A randomized block design is an experimental design in which experimental units are grouped into blocks on the basis of characteristics expected to affect the response, and treatments are assigned by randomization within each block. Comparisons are therefore made among units that are relatively similar, reducing the influence of known sources of background variation. In the common randomized complete block design (RCBD), every block contains every treatment, usually once. More generally, randomized block designs can be incomplete, with only a subset of treatments represented in each block.
Purpose and basic principles
Experiments often involve units that differ before any treatment is applied. Agricultural plots may differ in soil fertility, production runs in operating conditions, or participants in baseline characteristics. Such differences can obscure treatment effects.
Blocking separates this background variation from the variation used to assess treatment differences. A useful blocking variable predicts the response but is not itself the principal subject of the experiment. Blocks are constructed to make units within each block relatively homogeneous; substantial differences between blocks are acceptable and can make blocking particularly effective.
Blocking and randomization have distinct roles:
- Blocking accounts for specified sources of variation through the arrangement of the experiment.
- Randomization within blocks prevents treatment assignment from being systematically determined by other unit characteristics and supplies a basis for statistical inference.
Blocking does not replace randomization. Assigning treatments in a fixed order within each block can associate treatment with position, time, or another unrecognized factor.
A block is also different from a treatment group: a block generally contains units receiving different treatments. Its purpose is to support treatment comparisons under similar conditions.
Randomized complete block design
Suppose an experiment compares treatments using blocks. In the standard complete design, each block contains experimental units, and each treatment is assigned to exactly one unit in that block. The experiment therefore contains observations.
A typical procedure is:
- Identify the experimental unit—the smallest unit independently assigned a treatment.
- Choose a blocking characteristic before treatment assignment.
- Form blocks, each containing suitable units.
- Randomly permute the treatment assignments within each block, ordinarily independently across blocks.
- Measure the response and analyze treatment differences while accounting for blocks.
Each treatment has replications, one in each block. Repeated measurements on the same experimental unit do not automatically constitute additional independent treatment replications.
Example
An experiment compares four fertilizers on a field with a fertility gradient. The field is divided into six compact blocks, each chosen so that its four plots have similar underlying fertility. Within each block, fertilizers A, B, C, and D are randomly assigned to the four plots.
One possible allocation is:
| Block | Plot 1 | Plot 2 | Plot 3 | Plot 4 |
|---|---|---|---|---|
| 1 | C | A | D | B |
| 2 | B | D | A | C |
| 3 | A | C | B | D |
| 4 | D | B | C | A |
| 5 | B | A | C | D |
| 6 | C | D | B | A |
The repeated letters across blocks represent replication of treatments, not a fixed spatial pattern. The actual allocation must result from the specified randomization procedure.
Fertilizer comparisons are based on differences within blocks. An unusually fertile block does not favor a particular fertilizer merely because all its plots have higher yields: every fertilizer is represented there.
Statistical model
A standard additive model is
where:
- is the response to treatment in block ;
- is the overall mean;
- is the effect of treatment ;
- is the effect of block ;
- is the residual error.
For a fixed-effects formulation, constraints such as
make the parameters identifiable.
The classical normal-theory analysis assumes residual errors with statistical independence, zero mean, common variance , and a normal distribution. Normality is needed for the usual exact finite-sample tests and intervals, not for defining the design itself.
Blocks may instead be modeled as random effects when they represent a population of possible blocks. This changes the interpretation of block variation and the scope of inference, although standard treatment comparisons in a balanced complete design often retain the same algebraic form.
Treatment estimation
Let denote the mean for treatment across blocks. In the balanced complete design, the estimated difference between treatments and is
This is also the average within-block difference:
Under the additive model, the block effect cancels from each difference. If is the residual mean square, the estimated standard error is
A confidence interval can then be constructed using Student’s -distribution with the appropriate residual degrees of freedom. Simultaneous comparisons among many treatments require attention to multiple testing.
Analysis of variance
For a complete, balanced design with one observation per treatment–block combination, analysis of variance partitions total variation into treatment, block, and residual components.
Writing for the mean of block and for the overall mean,
and
The associated degrees of freedom are:
| Source | Degrees of freedom |
|---|---|
| Treatments | |
| Blocks | |
| Residual | |
| Total |
To test the null hypothesis that all treatment effects are equal, the classical statistic is
Under the additive independent normal-error model and the null hypothesis, this has an -distribution with and degrees of freedom.
