Why Missing-Value Recording Matters
A data table can look complete even when some entries do not contain measurements or answers. A blank cell, a dash, a code such as \(-9\), and the text “not applicable” might all be used to indicate that something was not recorded. But they do not necessarily mean the same thing. Some may indicate an unknown response, while others may say that the question did not apply. A number such as zero, meanwhile, may be a perfectly valid observation.
The goal is not to make every cell look filled. It is to represent what is known accurately and consistently, so that a reader or analyst can tell which values are observations and which entries have a different meaning. This is part of responsible data handling: the convention should be understandable, documented, and applied consistently.
Blank Cells, Codes, and Real Values
A blank cell is not self-explanatory. It could mean that no answer was provided, the question was not asked, the value was lost during transfer, or a person entered an empty field by mistake. If the data table contains only a blank, the reason may be unknowable. Do not infer a reason unless the study records support it.
Some data sources use visible codes for missing entries. For example, a survey export might use \(-9\) for “no response” or 999 for “not recorded.” These codes are not measurements merely because they look numeric. If a code is used for missingness, document its meaning and ensure it is not treated as a genuine observed value in calculations. A number such as \(-9\) could be a valid measurement in another context, so the code’s meaning depends on the variable and the documented convention.
Zero requires particular care. For a variable measuring the number of appointments attended, zero means that the person attended no appointments. It is an observed count, not an absent answer. Replacing zero with a missing-value code would erase information and could change summaries such as the mean.
| Entry in a field | Possible interpretation | What to check |
|---|---|---|
| Blank cell | No value displayed; the reason may be unknown | Was it unanswered, not asked, or lost? |
| \(-9\) | Could be a missing-value code | Does the data documentation assign it that meaning? |
| 0 appointments | A valid observed count | Is zero possible for this variable? |
| “Not applicable” | May indicate ineligibility, or may be a substantive answer category | What question and response options give it meaning? |
“Not Applicable” Is Not the Same as “Unknown”
“Not applicable” needs a context-specific interpretation. If a person was not eligible to receive a question under the study rules, the question may not apply to that person. That differs from an eligible person who was asked but did not answer: in the second case, the answer is unknown. Both situations lack an observed answer to the measurement question, but the reason matters and should not be obscured.
In another setting, “Not applicable” may be an actual response category that a participant selects, perhaps to communicate that none of the listed descriptions fits. If it is a meaningful answer offered by the instrument, it is a valid category for that variable. Do not automatically convert every appearance of those words into missingness. Determine whether the value is a response or a status describing why no response exists.
Worked Example: Zero Appointments and One Blank
A hypothetical clinic data table records the number of follow-up appointments attended by five people. The entries are 0, 2, blank, 0, and 1. Assume the blank indicates no recorded answer, and the data source does not specify why it is blank.
The four recorded values are 0, 2, 0, and 1. All four are valid nonnegative counts. In particular, each zero means that the person attended no follow-up appointments; neither zero is missing.
There are five records and four observed values, so the number of missing entries is:
The observed count total is \(0+2+0+1=3\) appointments across the four people with recorded values. If we calculate the mean among those with observed values, it is:
This mean describes the four people with recorded appointment counts, not all five people. The fifth person's count is unknown from these data. Treating the blank as zero would instead produce \(3/5=0.60\) appointments, but that calculation would assert an unrecorded answer that the table does not establish.
Choosing and Documenting a Consistent Convention
A consistent convention lets users distinguish observed answers from missing entries without guessing. The convention should fit the file format and preserve the information available in the source. In an analysis-ready table, a system-recognized missing value is often preferable to an ordinary number or text label that software might mistakenly calculate as data. If the original source uses codes such as \(-9\), retain the untouched source as required by the data-handling plan and record how that code is represented in the analysis-ready version.
When the reason for missingness is known and useful, record it separately rather than hiding several meanings behind one code. For example, the value field for a numeric measurement might be missing, while a separate status field records “not asked,” “not applicable,” or “no response.” Use only distinctions supported by the data collection process; do not invent a reason for a blank after the fact.
A useful written convention states which symbol or system value represents missingness, what each source code means, how “not applicable” is handled, and whether the reason for absence is recorded separately. If a categorical question includes “Not applicable” as a genuine response option, document that it is retained as a valid category. The convention should also make clear that valid zeros remain unchanged.
