How to Report ICC Results in APA Style

You've run the analysis. The output shows an intraclass correlation of 0.87 and now you need to turn it into a sentence a journal reviewer will accept. This guide gives you the exact reporting format — the one-sentence template, a worked Methods and Results paragraph, the benchmarks to cite, and the omissions reviewers flag most often. If you still need to choose which ICC to compute, start with our full ICC guide covering all six forms; this page assumes the analysis is done and the write-up is what's left.

The One-Sentence Template

A complete ICC report answers four questions in a single sentence: which ICC (form, following the model/type/unit conventions of McGraw and Wong, 1996), how big (the point estimate), how precise (the 95% confidence interval), and — where your software provides it — the significance test (the F-test from the underlying ANOVA). The template:

Filled in for a common case — three raters, thirty subjects, generalizing to other raters, single ratings used in practice:

Inter-rater reliability was assessed using a two-way random-effects
model with absolute agreement for single measures, ICC(2,1) = .87,
95% CI [.79, .92], F(29, 58) = 21.30, p < .001.

And for the average-measures, fixed-rater case, ICC(3,k):

The reliability of the mean rating across the three study coders was
estimated with a two-way mixed-effects model, consistency type, for
average measures, ICC(3,3) = .93, 95% CI [.88, .96].

Three details of APA number formatting matter here. First, ICC values and their confidence limits are written without a leading zero (.87, not 0.87), following the APA convention for statistics that cannot exceed 1 in absolute value. Second, the confidence interval goes in square brackets after "95% CI". Third, statistical symbols (F, p, df) are italicized, while "ICC" itself is set in roman type as an abbreviation.

The form label alone — "ICC(2,1)" — is not enough on its own, because the Shrout–Fleiss-style numbering is not universal across software and fields. McGraw and Wong (1996) defined ICC forms by three explicit choices: the model (one-way random, two-way random, or two-way mixed effects), the type (consistency or absolute agreement), and the unit (single measures or average of k measures). Koo and Li (2016) turned this into a reporting guideline: state all three in words, every time. The templates above do exactly that — the parenthetical form is a convenience, the prose description is the requirement.

Worked Example: Methods and Results Paragraphs

Here is how the reporting splits across a manuscript's Methods and Results sections for a typical inter-rater reliability study (three physical therapists rating shoulder range of motion in 30 patients).

Methods

Inter-rater reliability of the range-of-motion ratings was evaluated
using the intraclass correlation coefficient. Because the three raters
were drawn from a larger pool of clinicians and each rated all
patients, and because clinical decisions would be based on a single
clinician's rating, we computed ICC(2,1): a two-way random-effects
model, absolute-agreement type, single-measures unit (McGraw & Wong,
1996; Koo & Li, 2016). Analyses were conducted in R using the irr
package. Reliability was interpreted against the benchmarks of Koo
and Li (2016): values below .50 poor, .50 to .75 moderate, .75 to .90
good, and above .90 excellent.

Results

Inter-rater reliability was good, ICC(2,1) = .87, 95% CI [.79, .92],
F(29, 58) = 21.30, p < .001. Because the lower bound of the confidence
interval fell in the good range and the upper bound reached the
excellent range, the ratings were judged sufficiently reliable for
single-rater use in subsequent analyses.

Note what the Results paragraph does with the confidence interval: it interprets the interval, not just the point estimate. Koo and Li (2016) explicitly recommend basing the reliability judgment on the 95% CI, since the point estimate alone overstates certainty — an ICC of .87 with a CI of [.51, .96] is a very different finding from .87 with [.79, .92], even though the headline number is identical.

What Reviewers Check

Reviewers who work with reliability data tend to scan for a short, predictable checklist. Getting ahead of it saves a revision round.

• Which ICC form, and why. The single most common query. Your design determines the model: different raters per subject → one-way random; same raters for all subjects, treated as a sample from a larger population → two-way random; same raters, and only those raters matter → two-way mixed. The Methods paragraph should state the choice and the design fact that justifies it, as in the worked example above.
• Consistency vs. absolute agreement, stated explicitly. If raters need to be interchangeable — one clinician's score substitutes for another's — absolute agreement is the defensible type, because it penalizes systematic offsets between raters. Consistency ignores those offsets and only tracks rank-order correspondence. Reviewers will ask which you used if you don't say.
• Single vs. average measures matching actual use. If downstream decisions rest on one rater's score, report the single-measures ICC even if you averaged raters to estimate it. Average-measures ICCs (the ",k" forms) are always higher; reporting ICC(2,k) when practice uses single ratings inflates apparent reliability.
• A confidence interval. A bare point estimate reads as incomplete. Every mainstream implementation (R's irr and psych packages, Python's pingouin, SPSS) outputs the 95% CI — report it in brackets.
• An attributed benchmark. If you call .87 "good," say by whose scale. Koo and Li (2016) — below .50 poor, .50 to .75 moderate, .75 to .90 good, above .90 excellent — is the most commonly cited standard in current reliability work. It is a guideline, not a law: fields with high-stakes single measurements (clinical diagnostics, legal contexts) often demand more than the "good" band, and reviewers respect a sentence acknowledging the threshold appropriate to your application.
• Software and sample details. Name the package, and make sure n (subjects) and k (raters) are recoverable from the text — the F-test degrees of freedom help here, since for the two-way single-measures forms they follow (n − 1) and (n − 1)(k − 1).

Common Mistakes

Reporting a bare "ICC = .87" with no form or model

The same data yield six-plus different ICC values depending on model, type, and unit — differences that are routinely large enough to change the qualitative conclusion. An unlabeled ICC is unreplicable and uninterpretable, and it is the first thing a methods-literate reviewer will flag. Always attach the McGraw and Wong (1996) description: model, type, unit.

Mixing benchmark scales without citing either

Cicchetti's (1994) widely used scale calls .60–.74 "good" and ≥ .75 "excellent"; Koo and Li (2016) reserve "excellent" for values above .90. An ICC of .78 is therefore "excellent" on one scale and merely "good" on the other. Calling it "excellent (Koo & Li, 2016)" is an attribution error that reviewers do catch. Choose one standard, cite it in Methods, and use its labels verbatim.

Omitting the confidence interval

Especially with small samples, ICC estimates are imprecise, and the CI is where that shows. Reporting ICC = .81 without noting that the interval runs [.48, .93] hides the fact that the data are compatible with poor reliability. If the CI is wide, say so and temper the interpretation — that is more credible than hoping nobody computes it.

Mislabeling consistency as agreement (or vice versa)

These are different parameters, not synonyms. A consistency ICC can be high while raters disagree substantially in absolute terms — for example, when one rater scores everyone about ten points higher. Writing "raters showed strong agreement, ICC(3,1) = .91 (consistency)" conflates the two: a consistency estimate supports a claim about relative ordering, not about interchangeable scores. Match the word in your prose to the type you actually computed.

Reporting the average-measures ICC as if it described single ratings

ICC(2,k) and ICC(3,k) describe the reliability of the mean across k raters. Unless your protocol will always average that many raters, these values overstate the reliability of the measurement as used. State the unit and make sure it matches the design.

Preparing a full manuscript? See statistical analysis for researchers — citations and R code included with every analysis.

Frequently Asked Questions

References

Cicchetti, D. V. (1994). Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment, 6(4), 284–290.

Koo, T. K., & Li, M. Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for reliability research. Journal of Chiropractic Medicine, 15(2), 155–163.

McGraw, K. O., & Wong, S. P. (1996). Forming inferences about some intraclass correlation coefficients. Psychological Methods, 1(1), 30–46.