Can an item with no variation have Pearson correlation zero?

Its correlation is undefined, not automatically zero.

AI Glossary / Statistics and probability / Advanced

Item-total correlation

Definition

Item-total correlation measures the association between responses to one questionnaire or test item and a score formed from a set of items. The uncorrected version includes the item in that total. The corrected item-total correlation, often called item-rest correlation, excludes it. Reports should identify which version they use because inclusion creates a part-whole relationship.

What it means

If people who score highly on the rest of a scale also tend to score highly on an item, the item has a positive item-rest correlation. A weak or negative relationship invites examination of the item's wording, coding, variability, and relevance to the intended scale.
This is an internal relationship among recorded scores. It does not establish that the scale measures its intended construct, predicts a useful external outcome, or treats different groups comparably. Agreement need not indicate distinct or valid information.

How it works

For each respondent, identify the item's score X and calculate the total T. The item-rest score is R = T minus X. Then calculate the chosen correlation across respondents between X and either T or R.
Pearson correlation is commonly used for numerical scores. It divides the covariance by the product of the standard deviations. If either variable has zero variance, the correlation is undefined rather than zero.
Use consistent coding and a documented missing-data rule. Check software labels: a raw item-total correlation, a standardized version, an item-rest correlation, and a correction involving reliability are not necessarily the same statistic.

Three practical examples

An AI literacy questionnaire team checks whether an item about understanding model limitations aligns with the intended literacy subscale.
A training assessor examines a negative item-rest correlation to determine whether a question was incorrectly keyed or tests a different skill.
A survey designer compares statistical results with content coverage before deciding whether an item should be revised, retained, or removed.

Think of it this way

To see whether one musician follows the rest of an ensemble, listen to that musician alongside the others with their own sound removed. Including the musician in both recordings makes agreement partly automatic. Even agreement with the ensemble does not prove that the ensemble is playing the right piece.

Worked example

Consider an original four-person illustration. The item scores X are 1, 2, 3, and 4. The remaining-item totals R are 3, 1, 1, and 3. The complete totals T = X plus R are therefore 4, 3, 4, and 7.
The mean item score is 2.5 and the mean rest score is two. Centered item scores are minus 1.5, minus 0.5, 0.5, and 1.5. Centered rest scores are one, minus one, minus one, and one.
Multiplying the corresponding centered values gives minus 1.5, 0.5, minus 0.5, and 1.5. These sum to zero. The corrected item-total correlation is consequently zero: this item has no linear association with the rest score in the example.
The full total has mean 4.5 and centered values minus 0.5, minus 1.5, minus 0.5, and 2.5. The cross-products with the centered item scores sum to five. The sums of squared deviations are five for X and nine for T.
The uncorrected correlation is therefore five divided by the square root of forty-five, approximately 0.7454. The same item now appears strongly related to the full total simply because its own variation contributes to that total. This does not contradict the zero item-rest correlation.
The algebra explains the difference: covariance of X with X plus R equals variance of X plus covariance of X with R. Even when the second term is zero, the item's own variance remains in the uncorrected numerator.
For a separate coding illustration, keep X = 1, 2, 3, 4 but let the rest totals be 8, 6, 4, 2. Their relationship is perfectly decreasing, so the item-rest correlation is minus one.
Suppose the item's wording and scoring instructions establish that it should be reverse coded on a one-to-five scale. Replacing X by six minus X gives 5, 4, 3, 2. Those values now have correlation plus one with the unchanged rest totals. Reverse coding is justified here by the stipulated scoring design, not merely by a desire to obtain a positive result.
Changing a rest total to a rest average by dividing every person's score by the same positive number leaves Pearson correlation unchanged. For example, dividing 8, 6, 4, 2 by two gives 4, 3, 2, 1 and preserves the negative correlation with the original X.
That equivalence does not automatically hold when different respondents answer different numbers of items, because each person's divisor may differ. A missing-data policy can therefore change the relationship being measured.
Finally, if every respondent scores three on the item, its standard deviation is zero. Neither an item-total nor an item-rest Pearson correlation can be calculated normally. Reporting zero would incorrectly suggest an estimated absence of association rather than insufficient variation.
These small examples illustrate mechanics, not adequate samples for evaluating a real assessment.

Illustrative case study

Fictional company Alder Learning reviews an AI skills questionnaire. An item looks acceptable when correlated with the total including itself but has a weak relationship with the remaining items.
The reviewer checks wording, response variation, and whether the question belongs to a different skill area. The team considers a separate subscale rather than automatically deleting content to increase internal agreement. This fictional review illustrates an investigation; it does not claim that a particular correlation proves validity.

Common misconceptions and mistakes

Including an item in its comparison total can inflate apparent agreement.
Corrected does not mean corrected for every source of measurement error.
A negative coefficient is a prompt to investigate, not automatic permission to reverse the answers.
A high coefficient can reflect redundant wording.
There is no universally appropriate deletion rule based on one numerical cutoff alone.

Business use

Use item-total analysis during questionnaire development and review, alongside content expertise, response distributions, and appropriate scale validation. State the population, coding, score construction, correlation definition, and missing-data handling. Evaluate revised items with suitable new evidence so that changes are not justified solely by their fit to the development sample.

Check your understanding

What is the first example's item-rest correlation?

Answer: Zero.

What is its uncorrected correlation?

Answer: Approximately 0.7454.

Why do they differ?

Answer: The full total includes the item's own variation.

Related terms

Item analysis
internal consistency
Pearson correlation
reverse coding
Cronbach's alpha
scale validity
part-whole correlation
subscale

Learn next

Compare item-rest diagnostics with factor analysis, content validity, and reliability measures, keeping their different purposes explicit.

Sources

William Revelle, psych package alpha documentation, item-whole correlation definitions: https://www.personality-project.org/r/psych/help/alpha.html . Numerical examples and company scenario are original illustrations.

Previous: Item stem and leaf plot Next: Item tree analysis