Is this a theorem about prime numbers?

No. It is a methodological warning, not a universal mathematical result governing experimental validity.

Are equal group sizes enough?

They can improve allocation balance but do not fill a missing factorial combination.

Can the three observed groups still answer useful questions?

Yes. In the example they support particular conditional comparisons, but not the unrestricted two-factor interaction.

AI Glossary / Statistics and probability / Intermediate

Wike's law of low odd primes

Definition

Wike's law of low odd primes is a deliberately provocative experimental-design maxim associated with a 1973 article by Edward L. Wike. It warns that a small odd-prime number of treatment conditions can signal an incomplete factorial arrangement. It should be treated as a prompt to inspect the design, not a mathematical rule that every three-group experiment is defective.

What it means

The practical concern is whether a study includes the combinations needed to separate the effects it claims to estimate. Testing “baseline,” “AI tool” and “AI tool plus training” leaves out “training without the AI tool.” That missing condition matters if the question concerns how tool and training effects interact.

Counting groups can draw attention to the problem, but the scientific question determines whether it is actually a problem. A study comparing three distinct products under one factor can be perfectly sensible. The number three alone neither creates confounding nor establishes that a design is unbalanced in allocation size.

How it works

Identify the factors that vary across treatments and list their levels. Lay out all combinations relevant to the proposed analysis. Two binary factors create four combinations. Check which are observed and which effects can be estimated without adding unsupported assumptions.

A missing combination may prevent estimation of an interaction even when some conditional comparisons remain available. Equal numbers in the observed groups do not supply information about an unobserved combination. If only particular contrasts matter, specify them clearly. If a simpler model assumes no interaction, acknowledge that assumption instead of claiming the design has measured the interaction directly.

Three practical examples

1. AI adoption: A company compares standard work, an AI assistant, and the assistant with coaching. The missing coaching-only condition limits claims about whether coaching specifically enhances the assistant's effect.

2. Website testing: A business tests an old page, a new layout and a new layout with revised copy. Separating layout-copy interaction requires considering the missing old-layout/new-copy combination.

3. Product comparison: A team randomly assigns users to three alternative recommendation systems. If the aim is simply to compare those systems as complete packages, three arms need not imply a flawed factorial design.

Think of it this way

Think of evaluating two switches using a table with one corner missing. You can compare the corners you have, but you cannot directly observe what happens at the absent corner. The warning is about the missing comparison, not about a mysterious statistical property of the number of visible corners.

Worked example

Consider an original teaching example with binary factors A, an AI tool, and B, coaching. The observed average task scores are 50 for neither intervention, 70 for A alone and 80 for both A and B. The B-only condition is missing.

The tool contrast without coaching is 70 − 50 = 20. The coaching contrast among tool users is 80 − 70 = 10. However, the coaching contrast without the tool cannot be calculated from these three means.

If the missing B-only mean were 60, the interaction contrast would be (80 − 70) − (60 − 50) = 0. If it were 75, the contrast would instead be 10 − 25 = −15. Both possibilities agree with every observed mean. Thus the interaction is unidentified from these cell means without an additional assumption. In regression terms, the three observed treatment patterns provide rank three for an intercept, A, B and AB model with four coefficients. Repeating the same patterns more often does not create the missing independent pattern.

Illustrative case study

In this fictional case, Orchard Services wants to evaluate an AI scheduling assistant and an employee workshop. Its initial plan includes baseline, assistant-only and assistant-plus-workshop groups. During review, the analyst asks whether the business needs to know the workshop's standalone contribution or merely compare the three proposed packages.

Because the intended decision includes whether to offer workshops without the assistant, the team adds that condition and plans adequate replication. It documents outcomes and contrasts before running the study. The improvement comes from aligning the design with the decision, rather than treating a four-group count as a guarantee of experimental quality.

Common misconceptions and mistakes

Do not declare every design with three, five or seven treatments invalid. A single factor can legitimately have any of those numbers of levels. “Balanced” can also refer to equal allocation across groups, which is a different issue from including all combinations of several factors.

Do not assume that all incomplete factorial designs are accidental or unusable. Carefully designed fractions can answer specified questions under stated assumptions. Conversely, an even number of treatments does not guarantee identifiable effects, randomization, adequate power or unbiased measurement. The label is best used to trigger a design audit, not to replace one.

Business use

The maxim is useful in project planning because packaged interventions often change several things at once. Managers may later ask which component produced the result, even though the original comparison never isolated those components.

Before allocating participants or implementation effort, map each planned decision to an estimable contrast. Explain which claims the proposed design can support and where extra conditions or assumptions are necessary. This can prevent an expensive AI pilot from yielding an impressive overall difference while leaving its main business question unresolved.

Check your understanding

Which group is missing from baseline, tool-only and tool-plus-coaching?

Answer: Coaching without the tool.

Would doubling participants in the existing three groups identify the interaction?

Answer: No. It may improve precision for observed comparisons but does not reveal the missing cell mean.

Is a three-product randomized comparison automatically defective?

Answer: No. Its adequacy depends on its purpose and implementation.

Related terms

Factorial design
Confounding
Fractional factorial design
treatment combination
interaction
estimability
balanced design
experimental contrast

Learn next

Next, draw treatment-combination tables for a proposed experiment and write its intended comparisons. Study model-matrix rank and interaction contrasts. Then distinguish deliberate fractional designs from missing conditions that prevent the project from answering its stated question.

Sources

Wike's publisher issue contents and abstract preview confirm author, date and a methodological principle with limitations: https://journals.sagepub.com/toc/prxa/33/1 . Full article unavailable. The preview does not state the specific odd-prime/incomplete-factorial warning; that historical attribution remains unverified. NIST supports the modern factorial explanation: https://www.itl.nist.gov/div898/handbook/pri/section3/pri3331.htm . Example and fictional case are original; no claim to reproduce Wike's procedure.

Previous: Wiener filter Next: Wikification