Negative Evidence Weighting¶

Why It Matters¶

Human attention focuses on present features. We catalog symptoms, signs, and test abnormalities. But the absence of expected findings powerfully discriminates hypotheses. If a disease usually produces finding X, and X is absent, that's evidence against the disease. Negative evidence weighting corrects the systematic error of attention bias toward present features.

The Key Move¶

For each hypothesis, identify findings that should be present if the hypothesis were true. Then explicitly check for them. If absent, calculate the likelihood ratio for absence - often this is the reciprocal of the sensitivity. A finding present in 90% of cases, when absent, provides a likelihood ratio of ~0.1, which dramatically lowers probability. Make negative findings as explicit as positive findings in your reasoning.

Classic Application¶

Ruling out appendicitis. A finding strongly associated with appendicitis is right lower quadrant rebound tenderness (sensitivity ~80%). A patient with abdominal pain has no rebound tenderness. Naive reasoning ignores this (focuses on pain that is present). Negative evidence weighting: absence of rebound has LR ~ 0.25, reducing appendicitis probability substantially. Negative evidence actively rules out the diagnosis, not just fails to rule it in.

Surprising Application¶

Security threat assessment. Intelligence suggests a sophisticated cyber adversary. If sophisticated, you'd expect anti-forensics (log deletion, timestamp manipulation, lateral movement through multiple hops). These are absent. Negative evidence: absence of anti-forensics has LR ~ 0.1 for sophisticated actor. This actively suggests less sophisticated threat (opportunistic rather than targeted), changing defensive priorities. The absence of expected indicators is diagnostic.

Failure Modes¶

Confusing "absence of evidence" with "evidence of absence." Negative evidence only counts when you've actively looked for a finding and it's genuinely absent. If you didn't check, absence is uninformative. The tool fails when expected findings have low sensitivity anyway - their absence doesn't discriminate. Also fails when observers miss findings that are actually present (false negatives from poor examination). The worst failure is double-counting - using both presence of A and absence of B when they're the same information.

Go Deeper¶

Sox, H.C. et al. (2013). "Medical Decision Making," 2^nd ed., particularly on likelihood ratios for negative findings; Gill, C.J. et al. (2005). "Diagnostic Accuracy of Clinical Features and Laboratory Tests for Community-Acquired Pneumonia," Academic Emergency Medicine, 12(10), 894-899.