Cronbach's Alpha, Explained
After reading this you will be able to compute Cronbach's alpha by hand for a small dataset, read the "alpha if item deleted" table, and decide whether your scale holds together well enough to sum into a single score.
What Cronbach's alpha measures
Cronbach's alpha is a number, usually between 0 and 1, that tells you how consistently the items on a scale move together. If you ask five questions that are all supposed to measure the same thing (say, job satisfaction), you want a respondent who scores high on one item to score high on the others. Alpha summarizes that tendency in a single value.
Suppose you gave a four-item questionnaire to five people. Person A answered 4, 5, 3, 4 and person D answered 2, 3, 3, 2. Notice that both people are internally coherent: A is high across the board, D is low across the board. When that pattern holds for most respondents, alpha climbs toward 1. When answers within a person scatter randomly, alpha drops toward 0.
Alpha is a reliability coefficient. High values mean that the total score you compute by summing the items carries relatively little measurement noise. It does not tell you whether the scale measures the right thing. That is validity, a separate question entirely.
When to use it, and when not to
Use alpha when you have several items measured on the same numeric scale (for example a 1 to 5 Likert scale), all pointed at one underlying construct, and you plan to add or average them into one score. Alpha estimates how reliable that summed score is.
Do not use alpha in these situations:
- The items measure different things on purpose. A questionnaire with a satisfaction block and a demographics block should not be run through one alpha.
- You want to measure agreement between raters or coders. For that, use Cohen's Kappa calculator, which corrects agreement for chance.
- You have only two items. Alpha is defined, but it becomes unstable and hard to interpret.
A common error is treating a high alpha as proof that a scale is good. Alpha rises simply by adding more items, even weak ones. A 20-item scale can reach 0.9 while every single item correlates only modestly with the total. Report the number of items alongside alpha.
The formula and the intuition behind it
The most direct form of Cronbach's alpha uses item variances and the variance of the total score:
Here k is the number of items, \sigma_i^2 is the variance of item i across respondents, and \sigma_T^2 is the variance of the total score (the row sums). The sum \sum \sigma_i^2 adds up the individual item variances.
The intuition sits in the ratio \sum \sigma_i^2 / \sigma_T^2. When items are unrelated, the total variance is just the sum of item variances, so that ratio is close to 1 and the parenthesis is near 0. When items move together, their covariances inflate the total variance, so \sigma_T^2 grows larger than \sum \sigma_i^2. The ratio shrinks, the parenthesis grows, and alpha rises. The leading factor k/(k-1) is a small correction that keeps the maximum at 1.
A worked example with the demo data
Four items, five respondents
Load the demo data. Five people answered four items:
| Respondent | Item 1 | Item 2 | Item 3 | Item 4 | Total |
|---|---|---|---|---|---|
| A | 4 | 5 | 3 | 4 | 16 |
| B | 3 | 4 | 4 | 3 | 14 |
| C | 5 | 5 | 4 | 5 | 19 |
| D | 2 | 3 | 3 | 2 | 10 |
| E | 4 | 4 | 5 | 4 | 17 |
- Compute each item mean. Item 1:
(4+3+5+2+4)/5 = 3.6. Item 2:4.2. Item 3:3.8. Item 4:3.6. - Compute each item variance (population form, dividing by
5). Item 1: deviations are0.4, -0.6, 1.4, -1.6, 0.4, squares sum to5.2, variance1.04. Item 2:0.56. Item 3:0.56. Item 4:1.04. - Sum the item variances:
1.04 + 0.56 + 0.56 + 1.04 = 3.20. - Compute the total-score variance. The totals are
16, 14, 19, 10, 17with mean15.2. Deviations0.8, -1.2, 3.8, -5.2, 1.8, squares sum to47.2, variance9.44. - Plug in with k = 4: \alpha = \frac{4}{3}\left(1 - \frac{3.20}{9.44}\right) = \frac{4}{3}(1 - 0.3390) = \frac{4}{3}(0.6610) = 0.881.
So alpha is about 0.88. On the usual rules of thumb that counts as good internal consistency. The calculator reports the same value.
The example uses the population variance (divide by n). If you divide by n-1 instead, the constant cancels in the ratio and alpha is identical. Use whichever is convenient; the result does not change.
