---
name: missing-data-handling
description: >
  Decides and documents what happens to every gap in a research dataset:
  separating not asked (out of base) from not answered, refused, don't know and
  lost; diagnosing whether the missingness is random or related to the answer
  itself; choosing between listwise treatment with an honest base statement,
  retaining non-response as a category, and the rare cases where imputation is
  defensible; handling item versus unit non-response and partial completes; and
  disclosing all of it. Use for "how do I handle missing data", "should I impute",
  "what do I do with don't know", "can I include partial completes", "the base
  keeps changing", "too many blanks in this variable".
category: 04 Data Preparation
ref: "04.03"
tier: 1
inherits: [K2, K3, K4, K5]
---

# Missing Data Handling

## 1. One-line description
Establishes what each gap in a dataset actually is, diagnoses whether the pattern of gaps is likely to bias the analysis, chooses the most conservative defensible treatment (which in commercial research is almost always to report the missingness rather than fill it), and documents the decision so that the effect of missing data on every reported figure is visible rather than absorbed.

## 2. What this skill is used for

**The research problem it solves.** Missing data is handled badly in two directions and both are common. The first is silent absorption: an analyst calculates every percentage on whoever happened to answer, the base moves from question to question, nobody says so, and a satisfaction figure that describes the 71% of the sample willing to answer a satisfaction question is reported as describing the sample. The second is inappropriate filling: a gap is treated as a zero, a don't know is scored at the scale midpoint, a blank is imputed from the mean, and a dataset that was 18% empty becomes 100% full without anybody being told that a fifth of the visible numbers are the analyst's arithmetic rather than the respondents' answers.

Behind both sits a definitional failure that causes more damage than either. Five different things routinely arrive in a dataset as an empty cell: a respondent who was routed away and should never have been asked, a respondent who skipped the question, a respondent who explicitly refused, a respondent who said they did not know, and a response lost to a technical fault or a break-off. These have nothing in common except their appearance. The first is not missing at all: those respondents are out of base and including them in a denominator is a base error, not a missing-data problem. The third and fourth are substantive answers about the respondent's relationship to the question, and converting them to blanks destroys real information. Treating all five as one thing called "missing" and applying one treatment to it is the single most consequential error in this area, and it is usually made before anybody notices there was a decision to make.

**Where it sits.** After cleaning and before analysis, though the missingness map is produced during validation and the decision made here constrains every base, every figure and every model downstream.

**Typical use cases.**
- Deciding the base convention for a question with substantial item non-response.
- Deciding whether don't know sits inside or outside the base, and what that does to the headline.
- Judging whether partial completes can be included, and in which analyses.
- Diagnosing whether non-response on a sensitive question is related to the answer people would have given.
- Assessing whether a proposed imputation is defensible, which usually means concluding that it is not.
- Producing the missing-data disclosure a report or a client audit requires.
- Explaining why bases differ between questions in a set of tables.

**Who uses it.** Quantitative researchers deciding base conventions; analysts building models where listwise deletion is quietly discarding a third of the sample; research directors judging whether a non-response pattern undermines a finding; client-side reviewers asking why the base moves.

## 3. When to use it

- A variable that matters has meaningful non-response and a base convention must be chosen and stated.
- Don't know or refused options exist and their treatment will move a reported figure.
- Bases differ across questions and somebody needs to know whether that is routing, non-response or both.
- Partial completes exist and a decision is needed about whether and where to include them.
- An analysis requiring complete cases (a model, an index, a segmentation) is about to discard cases silently.
- Non-response is concentrated in one subgroup, one question type or one part of the questionnaire.
- Somebody has proposed imputation, or has already imputed, and the defensibility must be assessed.
- A report needs a missing-data disclosure that a reviewer would accept.

