How to Report Cronbach’s Alpha in APA 7

Nearly every manuscript with a questionnaire reports alpha, and nearly all of them go wrong in the same place: a bare number, in the wrong section, with .70 quoted as though APA had written it into the manual. Reliability is not a property of the instrument that you can copy from the original validation paper; it is a property of the scores you obtained in your sample, and it moves with the sample. Here is the exact sentence, the confidence interval almost nobody includes, and the omega that reviewers increasingly ask for. The general formatting rules live in how to report results in APA 7.

Where alpha goes: Measures, not Results

Alpha belongs in the Measures (or Instruments) subsection of your Method, attached to the questionnaire it describes. The reader has just been told how many items the scale has and what the response format is, which is exactly what they need in order to judge whether that .87 means anything or is arithmetic from having a lot of items.

The full sentence looks like this: “Depressive symptoms were measured with the PHQ-9 (9 items, 0–3 response scale). Internal consistency of the scores in this sample was good, α = .89, 95% CI [.86, .91], n = 150.” One instrument, one sentence, reliability estimated in your own data.

The second legitimate place is the diagonal of the correlation matrix, with the coefficients in parentheses and a note that says so: “Values in parentheses on the diagonal are Cronbach’s alpha coefficients.” That is what you do when you have six scales and do not want six near-identical sentences in Method, and it saves you a whole column of table.

What does not work is dropping alpha into Results with no context, or repeating it in three places with three different numbers of decimals. If you estimate reliability separately by subgroup or by wave, say which sample each figure came from — an alpha with no sample attached cannot be checked by anyone.

The exact sentence, with a 95% confidence interval

Alpha is not italicised. APA sets Greek letters in roman type, and α is one; the same goes for ω. Italics are for the Latin statistics: M = 3.42, SD = 0.81, t(148) = 2.34, r = .31. An italicised α is the fastest signal that nobody with the manual open has read the manuscript.

Two decimals and no leading zero, because the coefficient cannot exceed 1: α = .87, never 0.87. Same rule that governs p, r and R².

The confidence interval is what separates a serious reliability report from a box-ticking one. With n = 120 on a 10-item scale, α = .71 comes with a 95% CI of [.61, .78]: the point estimate clears the famous threshold, but the interval says that in this sample you cannot tell .61 from .78. Compute it with Feldt’s F-based approximation or by bootstrapping with 5,000 resamples; both are acceptable, and you say which one you used. The Cronbach’s alpha calculator returns the coefficient with its interval.

State the n the coefficient was estimated on whenever it differs from your main analytic sample. With listwise deletion a long battery quietly loses participants, and reliability can end up estimated on noticeably fewer people than your Participants section advertises.

Why .70 is a convention and not a rule

Nunnally (1978) proposed .70 for early-stage research, when a construct is still being explored, and asked for .90 or higher when scores are used to make decisions about individual people. Journals kept the number and threw away the context. The Publication Manual sets no cut-off at all: it asks you to report reliability and discuss it, not to pass a test.

How much reliability you need is decided by what the score is used for. To compare group means in a research study, .70 is defensible. For a clinical cut-off that labels a person as a case, .70 is a decision with consequences, because every unit of measurement error turns into someone classified incorrectly.

The ceiling matters too. An α above .95 is rarely good news: it usually flags redundancy, items that paraphrase one another and narrow the construct down to a single question asked ten ways. If your 30-item scale returns .97, the reviewer’s question will not be whether it is reliable, but why it needs 30 items.

And if you get .64? Report it, say what it limits — low reliability attenuates every correlation you estimate with that scale — and carry it into your limitations. What you do not do is delete items after seeing the data until you clear .70. If you drop an item, drop it for a reason about its content, and report alpha before and after.

McDonald’s omega, and why reviewers now ask for it

Alpha assumes tau-equivalence: every item measures the construct with the same weight, and the errors do not correlate. That condition is almost never met. When the loadings differ — and they always differ — alpha stops being the reliability and becomes a lower bound, so it underestimates. You are penalising your own scale for an assumption nobody tested.

McDonald’s omega (1999) uses the loadings a one-factor model actually estimates: ω = (Σλ)² / [(Σλ)² + Σθ]. With six items loading .90, .85, .80, .40, .35 and .30, alpha comes out at .76 and omega at .79. The gap looks small, but it always runs in the same direction and it widens as the standardised loadings become more unequal. Dunn et al. (2014) have been arguing for the switch for a decade, and it now shows up in review reports.

Omega has a cost: it needs a factor model that fits. If your CFA does not fit, omega is no more credible than alpha, it is merely worse justified. So the order is: fit the measurement model first — see how to report a CFA in APA 7 — and compute omega afterwards from that model’s standardised loadings. The McDonald’s omega calculator does the arithmetic once you have them.

