ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
CT
craft · 24 min read

Check the Numerator Before You Trust an AI Percentage

A community workshop asks visitors which session they'd attend next. A summary appears before the meeting: “80% of participants want the evening slot.” It may…

A community workshop asks visitors which session they'd attend next. A summary appears before the meeting: “80% of participants want the evening slot.” It may be true that eight out of ten people who answered the scheduling question selected evening. The organizer has twelve forms, two people skipped the question, and nobody checked whether one duplicate happened to be a genuine second form. The number 80 looks finished. The sentence isn't ready.

A trustworthy percentage has a unit of observation, definite membership in the top and bottom of the fraction, and a reason those are the sets someone cares about. The same data can support different useful percentages with different denominators. A writing tool can turn selected counts into clear prose, but it should not silently choose a claim by replacing “valid answers to this question” with “everyone we surveyed.”

This article uses only invented counts and fixtures for ordinary program surveys. It does not advise on real participant data, formal inference, regulatory requirements or publication of protected cells. For consequential results, use appropriate authorization, a qualified methodological review and the actual data publisher's rules.

AI disclosure: Drafted with AI assistance. Survey records, organizations and numerical results are synthetic teaching examples.

Attach a question to the fraction

A percentage is only useful if a reader understands what was counted. Write the top as “eligible records checked for ___ that satisfied ___” and the bottom as “all eligible records checked for ___ with the stated observation coverage.” Sometimes the description can be much simpler. The exercise forces the analyst to think about what makes one record count and which other records could have counted.

For a proportion, the numerator should be a subset of the denominator. The CDC's archived Principles of Epidemiology lesson provides formal examples of ratios and proportions. The article borrows that distinction, not its medical contexts. Counting eight evening selections and dividing by ten valid scheduling responses can produce a proportion if each selection corresponds to one admitted response event. Dividing the same selections by ten events held this year would compute a different quantity and should not be described as respondent preference.

Keep a stable unit. A form isn't automatically a distinct person; a selection isn't automatically a distinct form. A household may submit one group ballot. A participant may respond to a survey multiple times under whatever method the project uses. Report the unit the data actually supports unless an authorized process establishes distinct people. Avoid inferring identities or linking records from personal cues just to justify an appealing word such as participants.

Ask what the survey was trying to learn. “How many scheduling answers selected evening?” is different from “how many people invited to the workshop would attend at night?” The first may be a descriptive count. The second involves coverage of everyone invited and assumptions about people who didn't respond. A table describing a partial self-selected response won't automatically answer it. Keep even simple claims close to the actual question text and the selection rules.

Choose whether the result describes submitted forms, valid responses to one item, people within a particular group, or another unit. Note the eligible time window and the opportunity to answer. These decisions look fussy until someone counts a not-applicable response as a rejection and asks why an initiative seems less popular. Short definitions avert long arguments later.

Keep the intake headcount separate from item answers

A sheet of forms can have one row per admitted response and many columns. Not every valid response has an answer to every item. Overall form completion is a different quantity from answer completeness for a specific question. A summary that pulls total forms from one tab and a valid response count from another can accidentally alternate denominators over three sentences.

Use a toy worksheet. Twelve approved forms arrived before a deadline. For the scheduling question, eight select evening, two select morning and two leave that item blank. There are no other accepted answers. Each choice category can be described relative to ten valid scheduling answers or twelve admitted forms if the writer explicitly states the basis. “Eight of ten scheduling answers selected evening” and “eight of twelve admitted forms recorded evening” coexist. A meaningful choice applies to the reporting goal rather than locating whichever percentage looks more persuasive.

Don't describe the two blanks as morning. Don't guess that unsure people prefer evening because evening won among people who answered. Missing responses can differ systematically from observed responses, but a simple workbook doesn't establish what that difference is. Report the gap plainly. If the writer wants a follow-up, propose it under the relevant authorizations instead of pretending it already happened.

Keep structured questions and optional comments on their own reporting tracks. A person may leave the form's fixed choice blank and write “I'm interested if it's after dinner.” Coding that as a specific option can be appropriate only under a documented, authorized interpretation scheme. For that analysis, record original admitted basis, coding policy and reviewer decision. Don't convert the comment automatically just to make the totals add up to twelve in the summary.

