Statistical test selector
Answer a few questions about your design and get the test that fits, with the reasoning shown rather than hidden.
A statistical test selector turns three facts about your study, the type of outcome variable, what you are trying to do with it, and whether the assumptions hold, into the test that matches. The choice between a t-test, ANOVA, correlation, a regression, or a non-parametric alternative is not a matter of taste: it follows from the measurement level of your data and your research question. This tool walks that logic and links to the guide behind each recommendation.
Recommended test
Independent-samples t-test
Two independent groups on a continuous, roughly normal outcome are compared with the independent-samples t-test.
Read the reasoning
How the recommendation is reached
The selector follows the same decision tree a statistician uses. The first branch is the outcome variable: a continuous measurement points toward means and the parametric family, ordinal data toward rank-based tests, and unordered categories toward the chi-square family. The second branch is the goal, comparing groups, measuring an association, or predicting, and the third is whether the parametric assumptions hold.
Comparing two independent groups on a continuous outcome gives an independent-samples t-test when normality is reasonable, and the Mann-Whitney U test when it is not. Three or more groups move to ANOVA or the Kruskal-Wallis test. Paired measurements switch to the paired t-test, Wilcoxon signed-rank, or repeated-measures ANOVA. Association questions resolve to Pearson or Spearman correlation, or chi-square for two categorical variables, and prediction resolves to linear or logistic regression by outcome type. The recommendation is a starting point: it assumes you have checked the assumptions it names.
How to choose the right statistical test
Choosing a test feels overwhelming only because people start from the test name instead of from their own study. Begin with three questions in order. First, what is the measurement level of your outcome variable: is it continuous (a measured quantity), ordinal (ranked categories such as a Likert rating), or nominal (unordered categories)? Second, what is your research question: are you comparing groups, measuring an association, or predicting one variable from others? Third, how many groups or variables are involved, and are the observations independent or repeated on the same participants?
Those three answers narrow the field to one or two candidates before you ever think about a specific test. The final step is to check the assumptions the candidate test requires, normality and equal variances being the usual ones, and switch to a non-parametric alternative if they fail. Working in this order, from your data outward, is the whole skill, and our full guide on how to choose a statistical test walks every branch. Being clear about which variable is the predictor and which is the outcome, covered in the explainer on variable roles, keeps the decision honest.
Parametric vs non-parametric tests
Parametric tests, such as the t-test, ANOVA, and Pearson correlation, assume the data follow a roughly normal distribution and, for group comparisons, that the groups share a similar variance. When those assumptions hold, parametric tests have more statistical power, meaning they detect real effects with fewer participants. Non-parametric tests, including the Mann-Whitney U, Kruskal-Wallis, and Spearman correlation, work on ranks instead of raw values, so they make no normality assumption and tolerate skew, outliers, and ordinal data that parametric tests cannot handle cleanly.
The practical rule is to prefer the parametric test when its assumptions are reasonably met, especially with larger samples where the central limit theorem makes means well behaved, and to move to the non-parametric counterpart when the data are clearly skewed, the sample is small, or the outcome is ordinal. Check normality before deciding, rather than assuming it, using the steps in our note on how to check the normality assumption. The full comparison, including which nonparametric test pairs with each parametric one, sits in choosing between parametric and rank-based tests.
Tests for differences vs tests for relationships
Most analyses fall into one of two families, and naming the family first removes half the confusion. Tests for differences ask whether group averages differ: a t-test compares two groups, a one-way ANOVA compares three or more, and repeated-measures versions handle the same participants measured on several occasions. The outcome is continuous and the predictor is a grouping variable. These tests answer questions phrased as “is there a difference between…”
Tests for relationships ask whether two variables move together. A correlation coefficient measures the strength and direction of an association between two continuous variables, while a regression goes further and predicts one variable from one or more others. These answer questions phrased as “is there a relationship between…” or “does X predict Y?” Knowing which family your research question belongs to, before any software opens, is the single most reliable way to land on the correct test.
Matching your variables to a statistical test
Once the family is settled, the exact test follows from the number and type of your variables. With one continuous outcome and one two-level grouping variable, an independent-samples t-test fits; with a grouping variable of three or more levels, a one-way ANOVA does. Two continuous variables point to Pearson or Spearman correlation, and a continuous outcome predicted from several variables points to multiple regression. A categorical outcome changes everything: two categorical variables call for the chi-square test, and a binary outcome predicted from other variables calls for logistic regression.
A clean way to keep this straight is a short table mapping outcome type, predictor type, and number of groups to a named test, which is exactly the logic the selector above automates. Reading SPSS output correctly once the test is run is the next step, and our guide on reading the output SPSS produces covers it. When a binary outcome is involved, the worked example in logistic regression interpretation shows how the variable types drive the choice.
Common mistakes when choosing a statistical test
The most frequent error is running a parametric test on data that violate its assumptions, applying a t-test to badly skewed scores or to ordinal Likert items without checking normality or considering a rank-based alternative. A close second is treating an ordinal variable as continuous, averaging Likert responses as if the gaps between points were equal. A third is the multiple-comparisons trap: running many separate t-tests across several groups instead of a single ANOVA, which inflates the false-positive rate with every extra test.
Other recurring slips include ignoring the paired structure of repeated measurements and analysing them as independent groups, and confusing correlation with causation when a relationship test is mistaken for evidence that one variable causes another. Each mistake traces back to skipping one of the three setup questions, so the remedy is always to restate the measurement level, the goal, and the assumptions before committing. Our pieces on analysing Likert data and mistaking correlation for causation address the two errors that derail dissertations most often.
Frequently asked questions
- What statistical test should I use for categorical data?
For categorical (nominal) outcomes, the chi-square test of independence checks whether two categorical variables are associated, while a test of two proportions compares event rates between groups. When expected cell counts are small, Fisher's exact test replaces chi-square. Set the outcome to nominal in the selector and it points you to the right one.
- What statistical test should I use for ordinal data?
Ordinal outcomes, such as Likert ratings, are usually analysed with rank-based nonparametric tests: the Mann-Whitney U test for two groups and the Kruskal-Wallis test for three or more. Spearman's rank correlation measures the association between two ordinal variables. Setting the outcome to ordinal returns the match for your design.
- What test is used to compare two groups?
For a continuous outcome, an independent-samples t-test compares two separate groups and a paired t-test compares two measurements on the same participants, with the Mann-Whitney U test as the nonparametric alternative. The right choice depends on whether the groups are independent and whether the data meet the normality assumption, both of which the selector asks about.
- What statistical test shows the relationship between two variables?
To quantify the relationship between two continuous variables, use Pearson's correlation when the association is linear and both variables are roughly normal, or Spearman's correlation otherwise. When you want to predict one variable from another, linear regression is the tool. Choose association or prediction as your goal to get a recommendation.