How to Choose the Right Statistical Test: A Decision Flowchart for Beginners

The Question Every Researcher Dreads

Picture this: your data is collected, your questionnaire responses are sitting in a spreadsheet, and you open your analysis software with excitement. Then you see it — a menu with dozens of tests: t-test, ANOVA, chi-square, correlation, regression, Mann-Whitney, Wilcoxon… and your brain freezes.

You are not alone. Choosing the right statistical test is one of the most common anxieties in research, and it is also the question supervisors love to ask at a thesis defense: “Why did you use this test, and not that one?” A confident answer to that question comes from understanding a simple decision process — not from memorizing every test in existence.

This guide gives you exactly that: a five-question flowchart that narrows any research problem down to the right test. Work through it once, and you will never stare at that test menu with panic again.

Before You Start: Three Pairs of Words You Must Know

The flowchart asks about your data, so let’s get three distinctions straight. These are the only technical terms you need.

1. Continuous vs. categorical vs. ordinal data

• Continuous data can take (almost) any value along a scale: exam scores, reaction time in milliseconds, height in centimeters.
• Categorical data falls into distinct groups with no natural order: gender, yes/no answers, treatment vs. control group.
• Ordinal data has categories with a meaningful order, but the gaps between them are not equal: Likert-scale responses (strongly disagree → strongly agree), rankings, grades.

The most common confusion is the outcome variable’s type. Ask: what exactly did I measure, and what kind of number is it?

2. Independent vs. paired (related) groups

Independent groups contain different people: a treatment group and a control group, men and women, freshmen and seniors. Paired groups contain the same people measured twice (before/after), or people deliberately matched (twins, matched pairs). Getting this wrong is one of the most common mistakes beginners make — we will come back to it.

3. Parametric vs. nonparametric tests

Parametric tests (t-test, ANOVA, Pearson correlation) assume your continuous data is roughly normally distributed — the familiar bell curve. Nonparametric tests (Mann-Whitney U, Wilcoxon, Spearman) make no such assumption and are the safety net when your data is skewed, ordinal, or based on a small sample. Think of nonparametric tests as the sturdy backup plan, not a lesser option.

The 5-Question Decision Flowchart

Answer these five questions in order. Each answer narrows the options until only one or two tests remain.

Question 1: What is your research question really asking?

Strip away the academic wording and put your question into one of these buckets:

• Compare groups: “Do students who attend study groups score higher than those who don’t?”
• Measure an association: “Is study time related to exam scores?”
• Predict an outcome: “Can study time, sleep, and attendance predict exam scores?”

Your bucket points you at a family of tests: comparisons → t-tests/ANOVA family; association → correlation/chi-square family; prediction → regression family. If you are unsure, write your hypothesis as “There is a difference between…” (compare), “There is a relationship between…” (associate), or “X predicts Y” (predict).

Question 2: What type of data is your outcome variable?

Continuous outcome (a test score, a time, a measurement) keeps you in the t-test/ANOVA/correlation world. Categorical outcome (pass/fail, yes/no) pushes you toward chi-square tests or logistic regression. Ordinal outcome (Likert responses, rankings) usually calls for nonparametric tests or chi-square, depending on the question.

Question 3: How many groups or variables are involved?

Two groups? You are in t-test territory. Three or more groups? That is ANOVA’s job — running three separate t-tests here is a classic beginner error (see Common Mistakes below). For association, one pair of variables means a simple correlation; several predictors pointing at one outcome means regression.

Question 4: Are the groups independent or paired?

Same people measured twice (or matched pairs)? You need a paired version of the test: paired t-test instead of independent t-test, Wilcoxon signed-rank instead of Mann-Whitney U, repeated-measures ANOVA instead of one-way ANOVA. This distinction changes the answer at almost every branch, so double-check it before you commit.

Question 5: Is your continuous data (approximately) normally distributed?

Run a quick normality check — a Shapiro-Wilk test, or simply eyeball a histogram or Q-Q plot. If the data is roughly bell-shaped, parametric tests (t-test, ANOVA, Pearson) are fine. If it is clearly skewed or your sample is very small, switch to the nonparametric alternative in the same row of the table below. When in doubt, the nonparametric option is the safer bet.

The Decision Table: From Your Answers to Your Test

Here is the flowchart flattened into one lookup table. Find your research question type and follow the “if” conditions:

Bookmark this table. Ninety percent of student research questions land somewhere in it.

Worked Example: From a Real Question to the Right Test

Let’s walk a realistic scenario through all five questions.

The scenario: A psychology student wants to know whether a new study-skills workshop improves students’ test anxiety. She recruits 60 students, measures their anxiety with a questionnaire before the workshop, delivers the workshop, and measures their anxiety again afterward with the same questionnaire.

Question 1 — what is she asking? “Is there a difference in anxiety before vs. after the workshop?” → compare groups.

Question 2 — what type of outcome? Anxiety questionnaire scores are numbers on a scale → continuous.

Question 3 — how many groups? Before and after — that’s two measurements.

Question 4 — independent or paired? The same 60 students measured twice → paired.

Question 5 — normal data? She runs a quick histogram of the score differences; they look roughly bell-shaped → yes.

Result: two paired groups, continuous, normal data → paired samples t-test. If the differences had been badly skewed, the same row says Wilcoxon signed-rank test. Notice how one wrong answer at Question 4 (treating “before” and “after” as independent groups) would have led her to the wrong test entirely — that is why the pairing question matters so much.

Common Mistakes (and How to Avoid Them)

• Using a t-test for three or more groups. Running multiple t-tests inflates your chance of a false positive with every extra comparison. Three groups → ANOVA; then a post-hoc test tells you which groups differ.
• Using Pearson correlation on ordinal data. Likert-scale data usually isn’t continuous and normal. If your variables are ranked or skewed, reach for Spearman instead of Pearson.
• Treating paired data as independent. Before/after designs, matched pairs, and twins need paired tests. Paired tests are also more powerful, because each person acts as their own control — you’re throwing that advantage away with an independent test.
• Running every test until something is significant. Fishing through tests until one gives p < .05 is called p-hacking, and reviewers (and ethics committees) take it seriously. Choose your test from your research question before you run the analysis — ideally in your proposal. To understand why p-values demand discipline, see our guide on understanding p-values.
• Confusing association with causation. A significant correlation (or group difference) does not prove one thing caused the other — especially in non-experimental designs. Choose your wording as carefully as your test.
• Never checking assumptions. Running a parametric test on badly skewed data doesn’t make the data normal; it makes your result untrustworthy. Five minutes with a histogram is the cheapest quality control in research.

Frequently Asked Questions

What if my data isn’t normally distributed?

Don’t panic — and don’t transform your data into something unrecognizable just to force normality. Use the nonparametric alternative from the decision table (Mann-Whitney U, Wilcoxon, Kruskal-Wallis, or Spearman). They are perfectly respectable tests; journals publish them every day.

Can I report more than one test in the same analysis?

Yes — different research questions legitimately need different tests. What you should not do is try five tests on the same question and report only the one that “worked.” Pre-register your analysis plan, or at least write it down before running anything.

My data is from a Likert scale. Which test do I use?

Strictly speaking, Likert responses are ordinal, so the cautious choice is nonparametric (Mann-Whitney U, Wilcoxon, or Spearman). In practice, many researchers treat well-designed multi-item Likert scales as approximately continuous — but be ready to defend that choice to your supervisor. When in doubt, run the nonparametric test.

Does sample size affect which test I choose?

Mostly indirectly: with very large samples, even tiny departures from normality can trip a formal normality test, so lean on visual checks (histograms, Q-Q plots). With very small samples, nonparametric tests are usually the safer choice, and formal tests have low power anyway. Sample-size planning itself is a separate skill — the key point here is that your test choice follows your question and data, not your N.

How do I justify my test choice in my thesis?

Write one short paragraph: state the test, then justify it against the five questions — your research question, the data type, the number of groups, whether data were paired, and the assumption checks you ran. Examiners don’t expect a novel test; they expect a reasoned choice.

Your Turn: The 60-Second Habit

Before you run a single analysis, grab a piece of paper and answer the five questions in plain sentences: what am I asking, what did I measure, how many groups, paired or independent, normal or not. That one habit will prevent more statistical errors than any software tutorial ever could — and it gives you a ready-made answer when your supervisor asks, “Why this test?”

Want to go deeper? Once your test is chosen, our guides on ANOVA, the Pearson correlation table, and understanding p-values walk you through the next steps.