What Is Cronbach's Alpha (Video)

Written guide: Internal Consistency Reliability in Research: A Practical Guide with Examples

Explains Cronbach's alpha for measuring internal consistency reliability in tests.

Watch on YouTube

Chapters

  1. 0:25 Why a scale needs internal consistency
  2. 1:52 True score and error
  3. 2:44 Item variance, total variance and covariance
  4. 3:56 The alpha formula
  5. 4:40 Worked example
  6. 6:26 Interpreting alpha
  7. 7:10 Number of items and redundancy
  8. 8:13 What alpha does not show
  9. 9:11 Omega from factor loadings
  10. 10:35 Item analysis
  11. 11:38 Reading SPSS reliability output
  12. 13:25 Alpha and split-half reliability
  13. 14:25 Troubleshooting
  14. 15:29 Reverse-coding in practice
  15. 16:11 A scale-development case
  16. 17:40 Reporting alpha in a paper
  17. 18:23 Recap and quiz
Video transcript

0:25 Why a scale needs internal consistency

First, we will look at why a scale needs internal consistency and separate the true score from error. An anxiety questionnaire asks several questions, for example how often you feel restless or how often you worry about the future, and adds the answers into a single score. The total is useful only when the questions measure the same characteristic. A construct is the characteristic that a scale is meant to measure, such as anxiety or job satisfaction. Each question or statement on the scale is an item, and a Likert scale gives ordered response options from 1, strongly disagree, to 5, strongly agree. Internal consistency is the degree to which the items of one scale measure the same construct, so that their scores rise and fall together across respondents. The items of a consistent scale behave like the voices of a choir that hold the same note: when the voices agree, the sound is clear, and a singer who drifts off key weakens the sound in the same way that a poor item weakens a scale. Lee Cronbach published coefficient alpha in 1951 to measure this agreement. Alpha generalizes the Kuder-Richardson formula for items scored right or wrong.

1:52 True score and error

To see what alpha estimates, you need the classical model of a test score. Classical test theory writes an observed score as the true score plus error. The true score is the score a person would get with no random measurement error. Reliability is the proportion of the variance in observed scores that comes from true scores. Variance is the average squared distance of scores from their mean. Random errors on different items do not move together. When you add items, the true-score parts add up while the errors partly cancel, so a longer scale of good items is more reliable than a shorter scale. When several thermometers hang in the same room, each reading contains a small random error, and the average of the readings tends to lie closer to the true temperature than a single reading does.

2:44 Item variance, total variance and covariance

Second, we will define item variance, total variance and covariance, and build the alpha formula. Alpha compares the variance of single items with the variance of the total score, and the comparison makes sense once you see how the two kinds of variance differ. Here you use the sample variance, which divides the sum of squared deviations by n minus 1. Item variance describes how much respondents differ on one item, and total-score variance describes how much they differ on the sum. The variance of a sum equals the sum of the item variances plus twice the sum of all the item covariances. Covariance measures how two variables vary together, and it is positive when high scores on the first variable go with high scores on the second variable. When items measure the same construct, their covariances are positive, so the total varies much more than the item variances alone would produce. When items are unrelated, the covariances sit near zero and the total variance comes close to the sum of the item variances. Alpha rises as the share of the total variance that comes from covariance increases.

3:56 The alpha formula

The formula turns that comparison into a single number. Alpha equals k divided by k minus 1, multiplied by 1 minus the sum of the item variances divided by the total-score variance, where k is the number of items. The ratio inside the brackets is the share of the total variance that the items contribute separately. Subtracting that ratio from 1 leaves the share that comes from the items varying together. The factor k divided by k minus 1 corrects for the fact that the share from covariance can never reach 1 with a finite number of items. Alpha usually lies between 0 and 1, and it can fall below 0 when items vary in opposite directions.

4:40 Worked example

Third, we will work through an example and interpret alpha. To apply the formula, consider an illustrative data set in which six respondents answer four anxiety items on a 1 to 5 scale. The table lists their answers, from 5, 5, 5 and 5 for the first respondent to 2, 4, 2 and 4 for the sixth. You start with the variance of item 1. Its scores are 5, 3, 3, 4, 4 and 2, with a mean of 3.5. The squared deviations are 2.25, 0.25, 0.25, 0.25, 0.25 and 2.25, which sum to 5.5, and dividing by 5 gives 1.1. The same steps give 0.8 for item 2, 2.0 for item 3 and 1.1 for item 4, so the item variances sum to 5.0. You then add each respondent's answers and get totals of 20, 15, 10, 12, 15 and 12. The mean total is 14, the squared deviations sum to 62, and the total-score variance is 62 divided by 5, which is 12.4. Subtracting 5.0 from 12.4 leaves 7.4, the part of the total variance that comes from the items varying together. For the last step, k divided by k minus 1 is 4 divided by 3, about 1.333. 5.0 divided by 12.4 is 0.403, 1 minus 0.403 is 0.597, and 1.333 times 0.597 is 0.796, so alpha is about 0.80.

