When you see "alpha = 0.82" in a survey report, what does that number tell you? And if an analysis returns "alpha = −0.15", is your survey broken? Cronbach's alpha is one of the most frequently reported, yet least understood, statistics in employee surveys.
- Cronbach's alpha is a coefficient that shows how consistently people answer items that are expected to measure the same concept. For indices in organizational surveys, 0.70 is acceptable and 0.80–0.90 is considered good.
- When alpha comes out low or negative, the first thing to check is reverse-coded items. A single item left unreversed in the calculation can push a good scale's alpha below zero.
- The "alpha if item deleted" and "corrected item-total correlation" columns are the fastest way to find the item that is weakening a scale.
- Alpha shows that a scale is reliable, not that it measures the right thing. A very high alpha can also point to items that repeat each other.
In our post on survey reliability and validity, we covered the conceptual background of alpha, its formula and its relationship with the number of items. This post focuses on practice: what alpha should be, how to calculate it step by step with a small employee survey dataset, and what to do when it comes out low or negative.
What is Cronbach's alpha?
Cronbach's alpha is a coefficient that shows how consistently a group of items expected to measure the same concept are answered. For example, if an engagement index consists of four items and an employee answers "strongly agree" to one of them, they will most likely answer the others positively too. If the items really do measure the same thing, this consistency is high, and so is alpha.
Alpha usually takes a value between 0 and 1. That is not a requirement, however; as you will see below, it can also be negative.
What is a good Cronbach's alpha?
The short answer: for indices in organizational surveys, 0.70 is the floor and 0.80–0.90 is the target.
| Alpha value | Common interpretation |
|---|---|
| Below 0.60 | Weak; the items may not be measuring the same concept |
| 0.60–0.70 | Borderline; acceptable for subdimensions with few items |
| 0.70–0.80 | Acceptable |
| 0.80–0.90 | Good; the target range for organizational surveys |
| 0.90–0.95 | Very good |
| Above 0.95 | The items may be repeating each other |
These thresholds are not a statistical law but established conventions. Remember that alpha rises with the number of items: 0.68 is a reasonable result for a three-item subdimension, while 0.85 is unremarkable for a twenty-item scale. You can find the details of this relationship in the table in our post on reliability and validity.
A step-by-step example: a four-item index
The best way to see how alpha works is to calculate it on a small example. Suppose a team has four items measuring a "development" index, and eight employees have answered on a 1–5 Likert scale:
- I1: My manager takes an interest in my development.
- I2: I have opportunities to learn new skills in my job.
- I3: I have enough time to attend training programs.
- I4: I feel I cannot develop myself at this organization. (negatively worded)
| Employee | I1 | I2 | I3 | I4 |
|---|---|---|---|---|
| A | 5 | 4 | 3 | 1 |
| B | 4 | 4 | 2 | 2 |
| C | 2 | 3 | 4 | 4 |
| D | 3 | 3 | 5 | 3 |
| E | 1 | 2 | 3 | 5 |
| F | 4 | 5 | 2 | 2 |
| G | 2 | 2 | 4 | 3 |
| H | 5 | 4 | 3 | 2 |
A real analysis would need far more responses to calculate alpha. This small dataset is only meant to illustrate the logic of the calculation.
Step 1: Reverse the reverse-coded item
I4 is negatively worded: an employee who answers "strongly agree" (5) actually holds the most negative view of their development. To align this item with the others, you need to reverse its score. On a 1–5 scale, the formula is: new score = 6 − old score. So 5 becomes 1, and 4 becomes 2.
What happens if you skip this step? In the example data, the alpha calculated without reversing I4 comes out at −4.39. A negative alpha shows that the items are moving in opposite directions, and it almost always points to an item that was not reversed.
Step 2: Calculate the variance of each item
After reversing I4, the variance of each item across the eight responses is: 2.21 for I1, 1.12 for I2, 1.07 for I3 and 1.64 for I4. The total is 6.05. In Excel, you can use the VAR.S function for each column.
Step 3: Calculate the variance of the total scores
Each employee's total score across the four items is 17, 14, 11, 14, 7, 15, 11 and 16 respectively. The variance of these totals is 10.70.
Step 4: Plug the values into the formula
α = (k / (k − 1)) × (1 − sum of item variances / variance of total scores)
α = (4 / 3) × (1 − 6.05 / 10.70) ≈ 0.58
The result is 0.58: even with the reverse coding fixed, the index looks weak. So where is the problem?
Finding the problem item: "alpha if item deleted"
Most statistics software reports two more values for each item alongside alpha:
- Alpha if item deleted: Shows what alpha would be if that item were removed from the scale. If alpha rises noticeably when an item is removed, that item does not fit with the others.
- Corrected item-total correlation: Shows how strongly the item is related to the sum of the other items, excluding itself. Values below 0.30 are weak, and negative values are a sign of a serious mismatch.