## 4. When NOT to use it

- **The gap is a defect rather than a non-response.** A value lost through a merge error, a column that failed to export, a field wrongly typed and nulled on conversion: these are data faults, and the treatment is to recover the data, not to model around its absence. Go back to **04.01 Data Validation** and **04.02 Data Cleaning**. Modelling the absence of data that still exists somewhere is a waste of a defensible dataset.
- **The respondents were routed away.** This is the boundary that matters most and it is not a missing-data question at all. A respondent who was not asked a question is out of base for that question, permanently and correctly. They are never counted in a missing figure, never imputed, and never included in a denominator. Handling them belongs to base definition in **05.01 Descriptive Analysis**, which inherits the base register. If a request to "handle the missing data" turns out on inspection to be a request to handle routed-out respondents, the correct answer is that there is nothing to handle and the base is the base.
- **The intention is to fill gaps so that a table looks complete.** Per K4 §1, a slot in the output is not evidence that a number belongs in it. Imputing to avoid an empty cell, or to avoid explaining a moving base, converts an honest gap into a dishonest figure. The empty cell with its explanation is the better deliverable.
- **The missingness is so extensive that the variable cannot support analysis.** Above roughly a third missing on a variable, no treatment rescues it: complete-case analysis describes a self-selected two-thirds, and imputation is inventing the majority of the distribution from the minority. Report the response rate on the item, report the profile of who answered, and say the variable cannot carry the analysis, per K4 §1. That is a finding about the instrument and belongs in the methodology note.
- **The whole sample is missing on a construct the study promised to measure.** If a question was not asked, or was asked of the wrong people, no missing-data method fills the gap. The honest output is that the objective cannot be answered from this data. Route to **13.01 Research Quality Review** where the shortfall affects a commissioned deliverable.
- **Multiple imputation or maximum-likelihood methods are being considered on a non-probability commercial sample without the assumptions being stated and checked.** The methods are legitimate in the right setting. They rest on assumptions about why the data is missing that a commercial sample rarely allows anyone to check, and they produce complete-looking data whose completeness is a modelling artefact. Where such a method is genuinely warranted, it belongs with **05.06 Correlation, Regression and Causal Claim Control** and requires the assumption statement, the sensitivity analysis and the disclosure that go with it.
- **The decision is really about weighting.** If certain groups are under-represented because they did not take part at all, that is unit non-response and a composition problem, addressed in **04.05 Weighting and Base Management**. This skill covers unit non-response only insofar as it must be described and disclosed; it does not adjust for it.

## 5. Required inputs

**Required.** Without these the skill cannot run. If absent, ask. Where the work must proceed, choose the conservative option, state it where it bites, and log it, per K5 §5.

- **The cleaned dataset**, with 04.02's log, so that gaps created by cleaning (an invalid value recoded to missing) are distinguishable from gaps that arrived that way.
- **The missing-value code scheme**: the distinct codes used for not asked, not answered, refused, don't know and system missing. Where one code covers several states, that is the first finding, and it constrains everything that follows.
- **The routing and filter logic**, which is what separates out-of-base from missing. Without it, the primary distinction in this skill cannot be made and no base is trustworthy.
- **The questionnaire as fielded**, including which questions offered a don't know or prefer-not-to-say option, because a question without the option produces different non-response behaviour from one with it, and the two must never be compared or pooled.
- **The missingness map from 04.01**, by variable, by respondent and by position in the instrument.

**Optional, and what each one adds.**

- **Paradata (break-off point, per-question timing, device, mode):** distinguishes a deliberate skip from an abandonment, and is what allows a partial complete to be characterised rather than guessed at.
- **The full profile of every respondent including those who broke off:** allows the responders and non-responders on a given item to be compared on observed characteristics, which is the only empirical purchase available on the missingness mechanism.
- **The analysis plan:** determines which variables the decision must protect, and reveals in advance which multivariate analyses will trigger large listwise losses.
- **Fieldwork records including invitation and response counts:** allow unit non-response to be described honestly rather than left implicit in the achieved sample.
- **The previous wave's missing-data conventions:** required for tracker comparability, since a base convention changed between waves produces a movement that is not change.
- **Any sample frame information about non-responders:** the only way to say anything empirical about unit non-response bias rather than merely acknowledging the risk.

