Guides / likert scale
Likert scales: design, analysis and reporting
Likert data sit at the centre of the longest-running argument in applied statistics: are they ordinal or interval? This guide gives you a defensible answer, along with how to design the items, which tests to run, how to check reliability, how to display the results honestly, and what to write in your methods section.
Written and reviewed by Hafiz Ahmad Tariq, Senior BiostatisticianA Likert scale is a survey instrument in which respondents rate their agreement with a series of statements on an ordered response set, typically five or seven points from strongly disagree to strongly agree. Strictly, a single question is a Likert item and produces ordinal data; a Likert scale is the total or mean of several related items, which is often treated as interval data if the items are shown to be reliable.
Definition
Item or scale? The distinction that decides everything
A Likert item is one question. A Likert scale is several items combined into a total. Almost every argument about how to analyse Likert data dissolves once you are clear about which of the two you have.
This matters because the two carry different amounts of information. On a single item, the distance between ‘agree’ and ‘strongly agree’ is unknown and probably not the same as the distance between ‘neutral’ and ‘agree’. Summing eight items produces a score from 8 to 40 with many more possible values, and the summation tends to smooth out the unequal spacing of any individual item.
Analysing a single item? Treat it as ordinal — median, mode, frequencies, non-parametric tests. Analysing a multi-item scale with acceptable reliability? Treating it as interval — mean, SD, parametric tests — is standard practice and defensible.
Design
Designing the items
Most problems with Likert data are created at the design stage and cannot be fixed in analysis. These are the ones worth getting right.
Reverse-worded items
Mixing in negatively worded items is often recommended to disrupt acquiescence bias — the tendency to agree with everything. It works, but it has a cost: reverse items reliably load onto their own factor in analysis, and some respondents miss the reversal entirely.
If you use them, you must reverse-score before computing any total. On a 5-point scale, replace each response x with 6 − x. Forgetting this step is one of the most common serious errors in student survey work, and it quietly destroys your reliability coefficient — an alpha that comes back near zero or negative is almost always an un-reversed item.
Run a correlation matrix of your items before summing anything. If one item correlates negatively with all the others, it is reverse-worded and has not been reverse-scored.
Design
How many response points, and should there be a midpoint?
| Points | Best for | Trade-off |
|---|---|---|
| 4 | Forcing a direction | No neutral option; frustrates genuinely undecided respondents |
| 5 | General use, the default | Simple and familiar; slightly less discriminating |
| 7 | Scales needing finer discrimination | Better reliability; harder on mobile screens |
| 10 or 11 | Single-item ratings such as NPS | Closer to interval; risks false precision |
Reliability improves as you add points, but the gains flatten quickly — the improvement from 5 to 7 is modest, and beyond about 7 it is negligible while the cognitive burden keeps rising. Five and seven are the defensible defaults.
The midpoint debate
Include a midpoint
- Respondents genuinely may have no view
- Forcing a choice manufactures data that does not exist
- Removing it pushes people to abandon the survey
- Standard in most validated instruments
Omit the midpoint
- Some use the midpoint to avoid thinking
- It can absorb ‘prefer not to say’ responses
- Forces a directional answer for cleaner comparisons
- Common in customer satisfaction work
Keep the neutral option and add a separate ‘don't know / not applicable’ option outside the scale. That separates genuine neutrality from genuine ignorance, which the midpoint alone conflates — and lets you exclude the latter from analysis rather than treating it as a middling opinion.
Send your draft items. A named statistician reviews the wording, response options and scale structure before you collect data — which is the only point at which most of these problems are still fixable.
Get a fixed quoteMeasurement
The ordinal versus interval argument, settled
The purist position is that Likert responses are ordinal, so means are meaningless and only non-parametric methods are valid. The pragmatic position is that multi-item scales behave well enough that parametric methods are fine. Both are partly right, and the resolution is not a compromise but a distinction.
| What you have | Treat as | Central tendency | Typical tests |
|---|---|---|---|
| Single item | Ordinal | Median, mode | Mann-Whitney, Kruskal-Wallis, Spearman, chi-square |
| Multi-item scale, α ≥ .70 | Interval | Mean, SD | t-test, ANOVA, Pearson, regression |
| Multi-item scale, poor reliability | Ordinal | Median | Non-parametric, and fix the scale |
| Single item, large n, symmetric | Either, with justification | Report both | Either, stated in advance |
The empirical evidence supports this. Simulation studies have repeatedly found that t-tests and ANOVA applied to summed Likert scales produce accurate error rates across a wide range of realistic conditions, provided the distribution is not severely skewed and group sizes are not badly unequal. The parametric tests are robust here; they are much less robust for a single five-point item, particularly with small samples.
