4. Randomized complete block designs
In industrial experimentation, the assumption that all experimental units are homogeneous rarely holds. A pilot reactor may drift in temperature across a shift, a filling line may exhibit position-dependent variation, or raw material may arrive in distinct lots. When such systematic sources of variation exist and are not accounted for, they inflate the error term and mask real treatment effects. The randomized complete block design (RCBD) is the classical remedy: it partitions experimental units into homogeneous groups called blocks, and within each block every treatment appears exactly once, in random order.
Why Blocking Works: The Variance Decomposition Argument
The RCBD model for a single observation is:
y_ij = μ + τ_i + β_j + ε_ij
where:
μ = overall mean
τ_i = effect of treatment i (i = 1..t)
β_j = effect of block j (j = 1..b)
ε_ij = random error, assumed N(0, σ²)
The key structural feature is that the block effect β_j is removed from the residual. In a completely randomized design (CRD), that block-to-block variation would remain inside the error mean square, enlarging the denominator of the F-test. By isolating it, the RCBD reduces the error variance and increases the power to detect treatment differences — provided the blocking variable genuinely influences the response.
The ANOVA table for a balanced RCBD with t treatments and b blocks has (t−1), (b−1), and (t−1)(b−1) degrees of freedom for treatments, blocks, and error respectively. The treatment F-statistic is MStreatment / MSerror, tested against Fα, t−1, (t−1)(b−1).
Choosing a Blocking Strategy in Industrial Settings
A blocking variable must satisfy two conditions: it must be known or measurable before the experiment, and it must plausibly affect the response without interacting with treatments. Common industrial blocking factors include:
- Shift or operator: day vs. night shift, or three operators on a coating line.
- Batch or lot of raw material: incoming polymer lots with differing molecular weight distribution.
- Position or spatial location: furnace zone, oven shelf, or wafer position on a carrier.
- Time or run order: consecutive days when ambient humidity drifts.
- Equipment: multiple parallel machines producing the same part.
The practical rule is to block on the largest known nuisance factor that cannot be eliminated by randomization alone. If two nuisance factors are present and both matter, a Latin square design may be preferable; if blocks are too small to hold all treatments, an incomplete block design is required.
Worked Example: Adhesive Cure Strength
An engineer compares four adhesive formulations (A, B, C, D) on a bonding line. Five production batches of substrate are available, and batch-to-batch variation is suspected. Each batch is large enough to apply all four formulations, so a batch is a natural block. Within each batch, the four formulations are applied in random order. Bond strength (N) is measured after curing.
| Block (Batch) | A | B | C | D | Block mean |
|---|---|---|---|---|---|
| 1 | 42.1 | 45.3 | 40.8 | 44.0 | 43.05 |
| 2 | 41.5 | 44.8 | 39.9 | 43.6 | 42.45 |
| 3 | 43.0 | 46.1 | 41.5 | 44.9 | 43.88 |
| 4 | 42.6 | 45.7 | 41.0 | 44.4 | 43.43 |
| 5 | 41.9 | 45.0 | 40.3 | 43.8 | 42.75 |
| Treatment mean | 42.22 | 45.38 | 40.70 | 44.14 | 43.11 |
Computing sums of squares (illustrative values): SStreatment ≈ 62.4, SSblock ≈ 5.1, SStotal ≈ 71.8, so SSerror ≈ 4.3. With df = 3, 4, 12, MStreatment ≈ 20.8 and MSerror ≈ 0.358, giving F ≈ 58.1, far exceeding the critical value. The block effect is modest but real, and removing it sharpens the treatment comparison.
Efficiency Comparison: RCBD vs. CRD
The relative efficiency (RE) of an RCBD compared with a CRD using the same number of observations is estimated by:
RE = [(df_e(RCBD) + 1)(df_e(CRD) + 3)] / [(df_e(CRD) + 1)(df_e(RCBD) + 3)] × (MS_error(CRD) / MS_error(RCBD))
In practice, a simpler working estimate is RE ≈ MSerror(CRD) / MSerror(RCBD), where MSerror(CRD) is reconstructed as [SSerror(RCBD) + SSblock] / [dferror(RCBD) + dfblock]. If RE = 1.8, the RCBD achieves the same precision as a CRD with 80% more replicates. If RE is near 1.0, the blocking variable was ineffective and the lost error degrees of freedom may actually hurt.
Handling Nuisance Variables: Decision Logic
Not every nuisance variable should become a block. The decision depends on whether it is known, measurable, and free of treatment interaction. The diagram below summarizes the practical selection logic and the resulting design choice.
Practical Execution and Verification
Running an RCBD in a production environment requires discipline in three areas. First, block formation: assign units to blocks before any treatment is applied, using the nuisance variable as the grouping criterion. Second, randomization within blocks: generate an independent random permutation for each block, not one global order. Third, verification: after fitting the model, inspect residual plots for block-by-treatment interaction. If the interaction is significant, the additive model is violated and a factorial or split-plot approach is warranted.
A quick diagnostic is to plot treatment means against block means. Parallel profiles suggest additivity; crossing profiles indicate interaction. In the adhesive example, all four formulations rank consistently across batches, confirming the additive assumption.