## 6. Questions to ask before starting

1. **Are the five states distinguishable in the data, or has the file collapsed them?** Determines whether any of the analysis below is possible. *Default if unanswered:* reconstruct what can be reconstructed from routing, report which states cannot be separated, and treat the collapsed group conservatively (as non-response, outside the base) while stating that the treatment is forced by the coding rather than chosen.
2. **Which variables have missingness above the threshold at which it changes a conclusion?** Focuses effort and identifies what needs a stated convention. *Default:* investigate any analysed variable above roughly 5% non-response, and treat above roughly 10% as requiring an explicit disclosure next to the figure per K4 §4.3.
3. **Did the question offer a don't know option?** Determines what non-response means. Absence of the option pushes people who genuinely do not know into either a substantive answer they do not hold or a skip, and both distort the figure. *Default:* record the option's presence per question, never pool questions that differ on it, and note the instrument effect where a mid-point or skip is inflated.
4. **Is non-response on this variable plausibly related to what the answer would have been?** This is the mechanism question and it changes everything about how the figure should be read. *Default:* compare responders and non-responders on observed variables, report what that shows, and where non-response is on a sensitive topic, assume relatedness and disclose it rather than assuming randomness.
5. **What counts as a complete for this study, and was that defined before fieldwork?** Determines the partial-complete decision. *Default:* define a complete as having reached the end of the questionnaire, report partials separately, and include them only in analyses of the questions they actually answered, with the base stated.
6. **Is this a tracker, and what convention did previous waves use?** A base convention change is indistinguishable from real movement. *Default:* match the previous convention, and where a better one exists, report both for one wave and state the difference.
7. **Will any multivariate analysis be run on these variables?** Listwise deletion across ten variables each 5% missing can silently discard a third of the sample. *Default:* calculate and report the complete-case base for every planned multivariate analysis before it is run, not after.

## 7. Step-by-step methodology

**1. Separate the five states before counting anything.** For every variable, resolve each empty or special-coded cell into exactly one of: **not asked** (routed away, out of base, not missing); **not answered** (asked, skipped, no explicit refusal); **refused** (explicitly declined); **don't know** (explicitly selected); **lost** (technical failure, break-off, or a value destroyed in processing). Where the file uses one code for several states, reconstruct what routing allows and report the rest as unresolvable. *Correct result:* a five-state table per variable in which not-asked is reported separately and never entered into any missing total. If the only output of this skill were this table, it would already have prevented the commonest error in the area.

**2. Restate every base with routed-out respondents removed.** The denominator for a filtered question is those who should have been asked. Non-response then sits inside that denominator as a decision, not outside it as an accident. Reconcile per question: total sample, minus routed out, equals those asked; those asked, minus each non-response state, equals those giving a substantive answer. Both bases have legitimate uses and they are different numbers. *Correct result:* a base table with three columns (asked, substantive answers, and the non-response states in between) that reconciles for every question.

**3. Decide the status of don't know and refused as answers, not as gaps.** A don't know on an awareness, opinion or knowledge question is usually a substantive answer: not knowing is a real state and often the most important finding on the question. A don't know on a factual question about the respondent's own behaviour is more often a recall failure. A refusal on income, health or political questions is informative about sensitivity and frequently correlated with the answer. The general rule: **don't know belongs inside the base wherever not knowing is a possible true state of the world for that question**, and reporting it as a category is more honest than excluding it. Refusals normally sit outside the substantive base and are reported as a separate figure. The one error that has no defence, in either case, is arithmetic: **a don't know is never a scale midpoint and never a zero.** A midpoint asserts a neutral opinion the respondent did not express; a zero asserts an absence of the thing being measured. Both invent data, per K4 §2.1. *Correct result:* a written rule per question type, applied consistently, with don't know reported as its own category wherever it exceeds a few per cent.

**4. Diagnose the missingness mechanism informally, and say which of three situations you are in.** Formal terminology is available; what matters practically is which of three cases applies, because the consequences differ sharply.

- **Unrelated to anything.** Non-response looks random with respect to both the answer and observed characteristics: a scattered pattern, no concentration by subgroup, question or route. Consequence: the responding subset is a smaller version of the whole, complete-case analysis loses precision but not accuracy, and the honest treatment is simply the reduced base with the response rate stated.
- **Related to something else observed.** Non-response concentrates in an observed group: older respondents skipping a digital question, one market with a translation problem, one device type timing out. Consequence: the responding subset is compositionally different, subgroup figures remain usable, the total is biased toward the responding composition, and the disclosure must name the pattern. This is the case where a weighting adjustment can sometimes help, and it belongs to 04.05 with the assumption stated.
- **Related to the answer itself.** People who would have given a particular answer are the ones not answering: high earners skipping income, dissatisfied customers abandoning a satisfaction battery, non-users of a service unable to answer questions written for users. Consequence: **the observed figure is biased in a knowable direction and no treatment inside the dataset fixes it**, because the information needed is precisely what is missing. The only honest response is to report the figure with the direction of likely bias stated, and to say so next to the finding rather than in an appendix.

