How to Choose Statistical Tests for Dissertation Research – A Practical Decision Guide
Table of Contents
After twelve years of reading dissertation methodology chapters, I can tell you the test itself is rarely the hard part. Matching it to your exact question, data, and design is where students get stuck. This guide covers exactly how to choose statistical tests for dissertation research, four checks at a time.
Look at your research question, check your data type, count your groups, then confirm your assumptions. Four checks, in that order, get you to the right test almost every time. The rest of this statistical test selection guide walks through each one, gives you a selection table, and shows you how to defend the choice in your methodology chapter.
How to choose statistical tests for dissertation research, the 4-question framework
I use the same four questions with every student I advise, regardless of discipline.
Question 1, what is my research question actually asking
Every research question does one of three things. It compares groups, it tests a relationship between variables, or it predicts an outcome. Write your question using one of those three verbs before you touch a test name. This single step eliminates more wrong turns than any other step in the process.
Question 2, what type of data do I have
Nominal data has no order, like gender or country. Ordinal data has order but uneven gaps, like a Likert item. Interval and ratio data have equal, measurable gaps, like age or a test score. Get this right first, since choosing statistical tests based on variables is the foundation everything else in this guide builds on.
Question 3, how many groups, and are they independent or paired
Two groups need a different test than three or more. Independent groups, like men versus women, need a different test than paired groups, like the same participants measured before and after an intervention. Get this wrong and the test still runs, but the result answers a different question than the one you asked.
Question 4, does my data meet the test’s assumptions
Most parametric tests assume your data is roughly normal and your groups have similar variance. Check both before you commit. If either assumption fails, move to a non-parametric alternative, covered in the next section, rather than forcing a test your data cannot support.
The statistical test selection table
Think of this statistical test decision tree, laid out as a table, as what I actually hand my own students. Find your research question in the left column, confirm your data type, and read across.
| Research question | Data type | Groups | Parametric test | Non-parametric alternative |
|---|---|---|---|---|
| Compare 2 independent groups | Continuous | 2, independent | Independent samples t-test | Mann-Whitney U |
| Compare 2 paired groups | Continuous | 2, paired | Paired samples t-test | Wilcoxon signed-rank |
| Compare 3+ independent groups | Continuous | 3+, independent | One-way ANOVA | Kruskal-Wallis |
| Compare 3+ paired groups | Continuous | 3+, paired | Repeated measures ANOVA | Friedman test |
| Relate 2 continuous variables | Continuous | – | Pearson correlation | Spearman correlation |
| Relate 2 categorical variables | Categorical | – | Chi-square test of independence | Fisher’s exact test, small samples |
| Predict a continuous outcome | Continuous outcome | – | Linear regression | – |
| Predict a binary outcome | Binary outcome | – | Logistic regression | – |
| Compare observed vs expected frequencies | Categorical, 1 variable | – | Chi-square goodness-of-fit | – |
Comparing means or groups
The t-test and ANOVA family answers one question, is the average outcome different across groups. Pick a t-test for two groups, ANOVA for three or more, and add the paired or repeated measures version when the same participants are measured more than once.
Testing relationships between variables
Correlation tells you two variables move together, regression tells you by how much, and lets you predict one from the other. Use correlation when you only need the strength and direction of a relationship, and regression when you need a predictive model.
Analysing categorical data
Chi-square tests ask whether categories are associated or match an expected pattern. For paired categorical data, like the same group surveyed before and after, use McNemar’s test instead of a standard chi-square test.
When your data breaks the rules, choosing non-parametric alternatives
How to check normality, homogeneity of variance, and independence
Run a Shapiro-Wilk test for normality and a Levene’s test for equal variance before you run your main test, not after. SPSS, R, Stata, and Python all output both with a single command. Independence usually comes from your study design rather than a test, so check your sampling method directly rather than looking for a number.
Parametric to non-parametric swap table
| Parametric test | Non-parametric alternative | When to switch |
|---|---|---|
| Independent t-test | Mann-Whitney U | Non-normal data or a small sample |
| Paired t-test | Wilcoxon signed-rank | Non-normal differences between pairs |
| One-way ANOVA | Kruskal-Wallis | Non-normal data or unequal variance |
| Pearson correlation | Spearman correlation | Non-linear monotonic relationship or ordinal data |
What to do with Likert scale and ordinal survey data
A single Likert item is ordinal, and the safer choice is a non-parametric test. A composite score built from several Likert items, summed or averaged across a validated scale, behaves close enough to interval data that most committees accept parametric tests on it. State which one you have, and say why, in your methodology chapter.
Matching tests to common dissertation research designs
The table above is the general case. These are the specific designs I see most often in US and UK dissertations.
Comparing two groups
A business dissertation comparing employee engagement scores between two departments is a textbook independent samples t-test, or a Mann-Whitney U if the scores are skewed.
Comparing three or more groups
Comparing exam performance across three teaching methods calls for a one-way ANOVA, followed by a post-hoc test like Tukey’s, to see which pairs actually differ.
Testing a relationship between two variables
Testing whether study hours predict GPA is a Pearson correlation if both variables are continuous and roughly linear, or a simple linear regression if you want to predict one from the other.
Predicting an outcome from several variables
Predicting dropout risk from attendance, grades, and household income calls for multiple regression, or logistic regression if dropout itself is a yes or no outcome.