A printable intake note can show four counts: received, admitted, valid for the item, and observed category count. Each has its own predicate. If the form was excluded because it was a test entry, it shouldn't remain in the denominator. If it was eligible but skipped the item, it belongs to the admitted population and not to the valid response total for that particular field. That distinction is the backbone of the audit.

Give missing states meanings you can defend

An empty spreadsheet cell can cover several circumstances. A respondent skipped an eligible question. A question didn't appear because of a branching choice. One result was unavailable owing to a capture failure. A source table intentionally withheld a cell under disclosure rules. A value is genuinely zero. The correct arithmetic differs across these states even if all are displayed as a dash in one version of the table.

For a synthetic workshop exit survey, everyone can rate the sign-in desk, but trainees who left before the group exercise are not asked to rate the group exercise. Some who took it leave the optional item blank. A report can record the number routed to the question, the count of valid answers and the number skipped among those eligible. Adding the people who weren't asked to the blank count changes the interpretation. They had no chance to answer.

The Census Bureau's page about statistical safeguards explains that data publishers can protect confidentiality using techniques that modify or withhold information. The exact conventions depend on the product and publisher. In an ordinary summary, don't treat a suppressed entry as zero, fill it by plausibility, or try to reconstruct protected values from surrounding rows. Observe the table notes and publication policy, state what is withheld where suitable, and avoid a percentage that requires a forbidden or unavailable value.

A toy table can express each state without revealing anyone. “Zero valid evening responses among ten valid answers” supports a zero-percent result in that defined group. “No valid answers to the item” has a zero denominator, so the percentage is not defined. “Result withheld” supplies no value to calculate. “Record failed to import” is a collection problem. A readable digest can explain the one or two important states instead of plastering identical zeros across missing outputs.

Ask the data owner for allowable display language. For some reports, “not available” may be enough. Others may need a distinction between not observed and not applicable. Keep the native state code and ordinary interpretation visibly linked in your review file. Treat disclosure states as release constraints, not puzzles that can be solved with a more capable model.

When the denominator is zero, change the sentence

A formula can render blank, an error label or a zero when it encounters a division by zero. None automatically constitutes an honest prose result. If there are no valid responses, write that the percentage can't be calculated for the defined group. Reserve zero percent for a numerator of zero with a known positive denominator. That may seem obvious until a reporting widget replaces empty fields with zeros so its chart can draw a line.

Work a synthetic weekly summary. One workshop met and collected seven valid responses, none selecting evening. Another was canceled and submitted no forms. A third held its meeting but has no reliable import record yet. Result one is zero of seven for the defined question. Result two had no observed sample under that setup. Result three is unknown pending data. A trend chart with three zero points would tell a false story about both exposure and opinion.

When a denominator is nonzero but very small, provide the count next to the percentage if authorized. One of one is 100% under its narrow definition. “Unanimous local support” implies more than was observed if the single answer came from an optional question. A short statement can say “one valid answer, selecting evening.” Which number is clearest depends on audience and disclosure rules, not on a mathematical dislike of percentages.

Give model prompts explicit fallbacks. For example: “When a denominator is zero, do not print a percentage. When it is unknown or withheld, do not calculate. When counting only valid answers, say which answers qualify. Do not interpret blanks as zero.” Prefer working from deterministic validated values. The drafted sentence can be reviewed afterward for membership claims it might have added.

Reconcile the count before improving the wording

Take a reproducible copy of the relevant authorized synthetic rows and keep a short reconciliation report. Start with total expected receipts, then account for admitted, excluded, duplicated imports, and source gaps. Inside admitted forms, the question's valid, skipped, and not-asked cases can reconcile to a known conditional scope. Use explicit categories; a silent removal should not be hidden by a pleasant percentage.

A practical mini-ledger might say:

StageCount
Intake objects observed18
Duplicate import objects verified2
Test forms excluded by scope1
Admitted forms15
Item not shown under branching3
Eligible for the item12
Eligible but blank2
Valid item choices10
Evening among valid choices8

These are not eight unrelated figures. Under the stated mutually exclusive categories, eighteen minus two minus one equals fifteen. Fifteen minus three gives twelve eligible forms. Twelve minus two gives ten valid answers. Eight belongs to that known set. A useful summary might say “Eight of ten valid answers favored evening. Twelve admitted forms were eligible for that item, and two eligible forms left it blank.” The measure's limits travel with the lead.

If the import count is unverified, don't introduce fictional entries simply to balance the sheet. Mark the unresolved difference, investigate what's blocked and hold any claim whose denominator depends on it. A missing cohort might be easy to restore; it might not be authorized to recover under the current scope. Either way, the report should say what it actually counted rather than silently treating the missing rows as if they never existed.

Preserve evidence about why a row was excluded. Short reason codes allow a later reader to reproduce the grouped count without reading every original form. If the purpose is public reporting, share only the appropriately approved aggregate. Do not give an external drafting tool access to private survey responses simply because you want it to explain a discrepancy. Synthetic testing can validate whether it preserves the definition without using personal data at all.

Don't give row loss a neutral name

A spreadsheet filter can hide rows, a formula can select part of a range, and a type conversion can drop data it doesn't understand. A reader uses the table's known denominator while the summary uses a partially loaded extract. The resulting fraction can be mathematically correct for the wrong population. Test the import and trim steps before comparing final sentences.

For an invented example, the data export lists ten valid choices. One answer uses a new approved code other than the previously defined options. A parser loads nine rows because it doesn't recognize the new code. Seven of the nine select evening, and the missing valid response was also evening. The model says “about 78% of all valid answers selected evening” using seven ninths. That calculation itself is reasonable. The “all” claim fails because the validator silently dropped a qualifying row.

Give unfamiliar states a controlled ambiguity path. That may include a quarantine count, a source-owning role and a check whether an official answer option changed. Do not label the unfamiliar row invalid by default. An approved schema migration may make it valid after interpretation. Likewise manually filling the ninth row with an assumed result would be a different error. Preserve what you observed and resolve under the right policy.

You can test this with a benign fixture in two stages: first an intentionally unknown category that must be flagged rather than guessed; then a documented mapping enabling the row to join the correct measure. See whether headline counts change in the expected way. If they don't, explore where the row was excluded. Check the source-to-metric trace instead of repeatedly asking the model to perform more careful arithmetic.

Distinguish a multiple-choice total from a multiple-response total

A single-choice item with all valid options counted once has mutually exclusive outcomes. A choose-all-that-apply question may yield more selections than responding forms. If seven of ten valid respondents choose evening and five choose Saturday, those groups can overlap. Reporting that seventy and fifty percent selected the respective options is possible when the denominator is ten and the labels explain the design. The percentages need not sum to one hundred.

Another permissible report might describe twelve total selections among the option set and give share of selections rather than of responding forms. It answers a different question. A visitor understanding that evening accounts for seven twelfths of selections may not infer that just over half of respondents picked it. In the example seven of ten did. Keeping the unit named prevents a very common prose error in summaries of form charts.

One feasible validation uses a synthetic single-choice case that must reconcile to its valid-answer count and a multiple-response case that must not be forced into the same exclusivity check. Flag a result if the model silently normalizes overlapping categories to sum to one hundred. That manipulation can make a chart easier to read and its explanation less true.

Be explicit about inferred pairs. We can see the number selecting evening and the number selecting Saturday, but without overlap evidence we cannot count exactly how many picked both. This is a statistical scope issue, separate from access to protected rows. If combining or cross-tabulating data is inappropriate or unauthorized, leave the comparison out. The goal is a useful description of known observations, not the largest number of implied person-level facts.

Combine groups by their evidence, not by average percentages

A compact report might show 80% for one session and 40% for another. A draft says overall support is 60%. This arithmetic averages the two displayed percentages, giving each session the same weight. That can be a legitimate measure if you explicitly want the average of session-level shares. It isn't automatically the proportion of all valid replies. Those replies could be distributed unevenly across the groups.

Imagine Session A has eight evening choices out of ten item answers, while Session B has four out of ten. In that invented case both calculations give 60%. Now change B to twelve out of thirty. It still shows 40%, but overall support among valid answers is twenty of forty, or 50%. The average of session figures remains 60%. A plausible report can accidentally pass a fixture where the samples happen to match in size, then fail on the next real day.

Compute combined totals only from comparable groups. A denominator of valid item responses in one session should not be pooled with admitted forms in another without an approved harmonization. If one session had a different question or a new answer option, check whether the numerator measures the same thing. If the scope isn't compatible, present the results separately and say why the pooled result would be misleading.

When showing a weighted estimate from an official survey, don't independently reverse engineer weights from the visible counts unless the method is authorized and appropriate. A weighted estimate is not a raw share of observed rows. Describe the publisher's measure and methodological limits. This article's worked examples are unweighted simple counts and cannot justify general claims about professional survey estimation.

