How to Choose the Right Statistical Test for Your Thesis
T-test or ANOVA? Chi-square or regression? A plain-English decision guide to matching your research question and data to the right statistical test.
"Which test should I use?" is the most common question we hear from Masters and PhD students. The good news is that the choice is rarely mysterious. It comes down to three things: what you're trying to find out, what type of variables you have, and how many groups you're comparing.
Step 1: Identify your variable types
- Continuous (scale): measured on a numeric scale, such as income, age, test score or yield.
- Categorical (nominal): groups with no order, such as gender, county or type of business.
- Ordinal: ordered categories, such as education level or a single Likert item (Strongly disagree → Strongly agree).
- Binary: two categories, such as yes/no, adopted/did not adopt, or passed/failed.
Then decide which variable is your outcome (dependent variable) and which are your predictors or grouping variables (independent variables).
Step 2: Match your question to a test
| Your question | Data | Parametric test | Non-parametric alternative |
|---|---|---|---|
| Do two independent groups differ? | Continuous outcome, 2 groups | Independent samples t-test | Mann–Whitney U |
| Did the same people change (before/after)? | Continuous outcome, 2 measurements | Paired t-test | Wilcoxon signed-rank |
| Do three or more groups differ? | Continuous outcome, 3+ groups | One-way ANOVA | Kruskal–Wallis |
| Are two categorical variables associated? | Two categorical variables | Chi-square test of independence | Fisher's exact test (small samples) |
| Are two continuous variables related? | Two continuous variables | Pearson correlation | Spearman's rank correlation |
| What predicts a continuous outcome? | Continuous outcome, several predictors | Multiple linear regression | Robust or quantile regression |
| What predicts a yes/no outcome? | Binary outcome | Binary logistic regression (or probit) | — |
| What predicts an ordered outcome? | Ordinal outcome | Ordinal logistic regression | — |
| What predicts a count? | Count outcome (visits, children) | Poisson or negative binomial regression | — |
Step 3: Check the assumptions
Parametric tests like the t-test, ANOVA and linear regression rest on assumptions. Examiners often ask whether you checked them, so make sure you did:
- Normality of the outcome (or of the residuals, for regression). Use a histogram or Q–Q plot alongside a Shapiro–Wilk test. With large samples, small departures matter less.
- Equal variances between groups. Use Levene's test. If it fails, report Welch's t-test, which most software gives you automatically.
- Independence of observations. Repeated measures on the same people need paired or panel methods.
- For regression: linearity, no severe multicollinearity (check VIFs) and homoscedasticity. Use robust standard errors if variance isn't constant.
- For chi-square: expected counts of at least 5 in most cells. Otherwise, use Fisher's exact test.
If assumptions fail badly, use the non-parametric alternative in the table, or transform the variable (for example, log of income).
Special cases students often meet
Likert scale data
A single Likert item is ordinal, so analyse it with non-parametric tests or ordinal regression. A composite score built from several items that measure one concept is commonly treated as continuous. Check the scale's reliability first with Cronbach's alpha (0.7 or above is the usual benchmark).
Data over time
If you observe the same units (households, firms, counties) over several periods, you have panel data. Ordinary regression ignores that structure. See our guide to fixed vs random effects and the Hausman test.
Time series
One unit observed over many periods, such as GDP, inflation or exchange rates, calls for time series methods: unit root tests, cointegration, ARDL or VAR models. These are usually run in EViews or Stata.
Step 4: Report it properly
A good results section reports the test used, the test statistic, degrees of freedom, p-value and an effect size (Cohen's d, eta squared or an odds ratio), not just "significant". For example: "Mobile money users saved significantly more than non-users, t(398) = 4.21, p < .001, d = 0.42."
Still unsure which test fits your hypotheses? Book a consultation. We'll match your objectives and data to the right methods, run them with you in SPSS, Stata, R or EViews, and explain every result so you can defend it.
Want support with your analysis?
Thesis data analysis and defense prep in R, Stata, SPSS, EViews, Excel and Power BI, with code and explanations you can defend.
Get a free quote →