Design of Experiments Protocol¶
Protocol — instantiates Interaction Effect Mapping
A planning protocol that determines which factors, levels, combinations, assignment rules, and measurement windows will be used to detect interaction effects efficiently.
A Design of Experiments Protocol is the plan written before any data is collected — the document that decides which factors and levels to study, which combinations to actually run, how units are assigned, and when outcomes are measured, all chosen so the interaction effects of interest can be estimated with the fewest runs at the required precision. Its defining property is that it is upstream of execution: it does not measure anything or read off any interaction term; it commits to a coverage strategy and its statistical consequences in advance. Its central craft is coverage economics — deliberately choosing a fraction of the full grid (a fractional-factorial or optimal design) and knowing exactly which interaction effects that fraction can and cannot resolve. A good protocol makes the trade between runs and resolvable interactions explicit and pre-registered, so the run that follows is efficient and its blind spots are known.
Example¶
A manufacturing engineer needs to reduce defects in an injection-molding process. Six factors plausibly matter — melt temperature, injection pressure, cooling time, hold pressure, mold temperature, and material lot — each at two levels. The full grid is 2⁶ = 64 runs, and each run means retooling a production line: too expensive to do exhaustively. The protocol's job is to plan a fraction.
The engineer specifies a resolution-IV fractional-factorial design of 16 runs, chosen so that all main effects are clean and all two-factor interactions are estimable — while accepting that certain two-factor interactions are aliased with each other and that no three-way interactions can be resolved. The protocol fixes the level of each factor per run, the randomized run order (to keep a slow machine drift from masquerading as an effect), the number of replicate parts measured per run (setting the precision the interaction estimates will carry), and a follow-up rule: if two aliased interactions look active, a small fold-over set of extra runs is pre-planned to disentangle them. Nothing has been molded yet — but the coverage, the assignment, the measurement, and the escalation are all decided and their statistical consequences known.
How it works¶
- Choose factors and levels. Fix which factors enter and the discrete levels each takes — the design's dimensions.
- Select a coverage design. Pick full factorial, a fractional-factorial of a stated resolution, or an optimal design, so the runnable subset resolves the target interactions and its aliasing is known in advance.
- Set assignment and measurement. Specify randomization, blocking, replication, and measurement windows so nuisance variation does not confound the interaction estimates and the precision is planned, not accidental.
- Pre-plan the follow-up. Define the sequential rule — fold-over, augmentation, added replicates — that resolves ambiguities the first fraction cannot, before any data is seen.
Tuning parameters¶
- Fraction size / resolution — how large a fraction of the full grid to run. A higher-resolution design resolves more interactions cleanly but costs more runs; a leaner fraction is cheap but aliases effects together.
- Aliasing structure — which effects are deliberately confounded with which. Choosing the alias pattern decides which interactions you are willing to leave ambiguous.
- Replication / power target — replicates per design point, setting the precision (and confidence) the interaction estimates will carry.
- Randomization and blocking scheme — how assignment guards against nuisance variables and drift; more structure buys cleaner estimates at planning complexity.
- Sequential vs. one-shot — whether the protocol runs the whole design at once or in stages with pre-planned augmentation, trading up-front cost against adaptivity.
When it helps, and when it misleads¶
Its strength is coverage economics: it wrings the maximum resolvable interaction information out of a fixed run budget and — because the design is chosen in advance — the blind spots (which interactions are aliased, which orders are unresolvable) are known before, not discovered after.[n1] It is what gives an interaction study statistical discipline instead of an improvised set of runs.
Its failure mode is aliasing that goes unheeded: a lean fractional design confounds certain interactions, and a team that forgets its own alias structure will read an active alias as if it belonged to one specific interaction when it might belong to its confounded partner. The classic misuse is fractionating too aggressively — a design so lean the very interaction the study exists to find is aliased with a main effect — or treating the plan as the result and never running the pre-planned fold-over that would resolve the ambiguity. The guarding discipline is to keep the alias structure visible on the protocol itself and to commit to the sequential follow-up up front, so ambiguity is resolved by design rather than argued about after.
How it implements the components¶
factor_level_set— fixes which factors enter and the discrete levels each is set to across the design.interaction_matrix— the coverage plan is a structured statement of which factor combinations will be run and which interactions they resolve.review_or_iteration_rule— the pre-planned sequential follow-up (fold-over, augmentation) that resolves ambiguities the first fraction leaves.evidence_confidence_rating— replication and power targets set, in advance, the precision and confidence the resulting interaction estimates will carry.
It does not implement effect_measurement or interaction_type_classification — actually executing the runs, measuring each cell, and reading off the signed interaction term belong to Factorial Experiment; this protocol only specifies the plan those steps follow.
Related¶
- Instantiates: Interaction Effect Mapping — plans efficient, statistically-disciplined coverage of the interaction space.
- Consumes: Compatibility Screening — takes the screened, admissible combination set as the feasible region the design is drawn within.
- Sibling mechanisms: Factorial Experiment · Pairwise Combination Testing · Interaction Matrix Table · Compatibility Screening · Dependency Interaction Map · Treatment Interaction Analysis
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: Design of Experiments Protocol operates as a non-executable information artifact that externalizes static or prospective structure because it a planning protocol that determines which factors, levels, combinations, assignment rules, and measurement windows will be used to detect interaction effects efficiently.
Independent corroboration: The frozen evidence defines Design of Experiments Protocol as 'A planning protocol that determines which factors, levels, combinations, assignment rules, and measurement windows will be used to detect interaction effects efficiently', so its operative form is Representation, Specification & Plan.
Nearest alternative: Protocol, Workflow & Routine — The experiment design is a prospective specification of factors, assignments, and windows; it does not itself run exposures.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Experimental design cohered pre-data factor, level, assignment, measurement, aliasing, and resolution plans for estimating interactions efficiently.
Related originating lineages:
- Engineering & Design — Industrial engineering supplied factorial experimentation for product and process optimization.
Review resolution: Experimental design cohered pre-data factor, level, assignment, measurement, aliasing, and resolution plans for estimating interactions efficiently. The retained alternate lineages materially shaped the mechanism's form.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The protocol is the plan; Factorial Experiment is the run that executes it and returns the estimates. Keeping them separate lets a team revise coverage — add a factor, change the fraction, re-plan the fold-over — without re-interpreting an already-collected dataset, and lets the same protocol drive either a full factorial run or a reduced one.
[n1] In a fractional-factorial design, running only part of the full grid aliases (confounds) some effects with others: an estimated coefficient may reflect a main effect, an interaction, or their sum. The design's resolution states how severe the aliasing is; choosing it is choosing which interactions you are willing to leave ambiguous, and a follow-up "fold-over" run can de-alias effects that turn out to matter. ↩