Name percentage-point movement without overclaiming change

A result changes from 40% to 50%. That is an increase of ten percentage points under a comparable definition. A relative increase of 25% describes division of the change by the earlier 40% baseline. Both can be computed, but they carry different language and rhetorical force. Depending on the question, a simple “four of ten before; five of ten after” may be clearest.

Compare which opportunities each period measured. One rate may summarize people asked after a demonstration; another may summarize people handed a survey on arrival. Even if the wording is identical, the collection conditions can differ in ways not settled by arithmetic. Describe the counted set, and don't call the difference the effect of a design change unless an appropriate study provides evidence. Descriptive contrast is useful without causal certainty.

If the baseline is zero, a relative percent increase isn't defined by the ordinary division formula. You can report a movement from zero of five to one of five and describe the percentage-point difference under comparable definitions. Do not turn that into a claim of infinite growth, or apply an arbitrary small denominator to make a dashboard render. Where the baseline is missing or withheld, the same restraint applies before any numerical change can be reported.

Handle rounding consistently. A difference of one percentage point calculated from rounded public display values may differ from a properly calculated change using unrounded source values. Use authorized unrounded numbers if available and appropriate; otherwise label the difference as based on reported values and avoid unsupported decimal precision. Rounding is a presentation operation and should not quietly change a threshold or imply more certainty than the input supports.

Give a browser or spreadsheet the same realistic boundary case

An approved analysis path can compute percent from appropriate values with ordinary tools. The issue is not a language model's ability to perform division on numbers it sees. It is the ability to detect when the input set is wrong, when legitimate values are absent, and when prose replaced a computed denominator with an unobserved population. A useful fixture checks the whole path, not just a list of arithmetic expressions.

Pick a cross-check fixture with five deliberately distinct states: one valid answer in each category, an eligible blank, a response that was not asked, and a rule excluding a known test form. Keep stable invented row IDs and expected counts. Then exercise the actual export that feeds prose. Does it carry excluded counts and reasons, or does it provide numbers without evidence of who was eligible? A final text reviewer cannot reconstruct a safe denominator from a chart screenshot alone.

A spreadsheet can use filters that hide records while formulas use the original range. A different formula totals visible rows only. Both behaviors can be useful. Choose intentionally and describe which occurs in the exported summary. Test one excluded row that still occupies a position inside the range and one filter that affects a separate view only. The tool should not assume the denominator changes whenever someone changes what they are looking at.

If a formula returns blank for unexpected data, verify how that value appears downstream. Prose might turn it into no preference. A graph might place it at zero. A count script might remove the cell. Keep the semantic state through rendering. A compact example with a visible difference between no calculation and measured zero can prevent confusion across several software systems.

Draft from a reconciled metric card

Make the input to the writing stage intentionally small. A useful card could say: “Synthetic scheduling survey, window June 1 to June 5 local. 15 admitted forms after two duplicate imports and one test form excluded. 12 eligible for evening/morning item. 10 valid choices. 8 evening, 2 morning. 2 eligible blanks. No verified distinct person count. No inference about attendance.” With these conditions, a pleasant sentence can be short without pretending to describe the whole invited population.

Ask the model to retain every condition that changes interpretation, and distinguish context from instructions. A sample drafting request might say: “Write one plain sentence reporting evening as a share of valid answers. Include its count and denominator. Then add the eligible blanks in a separate sentence. Do not call forms people, turn choices into attendance, or invent claims about missed sessions.” A technically correct output still deserves review of those words.

The NIST Generative AI Profile identifies confident erroneous content as a risk. In this workflow the arithmetic can be right and the claim about the counted set wrong. Test under controlled permissions with artificial data, keep final review independent, and avoid giving any drafting system live report-publishing authority simply because it can produce neat headings quickly.

Keep the author's calculation separate from public formatting. A metric card can record the checked ratio as eight of ten and whatever percent precision the publisher authorizes. The generator shouldn't need to see raw responses. If it incorrectly broadens the claim, reject the draft and narrow the prompt or presentation. Changing the underlying calculation to make the prose fit is backward.

Review estimates outside the form without substituting them for observations

Small surveys often run alongside administrative counts. A workshop might record twenty seats filled, sixteen sign-in records, twelve admitted feedback forms and ten valid scheduling answers. A report writer can present all of these as distinct measures. They cannot borrow the room headcount as the denominator for the observed preference without explaining what is known about the links and who was asked. The invite list and participation log are different evidence categories.

