Nonparametric Tests in SPSS: 12 Essential Methods, Menus, Syntax and Interpretation

What nonparametric means

A nonparametric test generally avoids specifying a complete population distribution such as normality with a particular mean and variance structure. Instead, the procedure may convert observations to ranks, classify values around a median, count positive and negative paired differences or evaluate contingency-table frequencies. This can make the method more compatible with ordinal scales and heavily skewed outcomes.

The term does not mean the analysis has no assumptions. Nonparametric tests in SPSS still depend on independent observations or correct pairing, meaningful ordering, appropriate group definitions, transparent treatment of ties and a sampling design that supports the intended inference.

What nonparametric does not mean

It does not mean that every procedure tests medians. Mann–Whitney U and Kruskal–Wallis are pooled-rank tests and can respond to differences in location, spread or shape. Calling their results median differences requires additional conditions, especially similarly shaped group distributions. The median test is more directly connected to a common pooled median but uses less information because it reduces observations to two categories.

It also does not mean that nonparametric tests in SPSS should automatically replace parametric methods whenever a normality test rejects. Method selection should follow the design, measurement scale and estimand before it follows any diagnostic p-value.

Why SPSS is practical for this work

SPSS offers a modern Nonparametric Tests interface with objective-based choices, field roles, settings and automatically generated output. It also retains legacy dialogs that many textbooks and courses still use. Both routes can paste syntax, allowing the analyst to move beyond a one-time menu click and create a reproducible analysis record.

The software also places descriptive information, hypothesis-test summaries and detailed tables in the Output Viewer. That makes nonparametric tests in SPSS approachable for beginners while still allowing experienced analysts to verify group counts, ranks, expected frequencies and significance values.

Typical applications

Common applications include comparing satisfaction ratings between two departments, analyzing test scores across several independent classes, comparing pretest and posttest scores, testing repeated condition ratings, evaluating monotonic associations and analyzing small contingency tables. The unifying feature is not “bad data”; it is a design or question better represented by ranks, signs, medians or exact frequencies.

A strong SPSS workflow states the hypothesis first, chooses the procedure second, runs descriptives and charts third, and interprets the output only after checking that the grouping and missing-value rules are correct.

Why the design-first rule works

Suppose the same students provide G1 and G2 grades. Treating those values as two independent samples would ignore the matched structure. Conversely, comparing two different schools with a paired procedure would invent a relationship that does not exist. The correct nonparametric tests in SPSS begin with the sampling unit, not the appearance of the distribution.

The same principle applies to the estimand. Kruskal–Wallis and the median test can both be run on the same several-group outcome, but they do not use the same information. The worked example intentionally presents both so the reader can see the distinction.

Why generic rules fail

The rule “use nonparametric when normality fails” is too vague to be safe. It does not specify whether the data are paired, whether the outcome is ordinal, whether several groups are ordered, or whether the goal is to compare medians rather than general rank distributions. It also ignores the possibility that a robust parametric analysis remains appropriate.

A better rule is: write the null hypothesis in words, identify how observations are connected, then select the SPSS procedure whose statistic represents that hypothesis.

Independent or correctly related cases

Independent procedures require independent units. Related-samples procedures require valid pairings or repeated measurements from the same block.

Ordered outcome for rank methods

Mann–Whitney, Kruskal–Wallis, Wilcoxon and Friedman need an outcome that can be meaningfully ordered.

Representative sampling or assignment

Population inference still depends on the sampling or assignment design, even when the test itself does not assume normality.

Comparable shapes for median language

A rank-test result can reflect spread or shape. Interpreting it as a pure median difference requires comparable group shapes.

Symmetric paired differences

The signed-rank test has a cleaner location interpretation when paired differences are reasonably symmetric.

Transparent ties and exactness

Ties are common in grades and ratings. They affect exact calculations and should be acknowledged in the report.

How to check the assumptions in SPSS

Begin with Data View and Variable View. Confirm that each row represents the intended unit, that group values are labeled correctly, and that the outcome is numeric or ordinal as required. Then use Analyze → Descriptive Statistics → Explore to obtain group counts, medians, percentiles, boxplots and histograms. The goal is not to force the data to pass a checklist; it is to understand whether the planned interpretation is defensible.

For paired procedures, compare the number of valid pairs with the expected sample size and inspect the distribution of differences. For several independent groups, inspect whether one group is extremely small or has a radically different shape. In the worked example, the “other” guardian group has only 41 observations, much fewer than the mother group’s 455, so group summaries are essential context.

Assumptions that are often misstated

It is inaccurate to say that nonparametric tests in SPSS have no assumptions. They make fewer assumptions about a parametric distribution, but they may make stronger demands on independence, exchangeability, symmetry or comparable shapes depending on the method. A significance value cannot repair a miscoded grouping variable or a violated pairing structure.

