Your answers
Blanks, NA and the missing codes count as missing, and a respondent with any missing item is left out. Columns named like an identifier (id, respondent, timestamp, ...) and text columns are skipped unless you list them as items.
Result
Coefficients
Items
How honest are the intervals?
In 300 simulated studies per setting (6 items, one common factor, normal answers; eval/coverage.mjs), each interval was scored against the true reliability of the total score. With equal loadings every 95% interval covered it about 91 to 95% of the time. With unequal loadings (0.2 to 0.9) alpha aims at a lower number than the true reliability, so its Feldt interval covered it 93% of the time at n = 50, 73% at n = 200 and 11% at n = 1000, while omega's BCa interval held 93 to 95%. A bigger sample makes alpha's interval narrower around the wrong value.
Method and citation
Alpha (Cronbach 1951) is k/(k−1) · (1 − sum of item variances / variance of the total), from the item covariance matrix of complete respondents. Its interval is Feldt, Woodruff & Salih's (1987): (1 − alpha) follows an F distribution with n−1 and (n−1)(k−1) degrees of freedom. Standardised alpha uses the mean inter-item correlation instead, and is the alpha of the sum of z-scored items.
Guttman's lambda-2 adds the squared covariances to lambda-1 and is never below alpha; lambda-6 uses each item's error variance given all the others, 1/[C−1]ii (Guttman 1945). Both are lower bounds to reliability under the classical model (Sijtsma 2009).
Omega (McDonald 1999) fits one common factor to the item covariance matrix by unweighted least squares (minimum residual: principal axes iterated to convergence, from squared multiple correlations) and reports (Σλ)² / ((Σλ)² + Σψ). The standardised omega fits the correlation matrix instead. Its interval is the bias-corrected and accelerated bootstrap (Efron 1987) over respondents, which Kelley & Pornprasertmanit (2016) recommend; the acceleration comes from the jackknife. The same resamples give a BCa interval for alpha. The Spearman-Brown projection (Spearman 1910; Brown 1910) gives the number of similar items a target reliability needs.
Checked against other software on the bundled answers (eval/reference.py, eval/reference.json): alpha and its Feldt interval match pingouin 0.7.0 and scipy to 1e-9; standardised loadings match factor_analyzer 0.5.1's minres fit to 1e-5; omega matches a scipy least-squares fit to 1e-7; lambda-2, lambda-6, standardised alpha, item-rest correlations and alpha if deleted match numpy to 1e-10.
@article{cronbach1951coefficient, author = {Cronbach, Lee J.}, title = {Coefficient alpha and the internal structure of tests}, journal = {Psychometrika}, volume = {16}, number = {3}, pages = {297--334}, year = {1951}, doi = {10.1007/BF02310555}}
@article{guttman1945basis, author = {Guttman, Louis}, title = {A basis for analyzing test-retest reliability}, journal = {Psychometrika}, volume = {10}, number = {4}, pages = {255--282}, year = {1945}, doi = {10.1007/BF02288892}}
@book{mcdonald1999test, author = {McDonald, Roderick P.}, title = {Test Theory: A Unified Treatment}, publisher = {Lawrence Erlbaum}, address = {Mahwah, NJ}, year = {1999}, doi = {10.4324/9781410601087}}
@article{feldt1987statistical, author = {Feldt, Leonard S. and Woodruff, David J. and Salih, Fathi A.}, title = {Statistical inference for coefficient alpha}, journal = {Applied Psychological Measurement}, volume = {11}, number = {1}, pages = {93--103}, year = {1987}, doi = {10.1177/014662168701100107}}
@article{kelley2016confidence, author = {Kelley, Ken and Pornprasertmanit, Sunthud}, title = {Confidence intervals for population reliability coefficients: Evaluation of methods, recommendations, and software for composite measures}, journal = {Psychological Methods}, volume = {21}, number = {1}, pages = {69--92}, year = {2016}, doi = {10.1037/a0040086}}
@article{efron1987better, author = {Efron, Bradley}, title = {Better bootstrap confidence intervals}, journal = {Journal of the American Statistical Association}, volume = {82}, number = {397}, pages = {171--185}, year = {1987}, doi = {10.1080/01621459.1987.10478410}}
@article{sijtsma2009use, author = {Sijtsma, Klaas}, title = {On the use, the misuse, and the very limited usefulness of {Cronbach's} alpha}, journal = {Psychometrika}, volume = {74}, number = {1}, pages = {107--120}, year = {2009}, doi = {10.1007/s11336-008-9101-0}}
@article{spearman1910correlation, author = {Spearman, Charles}, title = {Correlation calculated from faulty data}, journal = {British Journal of Psychology}, volume = {3}, number = {3}, pages = {271--295}, year = {1910}, doi = {10.1111/j.2044-8295.1910.tb00206.x}}
@misc{nhanes2017dpq, author = {{National Center for Health Statistics}}, title = {National Health and Nutrition Examination Survey 2017-2018: Mental Health - Depression Screener (DPQ\_J)}, year = {2020}, url = {https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/DPQ_J.htm}}
Limits
- Reliable is not valid. A scale can measure something very consistently and still not measure what its name says. Reliability is necessary for validity, not proof of it.
- One factor. Alpha and this omega assume the items share one common factor. For a scale with subscales, compute each subscale separately; a hierarchical omega needs a multi-factor model that is not here.
- Listwise deletion. Respondents with any missing item are left out. If missingness is common or related to the trait, the estimate describes the respondents who answered everything.
- Ordinal answers. Coefficients come from Pearson covariances of the item scores, as the total score is computed. Ordinal (polychoric) alpha and omega describe a hypothetical continuous score instead and are not computed.
- Not modelled: the greatest lower bound, test-retest and inter-rater reliability (see the inter-rater agreement tool), correlated errors, item response theory, and survey weights.
Questions
What is a good Cronbach's alpha?
Common rules of thumb call 0.7 acceptable for research comparing groups, 0.8 good, and 0.9 or more what decisions about one person need. They are conventions, not laws. A very high alpha (above about 0.95) often means items that ask the same thing in different words, which makes a scale longer without making it better.
Alpha or omega?
Alpha equals the reliability of the total score only when every item measures the common trait equally strongly (tau-equivalence). When the items differ, alpha is a lower bound and omega, which lets each item have its own loading, is closer to the truth. Report both: when they agree the choice does not matter, and when they differ the gap tells you the items are unequal.
Does a high alpha mean my scale measures one thing?
No. Alpha rises with the number of items and with any correlation among them, so two unrelated clusters of items can still give a high alpha. Look at the eigenvalues and the one-factor residuals this page reports, or fit a factor model, to judge whether the scale is unidimensional.
Should I drop the item that raises alpha if deleted?
Not on that number alone. Removing an item that covers part of the construct can raise alpha and lower validity. Drop an item when its content is off-target, its wording is confusing, or it is reverse-keyed and was not reversed, and confirm the change on new data, because choosing items on the same sample inflates the result.
What do I do with reverse-keyed items?
Items worded in the opposite direction ("I rarely feel calm" next to "I often feel calm") must be reversed before computing reliability, or alpha falls and can go negative. List them under reverse-keyed items and give the scale's lowest and highest possible answers; the page recodes each answer as minimum plus maximum minus the answer, and warns about items that correlate negatively with the rest.