Simpson's Paradox, Stratification, and Causal Inference in Statistics
UC Berkeley Admissions Case and Simpson's Paradox
Historical Context and Lawsuit:
In the 1970s, a high-profile lawsuit was brought against the University of California, Berkeley alleging gender bias in graduate school admissions.
Aggregate university data revealed that male applicants were accepted at a significantly higher overall rate than female applicants.
The university administration had not set out to create discriminatory policies; admissions decisions were executed independently by departmental chairs.
Departmental Analysis and Resolution:
Berkeley hired statistician Betty Scott to perform a comprehensive subgroup analysis on the admissions data.
When admissions rates were evaluated separately across individual departments, nearly every department accepted female applicants at a equal or higher fractional rate than male applicants.
The apparent bias in the aggregate data vanished and reversed when conditioned on individual departments.
Definition of Simpson's Paradox:
First identified mathematically by Edward H. Simpson in 1951 during subgroup analyses.
Simpson's Paradox occurs when statistical trends, associations, or directional relationships observed in aggregate data completely reverse, disappear, or invert when the data is split into underlying subgroups or conditioned on a confounding variable.
Mathematical Framework of Simpson's Paradox
Departmental Acceptance Rates and Aggregation:
Let and represent the number of women applying to Department A and Department B, respectively.
Let and represent the number of women accepted to Department A and Department B, where w_A \begin{matrix} \frac{w_A + w_B}{W_A + W_B} \frac{w_A}{W_A} \frac{w_B}{W_B} \frac{m_A + m_B}{M_A + M_B} \frac{m_A}{M_A} \frac{m_B}{M_B} 10,000 21 2,100 100,000 10 10,000 110,000 12,100 \frac{12,100}{110,000} \times 100 \text{ per cent} \rightarrow 11 \text{ per cent} 2,000 20 400 100 10 10 2,100 410 \frac{410}{2,100} \times 100 \text{ per cent} \rightarrow 19.5 \text{ per cent} 80 \text{ per cent} 10 \text{ per cent} 100 50 50 50 \times 0.80 = 40 50 \times 0.10 = 5 45 100 \frac{45}{100} = 0.45 \rightarrow 45 \text{ per cent} 200 100 100 100 \times 0.80 = 80 100 \times 0.10 = 10 90 200 \frac{90}{200} = 0.45 \rightarrow 45 \text{ per cent} 100 90 10 90 \times 0.80 + 10 \times 0.10 = 72 + 1 = 73 100 \frac{73}{100} = 0.73 \rightarrow 73 \text{ per cent} 200 20 180 20 \times 0.80 + 180 \times 0.10 = 16 + 18 = 34 200 \frac{34}{200} = 0.17 \rightarrow 17 \text{ per cent} 10 80 \text{ per cent} 5 20 \text{ per cent} 10 \times 0.80 = 8 5 \times 0.20 = 1 9 15 \frac{9}{15} = \frac{3}{5} = 0.60 \rightarrow 60 \text{ per cent} 5 80 \text{ per cent} 10 20 \text{ per cent} 5 \times 0.80 = 4 10 \times 0.20 = 2 6 15 \frac{6}{15} = \frac{2}{5} = 0.40 \rightarrow 40 \text{ per cent} 0.80 0.80 0.20 0.20 \text{Weighted Average} = \frac{10 \times 0.80 + 5 \times 0.20}{15} = \frac{8 + 1}{15} = 0.60 \text{Weighted Average} = \frac{10}{15} \times 0.80 + \frac{5}{15} \times 0.20 = 0.60 \frac{10}{15} P(\text{Apply } A \text{ } M) \frac{5}{15} P(\text{Apply } B \text{ } M) P(\text{Accepted} \text{ } \text{Man}) = P(A \text{ } \text{Man}) \times P(\text{Accepted} \text{ } \text{Man}, A) + P(B \text{ } \text{Man}) \times P(\text{Accepted} \text{ } \text{Man}, B) p (1-p) M p \times M (1-p) \times M p \times W (1-p) \times W \text{Rate}_{\text{male, CF}} = \frac{p \times M \times P(\text{Accepted} \text{ } \text{Male}, A) + (1-p) \times M \times P(\text{Accepted} \text{ } \text{Male}, B)}{M} \text{Rate}_{\text{male, CF}} = p \times P(\text{Accepted} \text{ } \text{Male}, A) + (1-p) \times P(\text{Accepted} \text{ } \text{Male}, B) \text{Rate}_{\text{female, CF}} = p \times P(\text{Accepted} \text{ } \text{Female}, A) + (1-p) \times P(\text{Accepted} \text{ } \text{Female}, B) P(\text{Accepted} \text{ } \text{Male}, A) = P(\text{Accepted} \text{ } \text{Female}, A) P(\text{Accepted} \text{ } \text{Male}, B) = P(\text{Accepted} \text{ } \text{Female}, B) p \text{Rate}_{\text{male, CF}} = \text{Rate}_{\text{female, CF}} 80 \text{ per cent} \rightarrow 100 \text{ per cent} 60 \text{ per cent} \rightarrow 80 \text{ per cent} 20 \text{ per cent} 80 \text{ per cent} 95 \text{ per cent} 60 \text{ per cent} 90 \text{ per cent} 10 \text{ per cent} 85 \text{ per cent} 40 \text{ per cent} \text{Proficiency}_{\text{School A}} = 0.20 \times 0.95 + 0.80 \times 0.60 = 0.19 + 0.48 = 0.67 \rightarrow 67 \text{ per cent} \text{Proficiency}_{\text{School B}} = 0.90 \times 0.85 + 0.10 \times 0.40 = 0.765 + 0.040 = 0.805 \rightarrow 80.5 \text{ per cent} 95 \text{ per cent} > 85 \text{ per cent} 60 \text{ per cent} > 40 \text{ per cent} 80.5 \text{ per cent} 67 \text{ per cent} 1,000 200 800 500 50 \text{ per cent} 75 \text{ per cent} 200 150 50 \text{ per cent} 800 400 75 \text{ per cent} 50 \text{ per cent} 400 150 \begin{matrix}\text{end}\{cases}\right.