Before-after and repeated measures designs
Measuring the same employees’ stress scores before and after a wellness programme is a paired t-test for two time points, or a repeated measures ANOVA for three or more.
How to justify your statistical test choice in your methodology chapter
Picking the right test is only half the job. Your committee wants to see the reasoning on the page, not just the result.
The 3-part justification formula
State the test, the feature of your data that selected it, and the assumption check you ran. That is the whole formula. A committee is not looking for elegant prose here, it is looking for evidence the choice was reasoned rather than convenient.
Both of the peer-reviewed guides I return to most for this, a methodology review in the Indian Journal of Ophthalmology and a parallel guide in the International Journal of Ayurveda Research, make the same point. State your reasoning, not just your conclusion.
Sample justification sentences you can adapt
Both examples below follow APA 7 reporting conventions, the style most US and UK committees expect in the results chapter. An independent samples t-test was used to compare the outcome between the two groups, as both groups were independent and the outcome met the assumption of normality, confirmed by a Shapiro-Wilk test, p greater than .05.
A Kruskal-Wallis test was used in place of a one-way ANOVA because Levene’s test indicated unequal variances across the three groups.
What to do when a supervisor questions your choice
Bring your assumption test output, not just your opinion. A supervisor who sees a Shapiro-Wilk result and a Levene’s test alongside your chosen test has far less room to push back than one who only sees a test name on a page.
Running your chosen test in SPSS, R, Stata, or Python
Once you know the test, this is where to find it in each tool.
| Test | SPSS | R | Stata | Python |
|---|---|---|---|---|
| Independent t-test | Analyze, Compare Means, Independent-Samples T Test | t.test() | ttest | scipy.stats.ttest_ind() |
| One-way ANOVA | Analyze, Compare Means, One-Way ANOVA | aov() | anova | scipy.stats.f_oneway() |
| Chi-square | Analyze, Descriptive Statistics, Crosstabs | chisq.test() | tabulate, chi2 | scipy.stats.chi2_contingency() |
| Pearson correlation | Analyze, Correlate, Bivariate | cor.test() | pwcorr | scipy.stats.pearsonr() |
| Linear regression | Analyze, Regression, Linear | lm() | regress | statsmodels.OLS() |
Statistical test mistakes that get dissertations sent back
I see the same five mistakes in almost every revise-and-resubmit I review.
- Running a parametric test without checking assumptions first
- Treating a single Likert item as interval data without saying so
- Picking a test based on which one happens to give a significant result
- Reporting a test name with no justification sentence in the methodology chapter
- Mismatching paired and independent designs, like running an independent t-test on before-after data
Should you trust AI tools to pick your statistical test
AI tools are decent at naming a plausible test once you describe your data correctly. They are not reliable at checking whether you described it correctly in the first place, and that is where most dissertation students go wrong.
My own rule with students, use an AI suggestion as a second opinion, then verify it against the table earlier in this guide and run the assumption checks yourself. Do not cite an AI tool as your reasoning in the methodology chapter. A human reviewer will still ask you to defend the choice in your own words.
Frequently asked questions
How do I know which statistical test to use for my dissertation?
Answer four questions in order. What does your research question ask, compare, relate, or predict. What type of data do you have. How many groups, and are they independent or paired. Does your data meet the test’s assumptions. The selection table earlier in this guide maps the answers to a test name.
What is the difference between parametric and non-parametric tests?
Parametric tests assume your data follows a roughly normal distribution and require interval or ratio data. Non-parametric tests make no distribution assumption and work with ranked or ordinal data, which makes them the fallback when a parametric test’s assumptions fail.
What statistical test should I use for Likert scale data?
A single Likert item is ordinal, so use a non-parametric test like Mann-Whitney U or Kruskal-Wallis. A composite score averaged across several Likert items behaves close enough to interval data that most committees accept a parametric test, provided you state that reasoning.
What do I do if my data violates a test’s assumptions?
Switch to the non-parametric alternative. An independent t-test becomes a Mann-Whitney U, a one-way ANOVA becomes a Kruskal-Wallis, and a Pearson correlation becomes a Spearman correlation. The swap table earlier in this guide lists every common pair.
What is the difference between ANOVA and a t-test?
A t-test compares the means of exactly two groups. ANOVA compares the means of three or more groups in a single test, which avoids the inflated error rate you get from running several separate t-tests.
How do I justify my choice of statistical test to my committee?
State the test, the feature of your data that selected it, and the assumption check you ran, in that order. Attach the actual Shapiro-Wilk or Levene’s test output as evidence rather than asserting the assumption was met.
Can I trust AI to choose my statistical test?
Use it as a second opinion, not a final answer. Verify any AI-suggested test against your own data type, group structure, and assumption checks before you commit to it in your methodology chapter.
How many participants do I need for a t-test or ANOVA?
It depends on the effect size you expect, your significance level, and the statistical power you want, typically 0.80. G*Power, a free tool, calculates this directly for both tests. As a rough starting point, many two-group comparisons need around 30 participants per group to detect a medium effect reliably.
Still not sure your test fits your exact data? Send over your research question and variable types for a quick second opinion. Book a short consultation and get choosing statistical tests for dissertation research right before you run anything in SPSS, R, Stata, or Python.
Written by a dissertation statistics consultant with over a decade of experience helping doctoral candidates defend their analysis chapter. Connect on LinkedIn.