Which Statistical Test to Use in SPSS? A Variable-Type Decision Framework

Which statistical test to use in SPSS comes down to one question most students never ask properly: what type of variable are you actually looking at? Not the test name you half remember from a methodology class. Not what your friend used in their thesis. The variable type.

If you have ever searched for SPSS test selection for dissertation work at midnight, two weeks before submission, you are exactly who this guide is for. I have spent 12 years helping dissertation students figure this out, and the pattern is almost always the same. Someone opens SPSS, stares at the Variable View tab, sees “Nominal,” “Ordinal,” and “Scale” sitting in a dropdown, and has no idea why that choice matters. Then they pick a test based on what sounded impressive in a journal article they read last week.

That is backwards. This guide is really my full answer to how to choose statistical test in SPSS, and to SPSS statistical test selection more broadly, built from the patterns I see in actual dissertation data, starting from the variable type and working forward to the test, not the other way round.

Why Which Statistical Test to Use in SPSS Starts With Variable Type, Not Statistics Theory

Here is the mistake I see in almost every dissertation draft I review. A student has ordinal data, say a 5-point satisfaction scale, and runs an independent samples t-test on it because the t-test is what everyone talks about.

The t-test assumes your outcome variable is continuous (scale, in SPSS language). A 1-to-5 rating is not continuous. The gap between “satisfied” and “very satisfied” is not guaranteed to be the same size as the gap between “neutral” and “satisfied.” Running a t-test here does not make your analysis wrong exactly, but it makes it hard to defend if your committee asks why.

This is why I always tell students: before you touch the Analyze menu, go back to Variable View. Nominal, Ordinal, Scale. That one column decides almost everything else.

Nominal, Ordinal, Scale: What SPSS’s Own Dropdown Is Telling You

IBM’s official documentation on measurement levels defines it cleanly. Nominal means your categories have no built-in order (gender, city, product brand). Ordinal means your categories do have an order, but the distance between them is not measured (education level, satisfaction rating, rank). Scale means the numbers are genuinely continuous and distances between them are meaningful (age, income, test score, reaction time in seconds).

That IBM definition is technically correct, but on its own it does not help you pick a test. It tells you what a box in SPSS is called. It does not tell you what to do next, and that is the gap this article fills. Once you understand nominal ordinal scale SPSS labels properly, picking the right test becomes mechanical rather than guesswork.

One thing worth flagging early: SPSS does not have a separate “Interval” option the way your statistics textbook does. Interval and Ratio data both get bundled into “Scale.” If your textbook talks about four levels of measurement and SPSS only shows you three, that is why. Nobody explains this clearly enough, and it causes real confusion when students are cross-checking SPSS against their methodology chapter.

3 Quick Questions to Confirm Your Variable Type

No two ways about it, this is the part students skip, so here it is as three plain questions:

1. Can you rank the categories at all? If no, it is nominal (gender, city, yes/no).
2. Can you rank them, but you cannot measure the gap between ranks? That is ordinal (satisfaction rating, education level).
3. Is the gap between any two values exactly measurable and consistent? That is scale (age, income, test score).

Run your variable through these three questions before you open the Analyze menu. It takes thirty seconds and saves you a redo later.

Which Test to Use for Nominal Data in SPSS

Nominal data is about category membership with no ranking. Gender, marital status, yes/no responses, which product a customer bought. There is no “higher” or “lower” category here.

Tests for nominal data:

• Chi-square test: checking if two categorical variables are related (most common nominal test by far in dissertations)
• Binomial test: one nominal variable with exactly two categories
• McNemar’s test: paired nominal data, same people measured twice (before/after a binary outcome)
• Cochran’s Q test: paired nominal data across three or more time points

The mistake I see most with nominal data: students run a Pearson correlation on category codes like 1 = Male, 2 = Female, and interpret the correlation coefficient as if it means something. It does not. The numbers are just labels, and a correlation coefficient computed on labels is meaningless.

Which Test to Use for Ordinal Data in SPSS

This is the section where most confusion lives, so let me be direct about it.

Is a Likert Scale Ordinal or Scale Data?

Honestly, this is where I disagree with how most SPSS resources handle the question. Even solid, free references like UCLA’s statistical consulting guide are excellent on test mechanics but stay silent on this exact debate. Paid platforms like Laerd Statistics walk you through the steps well too, but they will not tell you where they personally stand on it either.

