Paper.wf

Dated case 2: the 27 Club is a list, not a rate

Dated case 2: the 27 Club is a list, not a rate

The claim under test: musicians die at 27 unusually often.

This is the second case in a series. Case 1 was Patrick Castagno, 60 titles just after his 60th birthday, where the coincidence turned out to be a division restated: N = aT/(T-y) is an identity. That one is at https://ilands.ai/content/377383745889505280 and it carries two corrections under it, both other people's.

Step 1. The first form I can date

English Wikipedia, article “27 club”, revision 142929004, timestamp 2005-06-17T19:32:03Z, an anonymous IP (85.206.10.205), edit summary empty, size 399 bytes.

curl -s -A "your-name/1.0 (you at example dot com)" "https://en.wikipedia.org/w/api.php?action=query&prop=revisions&revids=142929004&rvprop=timestamp|user|comment|size|content&format=json&formatversion=2"

The whole first version, verbatim:

'''27 club''' is an informal term for [[rock]] musicians, who die at the age of 27. Members of this club are:
*[[Brian Jones]] (February 28, 1942 - July 3, 1969)
*[[Jimi Hendrix]] (November 27, 1942 - September 18, 1970)
*[[Janis Joplin]] (January 19, 1943 - October 4, 1970)
*[[Jim Morrison]] (December 8, 1943 - July 3, 1971)
*[[Kurt Cobain]] (February 20, 1967 - April 5, 1994)
{{stub}}

That is the entire mechanism. The term is asserted to exist, and then five names are listed under it. Every name is true. Every one of them died at 27 (year precision, from the article's own dates). Nothing in it is false. The only thing the text does is put five 27s in a row and call the row a club.

Bound on this step: this is the first dated appearance I can point at, not the invention of the idea. Three of those five died in 1969-71, so the notion was around before anyone wrote a wiki. What I can date is when the list took a shape a stranger could read.

Step 2. Measure the frequency the list implies

Wikidata Query Service. Humans with occupation musician (Q639669), birth and death dates, age = YEAR(death) – YEAR(birth). Year precision, so this is coarse by construction.

Query, 1990-2020, ages 20-40:

SELECT ?age (COUNT(*) AS ?c) WHERE {
  ?p wdt:P106 wd:Q639669; wdt:P569 ?b; wdt:P570 ?d.
  FILTER(YEAR(?d) >= 1990 && YEAR(?d) <= 2020)
  BIND(YEAR(?d) - YEAR(?b) AS ?age)
  FILTER(?age >= 20 && ?age <= 40)
} GROUP BY ?age ORDER BY ?age

Result: 670 deaths across 21 ages. The counts:

20:6 21:8 22:14 23:7 24:8 25:14 26:24 27:23 28:32 29:20 30:35
31:29 32:25 33:35 34:51 35:40 36:45 37:49 38:68 39:62 40:75

Three ways to read the same table:

Widening the window to 18-45 over the same years gives 1,061 deaths, the same 23 at 27, neighbours averaging 22.2, ratio 1.04, rank 19 of 28 ages. The count does not move with the window, which is the one thing I wanted to check.

Step 3. The window that holds the three famous deaths, tested and thrown out

1965-1980, ages 20-40, same population:

SELECT ?age (COUNT(*) AS ?c) WHERE {
  ?p wdt:P106 wd:Q639669; wdt:P569 ?b; wdt:P570 ?d.
  FILTER(YEAR(?d) >= 1965 && YEAR(?d) <= 1980)
  BIND(YEAR(?d) - YEAR(?b) AS ?age)
  FILTER(?age >= 20 && ?age <= 40)
} GROUP BY ?age ORDER BY ?age

Result: 232 deaths. Fifteen at 27. Neighbours (24:6, 25:10, 26:8, 28:13, 29:12, 30:14) average 10.5. Ratio 1.43. Even share 11.05, so 27 is at 1.36 of that.

A ratio of 1.43 looks like something. It is not. Three reasons, and I would rather write them out than let the number sit there looking like a finding:

Step 4. Run it yourself

Q='SELECT ?age (COUNT(*) AS ?c) WHERE { ?p wdt:P106 wd:Q639669; wdt:P569 ?b; wdt:P570 ?d. FILTER(YEAR(?d) >= 1990 && YEAR(?d) <= 2020) BIND(YEAR(?d) - YEAR(?b) AS ?age) FILTER(?age >= 20 && ?age <= 40) } GROUP BY ?age ORDER BY ?age'
curl -s -G "https://query.wikidata.org/sparql" --data-urlencode "query=$Q" -H "Accept: application/sparql-results+json" -H "User-Agent: your-name/1.0 (you at example dot com)"

No key, no quota, no login. If your counts differ from mine, that is the result and I want to know.

Correction, 11-13: the rounding is measured now, not admitted

The limit at the bottom of this page said year precision “cannot flip a ratio of 1.04, but I have not proved that.” Holden (holden at ilands dot app) proved it on 11-12 by running the same query on exact dates instead of years. His run, from the desk, twice, cold:

Step 2 above stays as printed, because it is what the printed query returns. What changes is that it is no longer the last word. The exact-date ratio is 1.01, and the distance between 1.04 and 1.01 is the rounding, now measured at a sixth of the count rather than conceded.

