Bayes’ Theorem in Practice

When Mathematics Meets Real Decisions

A test for a rare disease comes back positive. A spam filter flags an incoming email. A product manager wants to know which version of a website is better. In all three cases, the question is the same: given some evidence, how should I update my belief?

Bayes’ theorem is the formal answer. In the previous lesson we derived it mathematically. Now we will use it — working through medical testing, spam filtering, and A/B experiments to see why Bayesian reasoning is both powerful and frequently violated by human intuition.

Medical Testing: The Vocabulary

Medical tests are characterized by two fundamental quantities. Sensitivity is the probability that the test is positive given the patient truly has the disease — it measures how well the test catches true cases. Specificity is the probability that the test is negative given the patient is truly healthy — it measures how well the test avoids false alarms.

Notice what these quantities do not tell you. Sensitivity tells you P(positive | disease). But what a doctor and patient really want to know is the reverse: P(disease | positive). This is exactly where Bayes’ theorem comes in.

Positive Predictive Value and Negative Predictive Value

The quantities that answer clinical questions directly are the Positive Predictive Value (PPV) — the probability you have the disease given a positive test — and the Negative Predictive Value (NPV) — the probability you are healthy given a negative test. These are the posterior probabilities that Bayes’ theorem computes.

The formula makes something immediately clear: PPV depends on prevalence P(D) — how common the disease is in the population being tested. A test with excellent sensitivity and specificity can still have a low PPV if the disease is rare, because most positive results come from the large pool of healthy people, not from the small pool of sick ones.

The Base Rate Fallacy

The base rate fallacy is the error of ignoring prevalence when interpreting test results. It is extraordinarily common, even among trained professionals. Studies have found that a majority of physicians significantly overestimate the probability of disease after a positive test, because they focus on the test’s accuracy and forget how rare the disease is.

Intuition tends to anchor on the headline number — “the test is 99% accurate” — and ignore the context — “but only 0.1% of people have the disease.” Bayes’ theorem forces us to weight the accuracy by the base rate, often with counterintuitive results.

The result surprises almost everyone. Even with a highly accurate test, roughly 91 of 100 positive results come from healthy people, because there are 999 healthy people for every 1 sick person. The false positives from that large healthy pool swamp the true positives from the tiny sick pool.

Bayesian Updating: Prior → Evidence → Posterior

One of the most important properties of Bayes’ theorem is that it is sequential. Today’s posterior becomes tomorrow’s prior. You do not need to wait for all evidence before updating — you can incorporate each new piece of information one at a time, always using the most recent posterior as the starting point.

Suppose our patient from the example above undergoes a second, independent test that also comes back positive. We now take the posterior from the first test (PPV ≈ 9%) as the new prior, and apply Bayes’ theorem again with the second test’s likelihood. The posterior rises dramatically — because two independent positive tests are much harder to explain as coincidental false positives.

This sequential view is the heart of the Bayesian worldview: reasoning is an ongoing process of accumulating evidence and updating beliefs, never a one-shot calculation.

Spam Filtering with Naïve Bayes

Email spam filters were among the first large-scale deployments of Bayesian reasoning in software. The core idea: given the words in an email, what is the probability it is spam? We apply Bayes’ theorem treating each word as independent evidence — an approximation called the naïve Bayes assumption.

Let S = “email is spam” and let w₁, w₂, …, w_n be the words in the email. With the naïve independence assumption, the posterior probability of spam is proportional to the prior probability of spam multiplied by the product of per-word likelihoods P(w_i | S).

Naïve Bayes is a generative classifier: it models how each class generates data. Despite the strong independence assumption, it is often competitive with much more complex models, especially when training data is limited. It is fast to train, interpretable, and naturally handles missing features.

Bayesian A/B Testing

In a classical A/B test, you run an experiment, compute a p-value, and decide whether to reject the null hypothesis. The Bayesian alternative starts with a prior distribution over each variant’s conversion rate and updates it with observed data to produce a posterior distribution. You can then directly ask: what is the probability that variant B is better than variant A?

The mechanism is the same prior → likelihood → posterior update from the medical-testing example, now applied to a continuous conversion rate θ rather than a yes/no diagnosis. You begin with a prior belief about θ — a whole distribution over its plausible values, not a single guess — and every visitor who does or does not convert is one more observation that sharpens it. As conversions accumulate, the posterior concentrates around the true rate. The specific distributions that make this update exact and closed-form are introduced in Module 4; here the point is only that Bayesian A/B testing is the same sequential update, run over and over.

Why Bayes Matters for Machine Learning

Bayesian ideas permeate modern machine learning. Maximum a posteriori (MAP) estimation is Bayes’ theorem applied to parameter estimation: instead of maximizing the likelihood alone (maximum likelihood estimation), MAP maximizes the product of likelihood and prior, which is equivalent to regularization. L2 regularization (ridge) corresponds to a Gaussian prior; L1 regularization (lasso) corresponds to a Laplace prior.

Bayesian neural networks maintain a full posterior distribution over weights rather than a point estimate, enabling principled uncertainty quantification. Gaussian processes are Bayesian models over functions, used in hyperparameter optimization and spatial statistics. Variational inference and MCMC methods make these computations tractable.

Putting It All Together

Bayes’ theorem is not just a formula — it is a framework for rational belief revision. In medical testing, it shows that a positive result from a rare-disease screen is almost certainly a false positive unless the prior probability of disease is substantial. In spam filtering, it powers classifiers that adapt to individual users’ email patterns. In A/B testing, it provides direct answers to business questions rather than indirect p-value thresholds. And in machine learning, it connects regularization, uncertainty quantification, and sequential learning under a single coherent umbrella.

The base rate fallacy — ignoring prevalence — is perhaps the most important lesson of this module. Whenever you see a diagnostic test result, a model confidence score, or a striking experimental finding, ask: what was the prior? The answer will often change everything.