JuliaStats/StatsBase.jl

> The descriptive-statistics workhorse of the Julia ecosystem — weighted > statistics, sampling, ranking, and histograms that nearly every Julia data > package depends on, still versioned 0.x after a decade.

GitHub repo · Documentation · License: MIT

Overview

StatsBase.jl provides the layer of statistics that Julia's Statistics standard library deliberately leaves out: weighted means/variances/quantiles, higher-order moments (skewness, kurtosis), counting (countmap), ranking (tiedrank, competerank), autocorrelation and partial autocorrelation, sampling with and without replacement, empirical CDFs, histograms via fit(Histogram, ...), and data transforms like zscore. It was started by Dahua Lin in February 2013[^1], making it one of the oldest continuously maintained packages in the Julia ecosystem, and it sits inside the JuliaStats organization alongside Distributions.jl, GLM.jl, and HypothesisTests.jl.

Its raw GitHub numbers (610 stars, 196 forks) badly understate its position. Julia packages are consumed through the General registry, not GitHub stars, and StatsBase is one of the most-depended-upon packages in that registry — a transitive dependency of most of the data/science stack. That centrality is also its defining tension: after 13 years and 70+ tagged releases the package is still at v0.34.x[^2]. Under Julia's semver conventions every 0.x minor bump is breaking, so a StatsBase minor release historically required a coordinated compat-bound update across thousands of downstream packages. A 1.0 release and a partial merge into the Statistics stdlib have been discussed for years without shipping.

The second structural shift worth knowing: the abstract modeling API that StatsBase originally defined (coef, nobs, r2, loglikelihood, predict, residuals, the StatisticalModel/RegressionModel types) was extracted into the dependency-free StatsAPI.jl[^3] so that lightweight packages can implement the interface without pulling in StatsBase itself. StatsBase now re-exports those functions, so old code keeps working, but new interface code should target StatsAPI.

Getting Started

using Pkg; Pkg.add("StatsBase")

using StatsBase

x = randn(1000)
w = Weights(rand(1000))

mean(x, w)                        # weighted mean
std(x, w; corrected=false)        # generic Weights define no bias correction
countmap([1, 2, 2, 3])            # Dict(1 => 1, 2 => 2, 3 => 1)
sample(1:100, 10; replace=false)  # sampling without replacement
h = fit(Histogram, x; nbins=20)   # histogram as a fitted object
F = ecdf(x); F(0.0)               # empirical CDF, callable
zscore(x)                         # standardize
corspearman(x, randn(1000))       # rank correlation

Architecture / How It Works

StatsBase is a pure-Julia library with a deliberately thin dependency stack (StatsAPI, DataStructures, SortingAlgorithms, and stdlibs). Its main design ideas:

Weights as types, not keyword arguments. AbstractWeights has concrete

subtypes — Weights, AnalyticWeights, FrequencyWeights, ProbabilityWeights, UnitWeights — and the type determines the statistical semantics. var/std apply a different bias-correction formula per weight type, and plain Weights intentionally has no corrected estimator (calling var(x, w; corrected=true) on it throws)[^4]. This is principled but routinely surprises users coming from R/pandas, where a weight is just a vector.

Extending stdlib generics. Weighted mean, var, std, quantile,

and median are implemented as new methods on the Statistics stdlib functions, so using StatsBase transparently upgrades functions you already call. The abstract model API comes from StatsAPI and is re-exported.

fit as the front door. Histograms and data transforms

(ZScoreTransform, UnitRangeTransform) follow the fit(T, data...) → object → transform/reconstruct pattern shared across JuliaStats, rather than one-shot functions.

Sampling algorithms selected by heuristic. sample and wsample

dispatch among several algorithms (alias tables, Fisher–Yates, reservoir-style sampling) depending on population size, sample size, replacement, and whether ordered output is requested[^5]. You get near-optimal algorithmic choices without naming them, at the cost of performance characteristics that shift with input sizes.