Identify blanks, special codes, and text labels, then consult the available data documentation or collection rules to determine what each represents.
Keep valid responses such as zero as observed values. Mark entries that do not contain an observed answer as missing under the chosen convention.
If the source distinguishes nonresponse from ineligibility, retain that distinction in a documented status field or equivalent record.
Write down the meaning of each code and the representation used in the analysis-ready data, including how meaningful “not applicable” responses are treated.
Worked Example: Translating a Special Code
Worked Example: Translating a Special Code
A hypothetical questionnaire export records weekly exercise sessions as 0, 3, \(-9\), and 2. The study codebook says that \(-9\) means “participant did not provide an answer”; zero means “provided an answer of no sessions.” Suppose the analysis-ready file uses a system-level missing value.
There are four records. Three have observed exercise counts: 0, 3, and 2. One entry, \(-9\), is a missing-value code according to the codebook, not a count of negative nine sessions. The analysis-ready values are therefore 0, 3, missing, and 2. The convention should state both that \(-9\) in the source means no answer and that it is represented as missing in the analysis-ready file.
The observed count total is \(0+3+2=5\) sessions. Among the three recorded answers, the mean is:
The denominator is three because there are three observed answers. Using four would implicitly treat the missing code as an observed fourth measurement, which contradicts the documented meaning. The zero remains in both the total and the denominator because it is a valid response.
Completeness Depends on the Right Denominator
A count or proportion of missing entries is useful only when its denominator matches the question being asked. One can describe the fraction of all records with an observed value, or the fraction of eligible people who answered. Those quantities need not be the same when some people were not eligible for a question.
For an eligible group, the response-completion proportion is the number of eligible people with an observed answer divided by the number of eligible people. This denominator excludes those to whom the question did not apply. The calculation does not turn missing answers into observed ones; it simply describes how complete the recorded answers are within the group for whom the question was relevant.
Worked Example: Eligible, Not Applicable, and Unanswered
In a hypothetical follow-up data set, 12 people were included. Ten were eligible for a question about the number of follow-up appointments attended and were asked it; two others were not eligible under the study rules, so the question was not applicable to them. Of the ten eligible people, nine provided an answer and one did not. The nine answers include valid counts, including any zero responses.
For the question's observed values, there are nine recorded answers and one eligible unanswered case. The two ineligible people have a not-applicable status, not an unknown answer to a question they should have answered. Across all 12 records, nine have an observed count and three have no count: one eligible nonresponse and two not-applicable cases.
The completion proportion among eligible people is:
Thus 90% of eligible people have a recorded answer. It would be misleading to report \(9/12=75\%\) as the response-completion proportion, because that denominator includes two people for whom the question was not applicable. The value \(9/12\) can describe the fraction of all included records with an observed count, but it answers a different question and should be labelled that way.
Common Mistakes and Why the Convention Matters
- Replacing blank cells with zero. This turns an unknown response into a specific observed answer. Keep it missing unless the source establishes that the person actually gave a zero.
- Treating every numeric code as a measurement. Check the documented meaning of codes such as \(-9\) before calculating. A code reserved for missingness must not enter arithmetic as if it were a response.
- Combining nonresponse and ineligibility without noting the difference. Both may have no measurement value, but they have different explanations. Retain the distinction when it is known and relevant.
- Assuming “not applicable” always means missing. It may be a valid response category in one question and an ineligibility status in another. Use the question and study rules to determine its meaning.
- Reporting a completion percentage with an unclear denominator. State whether the denominator is all records or only those eligible for the question.
A consistent missing-value convention protects the meaning of both missing and observed data. It prevents valid zeros from disappearing, prevents placeholder codes from becoming false measurements, and makes clear which cases were eligible to provide an answer. These distinctions are basic, but they affect every later summary that uses the recorded values.
Check Your Understanding
Answer each question by identifying what is observed, what is missing, and what the denominator should represent.
- A data column for number of missed classes contains 0, 1, blank, and 2. Which entries are observed, and how many values are missing?
- A codebook states that 999 means “not recorded.” Why should 999 not be included as a measured value?
- When might “Not applicable” be a valid category rather than a missing-value status?
- Eight people are eligible for a question, and six provide an answer. What is the response-completion proportion among eligible people?
- What information should a missing-value convention document when the source uses one code for nonresponse and another for ineligibility?