Why the distinction matters
A blank cell can have more than one meaning. It may represent a value that exists for the person but was not recorded, or it may mark a question that does not apply under the variable’s definition. Those cases should not automatically be treated alike. Missingness mechanisms such as MCAR, MAR, and MNAR describe why defined values are unobserved; they do not explain values that are not defined in the first place.
This distinction also affects the denominator for a missingness rate, the population to which an analysis refers, and whether a blank should be considered for imputation. The key question is not simply, “Is the cell empty?” It is: Is there a value that the variable is meant to have for this person?
The eligibility indicator \(E_i\) and response indicator \(R_i\) answer different questions. \(E_i\) concerns whether the variable applies. \(R_i\) concerns whether an applicable value was obtained. The usual missingness-mechanism question concerns \(R_i\) among records with \(E_i=1\). If a value is structurally absent, it is not a nonresponse to that value.
Define the variable before classifying the blank
Structural absence depends on how the variable is defined, not just on the questionnaire’s layout. For example, “number of cigarettes smoked per day among current smokers” is defined for current smokers; for nonsmokers, it is inapplicable under that definition. In contrast, “number of cigarettes smoked per day, with zero for people who smoke none” can be defined for everyone. A nonsmoker’s value is then zero, not structurally absent.
Questionnaire routing alone does not settle the classification. A question might be skipped because a respondent failed to follow instructions, an interviewer omitted it, or a survey system routed incorrectly. If the value is nevertheless defined for that person, the unanswered item is missing. Conversely, a valid skip rule may leave no value to record because the item applies only to a specified group.
The analysis target matters too. If the target is satisfaction with a service among people who used it, nonusers have no service experience to rate under that variable definition. If the target instead asks all residents to rate a clearly specified hypothetical service, a value is intended for everyone who can answer that question; a blank from a nonuser may then be missing. State the variable and target population explicitly before assigning a label.
Worked Example: Smoking Days and a Valid Skip
A hypothetical survey asks 120 adults whether they currently smoke. It then asks current smokers how many days in the past 30 days they smoked. Of the 120 adults, 30 report current smoking. Among those 30, 24 give a number of days and 6 leave the item blank. The other 90 are correctly routed past the item.
Suppose the variable is defined as “number of smoking days in the past 30 days among current smokers.” Then \(E_i=1\) for the 30 current smokers and \(E_i=0\) for the 90 nonsmokers. The six blank answers are missing because these respondents are eligible and their smoking-day values are defined. The 90 valid skips are structurally absent for this variable.
The eligible-item missingness rate is therefore \(6/30=0.20\), or 20%. Dividing the six missing answers by all 120 adults gives \(6/120=0.05\), or 5%, but that is not the missingness rate among people for whom the item applies. Counting all 90 structural absences as missing would instead produce \(96/120=0.80\), or 80%, which answers a different and misleading question.
MCAR, MAR, or MNAR can be considered for the six eligible nonrespondents, using the available information about them and the defined smoking-day values. The 90 ineligible records are not evidence for any of those mechanisms. The example does not establish which mechanism applies to the six blanks; that requires assumptions about how their response relates to observed and unobserved information.
Mechanisms apply within the eligible set
For a variable defined only when \(E_i=1\), the response probability of interest is \(P(R_i=1\mid E_i=1,\ldots)\). MCAR, MAR, and MNAR distinguish how response among eligible records may depend on data. In broad terms, MCAR makes eligible-record response independent of the data; MAR permits dependence on observed data after conditioning appropriately; MNAR allows dependence on the unobserved value even after that conditioning. The eligibility rule is not itself one of these mechanisms.
A skip indicator may be predictable from another recorded variable. That does not make structural absences MAR. The skip determines \(E_i\), while a missingness mechanism concerns \(R_i\) where a value is defined. If an eligible respondent’s value is missing and response depends on an observed predictor, that dependence can be relevant to MAR; if the same respondent was ineligible by definition, there is no missing \(Y_i\) to explain.
Worked Example: When a Skip Creates Missing Data
In a hypothetical adult health survey, annual income is defined for every adult, including adults with no income. A software error skips the income item for 18 of 200 adults whose recorded education level is “no high-school diploma.” The remaining 182 adults answer. The questionnaire intended to ask everyone.
Here all 200 adults are eligible: \(E_i=1\) for each person. The 18 blank income values are missing, not structurally absent. The skip rule’s appearance in the data collection system does not change the definition of annual income or the study’s intended target.
The overall eligible-item missingness rate is \(18/200=0.09\), or 9%. If the 18 skips occur only in the recorded education group, response is associated with an observed variable. That association is relevant when considering MAR, but it does not by itself prove MAR: mechanism classification also concerns whether response depends on unobserved income after conditioning on observed information. Calling the blanks structural would conceal genuine nonresponse and could wrongly remove records from the analysis.
Consequences for data handling and imputation
A structural absence should generally remain distinct from a missing value in the analysis data. It may be represented by a separate applicability code or eligibility indicator, provided the code cannot be mistaken for a measured value. The analyst should retain enough information to distinguish ineligible records from eligible nonrespondents.
