Mann–Whitney U is a rank-based test used to compare two independent groups. In a nursing dissertation, the important task is deciding whether its assumptions and interpretation match the research question before running it in SPSS.
This guide focuses on that one test. For an overview of alternatives, use our statistical test selection guide. For arranging a complete chapter, see the separate SPSS results chapter guide.
What does Mann–Whitney U test?
The test compares the ranks of observations in two independent samples. Its interpretation depends on the assumptions made about the underlying distributions. It is often described as a nonparametric alternative to an independent-samples t-test, but that description does not mean the two tests answer exactly the same question.
Under a common-shape location-shift interpretation, a difference can be discussed in terms of location, including medians. If the distributions differ substantially in shape or spread, a significant result should not automatically be described as a difference in medians. Fay and Proschan explain why assumptions matter to interpretation.
Start by writing the question in plain language. Are you comparing two groups on an ordered outcome? Are the observations independent? What kind of difference matters to the study? A software menu cannot answer those design questions.
When might it suit a nursing dissertation?
A suitable scenario could involve comparing an ordered response between two separate groups of participants. The outcome must support meaningful ranking. The grouping variable should identify two independent groups, and the analysis must respect how participants were sampled or allocated.
For example, an illustrative study might compare a questionnaire score between two independent groups of nursing students. That example does not establish that the test is appropriate for every questionnaire. The scale, distribution, research aim and design still need examination.
A small sample is not an automatic instruction to use this test. Nor should a single normality-test p-value determine the analysis on its own. Explain the choice in relation to the estimand, outcome characteristics and study design, seeking statistical supervision where needed.
What should you check before opening SPSS?
- The analysis concerns two independent groups.
- The outcome can be ranked meaningfully.
- Records follow the sampling design.
- Missing values and group codes are correctly defined.
- The interpretation planned for the result matches distributional assumptions.
- The comparison was selected for the research question rather than a desirable p-value.
Record these decisions in an analysis plan. A short, explicit plan reduces the risk of trying several tests and reporting only the one that appears significant. If the plan changes after examining data, explain the change transparently.
Why does independence matter?
Independent groups contain different observational units. Repeated measurements from the same person are paired, not independent. Using Mann–Whitney U for a before-and-after comparison of the same participants ignores that relationship.
For paired rank-based analysis, a different procedure may be appropriate. The correct choice depends on the question and assumptions. Do not select a test solely because its name appears beside another test in a nonparametric menu.
Also consider clustering. Participants from the same ward, class or service may share influences. Two group labels do not remove within-cluster dependence. If the design contains clustering, discuss an analysis that respects it rather than treating every row as unrelated.
How should the dataset be organised?
For a straightforward independent-group comparison, use one row per observational unit, an outcome variable and a clearly coded grouping variable. Keep a data dictionary explaining the coding and missing-value rules.
| Field | Purpose | Check before analysis |
|---|---|---|
| Study identifier | Distinguish records without using names. | Check duplicate entries and the unit of analysis. |
| Group | Identify the two comparison groups. | Confirm codes and value labels. |
| Outcome score | Provide the ordered measurement. | Confirm direction, range and scoring rules. |
| Missing-value indicator, if needed | Document unavailable or inapplicable data. | Ensure missing codes are not analysed as scores. |
This is a dataset-planning example, not a requirement to collect additional personal information. Use the minimum data needed and follow the project’s approved handling arrangements.
How do you inspect the outcome?
Summarise the data separately for each group. Examine sample size, missingness, plausible values and the distribution. A suitable plot can reveal patterns that a single summary statistic hides.
Check how the score was constructed. If it combines questionnaire items, follow the instrument’s scoring instructions, including any reverse coding or missing-item rules. Do not calculate a total using an improvised rule merely to obtain a variable for testing.
Record potential data errors before making changes. An extreme value may be a valid observation, a coding mistake or a sign that the variable has been misunderstood. Investigate using the available records and preserve a clear account of corrections.
Why should you examine distribution shapes?
The same median can accompany different spreads or shapes. Conversely, differences in ranked observations can reflect more than a simple shift in location. The descriptive evidence helps determine which interpretation is defensible.
If you intend to describe the result as a median comparison, justify the relevant distributional assumptions. If those assumptions do not fit, use language aligned with the actual rank-based comparison and discuss the limitation. Avoid adding “median” automatically to a reporting template.
Describe what you examined rather than writing only “assumptions were met.” Name the relevant checks and explain any uncertainty. A dissertation should show the reasoning behind the conclusion, not just announce a decision.
How do you run the test in SPSS?
IBM documents Mann–Whitney within its independent-samples nonparametric procedures. Depending on the version and interface, you may encounter the newer independent-samples workflow or the legacy two-independent-samples dialogue. Use the procedure available in your installation and consult the matching IBM documentation.
Select the intended outcome and the grouping variable, verify the group codes and choose the Mann–Whitney comparison. Check whether the software is applying an automatic selection or a customised test choice. Save the syntax or a clear record of settings so the analysis can be reproduced.
Do not copy values from whichever output table looks familiar. Confirm the test, included cases and group definitions in the actual output. If the output does not match the analysis plan, resolve the discrepancy before writing the results.
What about exact and asymptotic p-values?
Software can use different approaches to calculating p-values, including asymptotic approximations and, where available and appropriate, exact or permutation procedures. Ties and small samples can affect the choice. Record the method used rather than treating every reported p-value as interchangeable.
Check the documentation for the procedure and software version. Availability can depend on the installation. Do not label a result “exact” simply because a sample is small, and do not import another package’s defaults into an SPSS report without checking them.
If two outputs disagree, compare the data, missing-value handling, alternative hypothesis, treatment of ties and calculation settings. A difference in p-values can reflect those choices rather than a calculation error.
How do you read the output?
Begin with the number of valid observations in each group. Confirm that these counts agree with your cleaned dataset and missing-value decisions. Then examine group summaries and the test result.
Mean ranks describe the ordering of observations; they are not means on the original measurement scale. Use appropriate descriptive statistics to explain the actual outcome. Do not describe a higher mean rank as a specified number of points higher on the questionnaire.
Identify the reported U statistic, the p-value and its calculation method. Where relevant, record the standardised statistic and a justified effect-size estimate. Ensure the direction of any effect is consistent with the group coding and the statistic’s definition.
What should a reporting paragraph contain?
Report the question, groups, valid sample sizes, relevant descriptive statistics, test result and an interpretation matched to the assumptions. Follow the university’s style requirements and keep the result separate from broader literature-based discussion.
A template might read: “Scores were compared between [group A] and [group B] using a two-sided Mann–Whitney U test. Valid sample sizes were [nA] and [nB]. The groups’ descriptive summaries were [values]. The test produced U = [value], p = [value], using [calculation method]. These findings were interpreted as [appropriate statement].”
Replace every placeholder with verified output. The template deliberately contains no invented p-value or study result. Add an effect-size estimate and uncertainty interval where appropriate and supported by the selected method.
How should you handle effect size?
Select an effect-size measure that fits the analysis and reporting guidance. Explain its definition and direction. Different measures associated with rank-based comparisons are not interchangeable, so avoid applying a formula from an unrelated example without checking it.
If reporting a measure calculated from a standardised statistic, verify the denominator and any conventions used. If reporting a probability-based or rank-biserial measure, explain how higher and lower values correspond to the groups. Retain the calculation record.
Interpret magnitude in context. A label such as “small” or “large” should not replace discussion of the outcome, scale and practical meaning. A statistically significant result does not establish clinical importance.
What does a non-significant result mean?
A non-significant result does not prove the groups are identical. It indicates that the selected test did not provide sufficient evidence against the specified null under the analysis conditions. The sample, variability and design affect what differences the study can detect.
Report the descriptive evidence and uncertainty rather than describing the study as having “failed.” Do not change to a one-sided test after seeing the result to obtain a smaller p-value. Any directional hypothesis should be justified in advance.
Likewise, avoid repeating analyses across many outcomes without considering multiplicity. If the dissertation includes several comparisons, explain the analysis strategy and distinguish planned tests from exploratory work.
Can the test show that an intervention caused improvement?
The test does not establish causation by itself. A difference between independently observed groups may reflect selection, confounding or other design features. The causal interpretation depends on the study design, not merely on a p-value.
For observational data, use careful wording about differences or associations. If the project is an intervention study, explain allocation, baseline comparability and other design considerations. Do not claim that a rank-based test compensates for weaknesses in the study design.
Where the research question requires adjustment for important covariates, a simple two-group test may be insufficient. Discuss a suitable modelling approach with the supervisor or statistician rather than adding unsupported adjusted conclusions.
How can you check the analysis before submission?
Trace the final result back to the dataset and saved procedure. Confirm the valid sample sizes, group codes and outcome direction. Check that the table, paragraph and appendix all refer to the same analysis.
Then review language. Replace “proved,” “no difference exists” and “automatically compares medians” with statements supported by the design and assumptions. Check that descriptive statistics use the original scale and that mean ranks are labelled correctly.
Finally, compare the methods and results chapters. If the method describes a two-sided test with a particular missing-value rule, the result should reflect those decisions. Explain deviations rather than leaving contradictory accounts.
Worked planning example: an ordered questionnaire response
Imagine a student project comparing an ordered confidence response between two separate teaching groups. The purpose is to examine whether the response distributions differ. No intervention effect is assumed, and the example contains no participant data or calculated test result.
The first decision concerns the unit of analysis. If each student contributes one response and belongs to only one group, the individual records may fit a simple independent-group structure. If students contribute several responses across time, the analysis needs to account for repeated measurement. Where teaching is delivered in several classes, the student must also consider whether class membership creates dependence.
The second decision concerns the outcome. The student checks the response labels, their order and the direction of higher values. A code of zero used for “not answered” must not be treated as the lowest confidence category. The data dictionary should distinguish missing responses from genuine ordered responses before any ranking occurs.
From data checks to a defensible report
The third decision concerns interpretation. For this reason, the student examines the distribution in each group and writes the planned claim before looking for a significant p-value. If the groups differ in shape, the final explanation should not automatically describe a simple median difference. The plot and descriptive table help the reader understand what the rank comparison is summarising.
The fourth decision concerns reporting. A useful results table needs the actual number of analysed responses in each group and appropriate summaries. Next, the paragraph reports the verified U statistic, p-value and calculation method. A separate discussion considers the study’s sampling limits and possible explanations, including any differences in prior experience between groups.
This example shows why the analysis starts with a question and a data structure. Selecting “nonparametric” because the sample is modest does not resolve independence, scoring or interpretation. Those decisions remain the researcher’s responsibility even when SPSS produces a clean output table.
Questions to bring to a statistics supervision meeting
Bring the research question, data dictionary, anonymised variable list and a brief description of sampling. Ask whether the proposed comparison matches the intended claim, whether clustering or repeated observations matter, and which effect-size estimate is appropriate. Record the advice and revise the analysis plan before generating the final tables.
If the analysis has already been run, bring the saved syntax and output as well. A supervisor can assess a reproducible procedure more effectively than a paragraph containing only a p-value. Keep a record of corrections so the methods chapter describes what was actually done.
Frequently asked questions
Is Mann–Whitney U suitable for before-and-after data?
Not when the same participants provide both measurements and pairing is ignored. Select a method appropriate to paired data and the research question.
Does it always test medians?
No. A median or location interpretation needs additional assumptions. Examine the distributions and justify the wording used.
Can I use it for a single Likert item?
An ordered item may support a rank-based comparison, but suitability still depends on the design, question, ties and intended interpretation. Do not infer suitability from the response format alone.
Should I report SPSS output screenshots?
Follow the brief. Usually a clear results table and paragraph are more readable than raw screenshots. Retain the full output for checking and use appendices only as required.
Get feedback on your SPSS analysis
Our nursing SPSS data analysis support can help check test choice and reporting. Share anonymised data, the research question and your brief through the contact page. Get My Nursing Assignment Quote.
Conclusion
Use Mann–Whitney U when the independent-group design, ranked outcome and planned interpretation fit the question. Verify the data and settings, report more than the p-value and keep causal or clinical claims within the evidence.
