How to compute Cronbach's alpha coefficient: reliability measure (internal consistency) of a questionnaire or scale composed of k items, formula alpha = (k / (k - 1)) * (1 - Σ sigma^2_i / sigma^2_T), where sigma^2_i is variance of item i and sigma^2_T variance of total scores. Range [0, 1] (can be negative if dominant negative covariances). Typical thresholds: alpha >= 0.70 acceptable (Nunnally 1978), >= 0.80 good (clinical research), >= 0.90 excellent (individual tests). Fundamental coefficient in psychometrics, Likert scale validation, cognitive test batteries, customer satisfaction surveys.

Coefficient formula

Raw alpha: alpha = (k / (k-1)) * (1 - Σ sigma^2_i / sigma^2_T), where variances calculated with n-1 in denominator (sample variance). Equivalent: alpha = (k * mean_cov) / (mean_var + (k-1) * mean_cov), where mean_var is mean item variance, mean_cov mean inter-item covariance.

Standardized alpha (correlation-based): alpha_std = (k * r_bar) / (1 + (k-1) * r_bar), where r_bar is mean inter-item correlation. Preferred when items have different scales (e.g., mix Likert 5 and 7 points); convert to z-scores first. For homogeneous items alpha_std >= alpha_raw.

Confidence interval: Feldt 1965 proposed 95% CI based on F-distribution: lower = 1 - (1 - alpha) * F(alpha/2, n-1, (n-1)(k-1)), upper = 1 - (1 - alpha) / F(alpha/2, n-1, (n-1)(k-1)). R psych::alpha() implements, Python pingouin.cronbach_alpha. Useful for replicability: if CI includes 0.70 reliability uncertain.

Decomposition: alpha increases with (a) number of items k (Spearman-Brown), (b) high inter-item correlations, (c) high total variance. Item-total correlation r_it >= 0.30 indicates contributing item; r_it < 0.20 candidate for removal. Classic R psych::alpha() output includes "alpha if item dropped" for each item, identifies problematic items.

History and context

Lee J. Cronbach (1916-2001), Stanford psychologist, published "Coefficient alpha and the internal structure of tests" in Psychometrika 1951, formally defining the coefficient generalizing KR-20 formula (Kuder-Richardson 1937) for dichotomous items to polytomous items (Likert). Cronbach worked at Educational Testing Service Princeton and University of Illinois before Stanford. The formula rapidly became de facto standard in educational and clinical psychometrics. Today alpha is the most cited metric in psychological research (over 50,000 Google Scholar citations). Modern critiques: McDonald omega (1999) and Revelle GLB (2009) often more accurate for multiple dimensionality. APA Standards for Educational and Psychological Testing (AERA-APA-NCME 2014) require reliability >= 0.80 for high-stakes individual decisions, >= 0.70 for research. ICH E9 and regulators (FDA EMA) recommend alpha for PRO (Patient Reported Outcomes) in clinical trials. ISO 9001 customer satisfaction surveys use alpha to validate instruments.

Interpretation and threshold values table

• alpha >= 0.90: excellent, possible redundancy (review if some items are paraphrases)
• 0.80 <= alpha < 0.90: good reliability, recommended for publishable research
• 0.70 <= alpha < 0.80: acceptable (Nunnally 1978 minimum threshold)
• 0.60 <= alpha < 0.70: questionable, review items with low r_it
• 0.50 <= alpha < 0.60: poor, scale unreliable
• alpha < 0.50: unacceptable, rebuild scale from scratch
• negative alpha: negative covariances, reverse-coded items not corrected
• k effect: 5 items alpha >= 0.70 harder; 20 items alpha >= 0.90 frequent
• r_bar 0.30: alpha = 0.68 (k=5), 0.90 (k=20), 0.93 (k=30)
• Mean inter-item r: 0.15 - 0.50 optimal (Briggs-Cheek 1986)

Worked numerical examples

Example 1 - job satisfaction questionnaire JSS US. 5-item Likert 1-5 scale, n = 100 employees manufacturing company (BLS classification 31-33). Item variances: 0.82; 0.76; 0.91; 0.68; 0.85, sum 4.02. Total score variance 14.5. alpha = (5/4) * (1 - 4.02/14.5) = 1.25 * (1 - 0.277) = 1.25 * 0.723 = 0.904. Excellent reliability. Mean inter-item r = 0.65 high correlation.

