Regression-Discontinuity Designs at Membership-Tier Thresholds: Local Causal Identification on Either Side of a Spend-Based Upgrade Line: chapter answers
Comparing customers who reached the VIP spending line with those who did not is the commonest way membership reports measure "the effect of the benefits", and the one most prone to overstatement: customers who can reach the line already visit more and spend more freely. This book argues that only a comparison of customers who fall just on either side of the line identifies the causal effect of the benefits, which is the regression-discontinuity design introduced by Thistlethwaite and Campbell in the Journal of Educational Psychology in 1960 and set out as a practical guide by Imbens and Lemie…
DOI pending · monograph page goes live after publication · 5DL-B-303 · 中文版
- Chapter 1 · Either Side of the VIP Line: Selection Bias, the Naive Gap and a Local Comparison at the Threshold
Chapter 1 finds a naive gap of 19.005 on illustrative membership data against a true effect at the line of 2.5769, an overstatement of 7.37 times, while the local linear estimate with a bandwidth of 8 is 2.764. - Chapter 2 · Local Linear Regression: Triangular-Kernel Weighted Least Squares, Boundary Bias and an Effect That Holds Only at the Threshold
At the line of 60 the discontinuity estimate is 2.764 and the true value at the line is 2.5769. Yet the average benefit over all VIPs is 3.063, and over all customers (if everyone had the benefits) it is 2.001; all thre… - Chapter 3 · Choosing the Bandwidth: Trading Bias for Variance, a Monte Carlo Optimum and Boundary Cross-Validation
At a bandwidth of 2 only 1801 customers are used; the estimate is 2.723 with standard error 0.479, and the interval from 1.784 to 3.662 is wide. From 4 to 16 the estimate moves between 2.676 and 2.954, every interval co… - Chapter 4 · The McCrary Density Test: Local Lines on Either Side of a Histogram, Power and Blind Spots
Chapter 4 builds a McCrary-style density test by hand, whose z reaches 24.14 on data with top-ups, and shows the swapping kind of sorting it cannot see, where the predetermined covariate test gives z of 13.14. - Chapter 5 · Topping Up to Upgrade: How Precisely Manipulable Spending Breaks the Design
Chapter 5 shows the bias reaching 1.709 as top-ups grow and examines imprecise control and the donut. - Chapter 6 · Three Robustness Checks: High-Order Polynomials, Placebo Thresholds and the Fuzzy Design
If the benefits really act at the line of 60, then at other points where no rule changes, say 48 or 72, next year's spending should not jump. Program p62 places fake thresholds at 36, 42, 48 and 54 below the line and at… - Chapter 7 · Sources East and West: Rank for Count in the Book of Lord Shang and the Scholarship Discontinuity of 1960
Under rule A there are 666 units at a count of 24 and 675 at 25, a ratio of 1.014; the distribution is smooth. Under rule B the number at 25 jumps to 2656 while only 65 remain at 24, a ratio of 40.862; every count from… - Chapter 8 · Closing the Loop: Evaluating a Membership-Tier Threshold, the Distilled Tool and a Pre-Registered Plan
Chapter 8 gives a complete procedure for evaluating a membership-tier threshold, the distilled Threshold Discontinuity Effect Estimator and a pre-registered test. The book contains 23 runnable Python programs.
Illustrative model calculation, not a live quote