Blocking is a fundamental technique in experimental design used to control nuisance variation. That is, differences between experimental units that can mask the true effects of the treatments being studied. By isolating the treatment effects from this unwanted variation, blocking helps ensure that treatment comparisons are fairer and more precise.
The core idea behind blocking is simple: the experimental units are grouped into homogeneous blocks, such that the units within each block are as similar as possible. Once the blocks are formed, the treatments are then randomly assigned to the experimental units within each block.
For example, in an agricultural field trial, plots can be blocked based on factors such as soil fertility, wind exposure, and slope, to ensure that these environmental differences do not obscure the effects of the treatments.
One of the most used experimental designs is the Randomized Complete Block Design (RCBD). In an RCBD:
- The experimental units are grouped into homogeneous blocks
- Each block has the same number of experimental units, usually one for each treatment
- Within each block, the treatments are randomly allocated to the experimental units so that each treatment occurs in each block the same number of times
However, what if a block can’t accommodate all the treatments?
For example, suppose you’re conducting an agricultural field trial to compare the performance of 6 pesticide treatments and need to control for environmental factors like soil fertility, wind exposure, and slope. However, each block, defined as an area of land with relatively uniform environmental conditions, only has space for 3 plots (the experimental unit). This means you can’t include all 6 treatments in every block, as required by a RCBD.
This is where incomplete block designs are needed. These designs allow treatments to be distributed across blocks without requiring every treatment to appear in each block, while still enabling valid and statistically sound comparisons between all treatments. In this blog, we’ll focus on one of the most important incomplete block designs: the Balanced Incomplete Block Design (BIBD).
In an BIBD:
- Balanced: Every pair of treatments appears together in the same number of blocks, and each treatment is replicated equally across the entire design
- Incomplete Block: Each block includes only a subset of treatments (i.e., there are fewer treatments per block than the total number of treatments), and each treatment appears no more than once per block
- Treatments are randomly allocated to the experimental units within each block
Importantly, because not all treatments appear in every block, the treatment and block effects are partially confounded. This means it can be more difficult to distinguish the true effects of the treatments from differences due to block variability. However, the more often each pair of treatments appears together in the same block, the more information is available to separate treatment effects from block effects. Through balance (i.e., ensuring each pair of treatments occurs together in the same block the same number of times) BIBDs are statistically efficient, striking a compromise between practical constraints on block size and the need for precise treatment comparisons.
Let’s look at an example: 6 treatments arranged over 10 blocks of size 3

Replication: Over the entire design, each treatment is replicated five times
Balance: Every pair of treatments occur together in the same block exactly twice
In the above design, Treatment 1 never occurs in the same block as all other treatments. As a result, its effect is partially confounded with the block effects. However, we can still estimate Treatment 1’s effect through direct comparisons:
- with Treatments 3 and 5 in Block 2
- with Treatments 4 and 6 in Block 3
- with Treatments 2 and 5 in Block 6
- with Treatments 2 and 6 in Block 7
- with Treatments 3 and 4 in Block 10
and indirect comparisons:
- with the Treatments in Blocks 1, 4, 5, 8, and 9
The Genstat output below shows the skeleton ANOVA for our example design, along with a table giving the efficiency factors.