It is also inaccurate to say that Kruskal–Wallis always tests medians. The test evaluates rank distributions. The median test is the procedure in this article that explicitly classifies observations around the pooled median.

Variable View checklist

A grouping field may be string or numeric in the modern procedure. Legacy dialogs are often simpler with a numeric grouping code, so this guide also provides a transparent recode syntax.

Why pasted syntax matters

Menu selections are useful for learning, but a pasted command is the permanent analysis record. Syntax captures the test field, grouping field, missing-value rule and requested tests in a form that can be rerun after the data are updated. It also reduces the risk that two analysts click slightly different settings while believing they ran the same procedure.

For production-quality nonparametric tests in SPSS, save the syntax file together with the data and output. Use brief comments to identify the analysis purpose and the decision rule.

Modern versus legacy procedures

The modern Analyze → Nonparametric Tests interfaces organize analysis by one-sample, independent-samples and related-samples objectives. They can choose tests automatically or allow a custom selection. They also create a Hypothesis Test Summary and detailed model-view output.

The legacy dialogs remain useful because many textbooks show them and because the output is compact and familiar. Both approaches are SPSS workflows. The important requirement is to document which procedure and settings produced the reported result.

Missing data scope

Listwise or analysis-specific exclusion can change the valid sample size. Always compare the number of valid cases in output with the expected source count. The verified workbook contains 649 complete G3 and guardian records. If the SPSS output shows a different number, stop and investigate before interpreting the test.

When several outcomes are tested together, the missing-value scope becomes especially important because one setting can exclude cases based on all selected fields while another can analyze each field with its available cases.

Research questions

Rank-distribution question: Are the G3 distributions identical across the three guardian groups?

Median-classification question: Is the proportion of students above the pooled median the same across groups?

Variables and coding

What the descriptives suggest

The mother and father groups both have median 12, while the other group has median 11. The father group has the highest mean, and the other group has the lowest mean and the narrowest IQR. These differences provide a descriptive reason to perform the omnibus tests, but they do not replace the inferential analysis.

The unequal sample sizes also matter for visual interpretation. The mother group contains roughly eleven times as many cases as the other group, so the precision of group summaries is not equal.

Why two SPSS methods are reported

Kruskal–Wallis uses the full ordering of all 649 observations and is generally more informative about rank distribution differences. The median test discards within-category ordering and asks whether group membership is associated with being above versus at/below the pooled median. Agreement between the two methods strengthens the conclusion that the group pattern is not limited to one computational representation.

Disagreement would not automatically mean that one test is wrong; it could mean the methods are responding to different aspects of the data.

Why the median table is informative

The mother and father groups have nearly identical above-median rates, approximately 43.7% and 44.4%. The other group’s rate is only about 22.0%. That contrast explains the significant median-test chi-square. It also gives a direct, understandable interpretation that complements the pooled-rank result.

However, the median table should not be used to claim that mother and father are identical or that the other group differs significantly from both in separate adjusted tests. The table is descriptive within an omnibus procedure.

Why the Kruskal result adds value

Kruskal–Wallis uses every rank, not only whether a grade is above 12. It therefore captures ordering differences throughout the distributions. The significant H = 8.0177 indicates that the overall rank pattern is unlikely under a common distribution, even after accounting for ties.

When both procedures reject, the analyst can say that the guardian groups differ under two complementary nonparametric summaries, while still avoiding unsupported pairwise claims.

Mann–Whitney U

Use Mann–Whitney U for two independent groups with an ordinal or continuous outcome. SPSS ranks all observations together and evaluates whether the rank distributions differ. Report the U statistic, standardized test statistic if provided, p-value, group medians and an effect size.

Kruskal–Wallis H

Use Kruskal–Wallis for three or more independent groups. The result is omnibus: a significant H indicates that at least one distribution differs. Multiple comparisons or planned contrasts are needed to localize the pattern.

Median test

Use the median test when the explicit question concerns classification around a common median. It is robust and easy to explain but generally less efficient because it discards ordering information within the two median categories.

Two independent groups workflow

Choose Analyze → Nonparametric Tests → Independent Samples. On Fields, assign the continuous or ordinal outcome as Test Fields and the two-level group variable as Groups. On Settings, select Customize tests and request Mann–Whitney U. Use the Run button to create the output, then open the detailed view for ranks and significance.

When the group field has more than two levels, Mann–Whitney is not the correct omnibus test. Either filter to a planned two-group comparison or use a several-group method.

