Updating Beliefs with Evidence
A 48-year-old man: daytime sleepiness, loud snoring, morning headaches.
BLESS cohort (Pakhare, Joshi A, AIIMS Bhopal, 2024):
. . .
You administer the SONA questionnaire (Joshi A, Goyal A, Pakhare A, 2025):
What is the probability he has OSA now?
This is Bayesian reasoning — updating a prior belief with new evidence.
\[P(\text{Disease} \mid \text{Test}^+) = \frac{P(\text{Test}^+ \mid \text{Disease}) \times P(\text{Disease})}{P(\text{Test}^+)}\]
In words: Posterior = (Likelihood × Prior) / Evidence
| Component | Meaning | OSA Example |
|---|---|---|
| Prior | Belief before test | Prevalence = 30.5% |
| Likelihood | P(data | hypothesis) | Sensitivity = 73% |
| Evidence | P(data overall) | P(SONA+) = 37.5% |
| Posterior | Updated belief | 59.3% |
The same positive test gives different answers depending on the prior. In a low-prevalence population (5%), the post-test probability would be only 14.9%.
The posterior is ALWAYS a compromise. More data → posterior follows data. Less data → prior dominates.
| Frequentist | Bayesian | |
|---|---|---|
| Probability | Long-run frequency | Degree of belief |
| Parameters | Fixed but unknown | Random variables |
| Prior info | Not formally used | Explicitly incorporated |
| Result | p-value, CI | Posterior, credible interval |
| Interpretation | “If repeated infinitely…” | “Given the data…” |
The p-value problem: p = 0.04 does NOT mean “4% chance the null is true.”
A 95% CI does NOT mean “95% probability the true value is in here.”
A 95% credible interval DOES mean exactly that.
Uninformative = know nothing | Informative = prior studies | Strongly informative = large prior data
\[BF_{10} = \frac{P(\text{Data} \mid H_1)}{P(\text{Data} \mid H_0)}\]
| BF₁₀ | Interpretation |
|---|---|
| < 1 | Evidence favours null (H₀) |
| 1–3 | Anecdotal evidence for H₁ |
| 3–10 | Moderate evidence for H₁ |
| 10–30 | Strong evidence |
| > 100 | Extreme evidence |
Key advantage over p-values: A p-value can never support the null. A Bayes factor can.
BF₁₀ = 0.15 means BF₀₁ = 6.7 → moderate evidence for no treatment effect.
The Pralidoxime Story (Eddleston, Joshi R et al., PLoS Med 2009):
Being Bayesian means being willing to update — even when the data contradict your beliefs.
Region of Practical Equivalence — is the effect big enough to matter?
A new rapid diagnostic test for TB has sensitivity 92% and specificity 88%. In a tribal district of MP, TB prevalence is 500 per 100,000 (0.5%). A patient with chronic cough tests positive.
What is the post-test probability of active TB?
Bayesian calculation: P(TB | Test+) = (0.92 × 0.005) / (0.92 × 0.005 + 0.12 × 0.995) = 3.7%
Even with a positive test, the probability is only ~3.7% because the prior is extremely low. The prior dominates when prevalence is very low. This is why mass screening for rare diseases generates many false positives.
An RCT of a new antihypertensive (n = 40) finds mean SBP reduction of 6 mmHg (95% CI: −1 to 13, p = 0.09). A frequentist concludes “not significant.” A Bayesian analysis with an informative prior (from a related drug class showing 5 mmHg benefit) gives a posterior mean of 5.8 mmHg, 95% CrI [1.2, 10.4].
The Bayesian analysis borrows strength from prior knowledge and concludes there is a >97.5% probability of some benefit. The frequentist analysis, ignoring prior information, cannot reject the null. Both are valid approaches — but the Bayesian one incorporates more information and gives a more clinically useful answer for a small trial.
A pharmaceutical company tests whether a generic drug is equivalent to the branded version. The RCT (n = 500) yields p = 0.42 for the difference. The company claims “no difference — the drugs are equivalent.”
A p-value cannot prove equivalence — it only fails to reject a difference. A Bayesian analysis with BF₁₀ = 0.08 (BF₀₁ = 12.5) would provide strong evidence for equivalence. The Bayes factor directly quantifies evidence for the null hypothesis, which is exactly what’s needed here.
Warning
Biostatistics for Clinicians | Module 14