“The eight items formed a reliable composite (α = .86) and were summed to produce a total score, which was treated as continuous for analysis. Individual items were analysed non-parametrically.” That sentence pre-empts the question rather than leaving an examiner to raise it.
Analysis
Which test to use
| Research question | Single item | Multi-item scale |
|---|---|---|
| Compare two independent groups | Mann-Whitney U | Independent t-test |
| Compare two paired measurements | Wilcoxon signed-rank | Paired t-test |
| Compare three or more groups | Kruskal-Wallis | One-way ANOVA |
| Association between two measures | Spearman's rho | Pearson's r |
| Predict an outcome | Ordinal regression | Linear regression |
| Compare response distributions | Chi-square | — |
Note the pattern: the left column is the rank-based equivalent of the right. The non-parametric tests answer slightly different questions — Mann-Whitney compares distributions rather than means — which is why the interpretation wording differs, and why you should state which you used and why.
A worked decision
You have a 10-item job satisfaction scale with α = .89, and you want to compare satisfaction across three departments. Sum the ten items to give a score from 10 to 50, confirm the distribution is not severely skewed, and run a one-way ANOVA with post-hoc comparisons. If you also want to report on the single item “I would recommend this employer”, analyse that one with Kruskal-Wallis and report medians.
Our decision tool walks through the questions that determine the test — measurement level, number of groups, and whether the data are paired.
Use the test chooserReliability
Reliability: Cronbach's alpha and what it does not tell you
Before summing items into a scale you must show they measure the same underlying construct. Cronbach's alpha is the usual evidence.
| α | Conventional label | What to do |
|---|---|---|
| ≥ .90 | Excellent | Check for redundant items — very high alpha can mean you asked the same question repeatedly |
| .80–.89 | Good | Proceed |
| .70–.79 | Acceptable | Proceed, note it as a limitation |
| .60–.69 | Questionable | Examine item-total correlations |
| < .60 | Poor | Do not sum — the items are not measuring one thing |
Alpha is calculated from the average inter-item correlation and the number of items, and it estimates how consistently the items rank respondents in the same order. It is not a property of the instrument in the abstract: the same questionnaire can return a high alpha in one sample and a low one in another, because it depends on how much genuine variation exists in the group you measured. A homogeneous sample — all final-year students at one institution, say — restricts that variation and will depress alpha even for a well-constructed scale. Report it for your own sample rather than citing the figure from the original validation paper.
Three things alpha does not tell you
This is nearly always an un-reverse-scored item, not a genuine finding. Check your item correlations: a negative alpha means some items correlate negatively with the total, which for a properly scored scale should not happen.
Report alpha for each scale you construct, in the methods or results section, along with the number of items. Where a subscale structure is claimed, report alpha for each subscale rather than for the instrument as a whole. McDonald's omega is increasingly preferred where your software offers it, because it does not assume all items contribute equally.
Data quality
Missing data and response bias
Likert surveys attract several distortions that are invisible in a table of means but change what the numbers mean. All of them are worth checking before analysis.
| Pattern | What it looks like | What to do |
|---|---|---|
| Acquiescence | Agreeing regardless of content | Include reverse-worded items; check for flat response strings |
| Extreme responding | Only 1s and 5s | Note it; consider analysing the distribution rather than the mean |
| Central tendency bias | Everything at the midpoint | Check completion time; may indicate disengagement |
| Straight-lining | Identical answer down the whole page | Flag and consider excluding, with the rule stated in advance |
| Item non-response | Some items skipped | Report n per item; do not silently drop cases |
Handling missing items within a scale
If a respondent answers seven of your eight items, you have three options and they are not equivalent.
Whichever you choose, state the rule in your methods section and report how many cases it affected. A silent decision here is the kind of thing that unravels under viva questioning.
Responses completed far faster than the median are worth inspecting. A 40-item survey finished in 60 seconds has not been read. Set an exclusion rule before you look at the results, so the decision is not driven by whether excluding them helps your hypothesis.
