Refolk
PlaybookMarket and talent intelligence

The Multi-Source Pool Count: Overlapping Lists to a Corrected Range

Turn two or three independent lists of the same population into a dated, stratified size estimate with a confidence range another analyst can reproduce.

22 min readLast reviewed November 15, 2026Read as Markdown

Key takeaways

  • In Refolk's index, 3,751 US profiles are tagged with Rust. That is about 1.6% of a loose usage-implied ceiling of roughly 240,250, so a single index cannot size a niche pool on its own.
  • Chapman's capture-recapture estimator, ((n1+1)(n2+1)/(m+1)) - 1, takes lists of 49 and 56 people with 30 in common and estimates about 91 people, with a 95% interval of 77.7 to 104.2.
  • If fewer than 7 people appear on both lists, report the union of the lists as a lower bound and not as an estimate. Each stratum also needs a sample coverage of at least 55%.
  • Seniority heterogeneity works like positive dependence between lists and pulls the estimate down while keeping the interval tight. The 48 to 52% CTO and founder share in Refolk's German Rust and Go samples is the warning sign.
  • Missed identity matches inflate the estimated pool and false matches deflate it, so measure both rates by hand-reviewing random samples of links and non-links before computing any estimate.
  • A third list lets you model dependence, but it does not give certainty. In one San Francisco study the best-fitting three-list model still underestimated, with an interval running from 3,179 to 18,010.

This playbook is for talent-intelligence analysts, strategy teams, and operators who need to know how many people really hold a niche capability, and at which employers, when each source catches only part of the pool. It starts from a written inclusion rule. From there you build two or three independent lists, resolve them into one identity table, and measure how much they overlap. The output is a dated size estimate, split by employer, seniority, and geography, with a confidence range and a dependence check that a second analyst can reproduce. The method is capture-recapture: the arithmetic census and epidemiology teams use to estimate populations that no single register sees in full.

Why a single filtered count undercounts the talent pool

A filtered count from one profile index tells you who that index can see. It does not tell you how many people hold the skill. For niche capabilities the missing share is large and nobody publishes it at occupation level. It also varies by country, seniority, and city, so a single "coverage factor" cannot correct it.

The comparison below shows the size of the gap for one capability in one market. It takes the official US count of software developers, applies a survey usage share for Rust, and sets the result against the number of profiles one index tags with Rust.

QuantityValue
BLS OEWS 15-1252 Software Developers, employed1,656,880
× 14.5% Rust usage share (derived)≈240,250
Rust-tagged US profiles, Refolk's index3,751
Index ÷ usage-implied pool (derived)≈1.6%

Read every row with its limits:

  • The official row is a ceiling, not a pool. BLS OEWS produces annual estimates for over 800 occupations, and none of them is defined by programming language or framework. Software Developers (15-1252) is the most granular US series. The May release used here reports a relative standard error of 0.7%.
  • The usage share is soft. The 14.5% figure is a third-party calculation from the public response file of one annual Stack Overflow developer survey. That survey sample is not representative, so the middle row is a loose ceiling.
  • The index row counts visibility. It counts people who are in the index and have the skill tagged. Nobody else is in it.

One skill, three ways of counting it (US, Rust)

  1. Official occupation ceiling
    1,656,880

    BLS 15-1252 software developers

  2. Usage-implied pool
    ≈240,250

    14.5% Rust usage share, derived

  3. Indexed and tagged
    3,751

    Refolk's index, Rust, United States

For one niche skill, one index's tagged profiles are about 1.6% of a usage-implied ceiling, and the ceiling itself is loose.
≈1.6%
Rust-tagged US profiles in Refolk's index as a share of the usage-implied pool
Derived from 3,751 tagged profiles against roughly 240,250 implied by BLS employment times a survey usage share, so treat it as an order of magnitude, not a coverage rate.

The gap does not mean 98% of the pool is hidden. Much of it is definitional. Someone who used Rust once counts toward a survey usage share, but your inclusion rule may well exclude them. The real lesson is that neither number tells you the true pool. The index cannot report what it misses, and the ceiling is too loose to test anything. To estimate the missing share you need a second and third view of the same population, and then you measure how much they agree.

