How to Choose the Right Statistical Test: A Guide with Examples (2026)
“Which test do I run?” If I had a rupee for every time a student has asked me that while staring at a dataset, I could fund a small research grant. It is one of the most paralysing moments in quantitative research: your data is collected, your deadline is looming, and there are dozens of statistical tests with intimidating names, each seemingly a chance to get it wrong. The reassuring truth is that choosing a test is not guesswork, and you do not need to memorise all of them. You need to ask a few simple questions in the right order, and the correct test reveals itself.
This guide gives you that decision framework, explains the key distinction between parametric and non-parametric tests, and provides a plain-language reference table of the 16 tests you are most likely to need.
Start here: the 4 questions that choose your test
Almost every test-selection decision comes down to answering four questions in sequence (Scribbr, 2026; Field, 2024):
- What is your goal? Are you trying to compare groups, test an association/correlation between variables, or predict one variable from others?
- How many variables, and what are their roles? How many variables do you have, and which is the dependent (outcome) and which the independent (predictor)?
- What is the level of measurement? Is each variable categorical (nominal), ordinal, or continuous (interval/ratio)? This matters more than almost anything else.
- Does your data meet parametric assumptions? Chiefly, is your continuous data roughly normally distributed, with reasonably equal variances? If yes, you can use parametric tests; if not, you use their non-parametric equivalents.
Hold those four answers in mind and the table at the end of this guide becomes a simple lookup rather than a maze.
Parametric vs non-parametric: the distinction that trips everyone up
This single idea resolves half of all test-choice confusion.
Parametric tests — such as the t-test, ANOVA and Pearson’s correlation — assume your data meets certain conditions, most importantly that your continuous outcome is approximately normally distributed. When those assumptions hold, parametric tests are more powerful (better at detecting real effects).
Non-parametric tests — such as Mann-Whitney U, Kruskal-Wallis and Spearman’s correlation — make far fewer assumptions. You turn to them when your data is ordinal, your sample is small, or your continuous data clearly violates normality. They are the robust, lower-assumption alternatives.
A handy way to remember the pairings: Mann-Whitney U is the non-parametric cousin of the t-test; Kruskal-Wallis is the non-parametric cousin of ANOVA; Spearman’s is the non-parametric cousin of Pearson’s. Same job, fewer assumptions.
Comparing groups: t-tests and ANOVA
If your goal is to compare averages, the number of groups decides the test. To compare two groups, you use a t-test; to compare three or more, you use ANOVA. This is a rule students break constantly — running multiple t-tests across several groups instead of one ANOVA. Don’t. Each extra t-test inflates your chance of a false positive, which is precisely why ANOVA exists.
Within the t-test family, match the test to your design: an independent (Student’s) t-test for two separate groups, a paired t-test for the same people measured twice (before and after), and Welch’s t-test when your two groups have unequal variances or sizes.
Testing relationships: correlation and regression
If your goal is to see how variables move together, you want correlation or regression. Correlation measures the strength of a relationship (Pearson’s for linear, continuous data; Spearman’s for ordinal or non-linear monotonic relationships). Regression goes further and lets you predict: simple linear regression predicts an outcome from one variable, multiple regression from several, and logistic regression predicts a binary outcome (yes/no, success/failure).
Working with categorical data
When both your variables are categorical, you are usually testing for association, and the Chi-squared test is your default. For small samples or 2×2 tables, Fisher’s exact test is more appropriate because it gives an exact p-value rather than an approximation.
The reference table: 16 statistical tests at a glance
Use the four questions above, then find your situation here.
| Test | When to use it | Example |
|---|---|---|
| Descriptive statistics | To summarise and describe your dataset | Reporting the mean age and standard deviation of participants |
| Z-test | Comparing a sample mean to a known population mean (large sample, known variance) | Testing whether average customer spending differs from a known national average |
| T-test (Student’s) | Comparing the means of two groups (small sample, unknown variance) | Comparing average exam scores of two classes |
| Paired t-test | Comparing two related measurements on the same subjects | Comparing patients’ weight before and after an intervention |
| Welch’s t-test | Comparing two groups with unequal variances or sizes | Comparing scores from two classes of very different sizes |
| Chi-squared test | Testing association between two categorical variables | Whether admission to a programme is related to applicants’ region |
| Fisher’s exact test | Association in small samples or 2×2 tables | Testing a treatment effect in a small clinical trial |
| Mann-Whitney U test | Non-parametric comparison of two independent groups | Comparing satisfaction ratings between two products |
| ANOVA | Comparing the means of three or more groups | Whether glucose level differs at one, two and three hours after a meal |
| Kruskal-Wallis test | Non-parametric alternative to ANOVA (3+ groups) | Comparing satisfaction across several departments |
| Pearson’s correlation | Strength of a linear relationship between two continuous variables | Whether HbA1c relates to triglyceride levels in patients |
| Spearman’s rank correlation | Non-parametric / monotonic relationship strength | The relationship between education level and income |
| Simple linear regression | Predicting one continuous outcome from one predictor | Predicting lung function from height |
| Multiple regression | Predicting one outcome from several predictors | How age, body fat and sodium intake influence blood pressure |
| Logistic regression | Predicting a binary (yes/no) outcome | Predicting the likelihood of heart disease from risk factors |
| Factor analysis | Reducing many variables to a few underlying factors | Identifying personality traits underlying survey questions |
A worked example, start to finish
Suppose you want to know whether a stress-reduction workshop improved wellbeing scores among the same group of nurses, measured before and after. Walk the four questions: your goal is to compare (1); you have one continuous outcome measured twice on the same people (2); the outcome is continuous (3); and if the differences are roughly normally distributed (4), you use a paired t-test. If they are not normal, you use its non-parametric cousin, the Wilcoxon signed-rank test. That is the entire logic — four questions, one answer.
A few cautions from experience
First, choose your test before you collect data, not after. Your intended analysis should be decided at the design stage, so your sample size and measures actually fit the test. Deciding afterwards is how people end up with data that no test can rescue. (This is part of why your methodology chapter should name your analysis plan upfront.)
Second, check your assumptions rather than hoping. Before running a parametric test, look at normality and variance; software makes this quick, and skipping it can invalidate your results. Third, remember that a statistical test answers a narrow question — it does not interpret your findings for you. The meaning is still your job.
Finally, if you are new to running these in software, my SPSS for beginners guide walks through the clicks, and if you are still deciding whether your study is even quantitative, start with qualitative vs quantitative research.
Frequently asked questions
How do I choose the right statistical test?
Ask four questions in order: your goal (compare, associate, or predict), how many variables and their roles, each variable’s level of measurement, and whether your data meets parametric assumptions. Those answers point to the test.
What is the difference between parametric and non-parametric tests?
Parametric tests (t-test, ANOVA, Pearson’s) assume conditions like normal distribution and continuous data; non-parametric tests (Mann-Whitney U, Kruskal-Wallis, Spearman’s) make fewer assumptions and suit ordinal data or violated assumptions.
When should I use a t-test versus ANOVA?
Use a t-test for two groups and ANOVA for three or more. Running multiple t-tests instead of ANOVA inflates the risk of false positives.
What statistical test is used for categorical data?
Chi-squared for association between categorical variables, Fisher’s exact test for small samples or 2×2 tables, and logistic regression to predict a binary outcome.
Should I decide my test before or after collecting data?
Before. Choosing your analysis at the design stage ensures your sample size and measures fit the test you need.
Final thought
The names are intimidating; the logic is not. Behind all sixteen of these tests sits the same short conversation: what are you trying to find out, what kind of data do you have, and does it meet the assumptions? Answer those honestly and the right test is rarely in doubt. Keep this table beside you, decide your analysis before you collect a single data point, and the moment of panic in front of your dataset simply stops happening.
Need help choosing or running the right analysis for your study? That’s part of what I help researchers with — see the services page.
References
Bevans, R. (2026). Choosing the right statistical test: Types & examples. Scribbr. https://www.scribbr.com/statistics/statistical-tests/
Field, A. (2024). Discovering Statistics Using IBM SPSS Statistics (6th ed.). SAGE Publications.
Research Prospect. (2026). Which statistical test you should use in research. https://www.researchprospect.com/which-statistical-test-you-should-use/