Test empirically where you can: compare responders and non-responders on every observed variable, and check whether non-response on the item correlates with related answered items. Where the topic is sensitive, or where the question asks about something people have a reason to conceal, assume the third case unless the comparison shows otherwise. *Correct result:* a named diagnosis per material variable with the evidence for it, and an explicit statement that the third case cannot be fixed by any method, only disclosed.

**5. Apply listwise treatment as the default, with an honest base statement.** For most commercial research the correct handling is to analyse those who answered, state the base, state the non-response, and move on. It is transparent, it invents nothing, and its weakness (that the responders may differ from the non-responders) is disclosed rather than hidden by a method that makes the same assumption invisibly. The default is wrong in three identifiable situations, and only these: when the resulting base falls below a reportable threshold; when the missingness is diagnosed as related to the answer, in which case listwise analysis produces a biased figure that looks clean and must carry the bias statement; and when a multivariate analysis would compound small per-variable losses into a large and non-random complete-case subset. *Correct result:* a stated default with the exceptions identified by name, and for each exception a decision made explicitly rather than by omission.

**6. Calculate the complete-case base for every planned multivariate analysis before running it.** Ten variables at 5% missing each can leave a complete-case base anywhere between 50% and 95% of the sample depending on whether the same people are missing throughout. Compute it, compare the complete-case subset with the full sample on key characteristics, and report both. Where the loss is large or the subset is compositionally different, the choice is between reducing the variable set, analysing on a pairwise basis where the method allows it with the varying bases reported, or accepting the reduced base with its profile disclosed. Silently proceeding is the one option that is not available. *Correct result:* a table per planned analysis showing full n, complete-case n, and the profile difference between them.

**7. Decide imputation deliberately, and expect the answer to be no.** State it plainly: **in most commercial research, imputation is not defensible and should not be used.** The reasons are practical rather than doctrinal. Imputation produces values indistinguishable in the file from collected ones. It requires an assumption about the missingness mechanism that a non-probability sample rarely permits anyone to test. Simple methods (mean, mode, last observation) are actively harmful: they shrink variance, strengthen apparent relationships, and put a value at the centre of a distribution for the people most likely to sit at its edges. And the commercial benefit is usually cosmetic, since the alternative is a stated base and a disclosed response rate, which costs nothing but an explanation.

The narrow cases where it can be defended: a derived total where all components but one are present and the total is arithmetically determined; a longitudinal design carrying forward a value with a stated and tested stability assumption; a formal method applied by someone who states the mechanism assumption, runs the analysis with and without it, and reports both. In every permitted case, three requirements hold without exception: the imputed values are flagged in a separate variable, the analysis is reported with and without imputation, and the imputation is disclosed wherever the affected figures appear. *Correct result:* either a documented decision not to impute with the reason, or an imputation carrying its flag variable, its dual reporting and its disclosure.

**8. Handle unit non-response as description, not correction.** Item non-response is a gap inside a participating respondent's record; unit non-response is a person who is not in the dataset at all. The second cannot be diagnosed from inside the data, because the evidence left. What can be done is to describe it: report the invitation, start, break-off and completion counts; report the achieved composition against the intended quota; and where any frame information about non-responders exists, compare it. Where response rates are low or unknown, say so and state the non-response risk, per K4 §7. Composition adjustment belongs to 04.05, and it corrects for observed characteristics only, which is a point the disclosure must make rather than let a weighted figure imply otherwise. *Correct result:* a unit non-response paragraph in the methodology with real counts, and no implication that weighting has resolved it.

**9. Decide on partial completes against a definition set before fieldwork, and include them by question rather than by case.** A partial complete is a respondent who started, answered some questions, and stopped. The useful framing is not whether to include them but where: a respondent who answered the first two sections has given usable data for those sections and nothing for the rest. Include their answers in analyses of the questions they answered, exclude them from analyses of questions they did not reach, report the resulting base variation explicitly, and check whether break-off point correlates with anything, because break-offs cluster at hard, long, sensitive or badly rendered questions and that clustering is itself a finding about the instrument. Where partials are included, report the figures with and without them once, since break-off is rarely random. *Correct result:* a stated inclusion rule, a partial-complete count with break-off distribution by question, and a comparison of headline figures with and without partials.

