Written and maintained by CASRAI Editorial Board
Last updated
Item-total correlation is the diagnostic that tells you, item by item, whether a candidate question belongs in a scale you are still finalizing. During scale purification — the item-screening stage between piloting a draft item pool and locking the final instrument — it is the single most direct answer to “is this specific item pulling its weight, or is it noise the rest of the scale would be better off without.” This guide covers what item-total correlation actually measures, why the corrected version is the one you should be reading, the conventional 0.30 cutoff and how strictly to treat it, and how it relates to (and sometimes disagrees with) the companion diagnostic most reliability-analysis output prints alongside it: alpha if item deleted.
What Item-Total Correlation Measures
An item-total correlation is simply the Pearson correlation between one item’s scores across respondents and the total score on the scale it belongs to. A high value means the item moves the way the rest of the scale moves — respondents who score high on the total tend to score high on that item too. A low or negative value means the item is behaving independently of, or even opposite to, everything else the scale is trying to measure.
There are two versions, and the distinction matters:
- Uncorrected (raw) item-total correlation — the item correlated against the total score that includes that same item. Because the item is part of its own comparison total, this version is mechanically inflated: an item is guaranteed some correlation with a sum that contains its own values, even if it has nothing to do with the underlying construct.
- Corrected item-total correlation (CITC) — the item correlated against the sum of every other item, with the item itself excluded from the total. This removes the built-in inflation and is the version scale-purification decisions should actually be based on. Statistical packages that report “corrected item-total correlation” (SPSS’s Reliability Analysis, R’s
psych::alpha()underraw.rvsr.drop, and equivalents elsewhere) are already giving you this corrected form — look for that label specifically rather than a plain “item-total correlation” column, which may be the uncorrected version.
The Conventional 0.30 Cutoff
The heuristic most commonly taught for corrected item-total correlation is a minimum of 0.30: an item with a CITC at or above 0.30 is generally retained as contributing meaningfully to the construct the scale is purifying toward, while an item below 0.30 is flagged as a candidate for revision or removal. It is worth being precise about what kind of number this is — it is a widely taught convention from psychometric scale-development practice, not a derived statistical significance threshold, and different methods texts state it slightly differently: some treat anything below 0.30 as the flag point, others use a somewhat looser 0.20, and some recommend a stricter 0.40 when the scale is being purified for high-stakes or published use rather than early piloting. Treat 0.30 as the standard default and adjust toward the stricter end when the instrument’s downstream use raises the cost of retaining a weak item.
A negative corrected item-total correlation is a stronger and more specific signal than a merely low one: it means the item moves in the opposite direction from the rest of the scale. In practice this is overwhelmingly caused by a reverse-worded item that was not reverse-coded before analysis (see the pitfalls section below) rather than a genuinely construct-irrelevant item — check coding direction before concluding the item itself is the problem.
Item-Total Correlation and Alpha If Item Deleted
Reliability-analysis output almost always reports corrected item-total correlation alongside a second, closely related diagnostic: alpha if item deleted — the value Cronbach’s alpha would take if that specific item were dropped and the coefficient recalculated on the remaining items. The two answer related but distinct questions:
- Corrected item-total correlation asks: does this item correlate with the rest of the scale strongly enough to belong?
- Alpha if item deleted asks: does removing this item improve the scale’s overall internal consistency?
The two diagnostics usually agree: an item with a low CITC typically also has an alpha-if-deleted value higher than the scale’s current overall alpha, because an item that barely correlates with the rest of the scale is, almost by definition, adding more measurement noise than reliable signal. When alpha-if-deleted for an item exceeds the scale’s current overall alpha, that is the direct, numeric statement that the scale would be more internally consistent without it — CITC explains why. Because the two can occasionally diverge (an item can sit just above the 0.30 CITC line while still nudging alpha-if-deleted upward, or vice versa in short scales where alpha is unusually sensitive to item count), read them together rather than relying on either alone. For the full mechanics of running this analysis and reading the alpha-if-deleted column specifically, see CASRAI’s Cronbach’s alpha in SPSS guide.
Worked Example: Flagging a Weak Item in a 6-Item Pool
The table below comes from a seeded, reproducible simulation (Node.js, seed 20260829, N = 40 simulated respondents, Box-Muller-generated latent scores mapped onto a 1–5 Likert scale), built specifically for this page — it is illustrative synthetic data, not a real dataset or published study. Five items were generated to load on a common latent trait (loading 0.75); a sixth was deliberately generated as near-noise (loading 0.05) to demonstrate what a genuinely off-construct item looks like in the output.
| Item | Corrected item-total correlation | Alpha if item deleted | Flag? |
|---|---|---|---|
| Item 1 | 0.731 | 0.669 | Retain |
| Item 2 | 0.767 | 0.643 | Retain |
| Item 3 | 0.711 | 0.667 | Retain |
| Item 4 | 0.532 | 0.715 | Retain |
| Item 5 | 0.624 | 0.687 | Retain |
| Item 6 | −0.196 | 0.874 | Flag — below 0.30, negative |
Overall Cronbach’s alpha across all six items is 0.759. Item 6’s corrected item-total correlation is negative — well below the 0.30 cutoff — and its alpha-if-deleted (0.874) is substantially higher than the overall alpha, meaning the scale would be materially more internally consistent without it. The two diagnostics agree cleanly here, which is the expected pattern for a genuinely off-construct item: dropping Item 6 and re-running the analysis on the remaining five items would be the standard purification step, followed by a content review of Item 6’s wording before deciding whether to revise and re-pilot it or drop it outright.