Several independent groups workflow

With three or more groups, select Kruskal–Wallis. If the research question specifically concerns medians, select the median test as a separate procedure rather than pretending that the rank test is a direct median test. If the groups have a meaningful order, an ordered-alternative test may be more efficient than a general omnibus procedure.

Always compare the hypothesis-test summary with the descriptive group table. A significant omnibus p-value without group medians, IQRs and sample sizes is incomplete reporting.

Wilcoxon signed-rank

Wilcoxon signed-rank uses the ranks and signs of nonzero paired differences. It is appropriate when the paired outcome is ordinal or continuous and the difference distribution supports the intended location interpretation.

Sign test

The sign test counts positive and negative differences while ignoring magnitude. It is less powerful but requires fewer shape assumptions and is useful when paired differences are severely asymmetric or dominated by outliers.

Friedman test

Friedman compares three or more related conditions by ranking values within each unit. It is the repeated-measures counterpart to a several-group rank comparison.

Modern menu route

Choose Analyze → Nonparametric Tests → Related Samples. Place the repeated variables in Test Fields. Select an objective or customize the test selection. The interface can request Wilcoxon, sign, McNemar or Friedman-type procedures depending on the number and measurement level of the fields.

Before running the analysis, verify that the paired variables are in the same row for each unit. A row-shift or unmatched merge can invalidate the result while still producing apparently reasonable output.

What to report

For Wilcoxon, report the number of negative, positive and tied differences, the test statistic, p-value and a paired effect size. For the sign test, report positive and negative counts. For Friedman, report the chi-square statistic, degrees of freedom, p-value and a repeated-measures effect such as Kendall’s W when available or calculated.

Do not describe a paired result as an independent-group comparison. The repeated structure is central to both the test and the interpretation.

Kruskal–Wallis as the primary omnibus test

Kruskal–Wallis is often the most informative general several-independent-samples procedure because it uses all pooled ranks. When group shapes are broadly comparable, a significant result is often discussed as a location difference. When shapes differ, the result should be described more generally as a distributional rank difference.

In the guardian example, H(2) = 8.0177 and p = .018. That result is significant but does not itself specify that the “other” group differs from both mother and father. The descriptive pattern suggests that possibility, but formal pairwise inference requires an adjusted comparison strategy.

Median test as a complementary view

The median test produces a simple table showing how many cases in each group lie above and at/below the pooled median. It is easy to explain to nontechnical readers and is less sensitive to extreme magnitudes. Its cost is reduced power because grades of 13 and 19 are treated identically once both are classified above 12.

Here, the “other” group has an above-median rate around 22%, compared with approximately 44% in the other two groups. That contrast produces χ²(2) = 7.6042, p = .022.

Ordered groups

If groups have a prespecified natural order, an ordered-alternative procedure may use that information more efficiently than Kruskal–Wallis. The ordering must be substantive and decided before inspecting the outcomes. Sorting group labels alphabetically does not create a scientific order.

Post-hoc comparisons

When the omnibus result is significant, request or calculate pairwise comparisons with multiplicity control. The adjustment rule should be chosen before interpreting results. Report adjusted p-values, not a selective set of raw pairwise values.

Menu route

Choose Analyze → Nonparametric Tests → One Sample. Assign the field, then select an automatic objective or customize the test. The objective-based interface is convenient, but the analyst should inspect the Settings tab to confirm what hypothesis is being tested.

Reference values

A one-sample test is only as meaningful as its reference. A median of 10, a proportion of .50 or a specified category distribution should come from a theory, policy, historical benchmark or preregistered target—not from searching for a value that makes the result significant.

Chi-square through Crosstabs

For two categorical variables, use Analyze → Descriptive Statistics → Crosstabs. Request chi-square statistics and display observed, expected, row and column percentages. If expected counts are small, consider an exact procedure where appropriate.

Report the table, chi-square statistic, degrees of freedom, p-value and an effect size such as Phi or Cramér’s V. The phrase “chi square nonparametric test in SPSS” is common in search behavior, but the analysis should still be described accurately as a categorical association test.

Paired binary outcomes

When the same units are measured twice on a binary outcome, McNemar’s test evaluates discordant pairs. It is not interchangeable with an independent-samples chi-square test because the observations are paired.

The same design-first logic applies throughout nonparametric tests in SPSS: dependence and measurement scale determine the procedure before distribution shape does.

Modern independent-samples route

1. Choose Analyze → Nonparametric Tests → Independent Samples.
2. On the Objective tab, choose Customize analysis when you need specific procedures.
3. On Fields, assign G3 as the test field and guardian as the group field.
4. On Settings, request Kruskal–Wallis and the median test.
5. Set the significance level and missing-data options.
6. Use Paste to save syntax, then Run.