**10. Write the missing data section of the preparation log, in the shared structure.** Every decision recorded as original issue (which variable, which states, what proportion), action (the treatment chosen), reason (including the mechanism diagnosis and why the default was or was not appropriate), and impact (the base it produces and the direction of any likely bias). Include the decisions not to act: a decision to report a variable with 22% non-response and disclose it is a decision, and it belongs in the log next to the ones that changed something. *Correct result:* a log section from which a reader can see every base convention in the study and the reasoning behind each, without opening the data.

**11. Write the disclosure, and attach it to the figures rather than to the appendix.** Per question or per measure where it is material: the base asked, the base answering, the response rate on the item, the treatment of don't know and refused, and where the mechanism diagnosis suggests bias, its direction. A figure carrying 18% item non-response with no note on the page is a figure that will be quoted as though it had none, per K2 §7. *Correct result:* a disclosure that travels into every downstream deliverable attached to the number it qualifies.

## 8. Analytical framework

Every missing-data decision is written in the category's shared log structure, with a mechanism step inserted between issue and action:

    Original issue → State classification → Mechanism diagnosis → Action → Reason → Impact

**Original issue.** The variable, the proportion missing, and where in the instrument it sits.
**State classification.** How the gap resolves across the five states, with not-asked separated out and excluded from the missing total.
**Mechanism diagnosis.** Unrelated, related to something else observed, or related to the answer itself, with the evidence and the consequence of each.
**Action.** Listwise with a stated base, retain as a category, exclude the variable, flag, impute (rarely), or report and disclose.
**Reason.** Why this action follows from the diagnosis, and why the default was or was not right here.
**Impact.** The base it produces, the effect on the reported figure, and the direction of any bias that survives the treatment.

Against the K2 chain, this skill governs what the Finding is a finding about. A percentage on a variable with 30% non-response is a finding about the responders, and that qualification is part of the finding rather than a footnote to it.

## 9. Output format

**1. Missing data disclosure block.** Item non-response summary across the study, the treatment rules adopted, the don't know convention, the partial-complete rule, whether anything was imputed, and the study's unit non-response position.

**2. Five-state table**, per variable.

| Variable | Total sample | Not asked (out of base) | Asked | Not answered | Refused | Don't know | Lost | Substantive answers | Item response rate |
|---|---|---|---|---|---|---|---|---|---|

Not asked never enters a missing total, and the table is laid out so that it cannot.

**3. Mechanism diagnosis table**, for every variable above the materiality threshold.

| Variable | % non-response | Responder vs non-responder differences observed | Diagnosis | Evidence | Consequence for the figure | Direction of likely bias |
|---|---|---|---|---|---|---|

**4. Treatment decisions**, in the shared log structure: issue, state classification, mechanism, action, reason, impact.

**5. Complete-case table**, for every planned multivariate analysis: full n, complete-case n, percentage retained, and the profile difference between the two.

**6. Partial complete report.** Count, break-off distribution by question, inclusion rule applied, and headline figures with and without partials.

**7. Unit non-response statement.** Invitations, starts, break-offs, completes, achieved versus intended composition, and the non-response risk stated.

**8. Imputation record**, present only if imputation was used: method, variables, cases, flag variable name, assumption stated, and results with and without.

**9. What could not be established**, per K3 §5.2: variables whose missingness makes them unusable, and objectives affected.

**Where the evidence is thin**, the format does not get filled. A mechanism that cannot be diagnosed because no comparison variables exist is reported as undiagnosed with the consequence stated, not assigned a plausible label. A variable above roughly a third missing gets a response rate, a responder profile and a statement that it cannot support analysis, not a treated figure. An empty imputation section is the expected result, not a gap. Per K4 §1, a table with a column for a treatment is not a reason to apply one.

## 10. Quality checks

Run before any figure affected by missing data is presented. These sit on top of K4 §8.