6:26 Interpreting alpha

To judge whether 0.80 is high enough, you compare it with thresholds that are widely cited in the psychometric literature. These conventions place alpha at about 0.70 for early-stage research, 0.80 for basic research, and 0.90 or higher when scores guide important decisions about individuals. Treat these figures as guidelines, and judge them against the purpose of the scale. The example, at 0.80, meets the usual standard for research use. Alpha describes the scores from one particular sample, and its value can change in another sample, so you report alpha for your own data.

7:10 Number of items and redundancy

Fourth, we will look at the number of items and at what alpha does not show. The value of 0.80 in the example also reflects the length of the scale, because alpha depends on the number of items as well as on how strongly the items agree. The standardized form of alpha depends only on the number of items and the average correlation between the items. With an average inter-item correlation of about 0.50, standardized alpha is about 0.80 for 4 items, 0.89 for 8, 0.92 for 12 and 0.95 for 20. A long scale can therefore reach a high alpha even when its items correlate only moderately. Asking the same question five times in slightly different words raises alpha without widening what the scale measures. Values above about 0.95 often signal redundant items, which make the questionnaire longer while measuring the same content again.

8:13 What alpha does not show

The alpha of 0.80 in the example shows how consistently the four items vary together, but it gives limited information about the structure of the scale. Alpha does not show unidimensionality, which is the property of measuring a single construct. A scale made of two related clusters of items can still produce a high alpha. Alpha also assumes that the items are tau-equivalent, so that each item measures the true score on the same scale and the items differ only in their error. When items relate to the construct with different strengths, alpha is a lower bound on reliability. McDonald's omega is a reliability coefficient that you compute from a factor model, and omega does not require tau-equivalence. Researchers increasingly report omega with alpha or in place of alpha. To check dimensionality, you run a factor analysis of the items.

9:11 Omega from factor loadings

Fifth, we will cover omega from factor loadings. McDonald's omega uses the factor loadings that alpha ignores, and a six-item example shows the calculation. A factor loading is the correlation between one item and the factor the items share. Its square is the share of variance the factor accounts for, and the remainder is the item's unique variance. Take a one-factor model with six items, with standardized loadings of 0.80, 0.75, 0.70, 0.65, 0.60 and 0.40. Add the loadings and you get 3.90. Square that sum and you get 15.21. Square each loading and subtract it from 1 to get the unique variances: 0.36, 0.4375, 0.51, 0.5775, 0.64 and 0.84. They sum to 3.365. Omega is the squared sum of loadings divided by that squared sum plus the sum of unique variances. 15.21 divided by 18.575 gives 0.82. Standardized alpha for the same six items is 0.81. Alpha sits slightly below omega here because the loadings differ.

10:35 Item analysis

Sixth, we will run an item analysis and read SPSS reliability output. Alpha summarizes the whole scale, and item statistics show how each item contributes to that summary. The corrected item-total correlation is the correlation between one item and the total of the remaining items. In the example the four values are 0.50, 0.71, 0.81 and 0.50. Alpha if item deleted is the alpha of the scale without that item. For the four items in the example, alpha if item deleted is 0.795, 0.712, 0.635 and 0.795, so removing item 3 would lower alpha most, because item 3 agrees most with the rest of the scale. If removing an item raises alpha, or if its corrected item-total correlation is very low, check the wording of the item before you remove it from the scale. Removing items to raise alpha also narrows what the scale covers.

11:38 Reading SPSS reliability output

You read these item statistics from software output, and SPSS prints them for the four-item example in two tables. The Reliability Statistics table has three columns, which give Cronbach's alpha of 0.796, alpha based on standardized items of 0.799, and the number of items, 4. The two coefficients differ slightly because the second one standardizes each item to a variance of 1 before the calculation, which removes the effect of items measured on different scales. The Item-Total Statistics table has one row for each item and four columns. Scale mean if item deleted is the mean of the total score with that item removed: 10.50 for item 1, 10.00 for item 2, 11.00 for item 3 and 10.50 for item 4. Scale variance if item deleted is the variance of the same total: 8.300, 8.000, 5.200 and 8.300. Corrected item-total correlation is the correlation between the item and the sum of the other items: 0.496, 0.712, 0.806 and 0.496. Cronbach's alpha if item deleted is the alpha the scale would reach without that item: 0.795, 0.712, 0.635 and 0.795. Items 1 and 4 have the lowest corrected item-total correlations, 0.496, and removing either of them leaves alpha at 0.795, slightly below the 0.796 of the full set.

