mcaceresb / stata-gtools Goto Github PK

Overview | Installation | Examples | Remarks | FAQs | Benchmarks | Compiling

Faster Stata for big data. This packages uses C plugins and hashes to provide a massive speed improvements to common Stata commands, including: reshape, collapse, xtile, tabstat, isid, egen, pctile, winsor, contract, levelsof, duplicates, unique/distinct, and more.

This package provides a fast implementation of various Stata commands using hashes and C plugins. The syntax and purpose is largely analogous to their Stata counterparts; for example, you can replace collapse with gcollapse, reshape with greshape, and so on. For a comprehensive list of differences (including some extra features!) see the remarks below; for details and examples see the official project page.

Quickstart

ssc install gtools
gtools, upgrade

Some quick benchmarks:

NOTE: Stata 17 introduced massive speed improvements to sort and collapse. In the MP version, in particular with many cores available, the native collapse can be up to twice as fast. (YMMV; overall native collapses could still be slower in some use cases.) gcollapse remains faster in SE and older Stata versions.

Gtools commands with a Stata equivalent

(+) The upper end of the speed improvements are for quantiles (e.g. median, iqr, p90) and few groups. Weights have not been benchmarked.

(.) Only gegen group was benchmarked rigorously.

(-) Benchmarks computed 10 quantiles. When computing a large number of quantiles (e.g. thousands) pctile and xtile are prohibitively slow due to the way they are written; in that case gquantiles is hundreds or thousands of times faster, but this is an edge case.

Extra commands

(-) fastxtile from egenmisc and astile were benchmarked against gquantiles, xtile (fasterxtile) using by().

(+) While similar to the user command 'groups' with the 'select' option, gtoplevelsof does not really have an equivalent. It is several dozen times faster than 'groups, select', but that command was not written with the goal of gleaning the most common levels of a varlist. Rather, it has a plethora of features and that one is somewhat incidental. As such, the benchmark is not equivalent and gtoplevelsof does not attempt to implement the features of 'groups'

(.) Other than the dated 'hdfe' command, I do not know of a stata command that residualizes variables from a set of fixed effects. The 'hdfe' command, as far as I can tell, morphed into the 'reghdfe' package; the latter, however, is a fully-functioning regression command, while 'gstats hdfe' only residualizes a set of variables.

Regression models

WARNING: Regression models are in beta and are only intended as utilities to compute coefficients and standard errors. I do not recommend their use in production; various post-estimation commands and statistics are not availabe. (See gstats hdfe for residualizing variables net of fixed effects.)

All commands allow the user to optionally add:

• absorb() for high-dimensional fixed effects absorptions.
• cluster() for clustering (multiple covariates assume clusters are nested).
• by() for regressions by group.
• weights for weighted versions. Unlike other weights, fweights are assumed to refer to the number of observations.

Linear regression is computed via OLS (or WLS), IV regression is computed via two-stage least squares (2SLS), and GLM (poisson or logit) regression is computed via iteratively reweighted least squares (IRLS). See the TODO section for planned features, or the Missing Features section in the documentation for what is missing before the first non-beta release.

Extra features

Several commands offer additional features on top of the massive speedup. See the remarks section below for an overview; for details and examples, see each command's help page:

• gcollapse
• greshape
• gquantiles
• gstats sum/tab
• gstats transform/range/moving
• glevelsof
• gtoplevelsof
• gegen
• gdistinct
• gregress
• givregress
• gglm (poisson and logit)

In addition, several commands take gsort-style input, that is

[+|-]varname [[+|-]varname ...]

This does not affect the results in most cases, just the sort order. Commands that take this type of input include:

• gcollapse
• gcontract
• gegen
• glevelsof
• gtop (gtoplevelsof)

Ftools

The commands here are also faster than the commands provided by ftools; further, gtools commands take a mix of string and numeric variables, which is a limitation of ftools. (Note I could not get several parts of ftools working on the Linux server where I have access to Stata/MP; hence the IC benchmarks.)

(+) Only egen group was benchmarked rigorously.

Limitations

• strL variables only partially supported on Stata 14 and above; gcollapse, gcontract, and greshape do not support strL variabes.

