Biostatistics is the backbone of evidence-based medicine. As a medical student, you don’t need to become a mathematician, but you must know how to choose the right statistical test and correctly interpret the results. This article walks you through the practical steps, from understanding your data to reading p‑values, with real clinical examples you will see in your studies and future practice.
“Statistics is the grammar of science.” — Karl Pearson. Knowing the right test is like choosing the correct verb tense – it makes your clinical findings clear and trustworthy.
The table below summarises the most frequently used statistical tests in medical research. Use it as a quick reference when you read journal articles or plan your own project.
| Test Name | When to Use | Example | How to Interpret |
|---|---|---|---|
| Independent t‑test | Compare means of two unrelated groups (continuous outcome) | Mean LDL cholesterol in statin vs. placebo group | If p < 0.05, the difference is statistically significant; report mean difference and 95% CI |
| Paired t‑test | Compare means of the same subjects at two time points | Blood pressure before and after a diuretic | Significant p‑value suggests a real change over time |
| One‑way ANOVA | Compare means of three or more independent groups | HbA1c across three different diabetes medications | Significant F‑test indicates at least one group differs; post‑hoc tests show which pairs |
| Chi‑square test | Test association between two categorical variables | Smoking status (yes/no) vs. lung cancer diagnosis (yes/no) | p < 0.05 means there is a significant association; check odds ratio for strength |
| Pearson correlation | Measure linear relationship between two continuous variables | BMI vs. fasting glucose | r close to 1 or -1 indicates strong correlation; p‑value says if it’s real |
| Linear regression | Predict a continuous outcome from one or more predictors | Predict weight from height, age, and sex | Look at the coefficient (slope) and its p‑value; R² explains variance |
| Logistic regression | Predict a binary outcome (yes/no) from predictors | Predict 30‑day mortality based on age, creatinine, and infection | Odds ratio with 95% CI; if CI does not include 1, predictor is significant |
“Absence of evidence is not evidence of absence.” — Carl Sagan. Medical students must resist the temptation to declare a treatment ineffective just because p > 0.05.
Imagine a randomized controlled trial comparing a new antibiotic (Group A) to standard therapy (Group B) for hospital‑acquired pneumonia. The primary outcome is clinical cure rate (yes/no) at day 14. The secondary outcome is length of hospital stay (continuous). The researchers collect data from 120 patients per group.
Choosing the correct statistical test and interpreting results accurately are core skills for every medical student. Focus on your data type, the structure of your study, and the clinical question. Use the table and examples above as a starting point. Always pair statistical significance with effect sizes and confidence intervals. With practice, you will read research critically and design better studies yourself.
Parametric tests (e.g., t‑test, ANOVA) assume data follow a known distribution (usually normal) and require continuous data with homogeneity of variances. Nonparametric tests (e.g., Mann‑Whitney U, Kruskal‑Wallis) make fewer assumptions and can be used on ordinal data or when normality is violated. For small samples or skewed data, nonparametric tests are safer.
Use a two‑tailed test unless you have a strong theoretical reason to expect an effect only in one direction (e.g., a new drug can only be better, not worse). In medical research two‑tailed tests are standard because they protect against unexpected adverse effects.
An odds ratio (OR) greater than 1 means the exposure increases the odds of the outcome; OR less than 1 means it decreases the odds. The confidence interval tells you precision. For example, OR = 2.5 (95% CI 1.8–3.5) means the exposed group has 2.5 times the odds of the outcome, and the result is significant because the CI does not include 1.
It means there is a 6% probability of seeing the observed difference (or more extreme) if the null hypothesis were true. This is not statistically significant at the conventional α = 0.05 level. However, it does not mean “no effect.” It could be a small sample size or a clinically important trend. Always evaluate effect size and clinical context.
Pearson correlation measures linear relationships and assumes continuous, normally distributed data. Spearman correlation is based on ranks and works for monotonic relationships (not necessarily linear) and for ordinal data or non‑normal continuous data. If you see a curve in your scatter plot, use Spearman.
The Bonferroni correction divides the α level by the number of tests (e.g., α = 0.05/10 tests = 0.005). It is very conservative. Alternative methods like Holm‑Bonferroni or the false discovery rate provide more power. The choice depends on whether you want to avoid any false positives (Bonferroni) or tolerate a small proportion of false discoveries (FDR).
No. When expected frequencies are small, the chi‑square approximation is unreliable. Use Fisher’s exact test instead, which calculates the exact p‑value regardless of sample size. Many statistical software automatically do this for 2×2 tables.
A clinically significant result is one that makes a real difference to patient care – not just a statistically significant p‑value. For example, a drug that lowers systolic BP by 2 mmHg might be statistically significant with a large sample, but 2 mmHg is rarely clinically meaningful. Always ask: would this change my management?
R² tells you the proportion of variance in the outcome that is explained by the predictors. For example, R² = 0.45 means 45% of the variation in blood pressure is explained by age, weight, and diet. The remaining 55% is due to other factors and random variation.
Intention‑to‑treat (ITT) includes all participants as originally allocated, regardless of adherence or dropout. This preserves randomization and gives a pragmatic estimate of effect. Per‑protocol analysis includes only those who completed the treatment as assigned, which can introduce bias. ITT is preferred for primary analysis in superiority trials.