The same model can be fitted by ordinary least squares, using indicator variables for treatments and blocks. This formulation is particularly useful when extending the analysis to additional predictors or unequal numbers of observations.
Randomization-based inference
Inference can also be based directly on the assignment mechanism rather than on normally distributed errors. Under the sharp null hypothesis that treatment has no effect on any experimental unit, a randomization test compares the observed statistic with its values under the treatment allocations allowed by the design.
The allowed reallocations occur within blocks. Permuting treatments freely across all units would generally represent a different experiment. Randomization tests for a sharp null and model-based tests of average treatment differences do not necessarily test identical hypotheses.
Why blocking can improve precision
Blocking is effective when the blocking characteristic explains appreciable response variation. By accounting for between-block differences, it can reduce residual variation and increase statistical power for treatment comparisons.
This improvement is not automatic. Fitting block effects consumes degrees of freedom. If units do not differ meaningfully between blocks, a blocked analysis may offer little benefit and can have less precision than a suitable completely randomized design.
The choice of blocks is therefore made using substantive knowledge or pretreatment measurements, not by selecting a grouping after observing treatment outcomes. When a useful predictor is continuous, a regression-based adjustment may sometimes retain more information than dividing it into categories.
Blocking alone does not guarantee that treatment effects generalize to other populations or settings. Random assignment supports causal inference for the experiment under suitable assumptions; random sampling addresses a different question about population representation.
Related designs
Matched pairs
With two treatments and two units per block, a complete block design becomes a matched-pairs design. Units are paired using relevant pretreatment characteristics, and treatment is randomized within each pair. Under the usual model, analyzing the within-pair differences is equivalent to the treatment test from the blocked analysis.
Incomplete blocks
A complete block may become impractical when the number of treatments is too large to fit into a reasonably homogeneous group. Incomplete block designs place only some treatments in each block. Balanced incomplete block designs arrange for every pair of treatments to occur together equally often.
Unlike the balanced complete case, treatment and block effects are generally not orthogonal in incomplete designs. Comparisons must account for the allocation structure, and their precision depends on how treatments are connected through shared blocks.
Factorial and more restricted designs
Treatments may themselves be combinations of factors in a factorial design. A complete block can contain every treatment combination, allowing estimation of main effects and interactions while controlling block variation.
Designs such as Latin squares accommodate two blocking directions. Split-plot designs impose different randomization restrictions at different experimental scales. Their error structures differ from that of a simple randomized complete block design and require corresponding analyses.
Limitations and interpretation
Treatment–block interaction
The additive model assumes that treatment differences do not systematically change across blocks. A treatment–block interaction occurs when, for example, one treatment performs better in one environment but worse in another.
With only one observation per treatment–block combination, a full interaction cannot be separated from residual error without additional assumptions. Replication within combinations or a more structured model is needed to estimate interaction and pure error separately.
Missing observations
Missing outcomes destroy the simple balance that makes treatment and block effects orthogonal. A linear-model analysis can often accommodate the resulting unequal cell counts, but the standard balanced formulas no longer apply directly. The reason for missingness also matters: outcomes missing because of treatment-related events can affect interpretation.
Allocation and dependence
Treatments must vary within blocks. Assigning one treatment to all units in a block creates a different design and may produce confounding between treatment and block if there is inadequate replication at the assignment level.
Blocking also does not remove all dependence among observations. Spatial association, repeated measurements, or shared processing conditions may require a model with an appropriate covariance structure. The analysis must reflect the experimental unit and actual randomization, rather than treating every recorded measurement as an independent replicate.
Historical development and applications
Randomized block designs became central to modern experimental methodology through the work of Ronald A. Fisher and agricultural experimentation at Rothamsted in the early twentieth century. Field experiments offered a clear motivation: soil and location differences needed to be controlled without sacrificing random treatment allocation.
The same principle applies to industrial experiments blocked by production batch or day, laboratory studies blocked by assay plate or specimen source, and randomized studies grouped by relevant baseline characteristics. Across these applications, the defining feature is restricted random assignment within groups constructed to improve treatment comparisons.