Which Statistical Test Should I Use for My Research?
A simple guide to choosing the right statistical test for a medical thesis or paper, based on your outcome type, number of groups and study design.
Choosing a statistical test feels complicated, but most thesis questions can be answered by asking three things in order. This guide walks you through them and ends with a quick reference table.
Question 1: What type is your outcome variable?
- Categorical (yes/no, cured/not cured, stage I/II/III)
- Numerical (haemoglobin, blood pressure, pain score treated as numerical)
- Time to an event (time to recurrence, survival time)
Question 2: How many groups are you comparing, and are they independent?
- Two independent groups (cases vs controls, drug A vs drug B)
- Paired or repeated measurements (before and after treatment in the same patients)
- Three or more groups
Question 3: For numerical outcomes, are the data roughly normal?
If yes, use a parametric test. If the data are clearly skewed, ordinal or the groups are very small, use the non-parametric alternative. See our guide on parametric and non-parametric tests.
Quick reference table
| Your question | Outcome | Test |
|---|---|---|
| Compare proportions between independent groups | Categorical | Chi-square test (Fisher's exact test if expected counts are small) |
| Compare proportions before and after in the same people | Categorical, paired | McNemar's test |
| Compare means of two independent groups | Numerical, normal | Independent t-test |
| Compare two independent groups, skewed data | Numerical, non-normal | Mann–Whitney U test |
| Before–after in the same people | Numerical, normal | Paired t-test |
| Before–after, skewed data | Numerical, non-normal | Wilcoxon signed-rank test |
| Compare three or more groups | Numerical, normal | One-way ANOVA with post-hoc test |
| Three or more groups, skewed data | Numerical, non-normal | Kruskal–Wallis test |
| Relationship between two numerical variables | Numerical | Pearson (normal) or Spearman correlation |
| Predict a numerical outcome from several factors | Numerical | Multiple linear regression |
| Predict a yes/no outcome from several factors | Binary | Binary logistic regression |
| Accuracy of a diagnostic test | Binary vs reference | Sensitivity, specificity, predictive values, ROC curve |
| Agreement between two observers or methods | Categorical / numerical | Cohen's kappa / Bland–Altman analysis or ICC |
| Time to an event | Time-to-event | Kaplan–Meier with log-rank test; Cox regression |
The chi-square test is usually considered unreliable when more than 20% of expected cell counts are below 5. In that case, especially in 2×2 tables, Fisher's exact test is preferred.
Worked examples
- Is the proportion of wound infection different between two dressing types? Categorical outcome, two independent groups: chi-square test.
- Does mean systolic blood pressure fall after 8 weeks of yoga in the same participants? Numerical, paired, roughly normal: paired t-test.
- Which factors predict hypothyroidism in pregnancy? Binary outcome, several predictors: logistic regression, reported with adjusted odds ratios.
Common mistakes
- Choosing the test after looking at which one gives p < 0.05.
- Using an independent test for paired data.
- Running many separate tests instead of one test for three or more groups.
- Forgetting post-hoc tests after a significant ANOVA or Kruskal–Wallis test.
- Reporting correlation as if it proved causation.
Key takeaways
- Decide by outcome type, number and independence of groups, then distribution.
- Plan the test for each objective in your synopsis, before you see the data.
- Report effect sizes and confidence intervals, not just p-values.
References
- Kirkwood BR, Sterne JAC. Essential Medical Statistics. 2nd ed. Oxford: Blackwell Science; 2003.
- Altman DG. Practical Statistics for Medical Research. London: Chapman & Hall; 1991.
- Bland JM, Altman DG. Multiple significance tests: the Bonferroni method. BMJ. 1995;310(6973):170.
- Lang TA, Altman DG. Basic statistical reporting for articles published in biomedical journals: the "Statistical Analyses and Methods in the Published Literature" or SAMPL guidelines. In: Smart P, Maisonneuve H, Polderman A, eds. Science Editors' Handbook. European Association of Science Editors; 2013.
This article is for education. Always follow the requirements of your university, guide, ethics committee and target journal.