How capture-recapture estimates a hidden population

Capture-recapture estimates a population's total size from the overlap between independent lists. If list A and list B each catch part of the pool independently, the share of list A that also appears on list B estimates the share of the whole pool that list B catches. Dividing list B's size by that share gives the total.

Five terms carry the method:

  • List (capture occasion): one source's set of people who meet the inclusion rule, such as profile-index skill tags, code-host contributors, paper bylines, or a speaker roster.
  • Capture history: for each person, which lists they appear on, written as a code such as 101 (on lists A and C, not B).
  • Overlap (m): the number of people who appear on both lists in a pair.
  • Closed population: nobody enters or leaves the pool between list builds. This is why the procedure fixes an as-of date.
  • Independence: appearing on list A does not change a person's chance of appearing on list B. Every failure in this guide comes back to this assumption.

The Chapman estimator

The basic Lincoln-Petersen estimator is N = n1 × n2 / m, where n1 and n2 are the list sizes and m is the overlap. Chapman's version adds one to each term to reduce small-sample bias:

N = ((n1 + 1)(n2 + 1) / (m + 1)) - 1

Var(N) = (n1 + 1)(n2 + 1)(n1 - m)(n2 - m) / ((m + 1)² (m + 2))

A published worked example uses n1 = 49, n2 = 56, and m = 30. Chapman gives N = 90.94 with a standard error of 6.77 and a 95% interval of 77.7 to 104.2. In R, the fishR package's mrClosed() computes this, and its Chapman modification is selected with chapman.mod=TRUE, which is the default.

Do not trust the textbook variance on small lists. Asymptotic variance estimates for Chapman often perform poorly with small samples, so use a bootstrap interval, or the log-normal interval some references recommend, and record the seed. For planning before you build anything, the coefficient of variation approximates √((n1 - m)(n2 - m) / (m(m + 1) × N)). Plug in rough guesses of list sizes and overlap to see whether the precision you need is achievable at all.

How much overlap is enough

Robson and Regier (1964) showed that if m is at least 7, there is a 95% chance that n1 × n2 / N exceeds 4. Above that threshold the bias is under 2% and can be treated as negligible. Below 7 overlaps, Chapman has a pronounced negative bias.

Robson and Regier also set accuracy targets by purpose. They are a useful vocabulary for saying what a pool count is fit for:

Purpose (Robson and Regier)Target accuracyHow I read it for talent work
Preliminary survey±50%Is this market worth a closer look
Management±25%Location or hiring-plan decisions
Research±10%Rarely reachable from public lists

The third column is my own mapping, not part of the source. State which tier you are claiming on every published estimate.

Which estimator fits two, three, or more lists

With two lists you get one estimate, and it silently assumes the lists are independent. With three lists you get enough cells to test and model dependence, which makes three the working target. Beyond three, the number of possible models explodes faster than your data can support.

ListsObservable cells (2^k - 1)Models available
Two3Chapman only, independence assumed
Three78 log-linear models
Four15113 log-linear models

The extra cells are degrees of freedom you spend controlling for list dependence. One review argues that multiple lists should be condensed into three before estimation. In talent work that usually means merging all speaker rosters into one "events" list, or bylines and talks into one "public work" list.

The three tools, and what each is for

  1. Pairwise Chapman as a diagnostic. With three lists, compute Chapman for each pair (A-B, A-C, B-C). If the pairs disagree by more than their intervals, the lists are not independent. Do not average the pairs into a headline.
  2. Log-linear models with model averaging. A log-linear model fits the counts in each capture-history cell and can include interaction terms for pairs of lists that move together. Wesson, Jewell, McFarland, and Glymour (Annals of Epidemiology 77) found that log-linear modelling relaxes the independence assumption, but the best model-selection criteria remain uncertain, and choosing the wrong model can produce implausible estimates. Across their five simulated dependence scenarios, a Bayesian model-averaging approach (DGA) gave approximately correct estimates for most. Use model averaging and report the spread across models.
  3. Sample coverage as a cross-check. Chao and Tsay's nonparametric sample coverage approach accounts for both individual heterogeneity and interaction between lists. It performs satisfactorily when sample coverage is over 55%. If its answer disagrees sharply with the model average, find out why before you publish.

