Randomized Experiments Exam Workshop

The defining information in a randomized experiment is the assignment mechanism. Keep potential outcomes fixed and ask what would change under another allowed assignment. That perspective separates causal estimands, estimators, and randomization-based uncertainty.

Worked problem 1: potential outcomes and assignment imbalance

Consider six units with the following fixed potential outcomes:

Unit $Y_i(0)$ $Y_i(1)$ $\tau_i$ Assignment $W_i$
1 10 14 4 1
2 12 15 3 0
3 9 11 2 1
4 15 18 3 0
5 11 13 2 0
6 13 17 4 1

The finite-sample average treatment effect is

\[\tau_S=\frac{4+3+2+3+2+4}{6}=3.\]

The observed outcomes are selected by

\[Y_i^{obs}=W_iY_i(1)+(1-W_i)Y_i(0).\]

For the displayed assignment,

\[\bar Y_1=\frac{14+11+17}{3}=14, \qquad \bar Y_0=\frac{12+15+11}{3}=12.67,\]

so

\[\hat\tau=1.33.\]

The realized estimate differs from the finite-sample ATE because the assignment happened to place units with relatively high $Y(0)$ in control. This is randomization error, not evidence that random assignment failed. Across all allowed complete randomizations, the difference in means is unbiased for $\tau_S$.

Worked problem 2: Neyman repeated-sampling inference

A completely randomized trial assigns 50 units to each arm. The observed mean difference is $4.8$, with $s_1^2=100$ and $s_0^2=144$.

The conservative variance estimator is

\[\widehat{\operatorname{Var}}(\hat\tau)=\frac{100}{50}+\frac{144}{50}=4.88,\]

and

\[SE(\hat\tau)=\sqrt{4.88}=2.21.\]

The large-sample statistic is $4.8/2.21=2.17$, and a 95% interval is

\[4.8\pm1.96(2.21)=(0.47,9.13).\]

The exact finite-population variance also contains $-S_\tau^2/N$. Because each unit reveals only one potential outcome, $S_\tau^2$ is not identified. Omitting the nonnegative term makes the reported variance conservative.

Worked problem 3: Fisher’s tea experiment

In Fisher’s classic design, a participant receives eight cups: four prepared milk-first and four tea-first. She is told that exactly four belong to each group and identifies four as milk-first.

Under the sharp null, the labels have no effect on her judgments. There are

\[\binom{8}{4}=70\]

equally likely sets of four cups she could select. Only one selection identifies all four milk-first cups, so the one-sided randomization probability of a perfect result is

\[p=\frac{1}{70}=0.0143.\]

This is an exact statement about the randomization mechanism under a sharp null for every cup. It is not a large-sample $t$ approximation and it is not a test only of a zero average effect.

Worked problem 4: blocked estimator

A trial blocks participants by site. Site A contains 40% of the sample and has an estimated effect of 8; site B contains 60% and has an estimated effect of 2. The overall blocked estimate is

\[\hat\tau=0.4(8)+0.6(2)=4.4.\]

An unweighted average, $(8+2)/2=5$, answers a different question because it gives the smaller and larger sites equal influence. Analysis must use the design weights unless the estimand explicitly weights blocks equally.

Exercises

  1. For $N=10$ and $N_1=4$, calculate the number of allowed complete randomizations and the probability of each. Compare this with Bernoulli assignment using $p=0.4$.
  2. Using the six-unit potential-outcome table, enumerate or program all $\binom63$ assignments. Verify that the mean of $\hat\tau$ equals 3 and inspect the randomization distribution.
  3. A trial has $N_1=80$, $N_0=120$, $\hat\tau=-2.4$, $s_1=7$, and $s_0=9$. Compute the conservative standard error, test a zero average-effect null, and construct a 95% interval.
  4. In 12 matched pairs, the treated-minus-control differences have mean 1.8 and standard deviation 2.4. Calculate the paired standard error and test statistic. Explain why analyzing 24 units as independent discards the design.
  5. Explain how Fisher’s sharp null and Neyman’s zero-average-effect null can lead to different questions and potentially different conclusions.
  6. A blocked experiment has stratum weights $(0.2,0.3,0.5)$ and effects $(5,-1,4)$. Compute the overall effect, then explain why an unweighted average is inappropriate.

Solution checks

  1. $\binom{10}{4}=210$, so each allowed vector has probability $1/210$. Bernoulli assignment permits every vector and does not fix the treated count.
  2. The randomization-distribution mean is 3; individual realized estimates vary around it.
  3. $SE=\sqrt{49/80+81/120}\approx1.135$, $z\approx-2.11$, and the interval is approximately $(-4.62,-0.18)$.
  4. $SE=2.4/\sqrt{12}=0.693$ and $t=2.60$ with 11 degrees of freedom.
  5. Fisher imputes every missing outcome under no effect for any unit; Neyman concerns an average and allows heterogeneous unit effects.
  6. $0.2(5)+0.3(-1)+0.5(4)=2.7$.

See