AI summary by TwinMind
Odds Ratios and Log(Odds Ratios), Clearly Explained!!!
StatQuest with Josh Starmer
16m 20s
Introduction and Odds Recap
- Josh Stormer introduces StatQuest on odds ratios and log odds ratios
- Previous episode ended on cliffhanger about odds vs odds ratios
- Odds = ratio of something happening to not happening
- Example: team winning odds = blue circles/red circles = 0.5
Odds Ratio Definition and Calculation
- Odds ratio = ratio of odds (not just odds themselves)
- Example calculation: (2/4)/(3/1) = 0.17
- Range properties:
- Denominator > numerator: odds ratio 0 to 1
- Numerator > denominator: odds ratio 1 to infinity
- Log odds ratio makes values symmetrical
- Example: odds ratio 2/4 over 3/1 → log odds ratio = -1.79
- Opposite ratio → log odds ratio = +1.79
Cancer and Gene Mutation Example
- Study population: 356 people total
- 29 with cancer, 327 without cancer
- 140 with mutated gene, 216 without mutated gene
- Odds calculations:
- With mutated gene: cancer odds = 23:117
- Without mutated gene: cancer odds = 6:210
- Final odds ratio = 6.88 (odds 6.88x greater with mutated gene)
- Log odds ratio = 1.93
- Interpretation: larger values = better predictor, smaller values = poor predictor
Statistical Significance Testing Methods
- Three methods available (no consensus on best):
- Fisher's exact test
- Chi-square test
- Wald test
- Common approaches:
- Fisher's/chi-square for p-value + Wald for confidence interval
- Wald for both p-value and confidence interval (ensures consistency)
Fisher's Exact Test
- References previous StatQuest on enrichment analysis
- Treats people as bag of M&Ms:
- 29 red M&Ms (cancer), 327 blue M&Ms (no cancer)
- Computer calculation: p-value = 0.00001
Chi-Square Test
- Compares observed vs expected values assuming no gene-cancer relationship
- Cancer probability calculation: 29/356 = 0.08
- Expected values with mutated gene:
- Cancer: 11.2 people
- No cancer: 128.8 people
- Expected values without mutated gene:
- Cancer: 17.3 people
- No cancer: 198.7 people
- P-values:
- With continuity correction: 0.000001
- Without continuity correction: 0.0000004
Wald Test
- Used for logistic regression odds ratios and confidence intervals
- Based on normal distribution of log odds ratios
- Simulation method:
- 10,000 random samples (300-400 people each)
- Cancer probability: 0.08, mutated gene probability: 0.39
- Creates histogram centered on zero
- Standard deviation calculations:
- From simulation: 0.43
- From observed values: √(1/23 + 1/117 + 1/6 + 1/210) = 0.47
- Test statistic: 1.93/0.47 = 4.11 standard deviations from mean
- Rule: >2 standard deviations → p-value <0.05
- Precise two-sided p-value = 0.00005
Statistical Test Comparison
- Validation using 10,000 random samples (should get 5% false positives):
- Fisher's exact: 4% false positives
- Chi-square with continuity correction: 3% false positives
- Chi-square without continuity correction: 5% false positives
- Wald test: 5% false positives
- All methods performed well
- Recommendation: use field-standard method, consider multiple tests for borderline p-values
Summary and Conclusions
- Odds ratio = ratio of odds
- Log odds ratio = log of odds ratio
- Both indicate relationship strength between variables
- Statistical significance determined by Fisher's exact, chi-square, or Wald test
- Field-specific preferences for test selection