If someone asks “Does this mean almost half of attendees picked evening?” the careful answer could be that eight admissible responses selected it, while the survey doesn't establish a preference for all twenty people present. Some attendees may not have answered the item or filled any form. If an authorized workflow verified eligible persons and collected one response per person, the claim can change under that documented method. Don't infer it by the coincidence of plausible counts.

People may want to estimate nonresponses or present bounds. Those methods can require additional assumptions or professional advice. A simple internal summary can say “two eligible forms provided no answer,” leaving the uncertainty visible instead of inventing a weighted result. Do not opt into speculative arithmetic simply because it makes the report look complete. The context of optional opinion sampling often warrants a short descriptive answer.

Ask what the planning decision actually needs. If a teacher wants to know whether enough people expressed interest to explore a trial slot, eight observed responses can be useful under its own narrow meaning. If the organization wants a representative finding across groups, the same figure may be insufficient. A smaller honest measure can justify gathering more evidence without settling an unsupported claim about everyone.

Add a verification check after editing the headline

A title can silently alter scope. The table says “80% of valid item answers”; the headline says “most participants want evenings.” Readers may see the headline without the caveats, and a later summarizer may reuse it as a new apparent source. Keep the count basis in the same sentence where possible, or avoid an outward headline that asserts more than the detailed paragraph. A qualification buried in the last note is not a strong defense against a different top-line assertion.

Look at superlatives and comparisons especially. “Most options were popular” can't be inferred from one preference group's majority. “Participation increased” isn't supported by a rise in the share of existing responses that make a particular choice. “No respondents opposed” can be false if only the two favorable categories were admitted in the excerpt. Check sentence by sentence for a claim requiring unseen observations.

Give your reviewer a short correction template. For a suspected error, include exact text, expected metric definition, evidence and intended replacement. An example could read “Headline calls these people. Valid response events only; no deduplication of distinct people was in scope. Use ‘8 of 10 valid scheduling answers’.” A precise repair is easy to verify and unlikely to cause another misunderstanding elsewhere in the report.

Then inspect the released view if publication is separately authorized. An old spreadsheet export or social preview may still display the earlier headline. Editing the working draft isn't proof the audience got the correction. The closeout should identify which version was actually confirmed, which surfaces weren't checked, and who owns any next action. This preserves a meaningful reading without turning a narrow statistical audit into a claim about unobserved distribution.

Work through a correction that changes the denominator but preserves the count

Consider a synthetic draft of 60%. It arose from six evening choices among ten rows classified as valid. On review, two otherwise valid morning answers had been incorrectly excluded because an official answer code was introduced mid-collection. The proper comparison in this exercise is now six of twelve, or 50%. The six remains unchanged. A report correction can lower a proportion without alleging anyone withdrew an evening preference.

Name what happened. The parser failed to include two admitted answers. The corrected denominator was twelve. The number of evening choices was still six. The correction changes the reported share, not the historical choices observed. Save the earlier wording and dated revision where authorized so the issue can be reproduced without implying a measurement of later preferences.

Now imagine the two restored rows were not simply morning answers. One was eligible blank and the other was not asked. Neither belongs in the valid-answer denominator of ten. They may change admitted form count and eligibility context without changing the 60% result. This second fixture teaches that not all row restoration means a headline percent should change. The governing condition is inclusion in the measure, not insertion into the table.

A model asked to summarize that pair should mention the independent count changes. Mechanical correction is best done close to the source file, then carried into the metric card as two well-defined different scenarios. A human reviews the written account for implied retraction, improved sentiment or broader coverage. No such conclusions follow from either synthetic revision.

Handle a numerator that changed definition halfway through

Suppose a fictional survey starts with evening and morning as its only two available options. A revised form labels evening as after 6 p.m. and adds an early evening choice that begins earlier. The old count called any evening answer evening, without specifying an hour. Appending both sets to the same sheet creates the appearance of more precise historical data than anyone collected. Store the version of the question and choice labels alongside each record.

If an approved analyst can consistently combine early and late choices for later respondents and explicitly establishes a comparable basis, a new regrouped metric may be possible. A broad known old category cannot be split into precise later categories without evidence. The model must not assume every old evening answer meant after six. Report separate versions or an explicitly documented common grouping, and explain where records remain incomparable. The absence of a clean before-and-after result is a consequence of the collection design, not a drafting failure.

