How to Choose Statistical Test in Clinical Research: T-test, Mann-Whitney U / Ranksum, ANOVA, Kruskal-Wallis, Paired t-test, Wilcoxon Signed-Rank, Chi-square, Fisher’s Exact, Log-rank, Cox regression
Clinical Epidemiology ResearchUniqcret doctor knowledgesData Analytics or StatisticsSystematic Reviews & Meta-AnalysesMethodology and Research Design
On this page
Step 1 Identify the Dependent Variable Y
| Y (Outcome) Example | Data Type | Typical Scale / Notes |
|---|---|---|
| Blood-loss volume (mL), Hb level (g/dL) | Continuous | Numeric, theoretically infinite decimals |
| Gender (male / female) | Binary | 2 ordered or unordered categories |
| Mortality (yes / no) | Binary | — |
| Pain score (1–10) | Ordinal | Ranked but distances unequal |
| Satisfaction (low / med / high) | Ordinal | — |
| Survival time (days) | Time-to-Event | Event plus censoring |
| Blood group (A / B / AB / O) | Categorical > 2 (Nominal) | Unordered, >2 |
| Seizure count per month | Count | Non-negative integers, often skewed |
| Proportion of wound infections (%) | Proportion / Rate | Bounded 0–1; often with different denominators |
Step 2 Define the Independent Groups X
- How many groups? (2 vs > 2)
- Structure:
- Independent groups (different patients / subjects)
- Paired / repeated (same subject measured twice or more)
(If X itself is continuous—e.g., dosage in mg—see “When to use regression” below.)
Step 3 If Y Is Continuous → Check Normality
stata: hist postopbleeding, normal swilk postopbleeding- p > 0.05 → “looks normal” → parametric tests
- p < 0.05 → not normal → non-parametric tests (or transform / use GLM)
Step 4 Master Test-Selection Table
| Dependent Y Type | # Groups | Structure | If Normal / Large Sample | If Not Normal / Small n / Ordinal |
| Continuous | 2 | Independent | Student's t-test | Mann-Whitney U / Wilcoxon rank-sum |
| >2 | Independent | One-way ANOVA | Kruskal-Wallis | |
| 2 | Paired | Paired t-test | Wilcoxon signed-rank | |
| >2 | Repeated | Repeated-measures ANOVA | Friedman test | |
| Binary | 2 or >2 | Independent | χ2 test of independence | Fisher's exact test (if any expected cell < 5) |
| 2 | Paired | McNemar's test | Exact McNemar test (mid-p) | |
| >2 time points | Repeated | GEE for binary / Logistic mixed model / Multivariable Logistic Regression | Cochran's Q test (if available) | |
| Categorical (>2, Nominal) | 2 or >2 | Independent | χ2 test (RxC format) | Fisher's exact (if sparse) |
| 2 | Paired | Stuart-Maxwell or Bowker test | Exact symmetry test | |
| >2 | Repeated | Multinomial logistic / GEE | Repeated measures log-linear model | |
| Ordinal | 2 | Independent | — | Mann-Whitney U / Wilcoxon rank-sum |
| >2 | Independent | — | Kruskal-Wallis | |
| 2 | Paired | — | Wilcoxon signed-rank | |
| >2 | Repeated | — | Friedman test | |
| Count | 2 or >2 | Independent | Poisson regression / NB model | Negative binomial (if overdispersion) |
| Repeated | GEE or mixed model | — | ||
| Proportion / Rate | 2 | Independent | z-test for two proportions | Fisher exact |
| >2 | Independent | χ2 test or z-test for k proportions | Fisher exact (if sparse) | |
| Repeated | Poisson GEE / Rate ratio modeling | — | ||
| Time-to-Event | 2 or >2 | Independent | Log-rank test / Cox proportional hazards | Breslow / Tarone-Ware / Fleming-Harrington (if PH violated) |
| Repeated Events | Shared frailty / PWP / AG models | Flexible parametric or RMST-based models |
Step 5 When to Use Regression Instead of a Simple Test
| If you need … | Regression of choice |
|---|---|
| Adjust for ≥1 covariate (age, sex, baseline Hb, …) | Linear (continuous Y), logistic (binary Y), multinomial logistic (nominal Y > 2), ordinal logistic, Poisson/negative-binomial (counts), Cox (time-to-event) |
| Predict risk / odds / mean rather than test difference | Same as above |
| Model continuous X (e.g., dose in mg) | Include X as continuous term or use spline |
| Handle interaction terms (e.g., Tx × sex) | Any generalized linear model (GLM) |
Step 6 Example Walk-Throughs
- Post-op bleeding (mL) across 3 TXA dose groups
- Y: Continuous
- 2 independent groups
- Normality p < 0.05 → Not normal
- → Kruskal-Wallis
- Infection type (A/B/C) by TXA vs no TXA
- Y: Categorical (>2)
- 2 independent groups
- → χ² test (2×3 table)
- Seizure counts pre- vs post-drug in same patients
- Y: Count, paired
- Small n, skewed
- → Wilcoxon signed-rank or Poisson GLMM if covariates needed
- 30-day survival across 4 surgical centers
- Y: Time-to-event
- 2 independent groups
- → Log-rank test; follow-up Cox model for adjustment
Quick Recap
- Start with Y: Identify its scale.
- Count & classify X groups: 2 vs > 2, independent vs paired.
- If Y is continuous: Check normality.
- Plug into the table to pick the test.
- Need adjustment or continuous X? → Move to regression.
0