Cronbach's Alpha: What It Measures, What It Does Not, and How to Report It

Last updated October 30, 2026·4 min read

Cronbach alpha summarizes how consistently a set of items moves together. Formally it compares the variance of the total score with the sum of the item variances, scaled by the number of items: alpha = (k ÷ (k − 1)) × (1 − (sum of item variances ÷ total variance)). Substantively it estimates the share of total-score variance that is shared among the items, under the assumption that the items measure one construct on the same scale. Every qualification in that sentence earns its place, and the rest of this guide is those qualifications, one at a time.

The bands, with their health warnings

The conventions most fields cite:

• 0.90 and above: excellent, with the redundancy caution below
• 0.80 to 0.89: good
• 0.70 to 0.79: acceptable for research use
• 0.60 to 0.69: questionable, and worth a stated defense
• Under 0.60: the scale needs work before its scores carry decisions

The bands describe customs, not thresholds in nature. A short screening scale feeding low-stakes decisions can serve well at 0.72. A score that gates a high-stakes individual decision needs much more, and alpha alone never supplies it. Report the band your field uses and the stakes your score feeds, side by side.

The item-count effect, in numbers

Alpha rises with the number of items at a fixed average inter-item correlation. With an average correlation of 0.30:

• Four items: standardized alpha = 4 × 0.30 ÷ (1 + 3 × 0.30) = 0.632
• Eight items: 8 × 0.30 ÷ (1 + 7 × 0.30) = 0.774
• Twelve items: 12 × 0.30 ÷ (1 + 11 × 0.30) = 0.837

The items did not get better between the first row and the third. The scale got longer. Two consequences follow. Comparing alphas across scales of different lengths compares lengths as much as quality. And repairing a weak scale by adding near-duplicate items raises the number while adding respondent burden and no new information, which is why the redundancy caution attaches to the highest band: alpha past about 0.95 often means the scale asks the same question several times and charges the respondent for each asking.

The item-level evidence is where repair happens

Three statistics do the practical work:

• Corrected item-total correlation: each item against the sum of the others. Low or negative values name items that do not belong, including items that were worded in reverse and never recoded, which announce themselves with negative correlations.
• Alpha if item deleted: the alpha of the scale without that item. A deletion that raises alpha materially identifies a repair candidate.
• Inter-item correlations: the pattern behind the averages, where two clusters hiding inside one scale first become visible.

On the worked scale from the calculator page (four items, six teaching respondents), alpha is 0.955, corrected item-total correlations run from 0.863 to 0.946, and alpha if item deleted runs from 0.926 to 0.954: every item earns its place, and the weakest, item three, is weak inside a strong set rather than a problem to cut.

What alpha does not say

Alpha is not evidence of unidimensionality. A scale built from two clusters of items, each cluster tight within itself and the two moderately correlated, posts a strong alpha while measuring two things. The dimensionality question belongs to factor analysis, run on the same items, which is why Analysis Studio keeps the two in one module. Alpha also is not validity: a scale can consistently measure the wrong construct, and consistency will not notice. And alpha from the sample at hand is an estimate with an interval; on small samples the interval is wide, and reporting the point value alone overstates what a pilot of thirty respondents can establish.

A reporting paragraph that holds up

Report the number of items, the sample size, alpha with its interval, and the item evidence behind any repair: "The eight-item scale (n = 214) showed good internal consistency, Cronbach alpha of 0.84, 95 percent interval 0.80 to 0.87. Corrected item-total correlations ranged from 0.41 to 0.68; no item deletion improved alpha. A one-factor model fit the items adequately, supporting treatment of the total as a single score."

That paragraph takes four lines and answers every routine reviewer question before it is asked.

Compute yours

Module 13 of Analysis Studio reports alpha with its Feldt interval, standardized alpha, McDonald omega, the full item table, and the factor analysis that answers the dimensionality question:

Module 13: Scale Reliability and Factor AnalysisOpen in Analysis Studio →

Sources and method notes

Formulas follow Cronbach (1951) with the Feldt interval; standardized alpha values in the item-count table were computed from the Spearman-Brown form at drafting and are re-verified at build.