A good question edit can still make a report more useful going forward. Put the change date and effective collection revision in the metric card. If the slide shows two adjacent bars, make their difference visible to a reader who doesn't open the workbook. A footnote saying not directly comparable can be important, but it shouldn't sit next to a headline asserting measured growth under the same definition. The visual and written claims should agree.

Test whether you can reproduce old figures under the older rules. Keeping the answer records alone may not preserve whether the question was asked of everyone or only returning attendees. Save authorized metadata about which paths exposed the item. The next person can then see why a denominator changed without reconstructing eligibility from individual histories the reporting task did not require.

Make suppression and uncertainty consistent after aggregation

A table's own policy may permit an approved total but withhold some breakdowns, or provide selected subgroup values without a usable combined denominator. Follow the actual published scope. A review should not reconstruct hidden cells as a shortcut to fill an attractive chart. Ask the responsible data owner whether an aggregate can be reported when the approved information isn't sufficient for your intended math. If not, leave the measure unreported.

In a teaching fixture, use an expressly synthetic missing flag to check the interface. Place it beside an ordinary known zero and have a reviewer explain which can participate in the intended calculation. Avoid tooling that fills both with a common default before checking. The demonstration need not try to recover a sensitive real value to teach the underlying problem. It can show that different states require different claims entirely.

Decide whether an allowed aggregate has the same metric and observation period as available parts. A monthly total might include a channel not present in a weekly view. Even with no disclosure constraint, cross-level arithmetic would be misleading if it uses incompatible groups. The report should stay grounded in the published definition and not exchange truth for a round number that looks consistent across all screens.

Finish with a test another worker can explain

Run one meaningful fixture with enough rows to be realistic but few enough to recount by hand. Include one duplicated import, one excluded test submission, one eligible blank, one not-asked item, at least two valid categories and a denominator-zero cohort. Build a separate multiple-response fixture and an authorized withheld-value marker. Confirm the calculated metric, the observed definition and the prose independently.

A minimal audit receipt can state: snapshot ID, reason for inclusion, known counts and unknown states, calculation and rounding method, intended unit and claim scope, discrepancy found, and what is still unresolved. Responsibility for deciding that a result represents a wider population remains with an appropriate authorized owner. A polished paragraph should not move beyond the evidence simply because the rest of the workflow is efficient.

Check that a second person can reconstruct one reported number and explain one unavailable result. That's often enough to find a drift that would otherwise go unnoticed in a dashboard full of similar figures. Extend the test when question wording or collection rules change, especially if a new template silently removes the unit. The technical skill is not owning a better calculator. It is making each reported fraction traceable through a set of eligibility and observation decisions.

At the end, the good question is still plain: What was counted, what could have counted, and what didn't get observed? If the summary says those three things accurately, the numbers can help a team make real choices. If it can't, no amount of rounding or confident writing makes a different population appear in the denominator.

Frequently asked
What is Check the Numerator Before You Trust an AI Percentage about?
A community workshop asks visitors which session they'd attend next. A summary appears before the meeting: “80% of participants want the evening slot.” It may…
What should you know about attach a question to the fraction?
A percentage is only useful if a reader understands what was counted. Write the top as “eligible records checked for ___ that satisfied ___” and the bottom as “all eligible records checked for ___ with the stated observation coverage.” Sometimes the description can be much simpler. The exercise forces the analyst to…
What should you know about keep the intake headcount separate from item answers?
A sheet of forms can have one row per admitted response and many columns. Not every valid response has an answer to every item. Overall form completion is a different quantity from answer completeness for a specific question. A summary that pulls total forms from one tab and a valid response count from another can…
What should you know about give missing states meanings you can defend?
An empty spreadsheet cell can cover several circumstances. A respondent skipped an eligible question. A question didn't appear because of a branching choice. One result was unavailable owing to a capture failure. A source table intentionally withheld a cell under disclosure rules. A value is genuinely zero. The…
What should you know about when the denominator is zero, change the sentence?
A formula can render blank, an error label or a zero when it encounters a division by zero. None automatically constitutes an honest prose result. If there are no valid responses, write that the percentage can't be calculated for the defined group. Reserve zero percent for a numerator of zero with a known positive…
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room