Yes; treatments commonly appear in several blocks.

Does incomplete mean unfinished data collection?

No; planned blocks can deliberately contain only a subset of treatments.

Does equal replication guarantee balance?

No; pair occurrences must also meet the stated requirement.

AI Glossary / Statistics and probability / Advanced

Block design

Definition

A block design is an arrangement of elements into subsets, called blocks, with specified incidence properties. In combinatorial mathematics, the elements are often called points, and requirements may control how often individual points or pairs occur together. In experimental design, elements can represent treatments allocated within groups of experimental units. Balanced incomplete block designs are an important special case.

What it means

Blocks are not necessarily a partition: the same element can appear in several blocks.
For an experiment, blocking can organize comparisons within relatively similar conditions, such as a session or batch. An incomplete block contains fewer than all treatments.

How it works

For a balanced incomplete block design, let v be the number of treatments, b the number of blocks, k the number of distinct treatments per block, r the number of blocks containing each treatment, and lambda the number containing each pair. Here k is less than v.
Counting treatment appearances gives vr = bk. Counting pairings with a fixed treatment gives r(k-1) = lambda(v-1). These equations are necessary checks, not a universal existence guarantee for arbitrary proposed parameters.
The plan can be represented by an incidence matrix: rows identify treatments, columns identify blocks, and a one marks membership. To use the plan experimentally, specify the physical blocks, assignment randomization, observations and analysis. Connectivity matters when treatment comparisons must be separated from block effects.

Three practical examples

Evaluation scheduling: a team compares several candidate systems but can assess only a subset during each session, so it plans repeated overlaps.
Product testing: researchers arrange treatments within material batches that are too small to hold every treatment.
Combinatorial verification: an analyst checks whether a proposed collection of subsets has the required replication and pair-concurrence counts before using it as a schedule.

Think of it this way

Imagine arranging guests at several small discussion tables. Each guest may attend several sessions, and the plan controls how often guests share a table.

Worked example

Consider an original four-treatment schedule with labels A, B, C and D. Each session can accommodate three treatments. Use four blocks: ABC, ABD, ACD and BCD. Within a block, each listed treatment appears once.
Here v = 4, b = 4 and k = 3. A appears in ABC, ABD and ACD; B appears in ABC, ABD and BCD. C and D also each appear in three blocks. Therefore r = 3, and the total number of treatment appearances is vr = 4 times 3 = 12, equal to bk = 4 times 3.
Now count pair occurrences. AB appears in ABC and ABD. AC appears in ABC and ACD. AD appears in ABD and ACD. BC appears in ABC and BCD. BD appears in ABD and BCD. CD appears in ACD and BCD. Every pair occurs twice, so lambda = 2.
The second counting equation gives r(k-1) = 3 times 2 = 6 and lambda(v-1) = 2 times 3 = 6.
With columns ordered ABC, ABD, ACD, BCD, the incidence rows are A: 1, 1, 1, 0; B: 1, 1, 0, 1; C: 1, 0, 1, 1; D: 0, 1, 1, 1. Each row sums to three and each column to three. The dot product of any two different rows is two, directly checking their shared-block count.
Suppose each treatment appearance requires one evaluation. The plan needs twelve evaluations, compared with sixteen if all four treatments were evaluated in each of four complete blocks. This saves four evaluations in this particular comparison, but it also provides three rather than four observations per treatment. It is not a claim that the incomplete plan has equal statistical precision to the complete plan.
For a contrasting schedule, use four two-treatment blocks: AB, AB, CD and CD. Repeated block contents here represent separate sessions. Every treatment appears twice, so the schedule has equal replication. However, AB and CD each occur together twice, while AC, AD, BC and BD never occur together. Equal replication has not produced pair balance.
Its concurrence graph has one connection between A and B and another between C and D, with no path joining the groups. Under the usual additive treatment-and-block model with unrestricted fixed block effects, comparisons between those groups cannot be separated from block differences using this schedule alone. This is a connectivity issue, not something repaired merely by collecting the same disconnected blocks again.
Finally, return to ABC, ABD, ACD and BCD. Before running an experiment, the coordinator must map the abstract labels to actual treatments and sessions and randomize assignments as appropriate. The listed subsets verify a combinatorial structure; they do not specify scoring rules, order effects or the assumptions for estimating treatment differences.

Illustrative case study

Fictional company Cedar Comparisons has four candidate systems and room for three evaluations per session. Its analysts use the four-block schedule above, verify every membership count and prepare a separate randomized assignment procedure.
The report retains session identifiers so that the analysis can account for session differences. When an evaluation is missing, the team rechecks the realized design rather than continuing to call the observed data perfectly balanced.

Common misconceptions and mistakes

A block design need not be balanced, and equal treatment replication is weaker than equal pair concurrence. Blocks can overlap, so treating them as disjoint groups of treatment labels can miss the central structure.
The counting equations are necessary conditions for a balanced design, not an automatic construction method or sufficient proof of existence. A valid combinatorial plan also does not guarantee that practical observations follow the intended statistical model. Missing runs, treatment interactions with blocks or systematic presentation order can change what conclusions are supported.

Business use

Block designs can organize constrained comparisons while making the allocation structure explicit. Teams should evaluate workload, treatment overlap, connectivity and the analysis required by the actual experiment. The benefit comes from matching a sound schedule to the operating constraints, not from describing any grouped activity as statistically balanced.

Check your understanding

What was lambda in the first schedule?

Answer: Two.

Why did the second schedule fail pair balance?

Answer: Four treatment pairs never shared a block.

Why was randomization still needed?

Answer: The subset list specified incidence, not the actual assignment procedure.

Related terms

Randomized blocks design
Combinatorial design
Replication (statistics)
Balanced incomplete block design
Randomized block design
Incidence matrix
Replication
Connectivity
Experimental design

Learn next

Build the pair-concurrence table for another proposed schedule and check both replication and connectivity before considering its analysis.

Sources

Hao Zhang, Purdue University, Incomplete Block Designs: https://www.stat.purdue.edu/~zhanghao/STAT514/Lecture_Notes/LectureNotes15-Incomplete-Block-Design-.html . The schedules, counts, incidence rows and company scenario are original illustrations.

Previous: Blinded experiment Next: Block matching