Rules
Part of Can a reporting pack survive contact with its own audience?
Grading a reporting explanation properly before it ships out the door
Reporting commentary that earns its place: how to grade an explanation, decompose a movement before narrating it, and say what you cannot explain yet.
Most of the effort in a recurring report goes into the numbers. Most of the value comes from the paragraph beside them. This page is about writing that paragraph well. It assumes the arithmetic is already right.
What to take away
- A restatement of the number is not an explanation. Say what changed underneath it, or say that you do not know yet.
- Decompose before you narrate. A movement in a total is nearly always a mix of a rate change and a composition change, and they call for different responses.
- Write the same sentence shapes every period. Predictable prose is faster to read and much harder to spin.
- Grade three names the components that moved; grade four names a mechanism and its evidence. Send nothing below grade three.
Four grades of explanation
Grade the commentary you are about to send.
Before you send, run this checklist:
- Does every sentence tell the reader something the table does not?
- Can you name the components behind any total movement?
- Is any claim of cause resting on one period of data?
If any answer is no, revise.
Four grades of explanation
Grade
- One
- Restates the number
- Two
- Names a correlate
- Three
- Names components that moved
- Four
- Names mechanism and evidence
What it says
- One
- Volume fell nine percent
- Two
- Volume fell and so did staffing
- Three
- Two of nine regions account for fall
- Four
- One intake channel offline six days
Example shape
- One
- Nothing new
- Two
- Guess, probably wrong
- Three
- Look in the right place
- Four
- Decide whether to act
What the reader can do
- One
- Two
- Three
- Four
| Grade | What it says | Example shape | What the reader can do with it |
|---|---|---|---|
| One | Restates the number in words | "Volume fell nine percent" | Nothing they could not read off the table |
| Two | Names a correlate | "Volume fell and so did staffing" | Guess, and probably guess wrong |
| Three | Names the components that moved | "Two of nine regions account for the whole fall" | Look in the right place |
| Four | Names a mechanism and its evidence | "The fall is confined to one intake channel, which was offline for six working days" | Decide whether to act, and on what |
Grade two is the dangerous one, because it reads like an answer. Two things moving together in one period is the weakest possible evidence, and a reader who accepts it will build a plan on it. If all you have is a correlate, label it as a lead rather than a finding.
Decompose before you narrate
Take an invented figure to see the shape. Suppose an average handling cost per case rises from 40 to 44 units between two months, and you have two case types.
Mix shift raises average cost
Case type
- Simple
- 800
- Complex
- 200
Month 1 vol
- Simple
- 30
- Complex
- 80
Month 1 cost
- Simple
- 400
- Complex
- 600
Month 2 vol
- Simple
- Complex
| Case type | Month 1 volume | Month 1 cost per case | Month 2 volume | Month 2 cost per case |
|---|---|---|---|---|
| Simple | 75 | 30 | 65 | 30 |
| Complex | 25 | 70 | 35 | 70 |
| Total / average | 100 | 40 | 100 | 44 |
Neither cost changed; the average rose because the mix changed. A commentary saying costs increased is wrong in the way that matters, and its suggested fix targets the wrong thing entirely.
This is the everyday form of Simpson's paradox, and it appears whenever a total is an average over groups of different sizes.
The first move on any movement is arithmetic, not prose. Split the change into the part explained by rates within groups and the part explained by group size, and report both.
If you cannot split it because the report lacks the components, that is a finding about the report, and it belongs in the commentary too. The components behind a total are the analytics foundations questions to settle before any split is attempted.
Sentence shapes worth reusing
Repetition here is a feature. A reader who has learned your shapes can skim four pages in a minute and stop at the one sentence that is different this time.
Reusable sentence shapes
- Decomposed movementX moved from A to B, most sits in component
- Undecomposed movementruled out checked causes, remainder unexplained
- ThresholdX is inside or outside agreed band for nth period
- Revisionlast period's X restated from A to B because reason
- Partial periodX covers days of total, not comparable
- "The total moved from X to Y."
- "The movement is concentrated in [group]."
- "The rate within [group] changed by Z."
- "The mix shifted toward [group]."
- "We cannot yet explain [movement]; we have ruled out [list]."
The last one prevents more bad decisions than any of the others. The UK Government Analysis Function's guidance on writing about statistics is worth reading in full for phrasing that survives a hostile reader.
Saying that you do not know yet
The instinct is to fill the space. Resist it. An honest "we cannot explain this yet, here is what we have ruled out" is a stronger position than a guess, because it can be updated without anyone losing face, and because the ruled-out list is itself useful.
Give an unexplained movement an owner and a date. That converts an admission into a commitment, which is what stops the same sentence appearing unchanged for four months running.
Revisions belong in the commentary
Numbers change after publication. Sources arrive late, a definition gets corrected, a duplicate is removed. None of that is a scandal, and all of it is corrosive when it happens silently.
State the restatement in words next to the number, keep the old figure recoverable, and say why it changed. Statistical agencies treat this as routine practice rather than an embarrassment; the Office for National Statistics publishes its approach to revisions and the reasoning transfers directly to an internal pack.
The review before it goes out
Read your own commentary against three questions. Does every sentence tell the reader something the table does not? Would the sentence still be true if the movement reversed next month? Is there a claim of cause anywhere that rests on one period of data?
Send or hold the commentary
Does every sentence tell the reader something the table does not?
continue to next question
rewrite before sending
If the answers are yes, yes, and no, send it. The rest of the discipline lives in the production process, which is covered in the overview of building a report worth keeping, and in the habits that keep a number from being sent bare in nine analysis errors.
Related reading on this site
For the underlying question of whether a movement is real at all, see asking about a number in the room. For the definitions the commentary depends on, see writing a metric definition.
Common questions
How long should the commentary be?
Short enough that a reader finishes it. One paragraph per material movement, and no paragraph for a movement that is inside the normal range, is a workable rule. A report where every line gets a comment trains readers to skip all of them.
Should the same person write the numbers and the commentary?
Ideally yes, because the person who built the query knows where the joins are soft. The risk is that they explain what they expected rather than what happened, so the review should be done by someone who did not build it.
What if a stakeholder asks me to soften a sentence?
Ask what specifically is inaccurate. If something is inaccurate, fix it. If nothing is, offer to add context rather than remove the statement, and keep the original wording in the record. A softened sentence that later turns out to have been right costs the report its authority.
We have no components to decompose by. What now?
Say so in the commentary, and treat it as a requirement for the next version. A total with no available breakdown can be monitored but cannot be explained, and readers should know which of those two they are getting.
For deciding what promise the pack carries before any of this matters, see the three publication tiers.