Production Notes

The 0.x compat treadmill is the real operational cost. Because minor

bumps are breaking, one package in your dependency tree capping StatsBase = "0.33" can hold your entire environment back from 0.34 — and with it, transitively pin other packages to old versions. When Pkg.resolve produces mysteriously old manifests, a StatsBase compat cap is a common culprit. Pkg.status --outdated and Pkg.why are the diagnostic tools.

v0.33 → v0.34 (2023) removed long-deprecated functionality; code that

had been ignoring deprecation warnings for years broke on upgrade. The 0.34 series itself (2023–2026) has been patch-only and conservative — 12 patches in three years[^2], which reads as "stable and maintained" rather than "stagnant", but also means long-requested features move slowly. 221 open issues reflect a wide user base against a small reviewer pool.

Weight-type semantics are a correctness footgun, not just an API quirk.

FrequencyWeights and ProbabilityWeights give different corrected variances on the same numbers. Picking Weights "because it's the obvious name" and then hitting the corrected-variance error — or silently passing corrected=false — is the most common way to get subtly wrong uncertainty estimates out of this package.

Histogram edge semantics. fit(Histogram, ...) defaults to

closed=:left bins; code ported from systems assuming right-closed bins (or from older StatsBase versions) can miscount boundary values. Pass closed explicitly when reproducing results from other tools.

Performance is good but not vectorized-C good. Kernels are plain Julia

loops: fast, type-stable, and AD-unfriendly in places (in-place accumulation). For datasets that don't fit memory or need single-pass computation, StatsBase has nothing — that niche belongs to OnlineStats.jl.

When to Use / When Not

Use when:

You're in Julia and need anything beyond stdlib mean/var/cor —

weighted statistics, ranks, sampling, ECDFs, histograms.

You need statistically distinct weight semantics (frequency vs probability

vs analytic) handled correctly rather than by hand.

You're implementing a statistical model type and want ecosystem-standard

accessors (via the re-exported StatsAPI interface).

Avoid when:

Stdlib Statistics already covers you — StatsBase is light, but every

dependency is compat surface on the 0.x treadmill.

You need streaming/out-of-core statistics — use OnlineStats.jl.
You need distribution fitting, hypothesis tests, or regression — those live

in Distributions.jl, HypothesisTests.jl, and GLM.jl; StatsBase is descriptive-statistics only.

Alternatives

JuliaLang/Statistics.jl — use instead when unweighted mean/median/

var/cor is all you need; it's the stdlib with zero added compat surface.

joshday/OnlineStats.jl — use instead for single-pass, mergeable, streaming

statistics on data too large for memory.

JuliaStats/Distributions.jl — use instead when you want parametric

distributions and MLE fitting rather than empirical/descriptive statistics.

JuliaStats/HypothesisTests.jl — use instead (or alongside) when the goal is

significance testing, not summarization.

scipy/scipy — use instead when you're in Python; scipy.stats covers most

of the same descriptive ground.

History

| Version | Date | Notes | |---------|------|-------| | — | 2013-02 | Repo created by Dahua Lin[^1]. | | v0.25.0 | 2018-08 | Julia 1.0-era release during the 0.7→1.0 transition. | | v0.33.0 | 2020-03 | Start of the long-lived 0.33 series (21 patch releases through 2022). | | v0.34.0 | 2023-05 | Deprecation removals; breaking minor that required ecosystem-wide compat bumps. | | v0.34.12 | 2026-06 | Latest release; 0.34 series has been patch-only since 2023[^2]. |

References

[^1]: GitHub repository, created 2013-02-11; license credits "Copyright (c) 2012-2016: Dahua Lin, Simon Byrne, Andreas Noack, Douglas Bates, John Myles White, Simon Kornblith, and other contributors." https://github.com/JuliaStats/StatsBase.jl [^2]: StatsBase.jl releases (v0.34.0 2023-05-01 … v0.34.12 2026-06-17). https://github.com/JuliaStats/StatsBase.jl/releases [^3]: StatsAPI.jl — dependency-free statistics API definitions re-exported by StatsBase. https://github.com/JuliaStats/StatsAPI.jl [^4]: StatsBase.jl docs, "Weight Vectors" — per-type bias correction semantics. https://juliastats.org/StatsBase.jl/stable/weights/ [^5]: StatsBase.jl docs, "Sampling from Population". https://juliastats.org/StatsBase.jl/stable/sampling/

Wiki: JuliaStats/StatsBase.jl