1. Are the five states separated for every variable, with not-asked excluded from every missing total?
2. Does every base reconcile: total, minus routed out, minus each non-response state, equals substantive answers?
3. Is the don't know convention stated once and applied to every question of the same type?
4. Has any don't know been scored as a midpoint, a zero, or any other numeric value?
5. Has any routed-out respondent been included in a denominator or in a missing count?
6. Does every variable above the materiality threshold carry a mechanism diagnosis with its evidence?
7. Where missingness is diagnosed as related to the answer, does the direction of likely bias appear next to the figure rather than in an appendix?
8. Has the complete-case base been calculated and reported for every multivariate analysis, before it was run?
9. Are all imputed values flagged in a separate variable, with results reported both ways?
10. Is the partial-complete rule stated, with the break-off distribution and the with-and-without comparison?
11. Is unit non-response described with real counts rather than implied by the achieved sample?
12. Does every reported figure carry its item response rate where non-response is material?
13. For a tracker, does the base convention match the previous wave, or is the change documented and quantified?
14. Is any variable being reported that the missingness makes unreportable?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **Routed-out counted as missing** | A "missing" figure that roughly equals the proportion filtered out of the question | Five-state table with not-asked in its own column, excluded from every missing total by construction |
| **Don't know as midpoint** | A mean score that includes don't know respondents scored at the centre of the scale | Don't know is a category or it is out of base; it is never a number |
| **Don't know as zero** | A frequency or spend variable where non-response has become an absence of the behaviour | Zero means the respondent has none; missing means we do not know. They are not the same and cannot share a code |
| **Moving base, unexplained** | Tables where n changes question to question with no note | Base table showing asked, non-response states and substantive answers for every question |
| **Silent listwise deletion in a model** | A model reporting n=612 on a study of 1,000 with no mention of it | Complete-case base calculated and reported before the model runs |
| **Mechanism assumed random by default** | Non-response treated as noise on a sensitive question | Compare responders and non-responders on observed variables; assume relatedness on sensitive topics unless shown otherwise |
| **Imputation as tidying** | A dataset with no gaps and no imputation flag variable | Imputation requires a flag, a dual report and a disclosure, or it does not happen |
| **Mean imputation on a skewed variable** | Reduced variance, a spike at the mean, strengthened correlations | Do not mean-impute; the people missing on income are systematically not at its mean |
| **Partials included silently** | Bases that exceed the stated complete count | Inclusion rule stated; partials included by question, never by case; figures reported with and without |
| **Unit non-response described as weighted away** | A methodology implying weighting corrected non-response bias | Weighting adjusts observed composition only; state that unmeasured non-response bias remains |
| **Convention drift across waves** | A tracker movement coinciding with a base convention change | Match the previous convention; report both for one wave if changing |
| **AI: filling a cell because the format has one** | A complete table where the underlying data was not complete | Per K4 §1, format is not evidence; empty cells with explanations are valid output |
| **AI: labelling a mechanism without testing** | A confident "missing at random" with no comparison behind it | Diagnosis requires the responder versus non-responder comparison, or it is reported as undiagnosed |
| **AI: reporting a percentage on an unusable variable** | A clean-looking figure on a variable that is 40% missing | Above roughly a third missing, report the response rate and the responder profile, not the percentage |

## 12. AI guardrails

Skill-specific only. K4 applies in full and is not repeated here.

1. **Never include a routed-out respondent in a missing count or in a denominator.** Not asked is out of base and is a separate concept from missing.
2. **Never assign a numeric value to a don't know or a refusal.** Not a midpoint, not a zero, not a mean, not a neutral score. A don't know is a category or it is out of base.
3. **Never merge the five states into a single missing figure**, and never report a missing percentage without saying which states it covers.
4. **Never impute without a stated mechanism assumption, a flag variable, dual reporting and a disclosure.** In the absence of any one of these, do not impute.
5. **Never use mean, mode or last-observation imputation on a variable that will be analysed**, since these methods distort the distribution in a known direction and produce a false impression of precision.
6. **Never state a missingness mechanism without the evidence for it.** Where the comparison cannot be run, report the mechanism as undiagnosed and state the consequence.
7. **Never run a multivariate analysis without first reporting the complete-case base and its profile difference from the full sample.**
8. **Never present a figure from a variable with substantial non-response without the item response rate in the same place**, per K4 §4.3.
9. **Never describe weighting as having corrected non-response bias.** It adjusts observed composition. Unmeasured non-response bias survives it.
10. **Never fill a gap because an output format has a slot for it.** An empty cell with an explanation is a valid and often superior output, per K4 §1.
11. **Never change a base convention between tracker waves without quantifying the effect and reporting both.**