In MICE, the distinction is consequential. An ordinary imputation model treats a missing cell as an unknown value of a variable defined for that record. Structural absences do not meet that description. Treating them as ordinary missing values may ask an imputation model to generate values where the variable has no meaning, alter the population represented by a conditional model, and distort summaries or associations. Eligibility and the variable’s definition should therefore be resolved before interpreting blanks as imputation targets.
The converse error is also important: marking a blank as structurally absent merely because a question was routed away can discard a real, defined value. If the variable applies but the response was not obtained, the blank remains missing even if a questionnaire or data-processing rule produced it. The earlier material on missingness mechanisms provides the framework for considering the missing eligible values; it does not turn inapplicable values into missing data.
| Situation | Eligibility and response | Appropriate interpretation |
|---|---|---|
| Item applies only to a defined subgroup; person is outside it | \(E_i=0\) | Structural absence for that variable |
| Item applies; answer was not obtained | \(E_i=1, R_i=0\) | Missing value; assess its mechanism among eligible records |
| Item applies; respondent reports none | \(E_i=1, R_i=1\), with a recorded zero | Observed value of zero, if zero is allowed by the definition |
| Question was skipped by a faulty route but applies to the person | \(E_i=1, R_i=0\) | Missing value, not structural absence |
Checks, pitfalls, and limits
Before calculating missingness rates or fitting an imputation model, examine the item wording, skip logic, codebook, and intended estimand. Cross-tabulate eligibility against observed response status. A clean route pattern can help identify structural absences, while eligible blanks may reveal nonresponse or a data-processing error. These checks classify the records; they do not prove MCAR, MAR, or MNAR.
- Wrong: “Every blank is missing, so I will count and impute them all.” Correction: First determine whether the variable is defined for each record. Only eligible, unobserved values are missing values for that variable.
- Wrong: “The questionnaire skipped the item, so the blank is structural.” Correction: A skip is structural only when the variable is inapplicable under its definition. A technical or accidental skip of an applicable item creates missing data.
- Wrong: “Nonsmokers have zero smoking days, so their blank is just a missing zero.” Correction: Decide whether the variable is defined as smoking days among current smokers or as days smoked for everyone, including zeros. Those are different definitions and may imply different eligible populations.
- Wrong: “Structural absence is MAR because it depends on an observed eligibility variable.” Correction: Eligibility is not response to a defined value. Assess the mechanism for \(R_i\) among records with \(E_i=1\).
There is no universal rule that a particular questionnaire item must be structurally absent or must be defined for everyone. The decision follows the scientific meaning of the variable and the analysis target. If those are not explicit, even a correctly programmed survey can yield ambiguous codes. Documenting the definition and eligibility rule is therefore part of the statistical analysis, not merely a data-cleaning detail.
Worked Example: Separating Three Kinds of Values
A hypothetical survey asks 200 adults about current employment. Its analysis defines “weekly hours worked in a paid job” only for people with a paid job. Of the 200 adults, 80 have a paid job: 68 report hours, and 12 leave the hours item blank. The other 120 have no paid job and are routed past the item.
For this definition, the 80 employed adults are eligible and the 120 others have structural absences. Among the eligible adults, 68 values are observed and 12 are missing. The eligible-item missingness rate is \(12/80=0.15\), or 15%. The 120 routed records are not additional missing hours values.
Now suppose the research question changes to “weekly paid work hours for every adult,” with no paid work explicitly coded as zero. Under that alternative definition, all 200 adults have a defined value. The 120 nonworkers should have recorded zeros, and failure to record those zeros would be a data-construction problem; the 12 blank responses among employed adults remain missing. Thus, the same raw questionnaire pattern can be classified differently under different variable definitions. The definition must match the question being analyzed.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A survey asks current dog owners how many walks they took with their dog last week. A non-owner is routed past the item. Under the stated definition, how should that blank be classified?
- A questionnaire asks every adult for annual personal income, but 10 eligible adults do not answer. What is the missingness rate if 100 adults were eligible, and what population should be considered for the item’s missingness mechanism?
- A dataset codes “not applicable” as 0 for a variable defined only among current smokers. Explain the potential problem and name one distinction the dataset should preserve.
- A survey system accidentally skips a question about age for 7 of 150 eligible participants. Is that structural absence or missing data? What is the eligible-item missingness rate?
Solutions. (1) Structural absence: under the stated definition, walks with one’s own dog last week do not apply to non-owners. (2) The rate is \(10/100=0.10\), or 10%. The mechanism is assessed among the 100 eligible adults, not among a broader group containing people for whom income was undefined. (3) Coding a structural absence as zero can make it look like a measured value and can distort summaries. Preserve whether each record is eligible, separately from whether an eligible value was observed. (4) It is missing data because age is defined for all 150 and the skip was accidental. The rate is \(7/150\approx0.0467\), or 4.67% of eligible participants.