Recall that in an incomplete block design, not all treatments appear in every block. As a result, treatment effects are partially confounded with block effects. This partial confounding is reflected in the ANOVA structure, where the Treatments term is estimated across two strata:
- the Blocks stratum, which captures variation between blocks
and
- the Blocks.Units stratum, which captures variation within block.
The efficiency factor (e.f.) indicates how much information about the treatment effects is contributed by each stratum. In our design:
- 20% of the information comes from between-block comparisons (i.e., indirect comparisons between treatments that do not appear in the same block)
- 80% comes from within-block comparisons (i.e., direct comparisons between treatments that appear together in the same block and are thus not confounded with block effects)
Estimating Treatment Means in an RCBD versus a BIBD
An important property of the RCBD is that the estimated treatment means are the same as the raw treatment means (i.e., the simple averages of the observed responses for each treatment across all blocks). This is because each treatment appears the same number of times in every block, allowing the block effects (i.e., nuisance variation) to be separated from the treatment effects by subtracting the mean of each block from the observations within that block. As a result, the influence of between-block variation can be accounted for without altering the treatment means.
In contrast, in a BIBD, treatments occur in an incomplete but balanced way across the blocks. Since each treatment occurs in only a subset of blocks, the estimated treatment means must be adjusted for the specific blocks in which the treatments appear. Consequently, the estimated treatment means typically differ from the raw averages.
ANOVA versus REML
Data from a BIBD can be analysed using ANOVA or REML (i.e., a linear mixed model). However, these methods may give slightly different estimates of the treatment means and their standard errors due to differences in how block effects are handled.
In ANOVA, blocks are treated as fixed effects. This means that treatment comparisons are made conditional on the specific blocks observed in the experiment. In contrast, in a linear mixed model, blocks are usually treated as random effects, under the assumption that the blocks represent a random sample from a larger population. This approach allows for the estimation of a block variance component.
By modelling blocks as random, the linear mixed model can “borrow strength” across all blocks, leveraging shared information to improve the estimates of the treatment effects, even when treatments appear in only a subset of blocks. This enables partial recovery of information that might otherwise be lost in ANOVA due to the incomplete representation of treatments across all blocks. In practical terms, for unbalanced or incomplete designs like BIBDs, this means that the treatment means estimated using a linear mixed model are often more precise (i.e., have smaller standard errors). Although, if the between block variance is large, the ANOVA and linear mixed model results will be very similar, as there won’t be much information to recover between blocks.
Below is some Genstat output from an ANOVA and a linear mixed model analysis of a BIBD. Notice the difference in the estimated treatment means, the s.e.d., and the denominator degrees of freedom associated with the F-test of the treatment term.


BIBD Parameter Restrictions
It is important to note that, due to the balance requirement, only certain combinations of treatments, block numbers, and block sizes are feasible for a BIBD. In particular, the following two conditions must be satisfied:
where:
In our example, v = 6 treatments, r = 5 (each treatment appears in 5 blocks), b = 10 blocks, k = 3 (each block contains 3 treatments) and λ = 2 (each pair of treatments appears together in the same block twice). Checking the conditions:
Resolvable Designs
Incomplete block designs come in two flavours: resolvable and non-resolvable. In a resolvable block design, the blocks can be arranged into groups such that each group contains a complete set of treatments. For a non-resolvable design, no such grouping is possible.Below is a resolvable form of our example design, generated using CycDesigN.

If you’d like to learn more about resolvable and non-resolvable designs check out this blog post: Resolvable or Non-Resolvable Design: What Does It Mean and Why Does It Matter?
Alpha Designs
Because the parameter requirements for BIBDs are quite restrictive, a BIBD may not be feasible for your experimental situations. In such cases, alternative incomplete block designs, such as Alpha designs, can be used. Alpha designs are a flexible class of resolvable incomplete block designs, especially suited for experiments involving a large number of treatments where the block size must be kept small due to high variability between experimental units.
In general, Alpha designs are not balanced, meaning that different pairs of treatments may not appear together in the same block as equally often. As a result, while a balanced block design has a single standard error of difference (SED) for all treatment comparisons, an unbalanced design, such as most Alpha designs, produces a range of SEDs. Importantly, Alpha designs are constructed to give optimal efficiency by maximizing the proportion of information on the treatment effects obtained from within-block comparisons and minimizing the spread of SEDs.
Generating a Design
Need to generate an experimental design? At VSNi, we offer two tools for creating robust and efficient experimental designs:
- CycDesigN – A powerful software tool for generating both resolvable and non-resolvable block or row-column designs, as well as single- and multi-location partially replicated (p-rep) designs, and crossover designs. CycDesigN is highly flexible and can, for example, generate block designs with unequal treatment replication.
- Genstat – A comprehensive statistical package that includes design generation and analysis tools for a wide range of experimental setups, including block, factorial, stratified (e.g., split-plot), and spatial (e.g., row-column) designs.
To learn more or request a demo, contact us at vsni.co.uk/contact/
2 Responses
I DID A PROJECT AND AM CONFUSED OF MY DESIGN. HAD 60 GENOTYPES, TWO BLOCKS OF 30 GENOTYPES PER REPLICATION GIVING A TOTAL OF 6 BLOCKS AND 180 TREATMENTS. WHAT IS MY DESIGN
Hello Gladys,
It sounds like you’ve used a resolvable incomplete block design, with 3 replicates per genotype (aka “treatments”), arranged in 6 blocks of size 30. The design is resolvable, as the blocks are grouped into 3 pairs, with each pair forming a complete replicate (i.e., containing all 60 genotypes). In total, there are 180 experimental units (i.e., plots within blocks).
We also have a blog on resolvable designs , along with a short tutorial video showing how to generate them in CycDesigN, which you might find helpful.
Thanks for the question!