Work out Cronbach's alpha for a questionnaire scale ALPHA
Paste the item scores straight from a spreadsheet, a row per person. You get alpha, each item's correlation with the rest, alpha without it, and a flag on any item that looks as if it was never reversed.
The item scores
Paste from a spreadsheet: columns split on tabs, commas or spaces. A first row of names is used as the item names. An empty cell, or anything that is not a number, counts as unanswered.
The result
For the write-up
How it works
ALPHA reads a header row of item names if there is one, reverse-scores the items you name before it starts, and leaves out anybody with an item unanswered, saying how many. Every figure is held to the textbook formulas worked in NumPy, and nothing typed here leaves this tab.
What it does, and what it does not
Cronbach's alpha is the most reported figure for whether the items of a questionnaire scale hang together: whether people who score high on one item tend to score high on the others. It is worked out here from the items' variances and the variance of the total, the same formula SPSS and every textbook use, and the item table shows the corrected item-total correlation (each item against the total of the others, not including itself) and alpha with each item left out.
Alpha is not a measure of whether a scale measures one thing, and it is not validity. It assumes each item measures the same thing equally well; when they do not, McDonald's omega, from a factor analysis, is the better estimate, and this page does not work it out. It gives no confidence interval for alpha. It does not reverse items by itself: it flags one that looks reversed and leaves the decision with you.
The figures are held, in this site's own tests, to the same formulas worked in NumPy on four scales, one with an item deliberately left unreversed. To describe one item's answers, FIGURES does that.
Questions people ask about ALPHA
What is a good Cronbach’s alpha?
There is no line that makes a scale good. 0.70 is the figure most often quoted, from Nunnally, who also wrote that it depends on the purpose. Alpha rises with the number of items, so twenty mediocre items can beat five good ones, and a very high alpha (above about 0.95) often means items that ask the same question twice. Report the figure and the number of items, and look at the item statistics.
What does a negative item-total correlation mean?
That the item goes the opposite way to the rest of the scale. Almost always it is a reversed item ('I rarely feel calm' in a calm scale) that was not reverse-scored before alpha was worked out. Name it in the Reverse-score box and the page flips it, using the lowest and highest possible scores you give.
What is alpha if item deleted?
What alpha would be with that one item left out. If it is higher than the alpha for the whole scale, the item is pulling consistency down. That is a reason to look at the item’s wording, not an automatic reason to drop it: a scale built by deleting items until alpha is highest tends to measure something narrower than intended.
How are missing answers handled?
Anybody with any item unanswered is left out of the whole calculation, which is what SPSS does by default (listwise deletion). The page says how many people that removed. If a lot are removed, the answer may not describe the people who were.
What is the difference between alpha and standardised alpha?
Alpha uses the items’ variances and covariances as they are. Standardised alpha is worked out from the average correlation between items, as if every item had been put on the same scale first. They are close when the items have similar spreads; when they differ, report alpha unless the items are going to be standardised before they are added up.
Doing this often? OBSCURA Gnosis works out alpha on the items of a whole questionnaire file, beside factor analysis, with the output saved to Word or PDF. On its own, US$2 a month; with everything else, US$5 a month or US$48 a year.