• Due to a Stata bug, gtools cannot support more than 2^31-1 (2.1 billion) observations. See this issue

• Due to limitations in the Stata Plugin Interface, gtools can only handle as many variables as the largest matsize in the user's Stata version. For MP this is more than 10,000 variables but in IC this is only 800. See this issue.

• Gtools uses compiled C code to achieve it's massive increases in speed. This has two side-effects users might notice: First, it is sometimes not possible to break the program's execution. While this is already true for at least some parts of most Stata commands, there are fewer opportunities to break Gtools commands relative to their Stata counterparts.

  Second, the Stata GUI might appear frozen when running Gtools commands. If the system then runs out of RAM (memory), it could look like Stata has crashed (it may show a "(Not Responding)" message on Windows or it may darken on *nix systems). However, the program has not crashed; it is merely trying to swap memory. To check this is the case, the user can monitor disk activity or monitor their system's pagefile or swap space directly.

• The OSX version of gtools was implemented with invaluable help from @fbelotti in issue 11.

• Gtools was largely inspired by Sergio Correia's (@sergiocorreia) excellent ftools package. Further, several improvements and bug fixes have come from to @sergiocorreia's helpful comments.

• With the exception of greshape, every gtools command has been written almost entirely from scratch (and even greshape is mostly new code). However, gtools commands typically mimic the functionality of existing Stata commands, including community-contributed programs, meaning many of the ideas and options are based on them (see the respective help files for details). gtools commands based on community-contributed programs include:

  • gstats winsor, based on winsor2 by Lian (Arlion) Yujun

  • gunique, based on unique by Michael Hills and Tony Brady.

  • gdistinct, based on distinct by Gary Longton and Nicholas J. Cox.

I only have access to Stata 13.1, so I impose that to be the minimum. You can install gtools from Stata via SSC:

ssc install gtools
gtools, upgrade

By default this syncs to the master branch, which is stable. To install the latest version directly, type:

local github "https://raw.githubusercontent.com"
net install gtools, from(`github'/mcaceresb/stata-gtools/master/build/)

The syntax is generally analogous to the standard commands (see the corresponding help files for full syntax and options):

sysuse auto, clear

* gstats {hdfe|residualize} varlist [if] [in] [weight], [absorb(varlist) options]
gstats hdfe hdfe_price = price, absorb(foreign rep78)
gstats residualize price mpg, absorb(foreign rep78) prefix(res_)

* gstats {sum|tab} varlist [if] [in] [weight], [by(varlist) options]
gstats sum price [pw = gear_ratio / 4]
gstats tab price mpg, by(foreign) matasave

* gquantiles [newvarname =] exp [if] [in] [weight], {_pctile|xtile|pctile} [options]
gquantiles 2 * price, _pctile nq(10)
gquantiles p10 = 2 * price, pctile nq(10)
gquantiles x10 = 2 * price, xtile nq(10) by(rep78)
fasterxtile xx = log(price) [w = weight], cutpoints(p10) by(foreign)

* gstats winsor varlist [if] [in] [weight], [by(varlist) cuts(# #) options]
gstats winsor price gear_ratio mpg, cuts(5 95) s(_w1)
gstats winsor price gear_ratio mpg, cuts(5 95) by(foreign) s(_w2)
drop *_w?

* hashsort varlist, [options]
hashsort -make
hashsort foreign -rep78, benchmark verbose mlast

* gegen target  = stat(source) [if] [in] [weight], by(varlist) [options]
gegen tag   = tag(foreign)
gegen group = tag(-price make)
gegen p2_5  = pctile(price) [w = weight], by(foreign) p(2.5)

* gisid varlist [if] [in], [options]
gisid make, missok
gisid price in 1 / 2

* gduplicates varlist [if] [in], [options gtools(gtools_options)]
gduplicates report foreign
gduplicates report rep78 if foreign, gtools(bench(3))

* glevelsof varlist [if] [in], [options]
glevelsof rep78, local(levels) sep(" | ")
glevelsof foreign mpg if price < 4000, loc(lvl) sep(" | ") colsep(", ")
glevelsof foreign mpg in 10 / 70, gen(uniq_) nolocal

* gtop varlist [if] [in] [weight], [options]
* gtoplevelsof varlist [if] [in] [weight], [options]
gtoplevelsof foreign rep78
gtop foreign rep78 [w = weight], ntop(5) missrow groupmiss pctfmt(%6.4g) colmax(3)

