Factorial Experiment¶
Test or assessment — instantiates Interaction Effect Mapping
Tests the focal factor, potentiating factor, and paired condition so interaction effects can be separated from isolated effects.
A Factorial Experiment is the mechanism that estimates the interaction term directly by manipulating every factor at every level and measuring the outcome in each resulting cell. Its defining move is full crossing: if you have two factors at two levels each, you run all four combinations — not the two "one-at-a-time" arms a naive test would use. Because every cell is observed, the combined effect can be compared against the additive prediction (each factor's main effect summed), and the leftover — the part the additive model cannot explain — is the interaction. Where a reduced test infers that interactions are unlikely, the factorial experiment produces a measured, signed, confidence-tagged estimate of the interaction itself. It is the archetype's most rigorous instrument and its most expensive one.
Example¶
An agronomy station wants to know how to raise winter-wheat yield. Two factors are in question: fertilizer regime (none / nitrogen-only / full NPK) and irrigation (low / high). Rather than test fertilizer on well-watered plots and irrigation on unfertilized plots — the one-at-a-time trap — the station fully crosses them into a 3×2 grid of six treatments, replicating each across several field blocks to average out soil variation. At harvest it measures yield per plot in every cell.
The additive story would say: NPK adds so many bushels, high irrigation adds so many, and together they add both. The grid says otherwise. NPK barely moves yield under low irrigation but lifts it sharply under high irrigation — the two factors potentiate each other. The interaction term (observed combined lift minus the additive prediction) is large and positive, and because each cell was replicated the station can attach a confidence band to it rather than a hunch. The recommendation that follows is not "use NPK" but "use NPK only where irrigation is high" — a conclusion no isolated test could have reached.
How it works¶
- Cross the levels. Define a discrete level set for each factor and build the full grid of combinations; every factor appears at every level of every other factor.
- Randomize and replicate. Assign units to cells at random and repeat each cell enough times that the interaction term — usually smaller than the main effects — rises above noise.
- Decompose the outcome. Partition the measured variation into main effects, interaction effect(s), and residual, typically by analysis of variance, so the interaction is an explicit estimated quantity, not an impression.
- Read the interaction, not just the cells. A nonzero interaction term with a tight confidence interval is the deliverable; its sign and shape say synergy, antagonism, or threshold.
Tuning parameters¶
- Levels per factor — two levels screen for direction cheaply; three or more trace curvature and thresholds but multiply the cell count. More levels, more resolution, more runs.
- Replication depth — how many units per cell. Deeper replication powers the interaction term (not just the main effects), narrowing its confidence band at direct cost in samples.
- Highest order estimated — whether the design resolves only two-way interactions or reaches three-way and beyond. Each order added enlarges the grid geometrically.
- Full vs. augmented — plain crossing, or crossing plus center points / replicated controls to check for nonlinearity between the tested levels.
- Blocking — grouping units to absorb a nuisance source (field, batch, day); tightens precision but spends design complexity.
When it helps, and when it misleads¶
Its strength is that it is the cleanest way to earn an interaction claim: because factors are manipulated and fully crossed, the estimated interaction term is causal and quantified, with a confidence rating that separates a tested interaction from a suspected one. It is the mechanism that turns "we think A and B combine oddly" into "the A×B term is this big, this direction, this certain."
Its failure mode is combinatorial explosion: cells grow as the product of every factor's levels, so a handful of many-level factors becomes thousands of runs, and powering the interaction term needs still more.[1] The classic misuse is a study sized to detect main effects but far too small for the interaction — the interaction estimate is then mostly noise, and an underpowered null gets misread as "no interaction." Manipulability and ethics also bound it: some factor combinations cannot be assigned to people or systems at all. The guarding discipline is to power the interaction term deliberately (not inherit main-effect sample sizes), and to hand the coverage-reduction problem to a planned design rather than improvising which cells to drop.
How it implements the components¶
factor_level_set— sets the discrete levels at which each factor is fixed and crossed; the grid is built from this set.effect_measurement— measures the outcome in every crossed cell, the raw material for decomposition.interaction_type_classification— the estimated interaction term's sign and shape classify the combination as reinforcing, antagonistic, or threshold-dependent.evidence_confidence_rating— replication and significance testing attach a confidence band to each estimate, marking tested versus suspected.
It does not implement factor_list, interaction_matrix, or review_or_iteration_rule — enumerating factors for reduced coverage and its escalation-to-bundle rule belong to its nearest twin, Pairwise Combination Testing, which trades the full grid for guaranteed pair coverage.
Related¶
- Instantiates: Interaction Effect Mapping — supplies the measured interaction terms the map is built on.
- Consumes: Design of Experiments Protocol — executes the crossing, coverage, and measurement plan that protocol specifies.
- Sibling mechanisms: Pairwise Combination Testing · Interaction Matrix Table · Compatibility Screening · Dependency Interaction Map · Treatment Interaction Analysis · Design of Experiments Protocol
Editorial Notes¶
Form Classification¶
Form family: Experiment, Test & Rehearsal
Rationale: Factorial Experiment operates as a bounded trial, probe, simulation, or rehearsal that generates evidence from performance because it tests the focal factor, potentiating factor, and paired condition so interaction effects can be separated from isolated effects.
Independent corroboration: The frozen evidence defines Factorial Experiment as 'Tests the focal factor, potentiating factor, and paired condition so interaction effects can be separated from isolated effects', so its operative form is Experiment, Test & Rehearsal.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Universal
Rationale: Factorial experiments and interaction estimation were canonically formalized in Fisherian experimental design.
Review resolution: Both reviewers agree that statistics_experimental_design is primary. I retain no additional origin domain only as formative origin lineages; single_lineage is appropriate because the alternate domains informed practice without constituting independent ownership. Reach is universal because the structure is portable across essentially any domain with the stated problem, an applicability judgment kept separate from provenance. Encyclopedia synthesis is false because the artifact is already established enough that encyclopedia-specific synthesis is not required. No unresolved historical ambiguity remains after reconciling the secondary fields.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Keep the plan and the run separate: the Design of Experiments Protocol decides which cells to run and how; the factorial experiment is the execution that produces estimates. That split lets a team improve the design (add a level, drop a factor, fractionate) without re-litigating what the last run found.
References¶
[1] The factorial design and the "interaction effect" were formalized by R. A. Fisher (The Design of Experiments, 1935); the interaction term is the deviation of the observed combined response from the sum of main effects. Its cost is that the number of cells is the product of the factors' level counts, which is why fractional and pairwise reductions exist. registry ↩