Never Use Zero to Mean Not Applicable

Never use zero to mean not applicable if zero is a real score on the scale. The minute you do that, every mean, percentage, and benchmark that includes the field is mathematically wrong, because the code for missing data becomes part of the data itself.

Why zero is the wrong sentinel

A sentinel value is a stand-in that tells your system, “this response is not a real measurement.” That only works if the stand-in cannot be confused with a valid answer. On a 1 to 5 scale, 0 is a valid number in the spreadsheet, but it is not a valid response on the instrument. That makes it tempting, because it is easy to type and easy to sort. It is also dangerous, because software does not know your intention. It only sees numbers.

If you save “not applicable” as 0, any calculation that does not explicitly exclude 0 will treat absence as a real low score. The damage is silent. A chart still renders, a mean still prints, and the result may even look plausible.

The arithmetic problem

Suppose you are collecting ratings on a 1 to 5 scale, and five supervisors respond to one item:

• 5, 4, 4, 3, not applicable

If you keep the missing response missing, the average of the four valid ratings is:

(5 + 4 + 4 + 3) / 4 = 16 / 4 = 4.0

If you encode “not applicable” as zero and compute the average across all five records, the result becomes:

(5 + 4 + 4 + 3 + 0) / 5 = 16 / 5 = 3.2

That is not a small formatting issue. It moves the score by 0.8 points on a 5-point scale, a 20 percent drop from 4.0 to 3.2. If you are reporting benchmark attainment, comparing sections, or looking for change over time, that one coding choice can change decisions.

The problem gets worse when missingness is uneven. If one site has more “not applicable” responses than another, the site with more zeros will look weaker even if its valid ratings are identical.

Why the bias is systematic

Zero does not just lower the mean. It also changes the denominator.

With proper missing-data handling, the denominator is the count of valid responses. With zero coding, the denominator becomes the count of all rows, including rows that should not have been scored at all. That means the reported average is no longer the average of observed performance. It is the average of observed performance plus administrative nonresponses.

The same issue affects:

• percent favorable calculations, if zero is treated as an unfavorable response
• item-level benchmarks, if “not applicable” is counted as failure
• subscale scores, if one missing item drags down the entire composite
• longitudinal trend lines, if a changing number of NAs alters the denominator from term to term

If you later filter the data in a dashboard, the damage may remain hidden. A user may see a single clean number with no signal that the denominator changed.

A real spreadsheet example

Consider a small file with six responses to a competency item:

A spreadsheet average returns:

(4 + 5 + 4 + 3 + 0 + 0) / 6 = 16 / 6 = 2.67

If those two zeroes are actually “not applicable,” the correct result is:

(4 + 5 + 4 + 3) / 4 = 16 / 4 = 4.0

The same raw values can produce two very different narratives: one says the group is below average, the other says the group is strong and two records should be excluded.

Better ways to represent not applicable

Use a true missing value, not a numeric stand-in. In practice, that means one of these approaches:

• blank cell in a spreadsheet, paired with documentation that blanks are excluded from calculation
• NA, NULL, or similar missing marker in a database
• separate coded field for response status, such as valid, missing, not applicable, or declined
• text labels in the raw file only if they are converted cleanly before analysis

The safest pattern is to store response status separately from the numeric score. For example, one column can hold the rating, and another can hold the response type. That lets you distinguish “student answered 2,” “student skipped,” and “item did not apply.” Those are not the same thing.

A useful rule

If a value can change the mean, it is part of the data. If it only explains why the mean should not be computed, it is metadata or status, not a score.

What to do in analysis and reporting

First, exclude nonresponses from the denominator unless the method explicitly says otherwise. Second, document the rule in the codebook and on any reporting layer that users can export. Third, audit the file for impossible scores, such as 0 on a 1 to 5 scale, before any dashboard refresh or accreditation report.

If you have legacy files that already use zero for not applicable, do not pretend the problem is cosmetic. Convert the field before analysis, or recode 0 to missing only after confirming that zero is never a valid response. If zero can mean both “lowest score” and “not applicable,” you cannot recover the truth from the column alone.

When zero is acceptable

Zero is fine when the instrument truly includes zero as a valid measurement, such as hours completed, incidents observed, or counts of events. In those cases, zero means none, not missing. The issue is not the number itself. The issue is whether the scale assigns zero a substantive meaning.

On most survey rating scales used in internships, faculty evaluations, and employer feedback, zero is not a valid point. That makes it a poor choice for not applicable. A clean missing code protects the math and the interpretation.

How this shows up in competency reporting

In multi-rater competency assessments, a single coding mistake can spread through several layers of reporting. Student self-ratings, supervisor ratings, and institutional summaries may all use the same item-level averages. If not applicable is stored as zero in any one layer, the composite can inherit the error and make the group look weaker than it is.

That is why systems designed for competency reporting should preserve missingness explicitly and separate it from scale values. In a structured reporting workflow, this is the difference between a score that reflects actual feedback and a score that reflects the file format. Tools such as the Career Readiness Report are useful partly because they keep that distinction clear in the data model and downstream reports.

The simplest test

Before publishing any summary, ask one question: if I replace every zero with missing, does the average change? If the answer is yes, then zero is not a safe proxy for not applicable.

That test is simple, but it catches a lot of bad reporting. If a field is supposed to measure performance on a scale, missingness must stay out of the arithmetic. Otherwise, the report is not averaging answers, it is averaging answers and absences together.