588   6.356499   5.186966   3.4725766
# 2 c2 -19.092254 -10.72516  50.412427   8.7380006 -25.7119274 -17.958011 -36.853824 -30.5827161
# 3 c3  -3.904880 -29.72276   4.338254   4.2112875  13.8705938 -27.368230  -5.214542 -10.4867005
# 4 c4   0.639523  19.74757  -9.602120   9.7104112  42.0912878  17.583594 -27.915967 -18.1145784
# 5 c5   8.184523  18.93554  -5.333235   1.5243547  -0.3306585   8.682935 -15.678443  18.3991608
# 6 c6  12.606978  67.07558  19.270685  43.8243108 -25.0283737 -21.785028 -10.059702   0.7774246
# 7 c7  24.400344  48.56792  27.552571 -16.9311897  -6.6046775 -28.627885 -12.092345  24.5298895
# 8 c8 158.342022  17.99249 -61.857965  36.3372079   0.2085139  -2.178978 -18.666774 -40.5714063

Lastly, since the panel is unbalanced, we may wish to create an RCA index for only the last year, but balance the dataset a bit more by taking the last available trade within the last three years. This can be done using a single subset call

# Taking the latest observation within the last 3 years
exports_latest <- subset(exports, y > 12 & y == fmax(y, list(c, s), "fill"), -y)
# How many sectors do we observe for each country in the last 3 years?
with(exports_latest, fndistinct(s, c))
# c1 c2 c3 c4 c5 c6 c7 c8 
#  8  8  7  7  8  8  6  8

We can then compute the RCA index on this data

exports_latest |>
    mutate(RCA = fsum(v, c, TRA = "/") %/=% fsum(v, s, TRA = "/")) |>
    pivot("c", "RCA", "s", how = "wider", sort = TRUE)
#    c        s1        s2        s3        s4        s5        s6        s7        s8
# 1 c1 0.9038055 0.9073996 0.7608879 0.5752643 0.8558140 0.6619450 0.8820296 0.9617336
# 2 c2 1.1725178 1.1771805 0.9871092 0.7462973 1.1102578 0.8587493 1.1442677 1.2476687
# 3 c3 1.2072861 1.2120870 1.0163796        NA 1.1431799 0.8842135 1.1781982 1.2846653
# 4 c4 1.2438173 1.2487635 1.0471341 0.7916788 1.1777713        NA 1.2138493 1.3235380
# 5 c5 1.0014055 1.0053877 0.8430546 0.6373858 0.9482314 0.7334270 0.9772781 1.0655891
# 6 c6 1.0234618 1.0275317 0.8616232 0.6514245 0.9691166 0.7495810 0.9988030 1.0890591
# 7 c7 1.3447625 1.3501101        NA        NA 1.2733564 0.9849009 1.3123624 1.4309531
# 8 c8 1.1226366 1.1271008 0.9451155 0.7145483 1.0630252 0.8222164 1.0955882 1.1945903

To summarise, collapse provides many options for ad-hoc or limited grouping, which are faster than a full fgroup_by(), and also syntactically efficient. Further efficiency gains are possible using operations by reference, e.g., %/=% instead of / to avoid an intermediate copy. It is also possible to transform by reference using fast statistical functions by passing the set = TRUE argument, e.g., with(mtcars, fmean(mpg, cyl, TRA = "fill", set = TRUE)) replaces mpg by its group-averaged version (the transformed vector is returned invisibly).

Conclusion

collapse enhances R both statistically and computationally and is a good option for tidyverse users searching for more efficient and lightweight solutions to data manipulation and statistical computing problems in R. For more information, I recommend starting with the short vignette on Documentation Resources.

R users willing to write efficient/lightweight code and completely replace the tidyverse in their workflow are also encouraged to closely examine the fastverse suite of packages. collapse alone may not always suffice, but 99% of tidyverse code can be replaced with an efficient and lightweight fastverse solution.

Sebastian Krantz

2024-11-03

Namespace and Global Options

Using the Fast Statistical Functions

Writing Efficient Code

Using Internal Grouping

Conclusion