For the example data, this table looks like this:
| Item | Corrected item-total correlation | Alpha if item deleted |
|---|---|---|
| I1: My manager takes an interest in my development. | 0.88 | −0.12 |
| I2: I have opportunities to learn new skills. | 0.63 | 0.31 |
| I3: I have enough time to attend training. | −0.48 | 0.94 |
| I4 (reversed): I feel I cannot develop myself. | 0.86 | 0.00 |
The table makes the problem clear: I3's relationship with the other items is negative, and removing it raises alpha from 0.58 to 0.94. Removing any of the other items, on the other hand, lowers alpha; in other words, they are the load-bearing items of the scale.
This does not mean I3 should be deleted right away. First, think about why. In this example, I3 may be measuring workload rather than development opportunity: the employees most open to development may also be the busiest, and so unable to make time for training. In that case, I3 is valuable information in its own right; it just belongs to the workload category, not the development index.
Checklist when alpha is low
When an index's alpha comes out lower than expected, check the following in this order:
- Is there a reverse-coded item, and was it reversed? This is the most common cause.
- Is the data correct? Missing answers may have been coded as zero, or a "don't know" option may have been counted as a point on the scale.
- Do the items really measure the same concept? Look at the "alpha if item deleted" table and reread the content of the item that doesn't fit.
- Are there too few items? In subdimensions of two or three items, alpha naturally stays low. In that case, the average inter-item correlation is more informative.
- Is the group too homogeneous? In a small team where everyone answers similarly, variance is low, so alpha can come out low as well. Calculate alpha on company-wide data whenever possible.
- Is there a translation problem? In global surveys, if alpha in one language is noticeably lower than in the others, one of the items may be understood differently in that language. We cover this in our post on language and culture pitfalls in global employee surveys.
What does a very high alpha mean?
An alpha above 0.95 looks like good news at first glance. More often than not, however, it shows that the items repeat each other: items such as "I am satisfied with my job", "I am content with my job" and "I like my job" get almost identical answers. This lengthens the survey without adding new information to the scale. Removing one of the items both shortens the survey and reduces respondent fatigue. We discuss how survey length affects participation in our post on how many questions an employee survey should have.
What alpha does not tell you
Alpha is a useful indicator, but it has limits:
- It does not show validity. Four items can be highly consistent with each other and still all measure satisfaction instead of engagement.
- It does not prove unidimensionality. A scale measuring two different subconcepts can also produce a high alpha. Factor analysis is needed to tell.
- It does not reveal response bias. If employees answer "agree" to every item without thinking, alpha comes out high. We discuss these biases in our post on survey response bias.
- It does not show how accurate a single team's result is. Alpha is a property of the scale; how confident you can be about one team's score depends on the number of responses. For that, see our post on sample size and margin of error.
Common mistakes
- Reporting a single alpha for the whole survey. Alpha is meaningful for items that measure the same concept. Calculating it across engagement, manager and workload items all at once produces a meaningless number. Calculate it separately for each index.
- Deleting items blindly to raise alpha. The "alpha if item deleted" table is a clue, not a decision. Before removing an item, consider its content and its importance to the organization.
- Changing validated scales without rechecking alpha. When you shorten or translate the items of an established scale such as the Utrecht Work Engagement Scale, recalculate alpha with your own data.
- Interpreting alpha in small groups. An alpha calculated from a few dozen responses is quite unstable. At least 100–200 responses are generally recommended for a reliable estimate.
In short, Cronbach's alpha shows in a single number how consistently the items of a scale are answered. Around 0.80 is a good target for organizational surveys; when you see a low or negative value, first check for reverse-coded items, then look at the "alpha if item deleted" table. A high alpha shows that your scale is consistent, not that it measures the right thing.
Frequently asked questions
What is a good Cronbach's alpha?
By common convention, 0.70 and above is acceptable, 0.80–0.90 is good and above 0.90 is very good. For indices in organizational surveys, the target is around 0.80. Values above 0.95 may indicate that items repeat each other. For subdimensions with only a few items, values between 0.65 and 0.70 can also be considered reasonable.
What does a negative Cronbach's alpha mean?
A negative alpha shows that some items are being answered in the opposite direction from the others. The most common cause is a negatively worded (reverse-coded) item whose score was not reversed before the calculation. Once you check the items and reverse the reverse-coded ones, alpha usually returns to the expected range.
Can Cronbach's alpha be calculated in Excel?
Yes. Calculate the variance of each item with the VAR.S function and the variance of each respondent's total score, also with VAR.S, then plug them into the formula α = (k / (k − 1)) × (1 − sum of item variances / variance of total scores). SPSS, R, Python and most survey platforms report the value directly.
How many respondents do you need for Cronbach's alpha?
There is no hard rule, but an alpha calculated from a few dozen responses is quite unstable. At least 100–200 responses are generally recommended for a reliable estimate. Rather than interpreting the alpha of very small teams, it is better to calculate it on company-wide data.
If Cronbach's alpha is high, is the survey valid?
No. Alpha only shows that the items are consistent with each other. Whether the items measure the right concept is a question of validity, assessed through methods such as content review, factor analysis or comparison with other measures.