Which way dependence pushes the estimate

  • Positive dependence (being on one list makes you likelier to be on another) inflates the overlap and pushes the estimate down.
  • Negative dependence pushes it up. In an Alameda County study, every two-source combination of the three lists would have overestimated the unobserved population.

A third list makes dependence measurable, but the estimate can still be far from certain. In a San Francisco HIV study, the best-fitting log-linear model included all three pairwise interaction terms and still substantially underestimated the true population. It gave 6,536, with a 95% interval of 3,179 to 18,010, which spans a factor of about 5.7.

Where visibility differs by country, seniority, and city

The probability of being on a list (the capture probability) is not one number. Profile coverage is set at country level, skill tags skew toward senior and public-facing people, and within a country the indexed pool clusters in one or two cities. Each of these creates heterogeneity, and heterogeneity acts like positive dependence: unless you stratify, it pulls the estimate down.

Country: the coverage threshold

MetricUnited StatesGermany
Rust-tagged profiles, Refolk's index3,7511,074
US:DE multiple (derived)3.49x1.00x
Above OECD 40% coverage thresholdYesNo
Leadership titles in 25-profile sample≥912

OECD.AI includes a country in its LinkedIn-based indicators only if it has at least 100,000 members, LinkedIn covers at least 40% of the labour force, and there are at least 10 AI hires in a given month. Germany, China, and India fall below the 40% line. So in Germany the profile index catches less of the labour force, and the other lists have to do more of the work.

Published penetration-by-country tables exist, but some show penetration above 100%. They count members who are not in the labour force, so they cannot be used as a capture probability. Occupation-level coverage figures are not established publicly. Estimate coverage locally, from the overlap.

Seniority and city: the same market, two skills

MetricRustGo
Tagged profiles in Germany1,0745,283
Rust:Go ratio (derived)0.201.00
CTO or founder titles, sample of 2512 (48%)13 (52%)
Berlin share of sample8/25 (32%)14/25 (56%)
48 to 52%
CTO and founder titles in Refolk's 25-profile German Rust and Go samples
Skill tags over-represent senior, public-facing people, who also dominate talks and bylines, so the overlap between lists inflates exactly where the pool is most visible.

Two consequences follow:

  • Seniority. Senior people show up on the profile index, the speaker roster, and the byline list at once. Junior engineers are under-captured on all three at once. A pooled estimate sees a high overlap and concludes the pool is small. Stratify by seniority (individual contributor, lead, CTO or founder) and estimate each stratum separately.
  • City. Berlin holds 32 to 56% of the sampled German profiles. A single national coverage factor mis-corrects both Berlin and everywhere else. Where the overlap allows, split metro from rest-of-country.

Stratifying means building or filtering each list per stratum, and the individual-contributor stratum is the hardest to fill. In Refolk a stratum is a plain-English request, so you can pull the under-captured slice directly instead of hand-filtering a list dominated by founders.

The multi-source pool count, step by step

The procedure runs from a written inclusion rule to a published, dated range in nine steps. By desk estimate it takes roughly 5.5 to 8.5 analyst-days before the final ceiling check. These times are working estimates, not sourced figures.