While the field finishes moving, report both: “Internal consistency was acceptable, α = .76, ω = .79.” Alpha because your readers expect it and it lets them compare with the earlier literature; omega because it is the number that actually estimates the reliability of your scores. If the scale has a general factor plus specific ones, the coefficient to report is hierarchical omega (ω_h), which says how much of the total variance that general factor carries.

Alpha rises with the number of items

The formula makes no secret of it: α = k · r̄ / [1 + (k − 1) · r̄], where k is the number of items and r̄ the mean inter-item correlation. Hold the mean correlation constant and alpha grows just by lengthening the scale, even when the new items add nothing that was not already there.

The numbers are blunt. A 40-item scale whose items correlate on average at a feeble .18 reaches α = .90. A 4-item scale whose items correlate at .40 — more than twice the real coherence — sits at α = .73. Compare the two on alpha alone and you will pick the worse instrument.

That is why it is worth reporting the mean inter-item correlation as well: unlike alpha, it does not depend on length. Clark and Watson (1995) suggested keeping it between .15 and .50 depending on how broad the construct is — below that the items are not measuring the same thing, above it they are repeating each other.

With two items, alpha is the wrong tool. The recommended statistic is the correlation between them corrected by the Spearman-Brown formula (Eisinga et al., 2013), or simply the correlation with its interval: r = .62, 95% CI [.51, .71]. A two-item alpha is not miscalculated, it is misapplied, and a psychometrically minded reviewer spots it immediately.

What reviewers flag

The validation study’s alpha. Copying the .91 from the paper you took the scale from says nothing about your data: that figure was estimated in another sample, often another language and another decade. Cite it as context if you like, but the one that counts is yours.

One alpha for a multidimensional scale. If your EFA or CFA returned three factors, an alpha for the total score blends three things and describes none of them. Report one per subscale. If you also use the total score, justify it with a bifactor or second-order model and give hierarchical omega alongside it.

A negative or absurdly low alpha. Almost always it is reverse-worded items that were never recoded. Before you write in your limitations that the scale behaved poorly, check the recoding and look at the inter-item correlation matrix: if a block of items has flipped signs, there is your answer.

Alpha on three- or four-point response scales. With few response categories, Pearson-based alpha underestimates reliability. The right choice is ordinal alpha or ordinal omega, computed on polychoric correlations, and you say in the text that it is the ordinal version. The full map of reliability coefficients is in reliability analysis: alpha, omega and ICC.

Frequently asked questions

What is an acceptable Cronbach’s alpha?

It depends on what the score is used for. The .70 figure is a convention from Nunnally (1978) for early-stage research; established scales are expected to reach .80, and scores used to make decisions about an individual, .90 or above. Anything over .95 usually signals redundant items. APA sets no threshold at all: it asks you to report reliability and discuss it.

Is Cronbach’s alpha italicised in APA 7?

No. APA sets Greek letters in roman type, and alpha is one; the same applies to omega. Italics are reserved for Latin statistics such as M, SD, t, r and p. Alpha also takes no leading zero, because the coefficient cannot exceed 1: write α = .87, not 0.87.

Where do you report Cronbach’s alpha in a paper?

In the Measures or Instruments subsection of the Method, next to the description of each questionnaire. The accepted alternative is to place it in parentheses on the diagonal of the correlation table with an explanatory note. What does not belong is a bare alpha in Results with no context, or the same coefficient repeated in three places.

Should I report McDonald’s omega instead of Cronbach’s alpha?

Report both while the field is in transition. Alpha assumes every item carries the same weight, which is almost never true, so it underestimates reliability; omega uses the loadings your factor model estimates. Omega does require a one-factor model that fits, so if your CFA fits poorly, switching coefficients fixes nothing.

How do you report a confidence interval for Cronbach’s alpha?

Use Feldt’s F-based approximation or a bootstrap with around 5,000 resamples, and state which you used. It is written like this: α = .87, 95% CI [.83, .90]. In small samples the interval is wide, and that width is precisely the information a bare alpha hides.

Can you report Cronbach’s alpha for two items?

You can compute it, but it is not recommended. With two items the appropriate statistic is the correlation between them corrected by the Spearman-Brown formula (Eisinga et al., 2013), or simply the correlation with its confidence interval. A two-item alpha is not miscalculated, it is misapplied.

With alpha in Measures, its 95% interval beside it and omega alongside, reliability stops being the soft spot of the manuscript. If your scale is multidimensional, or alpha came out low and you cannot tell whether the problem is the items or the sample, that is what my statistical consulting is for.