Open the methods chapter of almost any thesis and you will find scale reliability reported with Cronbach's alpha — usually as a reflex rather than a choice. Yet alpha dates from 1951 and rests on assumptions that real questionnaire data rarely satisfy; in the measurement literature, McDonald's omega is now the recommended default. This guide explains what each coefficient assumes, the specific ways alpha misleads, and how to compute omega in R, JASP and SPSS.
What internal consistency actually means
Internal consistency is the degree to which items intended to measure the same construct produce consistent responses. Formally, it estimates the proportion of variance in the scale's total score attributable to true score variance rather than measurement error: a coefficient of 0.85 means roughly 85 per cent of observed variance reflects systematic differences between respondents. One caution: internal consistency is not temporal stability (test–retest reliability); a state-like construct can show a high alpha alongside a poor test–retest correlation, and the two claims must be evidenced separately. Low reliability is not merely a measurement blemish, either: it systematically attenuates correlations and effect sizes and quietly drains a study's statistical power.
What Cronbach's alpha assumes — and why it breaks
For alpha to be an unbiased estimate of reliability, three conditions must hold simultaneously:
- Tau-equivalence: every item has the same factor loading on the latent construct. In real data, loadings almost always differ; the best defensible model is congeneric (loadings free to vary).
- Unidimensionality: the items measure a single construct. Alpha does not test this — it presupposes it; a high alpha on a multidimensional scale tells you nothing about dimensionality.
- Uncorrelated errors: item error terms are independent. Shared wording stems and reverse-keyed items routinely violate this.
When tau-equivalence fails, alpha typically underestimates reliability; when errors are correlated, it can be inflated instead. In other words, alpha is biased in a direction you cannot even predict in advance.
Two common ways alpha misleads
The first trap is item-count inflation. Alpha rises mechanically as items are added, even when the average inter-item correlation stays fixed. A pool of weakly correlated items (mean r = 0.25) will eventually produce a 'good-looking' alpha if you simply make the scale long enough — the chart below illustrates the mechanic. An alpha of 0.85 from a 20-item scale therefore does not signal the same measurement quality as 0.85 from a 6-item scale.
The second trap is misuse of 'alpha if item deleted'. Dropping items one by one because SPSS shows a higher alpha without them capitalises on sampling chance and erodes the scale's content validity. Item removal should be justified by theory and factor-analytic evidence; a 0.01 gain in alpha is not, by itself, a reason to discard an item.
McDonald's omega: the modern default
Omega total is computed from a factor model: it uses the items' actual (unequal) factor loadings, so it works under the congeneric model. If the loadings truly were equal, omega would equal alpha; since they almost never are, omega is the more accurate estimate. For multidimensional scales, omega hierarchical isolates the share of total-score variance attributable to the general factor alone — the honest way to justify using a total score. The composite reliability (CR) figure reported in structural equation modelling outputs (AMOS, lavaan, SmartPLS) is the same logic under another name.
| Property | Cronbach's alpha | McDonald's omega |
|---|---|---|
| Measurement model | Tau-equivalent (equal loadings) | Congeneric (loadings free) |
| Unidimensionality | Assumed, never tested | Examined alongside the factor model |
| Sensitivity to item count | Inflates mechanically as items are added | Computed from loadings, far less sensitive |
| Behaviour under violated assumptions | Biased (usually low, sometimes inflated) | Small bias, robust |
| Software | SPSS default menu, R, JASP | R psych, JASP, SPSS OMEGA macro |
Thresholds and what to report
The practical benchmarks are the same for both coefficients: ≥ 0.70 acceptable, ≥ 0.80 good. The lesser-known warning concerns the upper end: values above 0.95 are not a badge of excellence but a redundancy signal — items so similar they narrow the construct — and the item pool should be reviewed. Current best practice is to report omega (ideally with a confidence interval) as the primary coefficient and to include alpha for comparability with earlier literature. If you also claim temporal stability, report a test–retest correlation separately; our APA 7 reporting guide covers the formatting details.
A practical decision flow: for a unidimensional scale, report omega total as the primary coefficient. For a scale built from subscales, compute omega separately for each subscale; if a total score is also used, justify it with omega hierarchical, which shows the general factor's share. When the analysis sits inside a structural equation model, reporting composite reliability (CR) from the measurement-model loadings is sufficient, and the same 0.70 benchmark applies. Whichever coefficient you choose, presenting the item count, sample size and average inter-item correlation alongside it lets the reader place the value in context.
Computing omega in practice
- R: the omega() function in the psych package returns omega total and omega hierarchical together; the MBESS package adds confidence intervals.
- JASP: free, and its Reliability module reports alpha and omega with confidence intervals in one click — the easiest transition for SPSS users.
- SPSS: no omega in the standard menus up to version 27; Hayes's OMEGA macro runs via syntax. From SPSS 28 onwards, omega is built into the Reliability dialogue.
- AMOS / SmartPLS: composite reliability (CR) is computed from confirmatory factor analysis loadings; SmartPLS reports it automatically.
A high alpha does not prove your scale is good; quite often it only proves your scale is long.
Frequently Asked Questions
My Cronbach's alpha is below 0.70 — what should I do?
Before deleting items, examine dimensionality: if factor analysis shows the scale is not unidimensional, the correct fix is to report reliability separately for each subscale. Compute omega as well; with unequal loadings, omega can show the true reliability is higher than alpha suggests. Remove items only with content-based and factor-analytic justification.
Can I report alpha and omega together?
Yes, and that is exactly the current recommendation. Report omega with a confidence interval as the primary coefficient, and add alpha so readers can compare with older studies. If the two values are nearly identical, the tau-equivalence violation is small.
What sample size does omega require?
Because omega is factor-model based, the usual factor analysis guidance applies: at least 5–10 observations per item and ideally more than 200 respondents overall. With small samples the confidence interval widens, so always report the interval rather than the point estimate alone.
What reliability support does Celsus offer?
Celsus provides end-to-end support: testing scale dimensionality with confirmatory factor analysis, computing alpha and omega with confidence intervals, item analysis, and writing an APA 7 compliant methods section. All analyses are delivered as reproducible R or JASP files.