Skip to main content
v2026.11,986 entries · CC-BY 4.0

Language Sampling in Linguistic Typology: Methods and Bias

Language sampling explained: probability, variety, phenomenon-based and convenience samples, the Genus-Macroarea method, and a worked sample-size calculation.

Written and maintained by CASRAI Editorial Board

Published

On this page

Language sampling is the process of choosing which languages to include in a crosslinguistic study, so that what you find about the sample can be generalised to languages more broadly. Because roughly 7,000 languages exist and many are poorly described, typologists cannot examine them all, and the way the sample is drawn decides which conclusions the data can support. This guide sets out the main sample types, the Genus-Macroarea method with its arithmetic, and how to report a sample.

What is language sampling?

In typology, a sample is a subset of the world’s languages examined for a research question. Bell (1978), described by Miestamo (2025) as the first paper devoted to sampling in typology, supplied the vocabulary still used: the universe is the class of objects under investigation, the frame is the means of access to that universe, and the sample is the set of objects actually observed. In typology the frame is, in practice, the set of languages for which usable data can be found (Miestamo, Bakker and Arppe 2016).

The gap between frame and universe is the central problem. For work on what is a possible human language, the universe includes languages that have disappeared without a trace, while the frame contains only attested ones. Miestamo et al. (2016) note that the size and distribution of language groupings is shaped by non-linguistic historical factors, so good sampling aims to stop the chance of a genealogical grouping being included from being dictated merely by its size.

Sampling in this sense belongs to research methods more generally. The statistical ideas it borrows are covered in the CASRAI guides to simple random sampling, stratified sampling and sampling bias; what is distinctive about languages is that the units are not independent. Related languages share structure through descent, and neighbouring languages share it through contact.

Probability, variety, phenomenon-based and convenience samples

Probability samples

Rijkhoff et al. (1993: 171), as summarised by Miestamo et al. (2016), distinguish two basic types. A probability sample is built to test tendencies and correlations statistically, which requires the sampled languages to be independent of one another. Such samples have to balance representativeness against independence: the larger the sample, the more genealogical and areal connections tend to exist among its languages, and any sample has to compromise independence to some extent. Perkins (1989: 312) recommends around a hundred languages for most linguistic samples to balance the two requirements.

Variety samples

A variety sample aims to display as much variety as possible and to reveal the rarest types. Dependence among sample languages is not a direct problem here, and the more languages a variety sample holds, the better it may capture the crosslinguistic range of a phenomenon. Smaller samples show what is common, but their coverage of rarer features is quite random. Two methods were designed specifically for variety sampling: the Diversity Value method (Rijkhoff et al. 1993; Rijkhoff and Bakker 1998) and the Genus-Macroarea method (Miestamo 2003, 2005).

Phenomenon-based samples

Some phenomena, such as egophoricity, vowel harmony or switch reference, have a clearly restricted distribution. Miestamo (2025) calls the practice of choosing languages among those known to show the phenomenon phenomenon-based sampling, rather than drawing a balanced world-wide sample. If the aim includes charting the world-wide distribution of the phenomenon, a balanced sample is still needed, and the two approaches can be combined in one study. The stated risk is that the study misses languages that have the phenomenon but are not yet known to.

Convenience samples

A convenience sample consists of languages whose data are easily accessible to the researcher, for example the grammars on the bookshelf or in the institutional library. Miestamo (2025) points out that this tends to introduce areal and genealogical bias, and recommends stating the actual sampling principles instead of only labelling a sample as a convenience sample.

Simple random and stratified random sampling

Simple random sampling would mean taking a list of all languages, for example from Glottolog, and choosing by random numbers. According to Miestamo (2025) this would not produce a balanced sample of the world’s languages and has not been used in typological studies. Stratified random sampling first divides the frame into areal or genealogical strata and then picks randomly within them. Once strata are defined, random choice within them is desirable, but because adequate sources are often missing, pure random selection is very rarely possible.

Which sample type fits which research aim?

The table maps a research aim to the sample type the two sources associate with it.

Research aim Sample type What the sources say
Test a tendency or correlation across languages Probability sample Requires independence of units; needs a balance between representativeness and independence (Miestamo et al. 2016)
Show the full range of variation, including rare types Variety sample (Diversity Value or Genus-Macroarea) Dependence among languages is not a direct problem; larger samples catch rarer features (Miestamo et al. 2016)
Study a phenomenon with a restricted distribution Phenomenon-based sample Choose languages known to show it; use a balanced sample if the aim is world-wide distribution (Miestamo 2025)
Work from the data you can readily reach Convenience sample Likely areal and genealogical bias; state the sampling principles (Miestamo 2025)

One development changes the first row. Miestamo (2025) observes that with recent quantitative typology (Bickel 2015), biases may be controlled more efficiently through statistical modelling on larger datasets without strict independence requirements, which has made probability sampling more or less obsolete in language typology, while variety sampling still remains relevant.

The Genus-Macroarea method step by step

Genera and macroareas

The Genus-Macroarea method stratifies genealogically at the level of the genus and areally at the level of the macroarea. In Dryer’s classification in the World Atlas of Language Structures (WALS), each language belongs to a genus and each genus belongs to a family; the classification has 521 genera and 215 families, plus two groups that are not genera in the genealogical sense: creoles and pidgins, and sign languages. Genera are intended to be comparable in time depth across the world. Macroareas are continent-size linguistic areas, and Dryer (1992) distinguishes six: Africa, Eurasia, Southeast Asia and Oceania, Australia and New Guinea, North America and South America.

The four sample levels

At its simplest the method means picking one language from every genus, preferably at random. Using the WALS list of genera this gives a Genus Sample of 521 languages. Where sources are missing for a chosen language, the researcher tries another language in the same genus; a genus with no usable sources is left out. The languages actually included form the Core Sample, which may be much smaller than the Genus Sample. Two further levels exist: a Restricted Sample, used to balance the macroareas, and an Extended Sample (Miestamo et al. 2016).