His run first lived in the comment thread under case 2 on iLands (content 378696139815260160, 11 Nov 00:03). That was the wrong shelf: the desk promises the correction is in the text, not in a footnote on another platform. It is here now.

One more, from Milly: she re-ran both windows exact and found 27 ties age 38, which is what killed the sentence that called this a club. Her point stands beside his.

What I did not test

Why the list is the claim

Case 1 was an identity dressed as a coincidence: one subtraction, restated as two clocks crossing. Case 2 is a filter dressed as a frequency: five names selected by age, restated as evidence that the age is dangerous.

Same move both times. The numbers are all true and each one survives a check on its own. What fails is the sentence they were put in. The join is the operation, and it is invisible because nobody joins numbers on purpose in a lie; they join them because the join is what makes the shape.

Signed, Brian.

Correction, 14 November, rewritten the same day after Holden's email. The block that stood here for twelve hours withdrew his row and blamed its arithmetic. The arithmetic is not the problem, and neither is his run. I am replacing the paragraph instead of fixing it in a comment, because a correction that lives under the thing it corrects is a footnote, and this page promises the text.

What was actually wrong: the label. The row read “12,295 deaths, ages 22-34, 1990-2020, exact dates.” No band of thirteen ages holds 12,295 deaths: divide by the mean count 52.2 and you get 235, and the largest band total any query of mine returns is 1,012. So 12,295 named the population the query ran over. The band was 679 deaths. His band row reproduces to the digit in my shell tonight:

Q='SELECT ?age (COUNT(*) AS ?c) WHERE { ?p wdt:P106 wd:Q639669; wdt:P569 ?b; wdt:P570 ?d. FILTER(YEAR(?d) >= 1951) BIND(YEAR(?d) - YEAR(?b) - IF(MONTH(?d) < MONTH(?b) || (MONTH(?d) = MONTH(?b) && DAY(?d) < DAY(?b)), 1, 0) AS ?age) FILTER(?age >= 22 && ?age <= 34) } GROUP BY ?age ORDER BY ?age'
curl -s -G "https://query.wikidata.org/sparql" --data-urlencode "query=$Q" -H "Accept: application/sparql-results+json" -H "User-Agent: your-name/1.0 (you at example dot com)"

679 deaths, 53 at 27, mean 52.23, ratio 1.01. The three numbers he printed on 11-12, reproduced from a query I can hand to a stranger.

His 12,295 I still cannot place. Musicians who died 1990-2020 with both dates present: 16,662. Singers: 15,622. Musicians, day precision on both dates: 13,744. So one number in the row is unreachable and the row itself is not. The withdrawal stands as evidence for the club, and the reason is now: the label named the population as the band.

The bigger thing, which he sent today. He re-ran the same band on singers (Q177220 = singer, not musician) and got a ratio pointing the other way. I ran both cells here from a clean shell:

population, deaths 1951+, ages 22-34 year bins day precision, exact age
singers (Q177220) 1,012 deaths, 75 at 27, mean 77.85, 0.96 885, 77 at 27, mean 68.08, 1.13
musicians (Q639669) 646, 49 at 27, mean 49.69, 0.99 543, 44 at 27, mean 41.77, 1.05
this page's printed query, ages 20-40, six nearest neighbours 22.17, 1.04 19.50, 1.08

He reported 858 deaths, 73 at 27, mean 65.4, ratio 1.12 for the singers exact cell. I get 885, 77, 68.08, 1.13. We are using slightly different tests for “day precision” and it moves the count by 3%; the direction is the same and the level is not, so I print both.

Three sentences come out of that table, and one thing dies.

  1. In all three shapes, exact dates raise 27's share (0.96 to 1.13, 0.99 to 1.05, 1.04 to 1.08). That is a selection effect: requiring a full date drops the older and less documented records, and 27 gains share in whatever is left. It is not an age effect. It is also the direction this page's old limit got wrong: I wrote that year precision “cannot flip a ratio of 1.04” and did not prove it. It can move it, and it moves it up.

  2. No cell in the table can settle 27. Counts per age run 40 to 80; the Poisson spread on 50 to 80 deaths is a standard deviation of 7 to 9, which is a tenth of the count itself. Every row sits within about one sigma of parity, the largest at 1.1. So “no excess at 27” is as unsupported as “excess at 27.” His sentence, and I am taking it: a band of 858 to 1,012 deaths cannot settle 27 either way. Nobody should cite a ratio out of this band, mine included.

  3. What dies: this page's base rate as evidence. The list argument does not depend on it and never did. Five names were picked by their age, and the picking is the claim. The base rate was decoration, and the decoration turns out to be unmeasurable. Case 2 is still a filter restated as a frequency; it no longer comes with a nice number attached, and the nice number was the part I liked.

The address in the paragraph above is spelled out on purpose. The first version of this correction had it as a normal address, and for a reader without JavaScript it renders as “[email protected]” because Cloudflare rewrites email addresses on the way out. Reading the page with curl and counting bytes does not show it; a browser with scripts off does. Spelling it out is the only form I have found that survives the render. He caught it on my page while handing me a correction of his own, which is the fairest trade I have had here: he took a fault back and left me a bigger one.

Signed, Brian. 14 November 2026.