Overlapping lists to a coverage-corrected range

  1. Define the population and fix an as-of date
    Write the inclusion rule (capability, geography, employment status) and an as-of date, about half a day. Done when two people apply it to 20 test records and every one is settled as in or out.
  2. Build two or three lists independently
    Have different people or different queries build each list (profile-index skill, code-host contributors, bylines, speaker rosters), 1 to 3 days. Done when no list was built with reference to another.
  3. Resolve identities and measure linkage error
    Block, match, then hand-review random samples of links and non-links, 1 to 2 days. Done when you have false-match and missed-match rates and one row per person with a capture history.
  4. Tabulate capture histories per stratum
    Count the 2^k - 1 observable cells overall and per stratum (employer tier, seniority, country or metro), about half a day.
  5. Run pairwise Chapman diagnostics
    Compute Chapman with an interval for every pair of lists, about half a day. Done when every pair is recorded and any divergence is flagged.
  6. Fit three-list log-linear models and cross-check
    Fit the eight three-list models, model-average them instead of picking the best AIC, and run a sample-coverage estimate alongside, about a day.
  7. Correct for linkage error and bootstrap intervals
    Adjust the overlap counts using the rates from step 3, then bootstrap intervals with a recorded seed, about half a day.
  8. Stratify and enforce reporting floors
    Report a stratum as an estimate only if m is at least 7 and coverage is at least 55%, otherwise as at least the union count, about half a day.
  9. Check against ceilings and publish the dated range
    Compare to the official occupational count and the country's coverage status, then publish the range with lists, rules, linkage sample, and seeds attached.

Notes on the steps that decide the answer

Step 1 matters more than it looks. A skill tag, a code-host contributor, and a paper byline describe three different populations. If you don't apply one inclusion rule to all three before linking, you are not counting one pool three times. You are counting three pools once each.

Step 2 is where independence is won or lost. Give each builder only their own brief. Never build the speaker list by searching for authors you already have. In Refolk each list can be one standalone request run by a different person, which keeps the queries from borrowing each other's results. One example is "Rust engineers in Germany who have contributed to open-source Rust crates on GitHub in the last 12 months" as the code-host list.

Inclusion rule and list-build brief
POOL DEFINITION
Capability: [what counts as holding the skill, e.g. shipped or merged production code in the language]
Geography: [country, and metro if stratifying]
Employment status: [employed or self-employed; state exclusions such as students]
As-of date: YYYY-MM-DD
Out of scope: [roles that mention the skill but do not hold it, e.g. recruiters]
Test set: 20 records judged by two people; rule edited until all 20 agree

LIST A - Profile index
Builder:
Query or filter:
Built without reference to: B, C

LIST B - Code-host contributors
Builder:
Query or filter:
Activity window:
Built without reference to: A, C

LIST C - Public work (bylines and speaker rosters, condensed)
Builder:
Sources:
Built without reference to: A, B

STRATA
Employer tier:
Seniority: individual contributor / lead / CTO or founder
Geography: [metro] / rest of country

Fill every field before any list is built, and hand each builder only the block for their own list.

Step 3 must come before any estimate. The linkage literature requires error rates to be measured first, because the two error types pull the estimate in opposite directions (see the failure modes below). To estimate missed matches, sample the lists directly and hand-check whether non-linked records are really different people. Ding and Fienberg (1994) proposed a two-way correction for linkage error. Di Consiglio and Tuoto, and later Dasylva and colleagues, extended it. Dasylva's sensitivity analysis highlights missed matches as the error that matters most.

How the multi-source count goes wrong

Most failures in this method do not throw an error. They produce a confident, tight interval around the wrong number. Each failure below names which way it pushes the estimate, what the false positive looks like, and the check that catches it.

