The Clinician’s Statistical Toolkit
The Deadly Assumption (1980s)
The CAST Result (1989)
Surrogate Endpoint Fallacy
Suppressing a risk marker is not the same as preventing the disease. The marker (PVCs) was treated instead of the patient.
| Case | What happened | Module |
|---|---|---|
| HRT | Observational studies said it prevented heart disease. RCT (WHI) showed it caused it. | 6, 14 |
| Hydroxychloroquine | Preliminary data drove global prescribing. RCTs showed no benefit. | 8 |
| Vioxx | Cardiovascular risk hidden in subgroup analyses. Withdrawn after ~88,000 excess MI. | 14 |
As a clinician, you will read research every week. This course gives you the tools to read it critically.
Three practical functions for clinicians:
When you look at lab reports over time or compare ward outcomes — that’s descriptive statistics.
A BP of 142/90 in one reading doesn’t confirm HTN. A p = 0.04 doesn’t prove a drug works. Statistics handles uncertainty.
Understanding study design, bias, and analysis lets you separate reliable evidence from noise.
Figure 1
Stevens’ Four Levels of Measurement
| Scale | Order? | Equal intervals? | True zero? | Example |
|---|---|---|---|---|
| Nominal | No | No | No | Blood group, Sex |
| Ordinal | Yes | No | No | NYHA class, GCS |
| Interval | Yes | Yes | No | Temperature (°C) |
| Ratio | Yes | Yes | Yes | Height, Weight, BP |
The Ratio Test
Can you meaningfully say “twice as much”? If yes → Ratio. If no → Interval or lower.
Figure 2
| Scale | Central Tendency | Spread | Tests |
|---|---|---|---|
| Nominal | Mode | Frequency | Chi-square, Fisher |
| Ordinal | Median | IQR | Mann-Whitney, Spearman |
| Interval | Mean, Median | SD, IQR | t-test, ANOVA, Pearson |
| Ratio | Mean, Geo. mean | SD, CV | All parametric tests |
Common Mistake
Computing the mean of NYHA class or Likert scale data. These are ordinal — use the median and non-parametric tests.
Figure 3
Figure 4
Figure 5
Confounders
A confounder is associated with both exposure and outcome, distorting the apparent relationship. Coffee → Lung cancer? No — smoking is the confounder.
Statistics is not optional — flawed reasoning has killed patients (CAST, HRT, HCQ)
NOIR classification determines which statistics and tests are valid
Interval ≠ Ratio — the “true zero” test (Temperature °C vs Height cm)
Descriptive → Inferential — from “what I see” to “what is true”
Independent → Dependent — beware of confounders
Videos:
Books:
Biostatistics for Clinicians | Module 1