13:25 Alpha and split-half reliability

Seventh, we will connect alpha to split-half reliability, troubleshoot low values and handle reverse-coding. Alpha also connects to an older method, split-half reliability, and the connection explains why researchers prefer alpha to a single split. A split-half coefficient compares two halves of a test, and the Flanagan-Rulon form uses the variances of the two half scores and of the total. Cronbach showed that alpha equals the mean of all possible split-half coefficients of this form. In the example, the three ways to split four items into two pairs give 0.94 for items 1 and 2 against items 3 and 4, 0.68 for items 1 and 3 against items 2 and 4, and 0.77 for items 1 and 4 against items 2 and 3. Their mean is 0.796, which equals alpha. Alpha therefore removes the arbitrary choice of one particular split.

14:25 Troubleshooting

When you calculate alpha on your own data, an unexpected value usually has one of a few common causes. A negative or very low alpha points first to reverse-worded items that you have not yet recoded. A reverse-worded item states the construct in the opposite direction, such as the statement that you feel calm on an anxiety scale. If item 3 in the example were left in reversed form, alpha for the same data would fall to minus 2.0. On a 1 to 5 scale you recode such an item with 6 minus the score. A two- or three-item scale often gives a modest alpha even when the items agree reasonably, so report the mean inter-item correlation as well. For a questionnaire with several subscales, you calculate alpha for each subscale separately, because alpha for the whole questionnaire mixes items that measure different constructs. Report alpha with the number of items, the sample size and, where possible, a confidence interval, and consider reporting omega.

15:29 Reverse-coding in practice

Four habits prevent the error that an unrecoded item causes. Identify reverse-worded items when you write the scale, and mark them in the codebook. Recode them before you score anything, by subtracting each score from the sum of the highest and lowest possible scores, which gives 4 minus the score on a 0 to 4 scale. Keep the original column and create a new recoded column beside it, so the raw responses stay available for checking. Then correlate each recoded item with the total score. A recoded item should correlate positively with the total, and a negative correlation means the recoding is wrong or missing.

16:11 A scale-development case

The last part follows a scale-development case and shows how to report alpha in a paper. A team that builds a new work-stress questionnaire applies these methods in sequence. A team writes ten items on work stress, each answered on a scale from 1 to 5, and 200 people respond. Alpha for the ten items is 0.83, with a 95 percent confidence interval from 0.79 to 0.86. The interval comes from the F distribution, and it shows how much sampling error surrounds the estimate. The item statistics show one problem. Item 7 reads 'I enjoy the variety in my tasks'. Its corrected item-total correlation is 0.14, and the other nine items lie between 0.51 and 0.62. Alpha if item 7 is deleted is 0.85. The team reads item 7 again and sees that it measures job satisfaction rather than stress. They remove it on content grounds as well as for the number. Alpha for the nine-item scale is 0.85, with a 95 percent confidence interval from 0.82 to 0.88. A factor analysis on a new sample tests whether the nine items load on one factor, and test-retest data on the same people after an interval tests whether scores stay stable when stress does not change.

17:40 Reporting alpha in a paper

The final step of the case is the sentence in the methods or results section. A reader needs the number of items, the sample and the coefficient with a confidence interval. A typical results section states that internal consistency of the nine-item work-stress scale was good in this sample of 200 people, with a Cronbach's alpha of 0.85 and a 95 percent confidence interval from 0.82 to 0.88, and with McDonald's omega reported alongside. State the reason for each removal in the text, so a reader can judge whether the scale still covers the construct it was written for.

18:23 Recap and quiz

Here is a short review of the main points. Alpha estimates reliability from the share of the total-score variance that comes from items varying together, corrected by k divided by k minus 1, and you compute it from the item variances and the total-score variance. The worked example gave 0.80 for four items. The size of alpha depends on the number of items, and a high alpha does not show that a scale measures a single construct. Item statistics show how strongly each item agrees with the rest of the scale, and the mean of all possible Flanagan-Rulon split-half coefficients equals alpha. Omega, calculated from the factor loadings, sits at or above alpha, and the work-stress case showed how you remove a weak item on content grounds and report alpha with its confidence interval. Test what you learned with this question. Consider a 10-item anxiety scale in which you forgot to recode one reverse-worded item. What happens to alpha, and what do you do before you calculate alpha again? Take a moment to work it out, then compare with the answer. The unrecoded item correlates negatively with the other items, so alpha drops, and the item shows a negative corrected item-total correlation. You recode the item, check the item statistics again, and then recalculate alpha.

Read more

All videos