How to Run Cronbach's Alpha in SPSS

If your questionnaire items do not hang together, every scale total you compute from them is noise, and every t-test or correlation you run on that total inherits the problem. That is why reliability analysis comes first, before descriptives, before hypothesis tests, before anything. Cronbach's alpha measures whether the items on a scale behave as measurements of the same thing, and SPSS computes it in a few clicks. Here is the full procedure: setup, reverse coding, the two output tables that matter, and the rules for deciding whether an item deserves to be dropped. For the theory behind the coefficient, see our conceptual piece on Cronbach's alpha; this tutorial is the click path.

When to Run Reliability Analysis

Any time your study uses a multi-item scale. That covers Likert questionnaires you built yourself, established psychological inventories (Beck Depression Inventory, Rosenberg Self-Esteem Scale, Perceived Stress Scale), and translated or adapted instruments, which need their reliability re-established in the new language and population regardless of how well they performed in the original.

Run it before everything else. The sequence in a typical thesis is: clean the data, recode reverse-scored items, run Cronbach's alpha per scale, compute the scale totals, and only then start on descriptive statistics and hypothesis tests. Students often bolt reliability on at the end because a supervisor asked for it, and occasionally discover at that point that one scale has an alpha of .48 and half the results chapter rests on it. Ten minutes of checking up front prevents that conversation.

One rule that trips people up: alpha is computed per construct, not per questionnaire. If your survey measures burnout with 9 items and job satisfaction with 6, you run reliability twice, once for each set. A single alpha across all 15 items answers a question nobody asked.

Setting Up Your Data

One column per item, one row per respondent. A 10-item anxiety scale occupies ten columns, named something like Anx1 through Anx10 (SPSS variable names cannot contain spaces). Each cell holds that respondent's raw answer, 1 to 5 on a standard Likert scale.

The running example for this tutorial: 142 respondents completed a 10-item anxiety questionnaire, each item scored 1 (strongly disagree) to 5 (strongly agree). Higher scores mean more anxiety. All ten items are meant to measure the same construct, and, after the recoding step below, all are scored in the same direction. Both conditions matter. Alpha assumes the items form one scale pointing one way; feed it a mix of constructs or a mix of directions and the number it returns is garbage.

Reverse-Coded Items: Fix These First

Most published scales include negatively worded items to keep respondents reading. In an anxiety scale where item 3 reads "I feel calm in most situations," a highly anxious person answers 1, not 5. Left as is, item 3 correlates negatively with the other nine items and drags alpha down, sometimes below zero. Recode it before running reliability, not after.

Go to Transform > Recode into Different Variables. (Recode into Same Variables also works, but it overwrites the original answers, and you cannot undo it after saving. The Different Variables route keeps the raw data intact, which you will be glad of the first time you mis-specify a recode.) Move Anx3 across, give the output variable a name like Anx3R, click Change, then click Old and New Values and enter the swaps for a 5-point scale:

• Old value 1 → New value 5
• Old value 2 → New value 4
• Old value 3 → New value 3
• Old value 4 → New value 2
• Old value 5 → New value 1

Click Continue, then OK, and use Anx3R instead of Anx3 in every analysis from here on. For a 7-point scale the same logic applies (1→7, 2→6, and so on). Check the scale's original publication or manual to find out which items are reverse-scored; guessing from the item wording usually works, but the manual settles it.

Running the Analysis

The menu path is Analyze > Scale > Reliability Analysis.

1. Move all ten items into the Items box: Anx1, Anx2, Anx3R (the recoded version), Anx4, and so on through Anx10. Only the items of this one scale, nothing else.
2. Leave Model set to Alpha, the default.
3. Click Statistics. Under "Descriptives for," check Item and Scale if item deleted. Click Continue.
4. Click OK.

The Statistics step is the one people skip, and it is the one that matters. Without "Scale if item deleted," SPSS prints the overall alpha and stops, leaving you no way to see whether a single bad item is dragging the coefficient down.

Reading the Output

The headline number sits in the Reliability Statistics table, which has exactly two cells: Cronbach's Alpha and N of Items. In our example, α = .84 across 10 items. The widely cited rules of thumb come from George and Mallery (2003):

Cronbach's Alpha	Interpretation
Below .60	Poor — the scale total is not trustworthy
.60 – .69	Questionable — sometimes tolerated for short or new scales
.70 – .79	Acceptable
.80 – .89	Good
.90 and above	Excellent

The .70 threshold is the one most committees enforce. One counterintuitive point: an alpha above .95 is not a triumph. Alpha rises with inter-item correlation and with the number of items, so a very high value on a long scale often means several items are near-duplicates of each other, adding length without adding information. Nobody fails a thesis over α = .96, but do not lengthen a scale just to push the number up.