Example 2 - Beck Depression Inventory APA-validated. 21 items, n = 200 outpatient psychology patients. R psych::alpha() output: alpha = 0.87, 95% CI [0.84; 0.90]. Item 13 r_it = 0.18 (problematic, "bodily preoccupation" doesn't discriminate), alpha if dropped = 0.89, suggests removal. Cross-cultural validation ICD-11 depression category.

Example 3 - SUS System Usability Scale. 10 standard Likert usability items, dataset 50 banking app users. Calculated alpha 0.82 (typical SUS range 0.80-0.93 in Bangor-Kortum-Miller 2008). ISO 9241-11 usability compliance. Decision: scale valid for UAT testing.

Example 4 - mini-mental state MMSE Medicare assessment. 11 dichotomous items (KR-20). alpha = 0.73 (literature 0.70-0.80 US, Folstein 1975). Cutoff dementia score 24/30. Medicare CMS coverage decisions, NIA-AA 2018 framework.

Example 5 - organizational climate OCS questionnaire. 15 items, 4 sub-scales (leadership, communication, development, equity). Sub-scale alpha 0.82-0.89, total 0.93. McDonald omega_h = 0.87 confirms general factor. BLS workforce study small-medium businesses USA.

Limits and considerations

1. Alpha assumes tau-equivalence (items measure same construct with same scale): violation common, McDonald omega more robust.
2. Sensitive to number of items k: high alpha with large k doesn't guarantee quality.
3. Items with different scales (mix Likert 5 and 7 points): standardize first.
4. Dichotomous items: use KR-20 (special case of alpha) or tetrachoric.
5. Reverse-coded items NOT recoded: alpha negative or low.
6. Multidimensionality: alpha doesn't detect, use factor analysis (EFA/CFA).
7. Item-rest correlation r_it >= 0.30 mandatory; <0.20 remove item.
8. Minimum n=100 for reliable CI (Streiner 2003).
9. For binary clinical trial outcomes: ICC more appropriate than alpha.
10. APA Standards 2014, COSMIN 2018 require confidence interval.

Common errors

1. Reporting alpha without confidence interval (non-compliant APA, COSMIN).
2. Not recoding reverse-coded items before computation.
3. Confusing alpha with validity (alpha measures reliability, not validity).
4. Computing alpha on scale with different sub-dimensions without preliminary factor analysis.
5. Mixing Likert 5 and 7 point items without standardization (inflated alpha).
6. Using alpha for scales with 1-2 items (not meaningful, formula requires k>=3).
7. Ignoring low item-total correlation r_it in output.
8. Removing items only to increase alpha (overfitting reliability).
9. Confusing raw with standardized alpha (report both).
10. Using alpha for ordinal data with few categories (prefer ordinal alpha or omega).

Practical applications

1. Clinical psychometrics: APA tests, Beck Depression Inventory, MMPI-2-RF, SCL-90.
2. Education: validation school tests, OECD PISA, NAEP, SAT-style.
3. Health: PRO (Patient Reported Outcomes), SF-36, EQ-5D-5L FDA.
4. Marketing: customer satisfaction surveys, NPS reliability, brand image scales.
5. HR: organizational climate questionnaires, engagement, 360° leadership.
6. Ergonomics: SUS System Usability Scale, NASA-TLX workload, ISO 9241-11.
7. Sociology: Census well-being scales, BLS work quality.
8. Sports: motivation scales, Borg RPE perceived exertion, USOPC psychology.
9. Customer experience: ISO 10004 survey design, SERVQUAL.
10. Research: scale validation for publication, FDA EMA regulatory submission.

Support

Questions: info@calcoi.com. Educational tool for scale reliability. For professional applications consult Streiner-Norman-Cairney "Health Measurement Scales" 5th ed (Oxford 2015), R psych::alpha(), Python pingouin.cronbach_alpha, SPSS Reliability Analysis, Mplus for omega.

Updated July 2026