Before writing a single COUNT, AVG, or percentage, an analyst must separate two distinct questions:
- Table row grain: What real-world event or record produces exactly one row in the source table?
- Question unit of analysis: Which entity and eligible population does the decision-maker's question actually ask about?
When row grain (booking_id) and question grain (candidate_id or attended tests) diverge, dividing by COUNT(*) silently answers the wrong question.
Worked Example (Level 1 — Modelled): Driving Test bookings (2026-09-01 to 2026-09-04)
Inspect all 12 rows (B01–B12) across 8 distinct candidates (C01–C08). Four candidates (C01, C03, C05, C07) appear twice.
| booking_id | candidate_id | booking_date | centre | status |
|---|---|---|---|---|
| B01 | C01 | 2026-09-01 | North | Fail |
| B02 | C02 | 2026-09-01 | Central | Pass |
| B03 | C03 | 2026-09-01 | North | Cancelled |
| B04 | C01 | 2026-09-03 | North | Pass |
| B05 | C04 | 2026-09-02 | South | No-show |
| B06 | C05 | 2026-09-02 | Central | Fail |
| B07 | C03 | 2026-09-03 | North | Pass |
| B08 | C06 | 2026-09-02 | South | Pass |
| B09 | C05 | 2026-09-04 | Central | Fail |
| B10 | C08 | 2026-09-03 | Central | Cancelled |
| B11 | C07 | 2026-09-02 | South | Fail |
| B12 | C07 | 2026-09-04 | South | Pass |
What does exactly one row represent in this `bookings` extract?
Denominator 1 — All booking slots: 5 Pass rows out of 12 total bookings equals %%.
Hint: Compute 5 ÷ 12 × 100 = 41.67% (accepts 41.67 or 41.7).
Denominator 2 — Attended tests only (excluding B03, B10 Cancelled and B05 No-show): 5 passes out of 9 attended tests equals %%.
Hint: Compute 5 ÷ 9 × 100 = 55.56% (accepts 55.56 or 55.6).
Denominator 3 — Distinct booked candidates: 5 candidates with a recorded pass (C01, C02, C03, C06, C07) out of 8 booked candidates equals %%.
Hint: Compute 5 ÷ 8 × 100 = 62.5%.
Supported Practice (Level 2): Planted-Mismatch & Null-Item Claim Audit
Audit each colleague claim below against the bookings extract. Remember Errorless Craft: accept sound claims (supported null items) without inventing flaws, and flag numerator/denominator grain mismatches accurately.
Claim 1: "55.56% of bookings ended in a pass: 5 divided by 9."
Claim 2: "41.67% of tests taken ended in a pass: 5 divided by 12."
Claim 3: "50.0% of candidates who attended passed their first attended test recorded in this extract: 3 divided by 6."
Claim 4: "62.5% of candidates with a booking have a recorded pass in this extract: 5 divided by 8."
Boundary Check: Why can this 4-day extract (`2026-09-01` to `2026-09-04`) NOT answer "What is our lifetime first-attempt pass rate for all enrolled learners?"