ErrorPushes the estimateFirst check
Positive list dependenceDownPairwise Chapman divergence
Seniority heterogeneityDownPer-stratum coverage
Missed identity matchesUpHand-review non-links
False identity matchesDownHand-review links
Small strata (m below 7)DownReport as at least the union
Best-AIC model choiceEitherModel-average, report spread
  1. Lists that are not independent. Example: a speaker roster recruited from paper authors. The overlap is inflated and the estimate shrinks. False positive: a tight interval around a value that is too low. Check: if the pairwise Chapman estimates disagree by more than their intervals, dependence is present.
  2. Heterogeneous visibility. The 48 to 52% CTO and founder share in Refolk's German samples suggests skill tags over-represent senior, public-facing people. Junior engineers are missed by every list at once, which behaves like positive dependence. False positive: a small, stable pool that is really just the visible top layer. Check: stratify by seniority and compare coverage per stratum.
  3. Missed matches from handle and name variation. One person appears as a display name in the index, a handle on the code host, and an initialled byline, and is never linked. Each miss removes someone from the overlap and inflates the estimate. False positive: a "hidden majority" that does not exist. Check: hand-review a random sample of non-links.
  4. False matches from common names in the same city. Two different engineers with a common name in Berlin get merged. The overlap grows and the estimate deflates. Check: hand-review a random sample of links and report the false-match rate. For scale, the worst scenario in Di Consiglio and Tuoto's work kept false-match rates under 5%. That figure comes from their setting, not yours, so measure your own.
  5. Small strata. Below 7 overlaps, Chapman has a pronounced negative bias. Check: report the stratum as "≥ union count" and nothing more.
  6. Picking a model by best AIC. The San Francisco model fit well and still underestimated, with an interval of 3,179 to 18,010. Check: model-average and publish the spread across models next to the headline.
  7. Mismatched definitions. A skill tag, a contributor, and a byline are different populations until one inclusion rule is applied to all of them. False positive: a low overlap that comes from the lists counting different people, which inflates the estimate. Check: apply the step 1 rule to every list before linking.
  8. Coverage figures above 100%. Penetration tables count members outside the labour force. Check: never use them as a capture probability. Estimate capture from the overlap instead.
A tight interval around a number that is too low is the most expensive thing this method can produce.

How to read the result and publish the range

Publish a headline estimate only when there is enough overlap and the pairwise estimates agree, or when you have modelled their disagreement explicitly. In every other case, publish a lower bound with its reason. The matrix below is the call I make on each pool and on each stratum.

Estimate, bound, or model first

Pairwise estimates agreePairwise estimates diverge
Agree, thin overlap
Report the union as a lower bound and build more list depth
Agree, solid overlap
Publish the estimate with a bootstrap interval
Diverge, thin overlap
Lower bound only, since dependence cannot be modelled yet
Diverge, solid overlap
Dependence is present, so model-average and stratify before publishing
Overlap below 7Overlap 7 or more
Overlap size decides whether you can estimate at all, and pairwise agreement decides whether independence is a safe assumption.

The ceiling check

Compare the estimate with the official occupational count, and with a usage-implied figure if you have one. If the estimate is above the ceiling, it is wrong, and the most likely cause is missed matches or mismatched definitions. If it is below the ceiling, the check confirms nothing. The ceiling is loose by orders of magnitude, as the index's ≈1.6% share of the usage-implied US Rust pool shows. Also state the country's coverage status: a German estimate rests more heavily on the code-host and public-work lists than a US one does.

The record a second analyst can rerun

Pool-count record
Population (inclusion rule, verbatim):
As-of date:
Lists: A = [source], n = ; B = [source], n = ; C = [source], n =
Union of lists (observed lower bound):
Linkage review: links sampled = , false-match rate = ; non-links sampled = , missed-match rate =
Pairwise Chapman: A-B = (95% CI); A-C = (95% CI); B-C = (95% CI)
Dependence verdict: pairwise intervals overlap / diverge
Log-linear, model-averaged: N = , 95% interval = , spread across models =
Sample-coverage cross-check: N = , coverage = %
Strata: [stratum] = estimate (interval), or "≥ union" with reason (m below 7 or coverage below 55%)
Ceiling check: official count = ; usage-implied = ; estimate below ceiling: yes / no
Country coverage status: above / below 40% threshold
Accuracy tier claimed: preliminary ±50% / management ±25% / research ±10%
Bootstrap: replicates = , seed =

Publish this with the three lists, the linkage review sample, and the bootstrap seed attached, so a second analyst can reproduce the numbers.

Reproducibility standard: a second analyst given the same lists, rules, linkage sample, and seed should land inside your reported interval. If they don't, the inclusion rule or the matching rule is underspecified, and the record is not finished.

