How to calculate the Odds Ratio (OR): from a 2×2 contingency table with cases (diseased) and exposed, OR = (a·d)/(b·c) where a=exposed cases, b=exposed controls, c=unexposed cases, d=unexposed controls. OR measures how many times exposure is more likely among cases versus controls. OR=1 no association, OR>1 risk factor, OR<1 protective.

Formula and key parameters

The Odds Ratio is the statistic of choice in case-control studies and logistic regression. From the 2×2 table: among cases (diseased), exposure odds are a/c. Among controls (healthy), exposure odds are b/d. OR = (a/c)/(b/d) = (a·d)/(b·c). The 95% CI uses Woolf's method: ln(OR) ± 1.96·√(1/a + 1/b + 1/c + 1/d), then exponentiate. When any count is zero, apply Haldane-Anscombe correction (+0.5 to all counts).

History and epidemiological context

The term "odds ratio" was introduced by Sir Ronald Fisher and Frank Yates in the 1930s. Cornfield (1951) showed OR from a case-control study approximates Relative Risk when disease is rare (<10%). Mantel and Haenszel (1959) introduced stratification formulas. Edwards (1963) proposed the logit of OR for regression. Woolf (1955) formulated the CI based on ln(OR) standard deviation. Today OR is the foundation of meta-analysis (Cochrane Collaboration), logistic regression, and all registrational epidemiological studies (FDA Guidance, EMA Reflection Papers, STROBE Statement).

2×2 table and related formulas

• Odds Ratio: OR = (a·d)/(b·c)
• Risk Ratio (RR): RR = [a/(a+c)] / [b/(b+d)] valid only for cohort and cross-sectional
• Risk Difference (RD): RD = a/(a+c) − b/(b+d)
• NNT (Number Needed to Treat): 1/|RD|; NNH if harm
• Sensitivity: a/(a+c) true positive rate
• Specificity: d/(b+d) true negative rate
• PPV: a/(a+b) positive predictive value
• NPV: d/(c+d) negative predictive value
• Yates chi-square: (|ad-bc|-N/2)²·N/[(a+b)(c+d)(a+c)(b+d)]

OR interpretation

OR = 1: no association. OR = 2: exposure doubles odds of disease. OR = 0.5: exposure halves odds. Conventional thresholds: OR 3 strong risk. 95% CI must exclude 1 for statistical significance: [0.8-2.5] not significant, [1.2-3.5] significant. Confounding requires stratified analysis (Mantel-Haenszel) or multivariate logistic regression.

Worked numerical examples

Example 1 — smoking and lung cancer (case-control). 200 cases and 200 controls; among cases 150 smokers (a=150, c=50); among controls 80 smokers (b=80, d=120). OR = (150·120)/(80·50) = 4.5. ln(OR) = 1.504. SE = 0.235. 95% CI = exp(1.504 ± 1.96·0.235) = [2.84; 7.12]. Statistically significant (excludes 1). Smokers have 4.5× higher lung cancer odds than non-smokers in this study.

Example 2 — vaccine and infection (cohort). 1000 vaccinated: a=20 infected, c=980 healthy. 1000 unvaccinated: b=100 infected, d=900 healthy. OR = (20·900)/(100·980) = 0.184. RR = 0.20. 95% CI = [0.11; 0.30]. Strong protective factor, significant. Vaccine efficacy = 1 - RR = 80%. NNT = 1/0.08 = 12.5 → 13.

Example 3 — drug and adverse effect (RCT). Treatment: a=15 events of 500 (3%). Placebo: b=10 events of 500 (2%). OR = 1.52. 95% CI = [0.67; 3.42]. Includes 1 → not significant. Underpowered trial for small adverse effect increase.

Example 4 — screening (sens/spec). Diagnostic test: 95 TP (a), 5 FP (b), 10 FN (c), 90 TN (d). Sensitivity = 0.905. Specificity = 0.947. PPV = 0.95. NPV = 0.90. OR = 171. Highly accurate test.

Example 5 — diet and pathology with zero count. a=0, b=5, c=10, d=15. Haldane correction: a=0.5, b=5.5, c=10.5, d=15.5. OR = 0.134. Without correction calculation diverges. Wide CI for small numbers.

Limits and assumptions

1. OR ≠ RR: for rare diseases (<10%) OR ≈ RR; for common diseases OR overstates the association.
2. Case-control studies do NOT allow direct RR (no population denominator).
3. Confounding: variables associated with both exposure and outcome bias OR. Use stratification or logistic regression.
4. Selection and recall bias typical in case-control.
5. Zero counts need Haldane correction (+0.5) or Fisher's exact test.
6. Multi-category tables (>2×2) use multinomial or ordinal logistic regression.
7. Matched data use McNemar test, not simple OR.
8. Effect modification: OR may vary across strata; assess with Breslow-Day.

Practical applications

1. Case-control epidemiology: cancers, rare diseases, risk factors.
2. RCT with binary outcome: efficacy, safety endpoints.
3. Meta-analysis: Cochrane Collaboration, forest plots.
4. Diagnostic tests: sensitivity/specificity, ROC analysis.
5. Genetics: GWAS, polymorphism-disease associations.
6. Occupational studies: workplace exposures (NIOSH, OSHA).
7. Marketing: conversion analytics, binary A/B testing.
8. Public health: intervention efficacy (CDC, WHO).
9. Veterinary medicine: animal epidemiology.
10. Psychology: binary questionnaires, symptom presence.

Common errors

1. Interpreting OR as RR for common diseases (overestimation).
2. Computing OR on case-control expecting RR.
3. Ignoring confounding and necessary stratification.
4. Not applying Haldane correction for zeros.
5. Reporting OR without confidence interval.
6. Confusing statistical significance (CI excludes 1) with clinical relevance.
7. Computing OR on matched data without McNemar.
8. Ignoring effect modification across strata.

Support

Questions: info@calcoi.com. Educational tool for OR calculation. For professional epidemiological analysis consult an epidemiologist or biostatistician per STROBE, Cochrane Handbook, FDA Guidance.

Updated July 2026