* gregress depvar indepvars [if] [in] [weight], [by(varlist) options]
gregress price mpg rep78, mata(coefs) prefix(b(_b_) se(_se_))
gregress price mpg [fw = rep78], by(foreign) absorb(rep78 headroom) cluster(rep78)

* givregress depvar (endog = instruments) exog [if] [in] [weight], [by(varlist) options]
givregress price (mpg = gear_ratio) rep78, mata(coefs) prefix(b(_b_) se(_se_)) replace
givregress price (mpg = gear_ratio) [fw = rep78], by(foreign) absorb(rep78 headroom) cluster(rep78)

* gglm depvar indepvars [if] [in] [weight], family(...) [by(varlist) options]
gglm price mpg rep78, family(poisson) mata(coefs) prefix(b(_b_) se(_se_)) replace
gglm price mpg [fw = trunk], family(poisson) by(foreign) absorb(rep78 headroom) cluster(rep78)

gglm foreign price rep78 [fw = trunk], family(binomial) absorb(headroom) mata(coefs)
gglm foreign price if rep78 > 2, family(binomial) by(rep78) prefix(b(_b_) se(_se_)) replace

* gcollapse (stat) out = src [(stat) out = src ...] [if] [if] [weight], by(varlist) [options]
gen h1 = headroom
gen h2 = headroom
local lbl labelformat(#stat:pretty# #sourcelabel#)

gcollapse (mean) mean = price (median) p50 = gear_ratio, by(make) merge v `lbl'
disp "`:var label mean', `:var label p50'"
gcollapse (iqr) irq? = h? (nunique) turn (p97.5) mpg, by(foreign rep78) bench(2) wild

* gcontract varlist [if] [if] [fweight], [options]
gcontract foreign [fw = turn], freq(f) percent(p)

* greshape wide varlist,    i(i) j(j) [options]
* greshape long prefixlist, i(i) [j(j) string options]
*
* greshape spread varlist, j(j) [options]
* greshape gather varlist, j(j) value(value) [options]

gen j = _n
greshape wide f p, i(foreign) j(j)
greshape long f p, i(foreign) j(j)

greshape spread f p, j(j)
greshape gather f? p?, j(j) value(fp)

* gstats transform (stat) out = src [(stat) out = src ...] [if] [if] [weight], by(varlist) [options]
* gstats range  (stat) out = src [...] [if] [if] [weight], by(varlist) [options]
* gstats moving (stat) out = src [...] [if] [if] [weight], by(varlist) [options]

sysuse auto, clear
gstats transform (normalize) price (demean) price (range mean -sd sd) price, auto
gstats range  (mean) mean_r = price (sd) sd_r = price, interval(-10 10 mpg)
gstats moving (mean) mean_m = price (sd) sd_m = price, by(foreign) window(-5 5)

See the FAQs or the respective documentation for a list of supported gcollapse and gegen functions.

Functions available with gegen, gcollapse, gstats tab

gcollapse supports every collapse function, including their weighted versions. In addition, weights can be selectively applied via rawstat(), and several additional statistics are allowed, including nunique, select#, and so on.

gegen technically does not support all of egen, but whenever a function that is not supported is requested, gegen hashes the data and calls egen grouping by the hash, which is often faster (gegen only supports weights for internal functions, since egen does not normally allow weights).

Hence both should be able to replicate all of the functionality of their Stata counterparts. Last, gstats tab allows every statistic allowed by tabstat as well as any statistic allowed by gcollapse; the syntax for the statistics specified via statistics() is the same as in tabstat.

The following are implemented internally in C:

(+) indicates the function has the same or a very similar name to a function in the "egenmore" packge, but the function was independently implemented and is hence analogous to its gcollapse counterpart, not necessarily the function in egenmore.