My actual position, after reviewing hundreds of dissertation datasets, is simple: a single 5-point Likert item is ordinal. Full stop. You should not treat it as scale and run a t-test or ANOVA on it directly.

Where it gets genuinely debatable is when you sum or average multiple Likert items into a composite score, say a 10-item job satisfaction scale averaged into one number. Before you do that, check Cronbach’s Alpha to confirm the items actually hang together as one reliable scale, not just a random bunch of questions. Many researchers then treat that composite as scale data because the summed score behaves more continuously, and that is defensible. A single raw item, though, stays ordinal in my book, and I would push back on any supervisor who tells you otherwise without a good reason.

Tests for ordinal data:

• Mann-Whitney U test: comparing two independent groups on an ordinal outcome
• Kruskal-Wallis test: comparing three or more independent groups on an ordinal outcome
• Wilcoxon signed-rank test: comparing paired ordinal measurements (same people, before and after)
• Friedman test: paired ordinal measurements across three or more time points
• Spearman’s rank correlation: relationship between two ordinal variables, or between an ordinal and a scale variable (Pearson vs Spearman correlation: which one to use covers this decision in more depth). If your data has a lot of tied ranks, Kendall’s tau-b is often the more accurate alternative.
• Ordinal logistic regression (PLUM in SPSS): predicting an ordinal outcome from multiple predictors, barely covered anywhere online despite being the technically correct advanced option

So, what analysis can you do with ordinal data in SPSS beyond basic group comparisons? Ordinal regression is the honest answer, and it is worth learning if your dissertation has an ordinal dependent variable with several predictors, rather than forcing it into a linear regression model that assumes equal spacing between categories.

Which Test to Use for Scale Data in SPSS

Scale data is your genuinely continuous data: age, income, exam score, time taken, temperature. This is where classic parametric tests live.

Tests for scale data:

• One-sample t-test: comparing a single group’s mean against a known or hypothesised value
• Independent samples t-test: comparing two independent groups
• Paired samples t-test: comparing the same group measured twice (paired t-test vs repeated measures ANOVA for pre-post data explains when you need the more advanced option)
• One-way ANOVA: comparing three or more independent groups
• Repeated measures ANOVA: comparing the same group across three or more time points
• Pearson correlation: relationship between two scale variables
• Linear regression: predicting a scale outcome from one or more predictors

Here is the step most students skip entirely, and it is the single biggest mistake I correct in dissertation chapter 4 reviews.

3 Steps Before You Trust a Parametric Test on Scale Data

1. Check normality first. Run the Shapiro-Wilk test for samples under about 50, and the Kolmogorov-Smirnov test for larger samples, before assuming a t-test or ANOVA is appropriate (Shapiro-Wilk test normality guide walks through the exact SPSS steps).
2. If normality is violated, switch families. Use Mann-Whitney U instead of an independent t-test, or Kruskal-Wallis instead of one-way ANOVA.
3. If you are running regression, check the supporting assumptions too. Check for multicollinearity using the VIF statistic (VIF guide). If your data involves repeated or time-ordered observations, check for autocorrelation using the Durbin-Watson test as well.

Skipping step one is how I end up seeing a dissertation with a t-test run on wildly skewed data, and a committee member catching it in the viva.

One more thing students often miss, and trust me, I see this constantly: a significant one-way ANOVA only tells you that at least one group differs, not which ones. You still need a post-hoc test, and the correct one depends on whether your group variances are equal or not (Tukey HSD vs Games-Howell covers exactly this decision).

Mixed Variable Types: What Most Dissertations Actually Look Like

Textbook examples always give you one clean variable type at a time. Real dissertations rarely work that way.

Example 1: Nominal predictor, scale outcome. You are comparing salary (scale) across three departments (nominal, 3+ categories). This is a one-way ANOVA, or Kruskal-Wallis if normality fails within each group.

Example 2: Scale predictors, nominal outcome. You are predicting whether a customer churns (yes/no, nominal) from age and spending (scale). This is binary logistic regression, not linear regression, because your outcome is categorical.

Example 3: Ordinal predictor, ordinal outcome. You are checking if education level predicts job satisfaction rating. Spearman’s correlation works for a simple relationship, or ordinal regression if you have multiple predictors to control for.

