AI summary by TwinMind
The Tea Lady Experiment - Numberphile
Numberphile
17m 6s
Classical Hypothesis Testing Framework
- Standard setup: null hypothesis (H0) vs alternative hypothesis (H1)
- Null typically "boring" scenario, alternative more "exciting"
- Example: testing average height of 4-year-olds
- Data samples x1 to xn from normal distribution with mean μ, variance 1
- H0: μ ≤ 36 inches, H1: μ > 36 inches
- Test statistic T(x) = sample average
- Asymptotic properties for large n allow use of normal distribution tables
- Reject H0 if T(x) > threshold c
- Don't reject H0 if T(x) ≤ c
Network Disease Spread Research Project
- Theoretical modeling of infectious disease spread on networks (pre-COVID)
- Problem: limited to single observation, not multiple samples
- Example: 5 individuals in classroom, observe who gets infected
- Can't use standard asymptotic properties with small sample size
- H0: completely connected graph (symmetric relationships)
- H1: specific network structure with interesting patterns
- Challenge: need statistical rigor for choosing between hypotheses
Fischer's Lady Tasting Tea Experiment
- Historical foundation of permutation testing
- Setup: lady claims ability to distinguish milk-first vs tea-first preparation
- H0: no special powers (random guessing), H1: special abilities
- Experimental design: 8 cups total (4 milk-first, 4 tea-first)
- Lady identifies which cups she thinks are milk-first vs tea-first
- Possible correct scores: 0, 2, 4, 6, or 8 (due to constraint of 4 each type)
Permutation Test Methodology
- Total possible arrangements: 8 choose 4 = 70 combinations
- Calculation: (8×7×6×5)/(4×3×2×1) = 70
- Distribution of correct scores under random guessing:
- 8 correct: 1/70 probability
- 6 correct: 16/70 probability
- 4 correct: 36/70 probability
- 2 correct: 16/70 probability
- 0 correct: 1/70 probability
- Observed result: 6 correct cups
- P-value calculation: P(T ≥ 6) = (16 + 1)/70 = 17/70 ≈ 24%
Statistical Significance Thresholds
- Typical significance levels: 5% or 2.5%
- For rare disease detection: may require 2.5% for high accuracy
- Fischer's experiment conclusion: 24% not sufficiently low to reject H0
- Only 8/8 correct (1.4% probability) would provide strong evidence
- More cups would create finer-grained distribution, enabling closer approach to 5% threshold
Application to Network Disease Modeling
- Test statistic: number of edges between infected nodes in alternative hypothesis graph
- Example: if 3 of 5 nodes infected, count connections between infected nodes
- Create reference distribution by considering all possible infection patterns
- Compare observed statistic to permutation-generated distribution
- Discrete distribution may not be monotonic but provides framework for hypothesis testing
