Does the design require equal sample sizes?

No, although balance can simplify interpretation and analysis.

What is an interaction?

A change in one factor's effect across levels of the other on the chosen scale.

Are repeated readings automatically independent replicates?

No; their dependence must be represented.

AI Glossary / Statistics and probability / Intermediate

Twoway factorial design

Definition

A two-way factorial design studies two factors by including every combination of their selected levels. If factor A has a levels and factor B has b levels, the full design contains a times b treatment combinations, or cells. It can estimate the effects associated with each factor and examine whether one factor's effect changes across levels of the other. Two-way refers to the number of factors, not to a requirement that each factor have only two levels.

What it means

Studying one factor at a time can miss combinations that work differently together. A factorial arrangement makes those combinations visible. A main effect summarizes a factor across levels of the other, while an interaction describes how its effect depends on that other factor. The design structure alone does not specify randomization, replication or whether measurements are independent or repeated. Those choices determine which comparisons and uncertainty calculations are justified.

How it works

Define the factors, their levels and the outcome. Include all chosen combinations, allocate suitable experimental units and randomize according to the actual operational constraints. Plan replication and any blocking or repeated measurements. Analyze using a model that represents both factors and their interaction, with errors matching the design. Examine cell means and relevant contrasts before relying on an average main effect. If some combinations are absent or impossible, acknowledge that the design is incomplete or follows a different structure rather than describing it as a fully crossed factorial.

Three practical examples

Example 1: A service experiment crosses two onboarding methods with three interface layouts. Example 2: An AI operations study crosses two model configurations with several workload levels. Example 3: A learning study crosses feedback type with practice duration. These examples can reveal whether a factor works consistently across contexts. If workload is measured rather than randomly assigned, causal claims about workload require additional assumptions even if the analysis includes it as a factor.

Think of it this way

Think of a grid with one factor defining rows and the other defining columns. Every cell represents a combination that must be studied. Comparing rows shows one factor's differences, and comparing columns shows the other's. If the row difference changes across columns, the factors interact on that outcome scale. The grid organizes the comparisons, but it does not itself provide independent data or eliminate confounding.

Worked example

Consider two onboarding methods, basic and enhanced, crossed with three layouts. The six illustrative mean scores are 50, 60 and 70 for basic onboarding and 55, 70 and 85 for enhanced onboarding. The enhanced-minus-basic contrasts are five, ten and fifteen points across layouts. Their equal-weight average is ten points, matching the difference between row means seventy and sixty. The contrast changes by ten points from the first to the third layout, demonstrating interaction in these stipulated means. Four independently randomized observations per cell would require 2 times 3 times 4 = 24 observations. The interaction in a conventional full model has (2 minus 1) times (3 minus 1) = two degrees of freedom. The six means alone provide no uncertainty estimate or significance conclusion; within-cell variation and the allocation design are still needed.

Illustrative case study

In this fictional case study, Rowan Onboarding tests training and interface combinations. The enhanced method appears beneficial on average, but its improvement varies substantially by layout. Analysts present the six cell estimates with uncertainty and investigate the interaction instead of recommending training based only on the overall average. The team also checks whether the planned rollout uses the same distribution of layouts as the experiment. Managers can then choose combinations supported by the study rather than assuming the average improvement applies everywhere.

Common misconceptions and mistakes

Common mistakes include confusing two-way with two-level, ignoring interaction while interpreting main effects, and counting repeated measurements as independent replications. A full factorial without replication can leave no pure-error estimate in a saturated cell-means model. Unequal cell sizes also make the weighting behind marginal comparisons important. Another error is assuming that two factors in an observational dataset imply randomized causal evidence. Factor labels organize an analysis, while the way units and conditions were assigned determines its causal support.

Business use

Businesses use two-way factorial designs to learn about combinations efficiently and avoid decisions based on isolated settings. Plan the combinations around realistic operating choices and identify the smallest interactions worth detecting. More cells spread a fixed budget across more conditions, which affects precision. Document whether factors are fixed or sampled, which units are randomized and whether execution restrictions create a split-plot structure. A clear design record helps analysts use the correct error terms and prevents an attractive table from concealing dependence or missing comparisons.

Check your understanding

How many cells are in the example?

Answer: Six.

What is the enhanced-minus-basic contrast for the third layout?

Answer: Fifteen points.

What is the equal-weight average contrast?

Answer: Ten points.

Why is reporting only that average incomplete?

Answer: The contrasts differ across layouts.

How many observations are needed for four independent observations per cell?

Answer: Twenty-four.

Learn next

Learn next how cell means decompose into main-effect and interaction components. Compare balanced and unbalanced designs and inspect the weighting behind marginal means. Then study randomization restrictions, repeated measurements and power for interaction contrasts so the analysis reflects how the observations were generated.

Main effect

Sources

Sources: Howell, Statistical Methods for Psychology Glossary, https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html; NIST, Sensitivity Analysis for Biometric Systems: A Methodology Based on Orthogonal Experiment Designs, https://nvlpubs.nist.gov/nistpubs/ir/2012/nist.ir.7855.pdf; NIST/SEMATECH, Nested variation and restricted randomisation, https://itl.nist.gov/div898/handbook/pri/section5/pri55.htm. These primary sources explain factorial combinations and design structure. The six-cell calculation is original and independently checked; the case is fictional.

Previous: Two-Tower Embedding Model Next: Twyman's law