Three good questions beat thirty mediocre ones

By Zak Fenton, MSc Workplace Health & Wellbeing, Alltoogether · Published 12 July 2026 · Last reviewed 12 July 2026 · Every claim is sourced; report errors and we fix them visibly.

1. Single items measure more than you think

The reflex objection to short surveys is that one question cannot measure a construct. For narrow, self-evident constructs, the evidence says otherwise: Wanous, Reichers and Hudy (1997, Journal of Applied Psychology) meta-analysed 17 studies covering 7,682 people and found single-item measures of overall job satisfaction correlated 0.67 (corrected) with full multi-item scales, and 0.72 against the best-constructed ones. Later single-item validations concur (Dolbier et al, 2005). The UK's own harmonised standard makes the same bet: the ONS-4 personal wellbeing questions, used across national statistics, are four single items.

What single items genuinely cannot do is decompose a construct into facets (which aspect of satisfaction, which source of demand). That is a real limitation, it is why deep-dive instruments exist as a separate opt-in layer, and no honest short survey should claim otherwise.

2. Long questionnaires pay for their length in quality

The length trade is not free in either direction. Galesic and Bosnjak (2009, Public Opinion Quarterly) randomised announced survey length and found participation fell as stated length grew, and that answer quality decayed within the questionnaire: later questions were answered faster, skipped more, and straight-lined more. A 30-item battery does not deliver 10 times the information of 3 items; it delivers diminishing information at compounding participation cost. Rolstad, Adler and Rydén (2011, Value in Health) add the honest caveat: content matters as much as length, so the three questions have to be the right three.

3. The design that makes 3 by 21 work

Covering 21 domains with 3 questions per person per fortnight sounds like a trick. It is a planned missing-data design, formalised by Raghunathan and Grizzle (1995) and developed through the three-form design literature (Graham et al, 2006): every respondent answers a common core (our 2 anchors, every cycle), and the remaining items rotate on a deterministic oldest-unseen-first schedule. Because the engine controls who sees which item when, the unasked cells are missing completely at random by design (Rhemtulla and Little, 2016). MCAR-by-design is the most benign missingness there is: the gaps carry no information about the answers that would have been given, so population-level estimates built across cycles are unbiased. A 2023 review in the industrial-organisational literature concluded planned-missing designs perform equivalently to short-form approaches while covering a broader construct space.

The estimates are then built the way the design requires: pooled across rolling cycle windows, fitted with models that use every observed answer (full-information maximum likelihood or multilevel estimation), never cycle-by-cycle snapshots joined with a line.

4. What we will not claim for it

Honesty about the design's limits, in public, on purpose.

• Individual trajectories on rotating items are out. A person sees a given rotator roughly once every 14 weeks; no method recovers a reliable individual trend from that, so the engine never reports one. (It never reports individual scores anyway: the n≥5 floor is enforced in the schema.)
• Organisation-level rotator trends need patience: they become trustworthy over multiple rotation cycles, not weeks. Anchor trends resolve faster, from roughly 6 fortnightly cycles.
• Facet-level diagnosis needs real instruments. Sustained signals route to validated deep-dive instruments with their own consent step, not to more pulse items.
• The design's guarantees do not cover self-selection. If the least-well people stop responding, no rotation scheme fixes that, which is why completion rate is reported beside every figure rather than kept backstage.

5. What to ask any vendor about a short survey

Which validated instrument does each item come from? Who decides the rotation, and is it deterministic and auditable? How do you estimate across cycles: one model over all observed answers, or cycle-by-cycle averages? What do you refuse to report because the design cannot support it? (The last one is the tell. Every measurement design has unsupportable claims; the honest vendor can list theirs, and ours are in the honesty rules.)

See it running

The interactive demo shows the pulse and its refusals live. Free for any employer at alltoogether.com · the method at openworkplacehealth.org · building it into your product is what this site is for.