Simpson's Paradox and the Back-Door Criterion: Reversals When Branch Data Are Pooled and De-Confounding with Causal Graphs: chapter answers
Every branch shows plan A converting better, yet the pooled table shows plan B ahead: should the owner believe the strata or the total? The two popular answers, that the larger pooled sample is more reliable and that finer is truer, both assume the answer is hidden in the data. The central claim of this book is that the causal graph, not the data, decides which table to believe. Before writing, the book checked the original sources and corrects the blueprint: Simpson's note of 1951 was not the first, since Yule discussed in 1903 how mixing universes creates association; Bickel and colleagues…
DOI pending · monograph page goes live after publication · 5DL-B-302 · 中文版
- Chapter 1 · Every Branch Wins, the Total Loses: the Reversal Condition for 2 x 2 x k Tables
Chapter 1 recomputes the reversal for Berkeley's six largest departments: pooled admission rates are 0.445 for men and 0.304 for women, while standardised to the applicant mix they are 0.387 and 0.430; in an illustrativ… - Chapter 2 · How Common Are Reversals? The Smallest Example, Random Tables and Sampling Noise
Chapter 2 finds the smallest reversal, which needs nine customers, and a seeded simulation gives a reversal frequency of 0.01648 for random tables, in line with the published figure of about one in sixty. - Chapter 3 · The Back-Door Adjustment Formula: Standardisation, Inverse Probability Weighting and a Structural Model with a Known Answer
Of all customers, the flagship holds 41/112, the community branch 5/16 and the outlet 9/28. With this mix, the conversion rate if everyone received plan A is 223/448, or 0.4978; if everyone received plan B it is 1003/22… - Chapter 4 · One Table, Three Causal Graphs: Confounder, Mediator and Collider
Chapter 4 pairs one table with confounder, mediator and collider graphs, whose correct answers are +0.0500 and -0.2493. - Chapter 5 · Enumerating Back-Door Sets: d-Separation, M-Bias and Counting Causal Graphs
Chapter 5 enumerates back-door sets and shows that adjusting for a pre-treatment collider creates a bias equal to 66.0 % of the effect. - Chapter 6 · Sampling Error in Stratified Estimates: Intervals, Power and Sparse Strata
Detecting an effect of 0.05 with 80 % power needs 393.5 customers per plan per branch, rounded up to 394, or 2364 customers across three branches and two plans. The table shows how power grows with sample size: 0.292 at… - Chapter 7 · Classics East and West: the Lüshi Chunqiu on Examining Hearsay, and Yule 1903
Chapter 7 examines the Lüshi Chunqiu line that many words seem wrong but are right, and many seem right but are wrong, together with Yule's paper of 1903. - Chapter 8 · Forced Stratification and Adjustment of Multi-Branch Reports: the Closed Loop, the Tool and a Pre-Registered Test
Chapter 8 sets out the closed loop of forced stratification and adjustment for multi-branch reports, the distilled Stratified Reversal Detector and a pre-registered randomised holdout: when the branch is the only confou…
Illustrative model calculation, not a live quote