## 13. Best-practice principles

- **Not asked is not missing.** Half the confusion in this area disappears the moment routed-out respondents are removed from the conversation. Do it first, every time.
- **Don't know is data.** On awareness, opinion and knowledge questions it is frequently the most important answer on the page. Treating it as an absence discards a finding and inflates every other category at the same time.
- **The honest handling of missing data is usually to report it, not to fill it.** A stated base and a disclosed response rate cost an explanation. A filled dataset costs the reader the ability to tell measurement from arithmetic, and they cannot recover it.
- **The mechanism determines the consequence, and only one of the three is fixable.** Random costs precision. Related-to-observed costs composition and can sometimes be adjusted. Related-to-the-answer costs accuracy and cannot be fixed by anything inside the dataset, only disclosed.
- **Where a question is sensitive, assume non-response is related to the answer.** People who decline to state income are not a random sample of earners. The burden of evidence sits with the claim of randomness, not against it.
- **Imputation makes invented values indistinguishable from collected ones.** That is its defining property and the reason for the flag variable. Any imputation scheme without a flag is a scheme for losing track of what is real.
- **Listwise deletion is an assumption, not a neutral act.** Analysing complete cases assumes the incomplete ones would have looked the same. State the assumption instead of letting the reduced base imply there was never a decision.
- **Small per-variable losses compound multiplicatively in multivariate work.** The complete-case base is the base of the analysis, and it needs the same disclosure as any other base.
- **Break-off is a finding about the questionnaire.** Where people stop is diagnostic of length, difficulty, sensitivity and rendering, and that information is worth more than the partial answers themselves.
- **A base that moves is fine. A base that moves unexplained is not.** Readers can handle varying denominators when they are told why.
- **Every decision not to act is still a decision.** Reporting a variable with 22% non-response and disclosing it belongs in the log next to the treatments that changed something.

## 14. Worked example

**INPUT.** A fictional university researches student wellbeing and service use across three campuses. 3,400 invitations, 1,180 starts, 940 completes, 240 partials. The instrument includes a sensitive section on financial hardship and a service-satisfaction battery routed only to service users.

**PROCESS.**

*Step 1.* Five-state resolution. The satisfaction battery shows 61% "missing", which resolves almost entirely to not asked: 574 of 940 completers never used a service and were routed away. Actual non-response among those asked is 4%. The apparent problem was a base problem wearing a missing-data costume, and the corrected statement is that satisfaction is based on 366 service users with a 96% item response rate.

*Step 2 to 3.* The financial hardship question offered "prefer not to say"; 14% selected it, 3% skipped. The wellbeing scale offered no don't know option and shows a mid-point pile-up of 29%. Rule adopted: prefer-not-to-say reported as its own category and excluded from the substantive base with both bases shown once; the wellbeing mid-point is reported as observed with a note that the absence of a don't know option inflates it, which is a finding about the instrument for the next wave.

*Step 4, and the judgement call.* Are the 14% prefer-not-to-say responses on hardship random? Comparison of responders and non-responders on observed variables shows non-responders are disproportionately in the highest and lowest self-reported income bands, and disproportionately international students. **Diagnosis: related to the answer itself, at both ends.** The consequence is stated plainly: the observed hardship rate among those who answered is 31%, and because non-response concentrates at both extremes, the direction of bias is not one-sided but the figure is unreliable as a population estimate. A draft that reported 31% as the student hardship rate is corrected to report it as the rate among the 83% who answered, with the non-responder profile stated on the same page. **Researcher review recommended** per K5 §2.4, since a hardship figure that under-represents the most vulnerable students has consequences for how the university uses it, and that is a judgement about use rather than analysis.

*Step 6.* The planned model uses eight variables. Complete-case base: 811 of 940, or 86%. The 129 excluded are disproportionately partial-section completers and disproportionately from one campus. Reported before the model runs, with the profile difference, and the campus skew disclosed as a limitation on the model rather than discovered afterwards.

*Step 7.* Imputation proposed for the hardship variable to preserve model n. **Declined**, and the reason logged: the missingness is diagnosed as related to the answer, so any imputation model would be built on exactly the wrong population, and the imputed values would be most wrong for the students the analysis exists to identify. The model is reported on 811 complete cases with the hardship variable's non-response disclosed, and the alternative of dropping hardship from the model is reported alongside so the reader can see what the variable contributes.

