What alpha can and cannot say
Alpha depends on the number of items, their covariance, and the score construction. A higher value means the items produce more internally consistent total scores under the coefficient’s assumptions. It does not by itself establish:
- that all items measure one dimension;
- that the scale is valid for its intended use;
- that scores are stable over time;
- that raters agree;
- that measurement is equally reliable in another population;
- that every item is useful.
Methodological work on internal-consistency estimation emphasizes that alpha is not a test of unidimensionality and can be misleading when its measurement assumptions do not fit the items.3 Factor structure and substantive validity require separate evidence.
The .70 rule and its provenance
| Value | Common label | Source trail | Context | Category |
|---|---|---|---|---|
| α = .70 | “Acceptable” | Often attributed to Nunnally; audited by Lance et al. (2006)1 | Early-stage psychometric research, not every use | DISPUTED |
Lance and colleagues traced several common cutoffs and argued that the blanket .70 claim misstates the cited guidance.1 Their provenance review reports that Nunnally’s guidance varied with purpose: early research could tolerate less precision than settings involving consequential decisions about individual scores. The primary Nunnally text was not independently verified for this page, so the historical interpretation is attributed to the verified review rather than presented as a confirmed primary-source quotation.
Hussey and colleagues later documented an abnormal abundance of reported alpha values at .70, consistent with a publication practice in which crossing a salient cutoff affects reporting.2 That finding does not make every alpha near .70 wrong. It does show why a threshold should not replace transparent computation and reporting.
The correct label is therefore not “universally acceptable.” It is “a widely repeated, disputed convention whose relevance depends on purpose and evidence.” Where statistical thresholds come from describes this provenance category.
How to interpret the value in front of you
Use this order:
- Identify the score and sample. Reliability belongs to scores in a population and setting, not permanently to a questionnaire name.
- Inspect the items. Confirm coding direction, missing-data handling, and whether a total score is substantively coherent.
- Read alpha with uncertainty. A point estimate alone hides sampling variation; report a confidence interval when possible.
- Examine dimensionality. A high alpha can arise from many redundant items or correlated dimensions.
- Match precision to purpose. Exploratory group research and consequential individual-score decisions have different stakes.
- Compare alternative coefficients. Omega may better match a congeneric measurement model when its assumptions and estimation are suitable.
What a low alpha may mean
A low alpha can reflect weak item relationships, too few items, multidimensional content, restricted score variation, coding errors, or a sample unlike the intended population. It is a diagnostic clue, not an automatic instruction to delete items.
Check item distributions and item-total relationships, but do not optimize alpha mechanically. Removing an item can raise alpha while narrowing content coverage and damaging validity. The construct definition remains the governing reason for including an item.
What a very high alpha may mean
A very high alpha can be compatible with precise scores, but it can also arise from repetitive or near-duplicate items. There is no universal point at which alpha becomes “too high.” Review item content, dimensional evidence, test length, and the intended breadth of the construct rather than applying an unattributed upper cutoff.
Why alpha can be negative
Negative alpha usually indicates that average item covariance is negative under the scoring used. Common causes include reverse-keyed items that were not recoded, incompatible items, data-processing errors, or a collection of items that does not support one composite score.
A negative value is a strong reason to audit coding and the covariance matrix before interpreting a scale. It is not a meaningful degree of “negative reliability.”
Alpha versus omega
McDonald-type omega coefficients use a factor-model representation rather than alpha’s more restrictive structure. Dunn and colleagues recommend omega as a practical alternative in many internal-consistency settings while also emphasizing correct model specification.3
Omega is not automatically valid because it has a different name. Report the version, estimation method, factor structure, and interval. PCA output and component loadings answer a different data-reduction question and do not substitute for a reliability model. For agreement among raters rather than item consistency, use kappa and ICC interpretation.
A defensible reporting sentence
“Scores from the eight-item scale had α = .74 in this sample. This exceeds the commonly repeated .70 convention, but adequacy was evaluated against the scale’s group-level research purpose, item structure, confidence interval, and dimensional evidence.”
That sentence attributes the rule, limits its scope, and keeps the actual measurement evidence visible.