Worked example: sizing a Restricted Sample

The Restricted Sample gives each macroarea the same percentage of its genera. In the original study that introduced the method (Miestamo 2003, 2005), which used an earlier list of 412 genera rather than the WALS list, a language could be included from 239 genera, plus one creole, so the Core Sample had 240 languages. Coverage was uneven: Eurasia reached 97% of its genera, while Australia and New Guinea reached only 43.2%, the lowest of the six macroareas.

Formula and symbols

For macroarea m, the number of languages in the Restricted Sample is nm = p × Gm, rounded to a whole number, where Gm is the number of genera in macroarea m, and p is the coverage percentage of the least well-covered macroarea in the Core Sample, expressed as a proportion. Choosing a smaller p is allowed if a smaller sample is needed. Languages are removed at random from over-represented areas until each reaches the same percentage.

The arithmetic

With p = 0.432 (43.2%), the figures given by Miestamo et al. (2016) are:

  • Africa: G = 66, so 0.432 × 66 = 28.512, which rounds to 29 languages.
  • Eurasia: G = 35, so 0.432 × 35 = 15.12, which rounds to 15 languages.
  • All genera: G = 412, so 0.432 × 412 = 177.984, which rounds to 178 languages.

The non-genus groups (creoles and sign languages) are outside these calculations; by default one creole and one sign language may be added, and in the original study one creole was included, so the Restricted Sample held 179 languages.

Interpreting a sample: what the numbers allow

The Core Sample is first and foremost a variety sample. Better-described areas are over-represented in it, because bibliographic bias tends to turn into areal bias. That is harmless for the aim of showing variety, but anyone who wants quantitative generalisations, such as global frequencies of a feature, should try to remove the bias, which is what the Restricted Sample does. In that sense the Restricted Sample is better suited to a quantitative component in a primarily qualitative study. The sample size depends heavily on the research question: topics that need thick, thorough grammars allow smaller Core Samples than topics for which sources are easy to find.

The data table below compares the sample designs and the sizes reported in the two papers.

Assumptions and common mistakes

  • Treating languages as independent. Independence is never absolute and is a matter of degree. Genealogical and areal links are the reason stratification exists.
  • Asking whether one language represents its family. Miestamo et al. (2016) describe the question of whether English is a good representative of Germanic as a misunderstanding of sampling: typologists draw crosslinguistic generalisations, not claims about individual languages.
  • Using random selection from the whole list. Simple random sampling over all languages would not produce a balanced sample (Miestamo 2025).
  • Ignoring bibliographic bias. Better-studied regions are over-represented unless the sample is restricted or modelled.
  • Assuming a small pilot sample needs no care. Miestamo et al. (2016) stress that even small pilot samples should be balanced areally and genealogically, so that untested hypotheses do not start a life of their own.
  • Relying on a genus list as an objective fact. Dryer himself admits that the criteria for a genus cannot be applied mechanically; the list rests on the literature and expert views.

How to report a language sample

State the sample type, the frame, the stratification and the selection rule for each stratum, the final number of languages, and the sources used per language. A methods sentence might read: “We drew a variety sample with one language per genus from the WALS genera list, choosing at random within genera where sources allowed, and restricted it to balance six macroareas.” Give the number of genera that could not be included and why. Where a sample is a convenience sample, say which convenience factors applied rather than leaving the label alone.

How to choose one language per genus in R or Python

Both snippets assume your own table with one row per language and a genus column. The six rows below are an illustrative toy table with placeholder labels, not real data.

R

set.seed(1)
langs = data.frame(language = c("L1","L2","L3","L4","L5","L6"),
                    genus    = c("G1","G1","G1","G2","G3","G3"))
picked = do.call(rbind, lapply(split(langs, langs$genus),
                  function(d) d[sample(nrow(d), 1), ]))
picked

Python

import random
random.seed(1)
rows = [("L1","G1"),("L2","G1"),("L3","G1"),("L4","G2"),("L5","G3"),("L6","G3")]
by_genus = {}
for lang, genus in rows:
    by_genus.setdefault(genus, []).append(lang)
picked = {g: random.choice(l) for g, l in by_genus.items()}
print(picked)

Each run returns exactly one language per genus. For a real study, replace the toy table with a genus list such as the one in WALS, and keep only languages for which you have adequate sources. See also the guide to random sampling in Python. For the discipline background, see what linguistics is.

Sources

Data: Language sampling designs and reported sample sizes

Sample designs used in language typology with the sizes and counts reported in Miestamo, Bakker & Arppe (2016) and Miestamo (2025): WALS genera and families, the Genus, Core and Restricted Samples, and the sample sizes cited for two earlier designs. Compiled from the two papers and the WALS genealogy page.

Language sampling designs and reported sample sizes
Design or classificationCount (languages or groupings)What the count refers to
WALS genera (Dryer 2013)521Genera in the WALS genealogical classification
WALS families (Dryer 2013)215Families in the same classification
Genus Sample521One language per genus using the WALS list
Core Sample (Miestamo 2003, 2005)240239 genera plus one creole, drawn from an earlier list of 412 genera
Restricted Sample (Miestamo 2005)17943.2% of genera in each macroarea (178 languages) plus one creole
Perkins (1989) recommendation100Approximate size recommended for most linguistic samples
Voegelin-based sample (Bybee et al. 1994), theoretical size94Planned size before missing sources
Voegelin-based sample (Bybee et al. 1994), actual size76Size after languages without adequate sources were left out

Sources

Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · free to try

Ask about Language Sampling in Linguistic Typology: Methods and Bias

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →