Choosing the right statistical test comes down to three things: the research question you are asking, the type of variables you measured, and how many groups or time points you are comparing. Get those three straight and the appropriate test is usually obvious. Most of the confusion in a dissertation analysis comes from skipping that step and reaching for a familiar test instead of the correct one.
Start with your research question, not the test
The most common mistake we see is a researcher who has already decided to run a t-test or a regression before looking at what the data can actually support. Work in the other direction. A question about a difference between groups points to one family of tests; a question about a relationship between variables points to another; a question about prediction points to a third. Write your research question in plain language first, then translate it. If you would rather answer a few questions and see the result, the interactive test selector walks the same decision tree and shows its reasoning.
Identify your variable types
Every test assumes a particular combination of dependent and independent variable types. The distinction that matters most is whether your outcome variable is continuous (a score, a measurement) or categorical (a group, a yes/no). If you are unsure which is which, read our guide on independent versus dependent variables before going further, because the rest of the decision depends on it.
A decision guide for the common tests
For most master's and doctoral projects, the test you need is one of the following. Match the row to your design:
| Your question | Outcome variable | Typical test |
|---|---|---|
| Difference between two groups | Continuous | Independent-samples t-test |
| Difference across three or more groups | Continuous | One-way ANOVA |
| Change in the same people over time | Continuous | Paired t-test or repeated-measures ANOVA |
| Relationship between two continuous variables | Continuous | Pearson correlation |
| Predicting a continuous outcome | Continuous | Linear regression |
| Predicting a yes/no outcome | Categorical | Logistic regression |
| Association between two categories | Categorical | Chi-square test of independence |
Parametric or not? Check the assumptions
The tests above are mostly parametric, which means they assume your data meets conditions such as normality and equal variance. When those conditions fail, the parametric result can be misleading and you switch to a nonparametric equivalent (a Mann-Whitney U in place of a t-test, a Kruskal-Wallis in place of ANOVA, a Spearman correlation in place of Pearson). This is not optional polish; a marker who spots an unchecked assumption will question the whole results chapter. Our guide to testing for normality walks through how to check the most important one and what to do when the data is not normal.
Watch the number of variables and groups
Two more details change the test. First, the number of groups: comparing two groups is a t-test, three or more is an ANOVA, because running several t-tests inflates your false-positive rate. Second, the number of predictors: one predictor is a simple regression, several is a multiple regression, and nested or grouped data (students within schools, measurements within patients) calls for a mixed-effects model. Doctoral work in particular often needs these more advanced models, which we cover under PhD statistics help.
A worked example, from question to test
Suppose your research question is whether students taught by three different methods, lecture, flipped classroom, and problem-based learning, end the semester with different exam scores. Work it through in order. The outcome is the exam score, a continuous variable. The grouping factor is teaching method, a categorical independent variable with three levels. You are asking about a difference between groups, not a relationship or a prediction, and there are more than two groups. Those facts land you on a one-way ANOVA, with a Tukey post-hoc test afterward to see which specific methods differ.
Notice that you never had to guess. Each property of the data narrowed the field by one step until a single test remained. If the same study followed the same students across three time points instead of three separate groups, the design would shift to a repeated-measures ANOVA, because the observations would no longer be independent. The question and the design decide the test, every time. Working a real dataset through this logic is the core of thesis statistics help.
Test-selection mistakes that cost marks
A handful of errors show up again and again in results chapters, and each is easy for an assessor to spot:
- Running multiple t-tests instead of ANOVA: comparing three or more groups with a string of pairwise t-tests inflates the false-positive rate that ANOVA is built to control.
- Ignoring the measurement level of the outcome: applying a test for continuous data to a categorical outcome, or the reverse, invalidates the result before you start.
- Treating ordinal Likert items as continuous without justification: a single Likert item is ordinal, and pooling it into a mean needs a stated rationale rather than a silent assumption.
- Choosing a test to chase a significant p-value: picking the analysis that happens to cross .05 rather than the one the design calls for is the fastest way to lose credibility in a viva.
When in doubt, plan the test before you collect data
The cheapest time to choose your analysis is before collection, when you can still change the design to fit the question. A documented analysis plan also protects you in your viva, because you can show the test was chosen on principle rather than picked to produce a significant p-value. If you are still at the design stage, that is exactly what statistical consulting is for. Once the test is settled, the rest of the work is reporting it correctly.