The Listening Cycle · episode 5

Plus or minus three

Composite scenario. Unofficial analysis for illustration. Not a named firm.

The organisation

A UK food manufacturer — 850 people, three sites: bakery, distribution, head office. They ran a 32-item Typeform census in April, n=184. The June board pack said ‘engagement 71%’ with no interval, and ‘eNPS +18’ from a 7-point ‘I would recommend this company as a place to work’. A slide of site pairwise tests was labelled Tukey. It was Welch with Bonferroni.

The constraint

Finite-population margin of error belongs on the headline the board can see. A 7-point recommend item is not eNPS. Method card: this is OLS, not causal; pairwise Welch + Bonferroni is not Tukey. Over-claiming the test is the People Scientist smell. Do not invent an industry corpus to make 71% look like a league table.

Three decisions

  1. 1. 71% without ± is a pretend census

    n=184 in an 850-person workforce is not a count of everyone. Finite-population MOE on that headline is ±3. The stamp the board sees is ‘71% ±3 (finite-population MOE) · n=184 · n-gate 30’. Hide the interval and you are selling a census you did not run.

  2. 2. A 7-point is not eNPS

    eNPS is 0–10, promoters 9–10, detractors 0–6. A 7-point Likert scored +18 and labelled eNPS is a different instrument with someone else’s name on it. The pack says ‘recommend (7-point), not eNPS’ or it does not ship.

  3. 3. Welch + Bonferroni is not Tukey

    The three sites were compared pairwise. The test is Welch with Bonferroni. Labelling it Tukey is how a People Scientist loses the room the moment a non-exec has a stats undergraduate. The method card sits on the slide, not in a footnote the projector crops.

The cycle

ChannelWhenWho fields
Typeform census (theirs)

People director

April, 32 items, all sitesTypeform — they already have it. We stitch the export
Site listening rounds

Site directors

May, bakery and distribution, sample of 24Offline — structured notes on shift, not a second survey
Board pack with method cards

People director + MD

June, company masterAnalysis here. Export blocked until the stamp carries ±, n, and the test it actually ran
  1. Sep — Sign the cycle. Write the banned labels first: fake eNPS, Tukey-on-Welch, 71% with no ±.
  2. Oct–Mar — Peak production into Christmas. No pulses. Exit conversations continue.
  3. Apr — They field in Typeform. 32 items, n=184.
  4. May — Site rounds on shift. Do not re-ask the 32 items as a clipboard survey.
  5. Jun — Board pack in their master. Headline 71% ±3. Recommend item not labelled eNPS. Pairwise card: Welch + Bonferroni, not Tukey. OLS is not causal.
  6. Jul — Four factory-level actions as WOOP. Re-measure in November on the items that have owners, not a fresh 32.
  7. Aug — Review the cycle. Decide whether n=184 was a fielding problem or a trust problem.

What we will not ask

  • eNPS on a 7-point Likert.
  • A headline without a finite-population interval.
  • Tukey on a Welch + Bonferroni pairwise.
  • Causal language on OLS.
  • An industry league table we do not have.

90-minute workshop

  1. 20 min · Put ±3 on the wall

    Same 71%, once naked, once with ±3, n=184, n-gate 30. The room has to say which number they would defend to a non-exec.

  2. 25 min · Un-name the eNPS

    Show the 7-point item next to a real 0–10 eNPS. Write ‘not eNPS’ on the slide. If the MD wants the brand name anyway, the pack does not ship.

  3. 25 min · Name the test

    Pairwise of three sites. Welch + Bonferroni. Not Tukey. OLS is not causal. The method card is on the slide, in their typeface.

  4. 20 min · What the June board is allowed to hear

    Four actions, owners, GAS dates. No league table. No borrowed test names. Sign the pack or stop.