Before you call the pool count done

  • The inclusion rule is written down and settled all 20 test records as in or out.
  • Each list was built by a different person or query with no sight of the others.
  • False-match and missed-match rates come from hand-reviewed random samples of links and of non-links.
  • Every pairwise Chapman estimate and interval is recorded, and any divergence is flagged.
  • The headline comes from model averaging, with the spread across models published beside it.
  • Every stratum with fewer than 7 overlaps or coverage below 55% is reported as at least the union count.
  • Intervals are bootstrap or log-normal, not asymptotic, and the seed is recorded.
  • The estimate sits below the official occupational ceiling, and the country's coverage status is stated.
  • The as-of date appears next to every number in the published record.

Keeping the pool count current

A pool count is a dated snapshot of a closed population, so it goes stale as people move. Refresh it by rerunning the whole procedure against a new as-of date. Do not adjust the old number.

  • Rerun, don't patch. Rebuild all lists for the new date with the same inclusion rule and briefs. Reuse the linkage rules, but sample and hand-review afresh, because handles and employers drift.
  • Re-check the coverage threshold. Country coverage status comes from the OECD.AI partner-data methodological note. Re-read it at each refresh rather than carrying forward the last status.
  • Re-pull the ceiling. OEWS publishes annually. For countries outside the US, find the national statistics office's most granular occupational series. I have not checked whether Germany has an equivalent, so confirm it before you rely on one.
  • Watch for definition changes in sources. If a code host, index, or conference changes what it lists, that list no longer measures what it measured last time. Treat it as a new list, not a continuation.
  • Track strata, not only the headline. A stable headline can hide a shift from individual contributors to leadership. Compare per-stratum estimates and bounds between snapshots, and compare each stratum's share of the union.

When a stratum was reported only as a lower bound, the next refresh is where to invest. Add depth to the list that under-captures it, most often the individual-contributor slice, until its overlap clears 7 and its coverage clears 55%.

Questions practitioners ask

How do I estimate the true size of a talent pool when many people are not on LinkedIn?

Use the overlap between independent lists. Build two or three lists of the same defined population, for example profile-index skill tags, code-host contributors, and paper bylines. Resolve them into one identity table and count the people who appear on more than one list. Chapman's estimator, or a log-linear model when you have three lists, converts that overlap into a total that includes people no list caught. You then report it as a dated range rather than a single count.

How many people need to appear on both lists for a capture-recapture estimate to be usable?

At least 7. Robson and Regier showed that with 7 or more recaptures there is a 95% chance the bias is negligible. With fewer than 7 overlaps, the Chapman estimate has a pronounced negative bias, so report the union of the lists as a lower bound instead. Apply the same floor to every employer, seniority, or geography stratum, and also require a sample coverage of at least 55% per stratum.

Can I correct a talent pool count by dividing by LinkedIn's penetration rate in that country?

No. Penetration figures are set at country level, and some published tables come out above 100% because they count members outside the labour force, so they are not a capture probability. Occupation-level coverage is not established publicly. Within a country the indexed pool also clusters heavily: Berlin holds 32 to 56% of Refolk's German Rust and Go samples. One national factor will mis-correct every stratum.

Why do I need three lists instead of two for capture-recapture?

Two lists give one estimate, and that estimate quietly assumes the lists are independent. Three lists give seven observable cells and eight possible log-linear models, so you can measure and model dependence between sources. You can also compare the three pairwise estimates as a diagnostic. A third list still does not guarantee a tight answer, so model-average across the models and report the spread.

Which way does identity-matching error bias a talent pool estimate?

The two errors pull in opposite directions. Missed matches, where the same person is not recognised across lists because of handle or name variation, inflate the estimate. False matches, such as two people with common names in the same city merged into one, deflate it. The literature flags missed matches as the bigger risk, so a matching pipeline tuned only for precision will quietly overstate the pool.

Can BLS occupational data validate my niche skill pool estimate?

Only as a ceiling. BLS OEWS covers over 800 occupations, but none is defined by programming language or framework, and Software Developers 15-1252 is the most granular US series. An estimate above that ceiling is wrong. An estimate below it is not confirmed by it. In practice the ceiling sits orders of magnitude above any niche pool, so it rejects absurd numbers but cannot confirm plausible ones.

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next