The percentile syntax mimics that of collapse and egen, with the addition that quantiles are also supported. That is,

gcollapse (p#) target = var [target = var ...] , by(varlist)
gegen target = pctile(var), by(varlist) p(#)

where # is a "percentile" with arbitrary decimal places (e.g. 2.5 or 97.5). gtools also supports selecting the #th smallest or largest value:

gcollapse (select#) target = var [(select-#) target = var ...] , by(varlist)
gegen target = select(var), by(varlist) n(#)
gegen target = select(var), by(varlist) n(-#)

In addition, the following are allowed in gegen as wrappers to other gtools functions (stat is any stat available to gcollapse, except percent, nunique):

Last, when gegen calls a function that is not implemented internally by gtools, it will hash the by variables and call egen with by set to an id based on the hash. That is, if fcn is not one of the functions above,

gegen outvar = fcn(varlist) [if] [in], by(byvars)

would be the same as

hashsort byvars, group(id) sortgroup
egen outvar = fcn(varlist) [if] [in], by(id)

but preserving the original sort order. In case an egen option might conflict with a gtools option, the user can pass gtools_capture(fcn_options) to gegen.

Differences and Extras

Differences from collapse

• String variables are not allowed for first, last, min, max, etc. (see issue 25)
• New functions: nunique, nmissing, cv, variance, select#, select-#, range, gini
• rawstat allows selectively applying weights.
• rawselect ignores weights for select (analogously to rawsum).
• Option wild allows bulk-rename. E.g. gcollapse mean_x* = x*, wild
• gcollapse (nansum) and gcollapse (rawnansum) outputs a missing value for sums if all inputs are missing (instead of 0).
• gcollapse, merge merges the collapsed data set back into memory. This is much faster than collapsing a dataset, saving, and merging after. However, Stata's merge ..., update functionality is not implemented, only replace. (If the targets exist the function will throw an error without replace).
• gcollapse, labelformat allows specifying the output label using placeholders.
• gcollapse, sumcheck keeps integer types with sum if the sum will not overflow.

Differences from reshape

• Allows an arbitrary number of variables in i() and j()
• Several option allow turning off error checks for faster execution, including: fast (similar to fast in gcollapse), unsorted (do not sort the output), nodupcheck (allow duplicates in i), nomisscheck (allow missing values and/or leading blanks in j), or nochecks (all of the above).
• Subcommands gather and spread implement the equivalent commands from R's tidyr package.
• At the moment, j(name [values]) is not supported. All values of j are used.
• "reshape mode" is not supported. Reshape variables are not saved as part of the current dataset's characteristics, meaning the user cannot type reshape wide and reshape long without further arguments to reverse the reshape. This syntax is very cumbersome and difficult to support; greshape re-wrote much of the code base and had to dispense with this functionality.
• For that same reason, "advanced" syntax is not supported, including the subcommands: clear, error, query, i, j, xij, and xi.
• @ syntax can be modified via match()
• dropmiss allows dropping missing observations when reshaping from wide to long (via long or gather).

Differences from regression models

gregress, givregress, and gglm do not aim to replicate the entire table of estimation results, nor the entire suite of post-estimation results and tests, that regress (reghdfe), ivregress 2sls (ivreghdfe), poisson (ppmlhdfe), or logit make available. At the moment, they are considered beta software and only coefficients and standard errors are computed.

• Results are saved either to mata (default) or copied to variables in the dataset in memory.
• by() and absorb() are allowed and can be combined.
• givregress does a small sample adjustment (small) automatically.
• givregress does not exit with error if covariates are collinear with the dependent variable.
• If the givregress model is not identified, standard errors and coefficients are set to missing instead of exiting with error.
• gglm runs with option robust automatically.
• If the givregress model is not identified, standard errors and
• If there are no non-linear covariates (i.e. all observations are numerically zero) then the coefficients and standard errors are both set to missing.

Differences from xtile, pctile, and _pctile

• Adds support for by() (including weights)
• Does not ignore altdef with xtile (see this Statalist thread)
• Category frequencies can also be requested via binfreq[()].
• xtile, pctile, and _pctile can be combined via xtile(newvar) and pctile(newvar)
• There is no limit to nquantiles() for xtile
• Quantiles can be requested via percentiles() (or quantiles()), cutquantiles(), or quantmatrix() for xtile as well as pctile.
• Cutoffs can be requested via cutquantiles(), cutoffs(), or cutmatrix() for xtile as well as pctile.
• The user has control over the behavior of cutpoints() and cutquantiles(). They obey if in with option cutifin, they can be group-specific with option cutby, and they can be de-duplicated via dedup.
• Fixes numerical precision issues with pctile, altdef (e.g. see this Statalist thread, which is a very minor thing so Stata and fellow users maintain it's not an issue, but I think it is because Stata/MP gives what I think is the correct answer whereas IC and SE do not).
• Fixes a possible issue with the weights implementation in _pctile; see this thread.

