A Likert scale is a survey question that asks people to rate their agreement on a fixed, ordered set of responses — typically 5 or 7 points from "Strongly Disagree" to "Strongly Agree" — and you analyze it by treating the responses as ordered categories, using medians and non-parametric tests for single items, or means and parametric tests when you sum several items into a scale. If you've collected a batch of survey responses and you're staring at a column of numbers from 1 to 5 wondering whether you're allowed to calculate an average, you're asking exactly the right question — and it's the one that trips up most first-time survey researchers.
Key Takeaways
- A Likert scale measures attitudes using ordered response options (e.g., 1 = Strongly Disagree to 5 = Strongly Agree); the data are ordinal, not interval.
- A single Likert item should be summarized with the median and mode, not the mean, because the spacing between options is not guaranteed to be equal.
- A Likert scale (several items summed or averaged into one score) is widely treated as continuous, allowing means, t-tests, and ANOVA.
- For single-item comparisons, use non-parametric tests: Mann-Whitney U (two groups), Kruskal-Wallis (three or more groups), or Spearman's rho (correlation).
- Roughly 5–7 response points is the sweet spot — enough variation to detect differences without overwhelming respondents.
What exactly is a Likert scale?
A Likert scale is a psychometric response format where respondents indicate how strongly they agree or disagree with a statement. Named after psychologist Rensis Likert (1932), a classic 5-point version looks like this:
"I feel confident using statistical software."
1 = Strongly Disagree · 2 = Disagree · 3 = Neutral · 4 = Agree · 5 = Strongly Agree
The critical detail: the numbers are ranks, not measurements. You know that "Agree" is more positive than "Neutral," but you cannot assume the psychological distance from "Disagree" to "Neutral" is exactly the same as from "Agree" to "Strongly Agree." That's why the data are called ordinal — ordered, but with unknown spacing.
Likert item vs. Likert scale: the distinction that decides your analysis
This is the single most important distinction for choosing a test, and it's the one most tutorials skip:
- A Likert item is one single question (one row of Strongly Disagree → Strongly Agree). Treat it as ordinal.
- A Likert scale is several related items combined into one total or average score (e.g., 10 anxiety questions summed into an anxiety score). This composite behaves much more like continuous data because summing smooths out the uneven spacing.
If you built a validated multi-item scale, you can generally use means, t-tests, and ANOVA. If you're analyzing one standalone question, lean toward medians and non-parametric tests.
Can you calculate a mean for Likert data?
For a single Likert item, report the median and mode rather than the mean, because the mean assumes equal intervals that ordinal data don't guarantee. For a multi-item Likert scale (summed or averaged), calculating a mean is standard practice and accepted in most peer-reviewed psychology and social science journals.
In practice, many researchers still report means for single items alongside the median — that's fine as a descriptive summary, but your inferential test (the one that produces a p-value) should match the data type. When in doubt for a single item, a non-parametric test is the defensible choice.
Which statistical test should I use for Likert data?
Match the test to (a) whether you have an item or a scale, and (b) how many groups you're comparing.
| Your situation | Data type | Recommended test | APA statistic |
|---|---|---|---|
| Compare 1 item across 2 groups | Ordinal | Mann-Whitney U | U, p |
| Compare 1 item across 3+ groups | Ordinal | Kruskal-Wallis H | H, p |
| Compare a scale score across 2 groups | Continuous | Independent t-test | t, p, d |
| Compare a scale score across 3+ groups | Continuous | One-way ANOVA | F, p, η² |
| Relationship between 2 items | Ordinal | Spearman's rho | rs, p |
| Before/after on same people | Ordinal | Wilcoxon signed-rank | Z, p |
If you're deciding between the Mann-Whitney U and the t-test for a two-group comparison, that choice hinges on exactly this item-vs-scale question — pick the non-parametric route for single items.
A worked example with real numbers
Suppose you surveyed 60 graduate students on a single item: "I feel confident choosing the right statistical test" (1 = Strongly Disagree to 5 = Strongly Agree). You want to know whether stats-course completers (n = 30) differ from non-completers (n = 30).
Because this is a single Likert item (ordinal) with two independent groups, the Mann-Whitney U test is the correct choice.
Step 1 — Describe each group. Completers: median = 4 (Agree). Non-completers: median = 3 (Neutral).
Step 2 — Rank all 60 responses together, from lowest to highest, and sum the ranks within each group. The Mann-Whitney U compares those rank sums.
Step 3 — Run the test. Suppose the output is:
U = 285.00, p = .008, r = .34
Step 4 — Interpret.
- U = 285 is the test statistic derived from the rank comparison.
- p = .008 is below .05, so the difference is statistically significant — course completers reported higher confidence than non-completers.
- r = .34 is the effect size (small–medium), telling you the difference is meaningful, not just detectable.
APA 7 write-up:
A Mann-Whitney U test indicated that course completers (Mdn = 4) reported significantly higher confidence than non-completers (Mdn = 3), U = 285.00, p = .008, r = .34.
Note the APA conventions: U, p, and r are italicized, and the leading zero is dropped from p because p-values cannot exceed 1.
How do I check reliability for a multi-item scale?
Before analyzing a multi-item Likert scale, report Cronbach's alpha to show the items reliably measure the same construct; a value of .70 or higher is the conventional threshold for acceptable internal consistency. If alpha is below .70, your items may not belong together as a single score, and combining them could distort your results. Alpha above roughly .95 can signal redundant items. This step is expected in any thesis or journal submission that uses a composite scale, so don't skip it.
How StatRyx analyzes Likert data for you
The friction with Likert data isn't the arithmetic — it's the decision-making: item vs. scale, ordinal vs. continuous, parametric vs. non-parametric. Get that wrong and a reviewer will send your manuscript back.
StatRyx handles that judgment automatically. Upload your survey data, and StatRyx detects the response structure, recommends the correct test (Mann-Whitney, Kruskal-Wallis, or an ANOVA on a validated scale), computes Cronbach's alpha for multi-item scales, and returns a copy-paste APA 7 write-up with the effect size already reported. It's the AI-powered alternative to clicking through SPSS menus or writing R code by hand — free, and built for researchers who need the right answer, not a