Block 1: Estimands, estimators and unbiasedness

Week 11 · Uncertainty I · Imai §7.1.1 · about 4 min

NoteHow this works

Run each cell with Run Code, look at the output, then answer the question at the end of the block. Every cell is editable; Start Over resets it. Errors are information: read them.

A simulated experiment where we know the truth. Potential outcomes: \(Y_i(0) \sim N(0, 1)\) and \(Y_i(1) \sim N(1, 1)\), so the true population average treatment effect (PATE) is 1. The sample of n = 100 people is created in the background; the SATE is the mean of \(Y_i(1) - Y_i(0)\) in this one sample.

1. One sample

2. Repeat the randomized experiment in the same sample

Assign half to treatment at random, 5,000 times, and compute the difference in means each time.

Conclusion: What do the SATE and the estimation errors show?
The SATE depends on which 100 people we drew (1.22 here; the PATE is exactly 1). The mean error is -0.001: random assignment makes the difference in means unbiased for the SATE. Individual errors still range from about -0.4 to +0.6.