Likelihood Ratio Reasoning¶

Why It Matters¶

Naive reasoning treats evidence as binary: "this symptom is present, therefore this disease." But many findings are non-specific - they appear in multiple conditions with different frequencies. Likelihood ratios correct the systematic error of treating presence as diagnostic when frequency is what matters. A finding that appears in 70% of disease A cases and 65% of disease B cases provides almost no diagnostic value, even though it's "associated with" both diseases.

The Key Move¶

For each piece of evidence, explicitly estimate: "If this hypothesis were true, what's the probability of seeing this evidence?" Then compare across hypotheses. A finding with high probability under one hypothesis and low probability under alternatives shifts your belief toward that hypothesis proportionally. Conversely, evidence that's equally probable under all hypotheses provides no discrimination. The power isn't in the evidence itself but in the probability ratio.

Classic Application¶

Chest pain diagnosis. "Chest pain radiating to the jaw" is taught as a classic heart attack symptom. But likelihood ratio analysis reveals: this finding appears in ~30% of actual heart attacks and ~5% of non-cardiac chest pain. The likelihood ratio is 30/5 = 6. Compare to "chest pain worse with deep breathing" (rare in heart attack, common in musculoskeletal pain or pulmonary embolism): likelihood ratio for heart attack is ~0.2. The second finding is more informative because of the larger ratio, even though it's a negative indicator.

Surprising Application¶

Feature selection in machine learning. You're predicting customer churn and considering features. "User logged in this week" appears in 80% of retained users and 75% of churned users - likelihood ratio near 1, low predictive value despite high frequency. "Downgraded subscription tier" appears in 40% of churned users and 5% of retained users - likelihood ratio of 8, highly predictive. The diagnostic reasoning structure transfers directly: you're not looking for correlated features, you're looking for discriminating ratios.

Failure Modes¶

Over-precision in ratio estimation. Treating rough clinical estimates as exact numbers creates false confidence. The tool fails when base rates are extremely different - even large likelihood ratios may not shift probability much if you started from very different priors. Also fails when evidence is not independent - combining likelihood ratios assumes findings don't interact, but many symptoms cluster. The worst misuse is calculating ratios post-hoc to justify a pre-existing belief rather than using them to discover discriminating evidence.

Go Deeper¶

McGee, S. (2012). "Evidence-Based Physical Diagnosis," 3^rd ed., which provides likelihood ratios for hundreds of physical findings; Fagan, T.J. (1975). "Nomogram for Bayes' Theorem," New England Journal of Medicine, 293(5), 257, showing how likelihood ratios update probabilities.