Where Purification Fits in Scale Development
Item-total correlation is a purification-stage tool, not a first-draft or a final-validation tool. It sits after an initial item pool has been generated and piloted on a sample, and before the scale is treated as finalized:
- Item generation — draft a pool of candidate items, typically several times larger than the intended final scale, covering the full breadth of the target construct.
- Expert/content review — check each item against the construct definition before any data is collected (see CASRAI’s content validity guide).
- Pilot administration — field the full item pool to a pilot sample.
- Item purification — this is where corrected item-total correlation, alpha-if-item-deleted, and (for larger item pools) exploratory factor analysis are used together to identify and remove weak or redundant items.
- Confirmation — the purified item set is validated on an independent sample, often with confirmatory factor analysis, before the scale is treated as final.
Item-total correlation is a reliability-side diagnostic: it tells you whether an item is internally consistent with the rest of the scale, not whether the scale as a whole measures the intended construct. A scale can purify to a clean set of item-total correlations and still have construct validity problems if the surviving items, taken together, do not actually capture the construct they were meant to. Purification is a necessary step in scale development, not a substitute for validity evidence.
Common Pitfalls
- Not reverse-coding first. A negatively worded item run without reverse-coding will show a low or negative corrected item-total correlation that has nothing to do with the item’s actual quality — it is a coding artifact, not a construct-fit problem. Reverse-code before purification, not after seeing a flagged item.
- Chasing the cutoff instead of reading the item. A CITC just under 0.30 is a prompt to review the item’s wording and content fit, not an automatic deletion order. An item can sit below the cutoff because it measures a genuinely distinct sub-facet of the construct, in which case the fix may be splitting the scale rather than discarding the item.
- Purifying on too small a pilot sample. Item-total correlations are correlation coefficients, and correlation estimates from small pilot samples are unstable — a weak-looking item in a pilot of 20 respondents may look very different at 200. Treat purification decisions from small pilots as provisional.
- Purifying and never re-checking content coverage. Repeatedly dropping the lowest-CITC item and re-running the analysis (a common iterative practice) tends to converge on a shorter scale with a higher alpha, but each deletion also narrows what the scale actually samples from the construct. Stop and check remaining content coverage, not just the reliability number, before finalizing.
Frequently Asked Questions
What counts as a good item-total correlation?
0.30 or higher on the corrected item-total correlation is the most commonly taught minimum for retaining an item during purification; 0.40 and above is a stronger, more conservative bar some methodologists apply when the scale is headed toward publication or high-stakes use. There is no single universally agreed number — treat 0.30 as the standard default rather than a fixed statistical law.
Is item-total correlation the same thing as a factor loading?
No, though they are related and often point the same direction. Item-total correlation is computed directly from raw scores against a simple sum; a factor loading comes from a factor model (see CASRAI’s exploratory factor analysis guide) and reflects the item’s relationship to a latent factor, which can differ from a simple total, especially when a scale has more than one underlying dimension.
Can corrected item-total correlation be negative?
Yes. A negative value means the item moves opposite to the rest of the scale. The most common cause by far is an unreverse-coded negatively worded item; a genuinely construct-irrelevant item is the other, less common cause. Check coding direction first.
How is corrected item-total correlation different from the uncorrected version?
The uncorrected version correlates an item against a total that includes that item’s own score, which mechanically inflates the correlation. The corrected version excludes the item from the total it’s compared against, removing that inflation — it is the version purification decisions should be based on, and the one most statistical software actually labels “corrected item-total correlation.”
Should I purify a scale using item-total correlation alone?
Not alone. Read it alongside alpha-if-item-deleted for the reliability side, and keep a content-validity check in the loop so items aren’t dropped purely for a marginal reliability gain at the cost of construct coverage. See CASRAI’s reliability in research guide for how item-total correlation fits among the broader family of reliability evidence.
Related CASRAI Resources
For the broader reliability coefficient this diagnostic is most often paired with, see CASRAI’s Cronbach’s alpha guide and the Cronbach’s alpha interpretation guide; for the step-by-step software output, see Cronbach’s alpha in SPSS. For the stages that surround purification, see content validity, construct validity, exploratory factor analysis, and confirmatory factor analysis. For the item format most scale-purification work is applied to, see Likert scale construction and questionnaire design. For measurement fundamentals, see psychometrics.