Also glance at the Case Processing Summary above the table. Alpha is computed only on respondents with complete data across all items, so if 142 people took your survey but the table says N = 121, listwise deletion has quietly removed 21 of them.

The Item-Total Statistics Table

This is the diagnostic table, and two of its columns do the work.

Corrected Item-Total Correlation shows how strongly each item correlates with the sum of the remaining items ("corrected" means the item itself is excluded from the total, so it cannot inflate its own correlation). Values above .30 indicate the item belongs to the scale. Values below .30 flag an item that respondents answer independently of the rest, and a negative value means either a reverse-coded item you forgot to recode or an item measuring something else entirely.

Cronbach's Alpha if Item Deleted answers a hypothetical: what would alpha be if this item were removed? Scan down the column and compare each value against your overall alpha. In our example, most items show .82 to .84 against the overall .84, which means removing any of them changes little. Item 7 is the exception: its corrected item-total correlation is .12, and alpha if deleted is .87. The item is pulling the scale down, not holding it up.

When to Delete an Item (and When Not To)

The temptation is to treat the "Alpha if Item Deleted" column as a menu and remove whatever pushes the number highest. Resist it. Chasing alpha item by item is overfitting: you tune the scale to the quirks of your particular 142 respondents, and the published version of the instrument stops being the instrument you actually used. Drop an item only when all three of these hold:

1. Its corrected item-total correlation is below .30.
2. Removing it raises alpha meaningfully, by at least .02 to .03, not by a rounding digit.
3. The deletion makes theoretical sense. Reread the item. Is it double-barreled, ambiguous, or arguably measuring a different construct? If the item looks fine and you cannot say why it misbehaves, note the weak statistic and think twice before cutting.

Item 7 in our example passes all three: correlation of .12, a jump from .84 to .87, and on rereading it turns out to be double-barreled ("I feel nervous and avoid social events"), tapping avoidance behavior rather than felt anxiety. Out it goes. After any deletion, re-run the analysis on the remaining items, because every statistic in the item-total table shifts once the scale changes. Delete one item at a time, never several at once. And keep in mind that deleting items from a validated instrument is a bigger decision than deleting them from a scale you wrote yourself; for established inventories, report the weak item and discuss it rather than quietly amputating it.

If you want a second opinion on the whole sequence, Academic Stats Agent runs the reliability analysis, flags weak items against these same criteria, and drafts the reporting sentence for you.

Reporting in APA Format

Reliability reporting is short. One sentence in the methods or results section per scale:

Internal consistency for the 10-item anxiety scale was good (α = .84).

If you deleted an item, say so and show your evidence:

After removing item 7 (corrected item-total correlation = .12), the final 9-item scale showed good internal consistency (α = .87).

The usual formatting rules: Greek letter α, no leading zero (α = .84, not 0.84), two decimals. If your study has several scales, a small table with one row per scale (scale name, number of items, alpha) reads better than a paragraph of repeated sentences. Report the alpha from your own sample, not the one from the original validation paper; you may cite the original value for comparison, but your data must carry their own coefficient. How you then analyze the resulting scale totals is covered in our guide to Likert scale data.

Frequently Asked Questions

What is an acceptable Cronbach's alpha for a thesis?

The conventional minimum is .70. Values between .60 and .69 are questionable but sometimes tolerated for short scales (3–4 items) or newly developed instruments, provided you acknowledge the limitation. Values of .80 and above are good, and .90 and above excellent, though anything over .95 may signal that your items are redundant rather than exceptionally reliable.

Why is my Cronbach's alpha negative?

A negative alpha almost always means one or more reverse-worded items were never recoded, so they correlate negatively with the rest of the scale. Recode those items first (1 becomes 5, 2 becomes 4, and so on for a 5-point scale) and re-run the analysis. If alpha stays negative after recoding, check for data entry errors or items that do not belong to the same construct.

Should I calculate alpha for the whole questionnaire or for each subscale?

For each subscale separately. Cronbach's alpha assumes all items measure a single construct, so running it across a questionnaire that mixes, say, anxiety items with self-esteem items produces a meaningless number. Report one alpha per subscale, each computed only from that subscale's items.