Differences from egen

• group label options are not supported
• weights are supported for internally implemented functions.
• New functions: nunique, nmissing, cv, variance, select#, select-#, range
• gegen upgrades the type of the target variable if it is not specified by the user. This means that if the sources are double then the output will be double. All sums are double. group creates a long or a double. And so on. egen will default to the system type, which could cause a loss of precision on some functions.
• For internally supported functions, you can specify a varlist as the source, not just a single variable. Observations will be pooled by row in that case.
• While gegen is much faster for tag, group, and summary stats, most egen function are not implemented internally, meaning for arbitrary gegen calls this is a wrapper for hashsort and egen.

Differences from tabstat

• Multiple groups are allowed.
• Saving the output is done via mata instead of r(). No matrices are saved in r() and option save is not allowed. However, option matasave saves the output and by() info in GstatsOutput (the object can be named via matasave(name)). See mata GstatsOutput.desc() after gstats tab, matasave for details.
• GstatsOutput provides helpers for extracting rows, columns, and levels.
• Options casewise, longstub are not supported.
• Option nototal is on by default; total is planned for a future release.
• Option pooled pools the source variables into one.

Differences from summarize, detail

• The behavior of summarize and summarize, meanonly can be recovered via options nodetail and meanonly. These two options are mainly for use with by()
• Option matasave saves output and by() info in GstatsOutput, a mata class object (the object can be named via matasave(name)). See mata GstatsOutput.desc() after gstats sum, matasave for details.
• Option noprint saves the results but omits printing output.
• Option tab prints statistics in the style of tabstat
• Option pooled pools the source variables and computes summary stats as if it was a single variable.
• pweights are allowed.
• Largest and smallest observations are weighted.
• rolling:, statsby:, and by: are not allowed. To use by pass the option by()
• display options are not supported.
• Factor and time series variables are not allowed.

Differences from levelsof

• It can take a varlist and not just a varname; in that case it prints all unique combinations of the varlist. The user can specify column and row separators.
• It can deduplicate an arbitrary number of levels and store the results in a new variable list or replace the old variable list via gen(prefix) and gen(replace), respectively. If the user runs up against the maximum macro variable length, add option nolocal.

Differences from isid

• No support for using. The C plugin API does not allow to load a Stata dataset from disk.
• Option sort is not available.
• It can also check IDs with if and in conditions.

Differences from gsort

• hashsort behaves as if mfirst was passed. To recover the default behavior of gsort pass option mlast.

Differences from duplicates

• gduplicates does not sort examples or list by default. This massively enhances performance but it might be harder to read. Pass option sort (sorted) to mimic duplicates behavior and sort the list.

Differences from rangestat

• Note that gstats range is an alias for gstats transform that assumes all the stats requested are range statistics. However, it can be called in conjunction with any other transform via (range stat ...). It was not intended to be a replacement of rangestat but it can replicate some of its functionality.

• flex_stats (reg, corr, cov) are not allowed (see gregress).

• Intervals are of the form interval(low high [keyvar]); if keyvar is missing then it is taken to be the source variable.

• Variables are not allowed in place of low or high. Instead they must be #[stat] where # is a number and stat is an optional summary statistic; e.g. interval(-sd 0.5sd x).

• Separate interval and interval variables can be specified for each target; e.g. gstats range (mean -3 3) x (mean -2 . time) y ....

• All statistics allowed by gstats tab are allowed by gstats range (except nunique or percent).

• Options casewise, describe, and local are not allowed.

There are two key insights to the massive speedups of Gtools:

1. Hashing the data and sorting a hash is a lot faster than sorting the data to then process it by group. Sorting a hash can be achieved in linear O(N) time, whereas the best general-purpose sorts take O(N log(N)) time. Sorting the groups would then be achievable in O(J log(J)) time (with J groups). Hence the speed improvements are largest when N / J is largest.