Reading the "alpha if item deleted" table
The calculator recomputes alpha four times, each time dropping one item. This tells you whether any single item is dragging the scale down. Read each row as: if you removed this item and kept the other three, alpha would become this value.
| Item removed | Items kept | Alpha of remaining |
|---|---|---|
| Item 1 | 2, 3, 4 | 0.834 |
| Item 2 | 1, 3, 4 | 0.829 |
| Item 3 | 1, 2, 4 | 0.923 |
| Item 4 | 1, 2, 3 | 0.834 |
The full-scale alpha is 0.881. Dropping Item 1, 2, or 4 lowers alpha, so those items help. Dropping Item 3 raises alpha to 0.923. That flags Item 3 as the weakest link: it correlates less with the others. Look back at the data and you can see why. When A scores high (Item 3 = 3, its lowest reading) and E scores high (Item 3 = 5), Item 3 does not track the total as cleanly as the rest.
How the number responds to your data
Two forces move alpha. The first is the strength of the correlations between items. The second is the number of items. The widget below lets you change both and watch alpha respond.
The formula the widget uses, \alpha = \frac{k \bar{r}}{1 + (k-1)\bar{r}}, is the standardized form of alpha, where \bar{r} is the average correlation between item pairs. Set \bar{r} = 0.3 and slide k from 4 to 20: alpha climbs from about 0.63 to about 0.90 without any single correlation improving. That is the trap to remember.
Common mistakes
- Forgetting to reverse-score negative items
- If one item is worded in the opposite direction (higher score means less of the trait), it will correlate negatively with the rest and crush alpha. Recode it before computing. A sudden negative "alpha if deleted" gap is a symptom.
- Mixing distinct subscales
- Running one alpha over items that measure two different constructs produces a middling number that describes neither. Compute a separate alpha for each subscale.
- Chasing a higher alpha by deleting items
- The deletion table is diagnostic, not a license to trim until alpha peaks. Removing Item 3 in the example lifts alpha to
0.923, but a three-item scale is shorter and may lose content you care about. Decide on substantive grounds first. - Reading alpha as a percentage
- An alpha of
0.88does not mean "88 percent of respondents agree". It is a reliability coefficient on a 0 to 1 scale, closer to a squared correlation in spirit.
Related tools
Alpha handles internal consistency of a numeric scale. When your question is whether two people (raters, coders, judges) assign the same categories, that is agreement, not consistency. Use the Cohen's Kappa calculator, which measures agreement between two raters and subtracts the agreement expected by chance.
Frequently asked questions
What is a good Cronbach's alpha value?
The common rules of thumb are 0.9 or above excellent, 0.8 or above good, 0.7 or above acceptable. Values below 0.7 suggest the items do not hang together well enough to sum into one reliable score. Very high values above 0.95 can signal redundant items that ask nearly the same question.
Can Cronbach's alpha be negative?
Yes. A negative alpha usually means at least one item is reverse-scored and correlates negatively with the others, or that the items simply do not measure a common trait. Recode any reversed items and check the deletion table before concluding anything.
How many respondents do I need?
Alpha can be computed from five respondents, as in the demo, but the estimate is unstable with small samples. For a trustworthy value aim for at least 100 respondents, and more when the scale has many items. Small samples make the "alpha if deleted" ranking especially noisy.
Does adding more items always raise alpha?
Almost always, as long as the new items correlate positively with the existing ones. At an average inter-item correlation of 0.3, going from 4 to 10 items raises alpha from about 0.63 to 0.81. This is why alpha should always be reported with the item count.
Is a high alpha proof my scale is valid?
No. Alpha measures reliability, meaning consistency of measurement. A scale can measure the wrong thing very consistently. Validity requires separate evidence, such as correlation with an outside criterion.
Related tools
Cohen's Kappa Calculator
Inter-rater agreement corrected for chance.
Binomial Distribution Calculator
Probabilities of k successes in n independent trials.
Chi-Square Test of Independence
Test association between two categorical variables from a contingency table.
Combinations & Permutations Calculator
nCr and nPr, with and without repetition.
Confidence Interval Calculator
CIs for a mean (z or t) and for a proportion (Wald & Wilson).
Correlation & Linear Regression Calculator
Pearson r, Spearman ρ, R² and the least-squares line.