*Step 9.* The 240 partials break off at two points: 108 at the hardship section, 79 at question 34 of a 41-question instrument. The hardship cluster corroborates the mechanism diagnosis at step 4 independently. Partials are included for the sections they completed, bases stated, and headline figures reported with and without: wellbeing mean moves 0.1 of a scale point, service use moves 2 percentage points because partials skew toward non-users.

**OUTPUT.** A five-state table showing the satisfaction base was never a missing-data problem; a don't know convention stated once; a mechanism diagnosis on hardship with its evidence and its direction-of-bias statement carried onto the page with the figure; a complete-case table for the model with its campus skew; a logged and reasoned refusal to impute; a partial-complete report with break-off clustering that corroborates the hardship diagnosis; a unit non-response statement (3,400 invited, 1,180 started, 940 completed); and one review point on the use of a hardship figure that under-represents the most vulnerable.

## 15. Advanced usage

**Missingness as substantive evidence.** Where people stop answering, and which questions they decline, is frequently the most informative pattern in an instrument. A cluster of refusals on one topic, a break-off spike at one question, or a demographic gradient in item non-response are findings about sensitivity, comprehension and burden. Report them in the methodology note and carry them into the next instrument via **02.01 Survey Questionnaire Design**.

**Trackers and comparability.** Base conventions and don't know treatment must match across waves or the series is not a series. When a convention must change, run one wave both ways, publish the difference, and record the version. A change of convention that coincides with a real market movement is unrecoverable afterwards, because the two effects cannot be separated retrospectively.

**Multi-market missingness.** Item non-response varies systematically by culture, mode and translation quality. Refusal rates on income and political questions differ by market for reasons that have nothing to do with the respondents' finances or politics. Diagnose per market before pooling, and never let one market's high non-response be absorbed into a global figure without a note, per K5 §2.2.

**Where imputation is genuinely warranted.** Longitudinal designs with a measured stability assumption, arithmetically determined totals, and formal multiple-imputation frameworks applied by an analyst who states and tests the mechanism assumption. In those cases the standard is higher, not lower: report the analysis with and without, report the between-imputation variability, and coordinate with **05.06 Correlation, Regression and Causal Claim Control**.

**Interaction with weighting.** Where the mechanism is related to an observed characteristic, a weighting adjustment can improve the composition of the responding set on that characteristic. It does nothing for missingness related to the answer, and applying it there produces a figure that is both adjusted and biased, with the adjustment providing false reassurance. Coordinate with **04.05**, and make sure the disclosure says which of the two situations the weighting addresses.

## 16. Skill chain

**Recommended previous skills:**
- **04.01 Data Validation.** Hands over the missingness map by variable, respondent and questionnaire position, with the four states already separated, which is what this skill resolves into treatments.
- **04.02 Data Cleaning.** Hands over a cleaned file in which invalid values have become defined missing codes, so that gaps created by cleaning are distinguishable from gaps that arrived.
- **02.05 Survey Logic and Flow Review.** Hands over the routing that separates out-of-base from missing, which is the first distinction this skill makes.

**Recommended next skills:**
- **04.05 Weighting and Base Management.** Takes the composition of the responding sample and the diagnosis of which non-response relates to observed characteristics, which is the only kind weighting can address.
- **04.04 Data Transformation and Dataset Preparation.** Takes the missing-value conventions, which every derived variable, net and index must respect, since a derived variable built across items with different missing treatments inherits all of them.
- **05.01 Descriptive Analysis.** Takes the base conventions and the item response rates that every reported figure must carry.
- **05.06 Correlation, Regression and Causal Claim Control**, where complete-case losses in multivariate work need handling with the assumptions stated.

**Runs well alongside:**
- **K3**, since missingness is one of the main things that should cap the confidence attached to a figure, and **K4 §7**, for the disclosure obligations on response rates.
- **13.04 Bias Detection**, where non-response related to the answer is a bias in the finding and not merely a gap in the data.
- **02.01 Survey Questionnaire Design**, which is where the instrument-level causes of non-response get fixed for next time.

---
A Yazi Supplied Skill and resource.