2. Compiled C code is much faster than Stata commands. While it is true that many of Stata's underpinnings are compiled code, several operations are written in ado files without much thought given to optimization. If you're working with tens of thousands of observations you might barely notice (and the difference between 5 seconds and 0.5 seconds might not be particularly important). However, with tens of millions or hundreds of millions of rows, the difference between half a day and an hour can matter quite a lot.

Stata Sorting

It should be noted that Stata's sorting mechanism is hard to improve upon because of the overhead involved in sorting. We have implemented a hash-based sorting command, hashsort, which should be faster Stata's sort for groups, but not necessarily otherwise:

The overhead involves copying the by variables, hashing, sorting the hash, sorting the groups, copying a sort index back to Stata, and having Stata do the final swaps. The plugin runs fast, but the copy overhead plus the Stata swaps often make the function be slower than Stata's native sort.

The reason that the other functions are faster is because they don't deal with all that overhead. By contrast, Stata's gsort is not efficient. To sort data, you need to make pair-wise comparisons. For real numbers, this is just a > b. However, a generic comparison function can be written as compare(a, b) > 0. This is true if a is greater than b and false otherwise. To invert the sort order, one need only use compare(b, a) > 0, which is what gtools does internally.

However, Stata creates a variable that is the inverse of the sort variable. This is equivalent, but the overhead makes it slower than hashsort.

Planned features:

• Things to add to gcollapse:
  • prod
  • geomean pos: exclude negative numbers and zero.
  • geomean abspos: ibid but take absolute value first.
  • Generally should you add an abs option to everything?
• Flexible save options for gregress
  • predict(), including xb and e.
  • absorb(fe1=group1 fe2=group2 ...) syntax to save the FE.
  • Choose which coefs/se to save.
• Improve formula documentation for summary statistics (e.g. gini)
• Internal consistency test for various parts of gquantiles. Each function section does cases but they should be consistent!

These are options/features/improvements I would like to add, but I don't have an ETA for them (i.e. they are a wishlist because I am either not sure how to implement them or because writing the code will take a long time). Roughly in order of likelihood:

• gregress missing features
  • Non-nested multi-way clustering.
  • HDFE collienar categories check.
  • HDFE drop singletons.
  • Detect separated observations in gglm, family(poisson).
  • Guard against possible overflows in X' X
  • Accelerate HDFE corner cases (e.g. very dense multi-way HDFE)
  • Include quick primers on OLS, IV, and IRLS in docs.
• Some support for Stata's extended syntax in gregress
• Update benchmarks for all commands. Still on 0.8 benchmarks.
• Dropmissing vs dropmissing but not extended missing values.
• Allow keeping both variable names and labels in greshape spread/gather
• Implement selectoverflow(missing|closest)
• Add totals row for J > 1 in gstats
• Improve debugging info.
• Implement collapse() option for greshape.
• Rolling (interval) and moving options for gregress.
• Add support for binary strL variables.
• Minimize memory use.
• Add memory(greedy|lean) to give user fine-grained control over internals.
• Create a Stata C hashing API with thin wrappers around core functions.
  • This will be a C library that other users can import.
  • Some functionality will be available from Stata via gtooos, api()
  • Improve code comments when you write the API!
  • Have some type of coding standard for the base (coding style)
• Implement gmerge
  • Integration with ReadStat?

Hi! I'm Mauricio Caceres; I made gtools after some of my Stata jobs were taking literally days to run because of repeat calls to egen, collapse, and similar on data with over 100M rows. Feedback and comments are welcome! I hope you find this package as useful as I do.

Along those lines, here are some other Stata projects I like:

• ftools: The main inspiration for gtools. Not as fast, but it has a rich feature set; its mata API in particular is excellent.

• reghdfe: The fastest way to run a regression with multiple fixed effects (as far as I know).

• ivreghdfe: A combination of ivreg2 and reghdfe.

• stata_kernel: A Stata kernel for Jupyter; extremely useful for interacting with Stata.

• stata-cowsay: Productivity-boosting cowsay functionality in Stata.

Gtools is MIT-licensed. ./lib/spookyhash and ./src/plugin/common/quicksort.c belong to their respective authors and are BSD-licensed. Also see gtools, licenses.