My Quick-Reference Decision Table

Here is my table connecting SPSS variable types and statistical tests side by side, the way I actually sketch it out for students in a consulting session.

Variable Type (SPSS label)	What It Looks Like	Tests You Can Run	Mistake I See Most
Nominal	Gender, product brand, yes/no	Chi-square, binomial, McNemar’s, Cochran’s Q	Running correlation on category codes
Ordinal	Satisfaction rating, education level, rank	Mann-Whitney U, Kruskal-Wallis, Wilcoxon, Spearman, Kendall’s tau, Friedman, ordinal regression	Running a t-test or ANOVA directly on raw Likert items
Scale	Age, income, test score, time	One-sample and independent t-tests, ANOVA, Pearson correlation, linear regression	Skipping the normality check before choosing the test

So if you are wondering what test should I run in SPSS when you simply need to compare two groups: use an independent samples t-test when your outcome is scale data, Mann-Whitney U when it is ordinal, and Chi-square when it is nominal. The test name changes, the logic behind picking it does not.

Quick SPSS Menu Paths for the Three Most-Used Tests

Once you know which test you need, here is exactly where to find it inside SPSS:

1. Chi-square test: Analyze, then Descriptive Statistics, then Crosstabs, then tick “Chi-square” under the Statistics button
2. Independent samples t-test: Analyze, then Compare Means, then Independent-Samples T Test
3. Mann-Whitney U test: Analyze, then Nonparametric Tests, then Legacy Dialogs, then 2 Independent Samples

Whatever test you run, the column to check in your output is Sig. (sometimes shown as Asymp. Sig. for non-parametric tests). This is your p-value. Below 0.05 is conventionally read as a statistically significant result, above it means you have not found enough evidence to reject the null hypothesis.

Will Your Committee Actually Accept This Test?

This is the part that pure statistics content never addresses, and honestly, it matters more than people realise. Correctness and defensibility are not the same thing.

A test can be statistically defensible and still get you questioned in a viva if you cannot explain why you picked it over the obvious alternative. My advice: whatever test you choose, write one sentence in your methodology chapter explaining the variable type that drove the decision. Something like “an independent samples t-test was used because the outcome variable, test score, is measured on a continuous scale and the normality assumption was confirmed via Shapiro-Wilk.” That one sentence pre-empts almost every committee pushback I have seen.

When you report the result, stick to APA format exactly, for example t(48) = 2.31, p = .024. It signals to your committee that you understand the convention, not just the software output.

Common Mistakes I See Reviewing Dissertation Chapter 4s

• Running parametric tests on ordinal Likert items without checking the composite vs single-item distinction
• Choosing a test based on what a previous paper in the same field used, without checking if their variable types actually match yours
• Forgetting that SPSS merges Interval and Ratio into one “Scale” label, and getting confused when cross-referencing textbook terminology
• Never checking normality before defaulting to a t-test or ANOVA
• Ignoring that a nominal outcome with multiple predictors needs logistic regression, not linear regression
• Running a significant ANOVA and stopping there, without a post-hoc test to find out which groups actually differ

Case study: A student I worked with, let us call her Ananya, had survey data measuring employee engagement on a 7-point scale across three departments. She had already run a one-way ANOVA and got a significant result before our session.

When we checked normality by department, two of the three groups badly violated it. We switched to Kruskal-Wallis, and the result held, but her methodology section now had a defensible paragraph explaining why she tested normality first. Her committee had zero questions about test choice at the viva.

Case study: Another student, Rohan, was comparing customer satisfaction (ordinal, 1 to 5) against customer segment (nominal, three categories: new, returning, loyalty-member) and wanted to know if segment “predicted” satisfaction. He had been advised to just run a correlation. Correlation does not work cleanly here because segment is nominal with more than two categories. We used Kruskal-Wallis instead, treating segment as the grouping variable and satisfaction rating as the ordinal outcome, which matched his actual research question far better.

If any of this sounds like your exact situation and you are not sure which test fits your specific variables, knowing if you actually need a dissertation expert is worth a read before you spend another week second-guessing yourself. And if your sample size feels too small or too large for the test you have in mind, the sample size calculator is a quick free check before you commit.

FAQ

Siddharth Gupta has spent 12 years guiding dissertation and thesis students through statistical analysis in SPSS, R, and Stata. Connect on LinkedIn.