Blog / What Cronbach’s Alpha Tells You—and What It Does Not

What Cronbach’s Alpha Tells You—and What It Does Not

An alpha of .80 summarizes a sample’s item covariance pattern. It is not proof that a scale is one-dimensional, valid, stable over time, or 80% accurate.

September 15, 2026 · 5 min read · By Cepingku Editorial Team

Cronbach’s alpha is often reported beside a questionnaire as though one number settled every quality question. It does not. Alpha is calculated from a group’s responses to a set of items on one administration. It summarizes a pattern in those data; it is not a person’s score, a percentage of correct answers, or this site’s ‘80% accuracy.’ Tavakol and Dennick note that alpha changes with item interrelationship and test length, and that it is a property of scores from a particular sample rather than a permanent label attached to an instrument.

A four-item calculation needs a sample

Consider a deliberately fictional four-item scale completed by 40 people. After all items have been coded in the same direction, suppose the four item variances add to 4 and the variance of the total score is 10. For k items, raw-score alpha is α = k/(k−1) × (1 − sum of item variances/total-score variance). Here, α = 4/3 × (1 − 4/10) = 4/3 × 0.6 = .80. The total-score variance includes covariance among items, which is why it need not equal the sum of their separate variances.

This is an arithmetic example, not a real scale or this site’s scoring method. The variances and covariances come from many respondents. One person’s four choices have no between-person variation from which to estimate them, so it is not meaningful to insert one person’s answers and call the result that person’s alpha.

QuestionEvidence that addresses it
Do the items show a useful covariance pattern in this sample?Alpha, reported with the sample, version, scoring, and uncertainty.
Do the items support one intended dimension?A stated construct, item content, and an appropriate structural analysis; alpha alone cannot decide.
Would scores be stable across time?A retest design that states the interval, conditions, and instrument version.
Is the proposed score interpretation and use supported?Validity evidence, such as content coverage and predicted relationships with external variables.

Why .80 is not a pass stamp

A value of .80 says that this particular covariance calculation produced .80. It is a reason to inspect the scale, not a universal approval label. Fixed cutoffs such as .70 ignore the purpose of measurement, the number and content of items, the consequences of a decision, the sample, and sampling uncertainty. Adding related items can raise alpha; near-duplicate items can therefore make a long scale look impressive without adding much new information.

Sijtsma’s analysis makes the core limitation concrete: alpha depends on item count and the ratio of summed item variances to total-score variance, rather than preserving the full covariance pattern. Different factor structures can share the same alpha, and a one-factor set can have high or low alpha as the common signal changes. That is why alpha by itself does not establish internal structure or unidimensionality. Read ‘internal consistency’ carefully as a shorthand for item interrelatedness in a sample, not as proof that every item measures precisely one thing.

Dimensionality and validity ask different questions

Unidimensionality concerns whether a single latent dimension is an adequate account of the items’ common variation under a specified model. It calls for a construct definition, content review, and structural evidence. Imagine an eight-item questionnaire that combines four items about social initiative with four about emotional recovery. A respectable overall alpha could still conceal two related dimensions. If theory distinguishes them, examine and report them as subscales instead of forcing one total score.

Validity concerns the interpretation and use of scores: for example, whether item content represents the intended trait, whether relations with outside measures follow predictions, and whether the interpretation holds in the relevant groups. Repeatability is useful for a measurement claim, but it does not establish that the claimed construct was measured. Alpha therefore cannot prove that a website, a questionnaire, or a single result is valid.

Interpreting alpha as a total-score reliability estimate also requires model assumptions. Under essential tau-equivalence, item true scores differ only by constants; a common-factor representation gives them equal contributions. Item errors should also be uncorrelated. Violations can bias the estimate. The fictional .80 calculation therefore does not, by itself, establish that 80% of observed differences represent true trait differences.

Alpha is not test–retest reliability

Test–retest evidence is about stability across occasions. Its interpretation depends on how long the interval was, whether the person’s situation plausibly changed, and whether the same version was used. A one-time alpha contains none of that timing information. It should not be used to claim that a score will be stable next week or next year.

Nor does switching automatically to omega make the reasoning disappear. Omega is model-based: its usefulness depends on the measurement model, the fit and plausibility of that model, and the same sample and scoring choices that need reporting for alpha. The useful question is not which coefficient wins by default, but which claim is being made and what evidence can actually bear on it.

A practical reporting sequence

  1. Define the construct and any intended subscales before looking for a single coefficient.
  2. Document item coding, missing-data handling, sample, and instrument version.
  3. Report alpha as a sample-specific descriptive result, then inspect item behavior and structural evidence.
  4. Add validity evidence for the intended interpretation and retest evidence only when making a time-stability claim.

Used this way, alpha remains a transparent description of one feature of a response dataset. It becomes misleading only when it is asked to certify a person, a dimension, a construct, a future score, or a product claim on its own.

Want to see where you land?

Take the Big Five Personality Test and get your results in minutes.

Take the Big Five Personality Test →

← Back to the blog