Blocked and Paired Randomized Experiments

Blocking groups similar units before treatment is assigned. Randomization then occurs within each block, ensuring treatment-control comparisons within important baseline groups.

Stratified or blocked estimator

For stratum $k$, define

\[\hat\tau_k=\bar Y_{1k}-\bar Y_{0k}.\]

With $N_k$ units in stratum $k$ and $N=\sum_{k=1}^K N_k$, the sample-average treatment-effect estimator is

\[\boxed{ \hat\tau= \sum_{k=1}^K\frac{N_k}{N}\hat\tau_k }.\]

The weights follow the target population. Equal weights would target an average stratum effect and can answer a different question when strata have different sizes.

Paired randomized experiments

A matched-pair design is blocking with two units per block and one treated unit in each pair. For pair $j$, define the treated-minus-control difference

\[D_j=Y_{j,T}-Y_{j,C}.\]

For $J$ pairs,

\[\boxed{ \hat\tau=\frac{1}{J}\sum_{j=1}^J D_j }.\]

The sample variance of pair differences is

\[s_D^2=\frac{1}{J-1}\sum_{j=1}^J(D_j-\bar D)^2,\]

and the standard error is

\[SE(\hat\tau)=\frac{s_D}{\sqrt J}.\]

Why blocking can improve precision

When outcomes are similar within blocks, within-block comparisons remove predictable baseline variation. Poorly chosen blocks need not improve precision, but valid within-block randomization still supports unbiased estimation when analyzed according to the design.

Analysis must follow design

  • Weight block effects according to the estimand.
  • Preserve pair labels in randomization tests and standard errors.
  • Do not analyze a paired experiment as if all treated and control observations were independently assigned across the entire sample.

Paired observations versus matched pairs

A paired $t$-test can analyze before-after measurements, naturally paired units, or randomized matched pairs. The algebra is similar, but the causal interpretation depends on how the pairs and treatment assignments were created.

See