Choosing the right statistical test — a practical guide for clinicians

Picking a statistical test feels complicated, but most thesis analyses come down to three questions:

1. What type is my outcome variable? Continuous (age, HbA1c), categorical (cured / not cured), or time-to-event (time to relapse).
2. How many groups am I comparing? Two, or three or more.
3. Are the observations independent or paired? Different patients in each group, or the same patients measured more than once (before and after, left eye vs right eye).

Answer those, and the table below points you to the test.

The decision table

Question	Normally distributed data	Skewed data / ordinal scores
Two independent groups, continuous outcome	Independent (unpaired) t-test	Mann–Whitney U test
Same subjects measured twice	Paired t-test	Wilcoxon signed-rank test
Three or more independent groups	One-way ANOVA (+ Tukey post hoc)	Kruskal–Wallis (+ Dunn's post hoc)
Same subjects at three or more time points	Repeated-measures ANOVA	Friedman test
Relationship between two continuous variables	Pearson correlation	Spearman correlation

For categorical outcomes:

• Two independent groups or more → Chi-square test. Use Fisher's exact test when expected counts are small (more than 20% of cells with an expected count below 5).
• Paired categorical data (e.g. test positive before and after treatment in the same patients) → McNemar's test.

For adjusting for confounders, move from tests to models:

• Continuous outcome → linear regression
• Binary outcome → logistic regression (reports adjusted odds ratios)
• Counts or rates → Poisson or negative binomial regression
• Time-to-event → Kaplan–Meier curves, log-rank test and Cox regression (reports hazard ratios)

Checking normality

"Normally distributed" decides between the two columns above. Check it with:

• Histograms and Q–Q plots — the most informative option.
• Shapiro–Wilk test — suitable for small to moderate samples. With large samples it flags trivial departures, so read it alongside the plots.
• Mean vs median — if they're far apart, the data are probably skewed.

Report normally distributed variables as mean ± SD, and skewed ones as median (IQR). Mixing these up is one of the most common comments on results chapters.

Beyond the p-value

Examiners and journals increasingly expect:

• Effect sizes with 95% confidence intervals — the mean difference, odds ratio or hazard ratio, not only "p < 0.05".
• Clinical relevance — a statistically significant 1 mmHg difference in BP rarely matters clinically.
• Correction for multiple comparisons when you run many tests — use proper post hoc tests after ANOVA rather than repeated t-tests.
• A statistical methods paragraph naming the software and version, the tests used for each objective, and the significance level.

Common mistakes

• Running a t-test on Likert-scale scores — ordinal data usually need non-parametric tests.
• Using an unpaired test on paired data (before/after in the same patients) — this throws away power.
• Applying chi-square with tiny expected counts instead of Fisher's exact test.
• Testing many outcomes and reporting only the significant ones.
• Treating correlation as causation.

Plan it before you collect data

The best time to choose your tests is while writing your synopsis: every objective should map to a specific analysis. That makes the results chapter much easier to write — and to defend at your viva.