Presentation
Displaying Likert data honestly
A bar chart of means is the most common and least informative way to present Likert results. Two items can share a mean of 3.0 when one had everybody answering neutrally and the other split the room entirely — and those are opposite findings.
| Display | Shows | Use when |
|---|---|---|
| Diverging stacked bars | Full distribution and the balance of opinion | Default choice for item-level results |
| Stacked bars (100%) | Proportions per category | Comparing many items compactly |
| Frequency table | Exact counts and percentages | Appendices, and any formal report |
| Bar chart of means | Central tendency only | Multi-item scales, not single items |
Percentages without the underlying counts are unreadable. “67% agreed” means something very different from 6 of 9 than from 402 of 600. Give both, and state how many people skipped the item.
Pitfalls
Seven mistakes that cost marks
1. Reporting a mean for a single item without justification
A mean of 3.4 on a five-point item implies the gaps between response options are equal. If you report it, say why that is reasonable — or report the median instead.
2. Forgetting to reverse-score
It corrupts every total and every reliability estimate that follows. Check the item correlation matrix before summing.
3. Summing items that do not belong together
Adding items from different constructs produces a number with no meaning. Establish reliability, and ideally factor structure, before summing.
4. Treating ‘not applicable’ as a midpoint
If your scale conflates “no opinion” with “moderate opinion”, your midpoint is uninterpretable. Separate them at the design stage.
5. Running a t-test on a single five-point item with a small sample
This is exactly the case where the ordinal objection bites hardest. Use Mann-Whitney.
6. Reporting alpha and stopping there
Alpha does not establish that the scale is unidimensional or that it measures what you claim. Where the structure matters, run a factor analysis.
7. Charting means without the distribution
It hides polarisation, which is frequently the most interesting thing in the data.
Reporting
Reporting Likert data in APA style
| Element | How to write it |
|---|---|
| The instrument | Responses were recorded on a 5-point scale (1 = strongly disagree, 5 = strongly agree). |
| Reliability | The eight items showed good internal consistency (α = .86). |
| A scale score | M = 31.4, SD = 5.9 (possible range 8–40) |
| A single item | Mdn = 4, IQR = 3–5 |
| A non-parametric test | U = 412.5, z = −2.31, p = .021, r = .28 |
| A parametric test | t(78) = 2.94, p = .004, d = 0.66 |
Send the dataset and your draft results. A named statistician verifies the reliability analysis, confirms the tests suit the measurement level, and returns annotated output with APA tables.
See SPSS data analysisAnswers
Frequently asked questions
Is a Likert scale ordinal or interval?
A single Likert item is ordinal — the response options are ordered but the distances between them are unknown. A Likert scale made by summing several related items is commonly treated as interval, provided the items are shown to be internally consistent, typically with Cronbach's alpha of .70 or above. State which you are doing and why in your methods section.
Can I calculate a mean for Likert data?
For a multi-item scale with acceptable reliability, yes — this is standard practice and simulation studies support it. For a single item, the median and the full frequency distribution are more defensible, because a mean assumes the gap between every pair of adjacent options is identical.
How many points should a Likert scale have?
Five or seven are the usual defensible choices. Reliability improves as you add points but the gains flatten beyond about seven, while the burden on respondents keeps rising. Four points forces a direction by removing the neutral option, which suits some purposes and frustrates genuinely undecided respondents.
Should I include a neutral midpoint?
Usually yes, because some respondents genuinely have no view and forcing a choice manufactures data. The stronger design keeps the neutral point and adds a separate 'don't know / not applicable' option outside the scale, so genuine neutrality is not confused with genuine ignorance.
What test should I use for Likert data?
For a single item, use non-parametric tests: Mann-Whitney U for two groups, Kruskal-Wallis for three or more, Spearman's rho for association. For a reliable multi-item scale, the parametric equivalents — t-test, ANOVA and Pearson's r — are appropriate and generally more powerful.
What is a good Cronbach's alpha?
Above .70 is conventionally acceptable and above .80 is good. Below .60 you should not sum the items into a single score. Be cautious about alpha above .95, which often means items are near-duplicates. Alpha rises simply with the number of items, so a long scale can score well on mediocre items.
Why is my Cronbach's alpha negative?
Almost always because a reverse-worded item has not been reverse-scored. On a 5-point scale, replace each response x with 6 − x before computing totals. Check the item correlation matrix: an item correlating negatively with all the others is the culprit.
How should I chart Likert responses?
A diverging stacked bar chart centred on the neutral midpoint is the clearest option, because it shows the full distribution and the balance of opinion. A bar chart of means hides polarisation entirely — two items can share a mean while one is unanimous and the other splits the sample.
Keep reading
Related guides and services
Send the data. Get a fixed quote.
Attach your dataset or just describe the project. A named statistician replies with a price and a deadline, usually within one working day.