Open the detailed output rather than relying only on the hypothesis-test summary. Verify valid cases, group values and significance values.

Legacy K independent samples route

1. Create a numeric grouping code if the grouping field is string.
2. Choose Analyze → Nonparametric Tests → Legacy Dialogs → K Independent Samples.
3. Move G3 to Test Variable List.
4. Move the numeric guardian code to Grouping Variable and define range 1 through 3.
5. Select Kruskal–Wallis and Median.
6. Paste or run the command.

Legacy output usually presents ranks and test statistics in compact tables, which can be convenient for teaching and manual verification.

Modern related-samples route

Choose Analyze → Nonparametric Tests → Related Samples. Add the paired variables, select the desired test and inspect the paired sample count. For three or more repeated variables, request Friedman. Use a post-hoc plan if the omnibus result is significant.

One-sample route

Choose Analyze → Nonparametric Tests → One Sample. Confirm whether the objective is a reference median, proportion, category distribution, distributional form or sequence property. Enter the reference before examining the result.

Why recoding may be needed

The modern NPTESTS procedure can work directly with categorical group fields, including labeled fields. Legacy commands often require a numeric grouping range. The explicit RECODE command maps father, mother and other to 1, 2 and 3 so the legacy K-W and MEDIAN subcommands can define the full group range.

Never assume that an automatic recode uses the order you intended. Explicit mapping preserves meaning and prevents group-code mistakes.

Syntax comments and versioning

Use an asterisk followed by a period for block comments. State the purpose of the test, the outcome, the grouping field and the decision threshold. Save a dated syntax version when changing the analysis. A small amount of documentation makes the command file much more useful than a sequence of unexplained commands.

When the output is updated, rerun the whole syntax file rather than manually replacing individual tables. That keeps charts, values and interpretation synchronized.

Exact and Monte Carlo options

Some installations provide additional exact-test capabilities. When exact or Monte Carlo inference is requested, report that method explicitly. Exact calculations may be computationally expensive for larger samples, while Monte Carlo values depend on the number of samples and randomization settings.

The verified guardian example uses large-sample chi-square reference probabilities with explicit tie-aware and contingency calculations.

Batch output management

For repeatable production, use OMS or OUTPUT EXPORT commands to route selected tables to files and save the full Output Viewer. Give output objects short titles because extremely long subtitles can be truncated or difficult to read. The public report should reproduce the exact values in the saved output.

Hypothesis Test Summary

The modern output summarizes the null hypothesis, test name, significance value and decision. This is useful for orientation but should not be the only reported table. Open the detailed view to inspect group ranks, pairwise comparisons or median classifications.

A decision of “Reject the null hypothesis” means the p-value is below the selected alpha. It does not measure effect size and does not prove that every group differs from every other group.

Ranks and test-statistics tables

For rank tests, inspect group n and mean rank. The mean-rank ordering helps describe which groups tend to have higher observations. The test-statistics table provides H, U, Z or chi-square, degrees of freedom where applicable and significance.

Do not infer practical importance from rank order alone. Combine ranks with group medians, IQRs and an effect-size estimate.

Median-test output

Confirm the pooled median and the above versus at/below counts. Check expected counts when recreating the contingency chi-square. In this example, values equal to 12 are included in the at-or-below category, matching the verified workbook rule.

The significant table tells you that the above-median proportion is not the same across all groups. The cell counts show that the other group contributes the clearest departure.

Footnotes and warnings

Footnotes may describe continuity correction, exact significance, ties, small expected frequencies or unavailable calculations. Warnings should never be deleted from the reasoning merely because they are inconvenient. If an exact method is unavailable, report the method that was actually used.

When SPSS output and a manually verified workbook disagree, pause the article production and reconcile group coding, missing cases, threshold rules and statistic conventions before publishing.

APA-style worked-example report

“Final grades differed significantly across guardian groups according to a Kruskal–Wallis test, H(2) = 8.02, p = .018. A complementary median test also indicated unequal above-median proportions across groups, χ²(2, N = 649) = 7.60, p = .022. Median G3 was 12 for the mother and father groups and 11 for the other guardian group.”

This wording reports both methods, gives the main descriptive pattern and avoids claiming untested pairwise significance.

Effect-size options

For two-group rank tests, standardized r, rank-biserial correlation or probability of superiority can be reported. For Kruskal–Wallis, epsilon squared or eta-squared-type rank effects are common. For contingency-table procedures, Cramér’s V can summarize association magnitude. For Friedman, Kendall’s W is commonly used.

State the formula or software source because effect-size conventions can differ. Do not label a small p-value as a “large effect.”

Minimum reporting checklist

Common reporting errors