Randomised Complete Block Design (RCBD)

Description

Group units into homogeneous blocks, run every treatment once in each block with fresh randomisation, and remove block-to-block variation from the error.

Blocking is Fisher's principle of local control: compare treatments within sets of units that are alike, so nuisance variation between sets does not blur the comparison. An RCBD has b blocks of t homogeneous units, each treatment appearing once per block and randomised independently within each block. Blocks may be non-specific (days, batches, recruitment waves) or intrinsic (sites, devices, suppliers). Analysis is a two-way ANOVA of blocks plus treatments, but blocks are not F-tested because they were not randomised. Relative efficiency against a completely randomised design is estimated from the block and error mean squares; Hinkelmann and Kempthorne suggest blocking is worth repeating only if it exceeds about 125 percent. Missing values can be estimated with Yates' formula at the cost of one error df each. The generalised randomised block design runs each treatment more than once per block, allowing a valid test of block-by-treatment interaction.

Attributes

Loss type
Siloed
Stage
Evaluation
Effort
Weeks

Steps

  1. Identify a nuisance source of variation (device, day, site, recruitment wave) and form blocks of t units as alike as possible, or rt units for a generalised design.
  2. Form blocks before randomising, then randomise treatments to units separately within each block.
  3. Analyse with a two-way ANOVA of blocks plus treatments and do not F-test the blocks.
  4. Check for nonadditivity with Tukey's one-df test and transform the response if needed.
  5. Estimate relative efficiency against a completely randomised design from MS(blocks) and MS(error); keep blocking in future studies if it exceeds about 125 percent.
  6. If values are missing, estimate them with Yates' formula or covariance and reduce the error df by one for each.
  7. If block-by-treatment interaction is expected, use a generalised design with each treatment repeated within each block.

Sources

  • Design and Analysis of Experiments, Vol. 1 (Hinkelmann & Kempthorne, 2008)
  • Statistical Design and Analysis of Experiments (Mason et al., 2003)

Also known as

  • RCBD
  • Randomized complete block design
  • Blocking
  • Generalised randomised block design
  • GRBD
  • Local control

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference