Interactive demo · Probability and Statistical Inference · Chapter 16

Potential outcomes: when does the difference between treated and untreated measure the effect of the treatment?

Each subject has two potential outcomes, C0 (untreated) and C1 (treated), and we observe only Y = CX. With binary outcomes there are four types of subject: survivors (1, 1), responders (0, 1), anti-responders (1, 0) and the doomed (0, 0). The causal effect θ = E(C1) − E(C0) is the share of responders minus the share of anti-responders. The association α = E(Y | X = 1) − E(Y | X = 0) compares two different groups of people. Panel (a) shows who ends up in each group, panel (b) moves from random assignment (λ = 0) to the chosen self-selection (λ = 1), and panel (c) shows what the difference in sample means estimates.

Share of each type (normalised)

survivors (1, 1)

responders (0, 1)

anti-responders (1, 0)

doomed (0, 0)

P(take the treatment | type)

survivors

responders

anti-responders

doomed

Self-selection λ (0 = random assignment)

Subjects per study n

Simulated studies

Try these

With λ between 0 and 1 each type is treated with probability (1 − λ)·P(X = 1) + λ·(its slider), which keeps the share of treated subjects fixed. Each simulated study draws n subjects (type, then treatment) with the seeded generator; studies in which an arm is empty are left out of the histogram.

In the lecture notes