Interaction Term Gate¶
Protocol — instantiates High-Dimensional Tractability Control
Requires evidence, rationale, and validation capacity before adding cross-feature interactions or segment combinations.
Interactions are where dimensionality explodes fastest: main effects grow linearly, but their cross-products grow combinatorially, and each interaction cell is thinner in evidence than the margins it came from. Interaction Term Gate is the admission protocol that refuses to let any cross-feature interaction or segment combination into a model unless it arrives with three things: a stated rationale for why the interaction should exist, pre-registered evidence or a prior hypothesis rather than a post-hoc discovery, and spare validation capacity — enough held-out data and statistical power to actually test it within the validated regime. Its defining move is that it governs interactions specifically, at the moment of admission, as a checkpoint that must be cleared with justification. It does not set the model's overall complexity budget and does not rank main-effect variables; it stands at the door where interactions try to enter and demands their papers.
Example¶
A phase III clinical trial has a pre-specified primary analysis, but as results come in, investigators propose a flurry of subgroup-by-treatment interaction tests: does the drug work better in patients with diabetes? In those over 65? In a particular genotype? In women? Each question is scientifically tempting, and with enough subgroups, some interaction will look significant by chance alone.[n1] The statistical analysis plan runs every proposed interaction through the gate.
For each one, the gate asks: is there a pre-registered biological rationale, or is this a fishing expedition prompted by the data? Is the subgroup powered — are there enough events in the interaction cell to detect a real effect, or is the estimate built on a handful of patients within the validated analysis regime? An interaction between treatment and a pre-specified genotype, hypothesized before unblinding and adequately sized, clears the gate and enters as a confirmatory test. A treatment-by-region interaction dreamed up after seeing the data, with only a dozen events in the relevant cell, is refused entry to the confirmatory analysis and demoted to explicitly exploratory, hypothesis-generating status. The gate does not compute the interaction; it decides whether the interaction is allowed to make a claim.
How it works¶
- Intercept at admission. Every proposed interaction or segment combination is stopped before it enters the model and must clear the gate rather than being added by default.
- Demand rationale and provenance. Require a stated mechanism for why the interaction should exist and evidence that it was hypothesized in advance, separating confirmatory tests from post-hoc discoveries.
- Check validation capacity. Verify there is enough held-out data and power to test the interaction inside the validated regime — an interaction whose cell is too thin to validate is refused or demoted.
- Charge against the standing budget. Admitted interactions are counted against the complexity budget the review process maintains, so the gate enforces rather than sets that ceiling.
Tuning parameters¶
- Evidence bar — how strong a prior rationale an interaction needs to enter; a high bar curbs false discoveries but can suppress real, unexpected effects.
- Confirmatory vs. exploratory split — whether failing the gate blocks the interaction entirely or merely demotes it to hypothesis-generating; demotion preserves discovery while protecting claims.
- Power/capacity threshold — the minimum events or held-out coverage per interaction cell required to admit it; stricter thresholds prevent thin-cell noise but exclude rarer combinations.
- Multiplicity discipline — how the count of tested interactions adjusts the significance bar; without it, enough gated tests still manufacture false positives.
- Scope of the gate — which interactions require review (all, or only beyond a certain order); gating high-order interactions matters most, since they are the thinnest.
When it helps, and when it misleads¶
Its strength is throttling the single largest source of high-dimensional false discovery — unconstrained interaction search — by forcing every cross-term to justify itself and to prove it can be validated, which keeps a model's confirmatory claims honest.
Its central failure mode is theater: a gate that demands a "rationale" but accepts any plausible-sounding story admits nearly everything, and one that ignores how many interactions were tried lets multiplicity smuggle false positives through one at a time.[n1] A classic misuse is treating a data-discovered interaction as confirmatory because it was written into the plan after it was seen. The guarding discipline is to require genuine pre-registration, to enforce a power/capacity floor per interaction, and to adjust the evidence bar for the number of interactions examined.
How it implements the components¶
interaction_term_policy— it is the policy: an explicit admission rule that every cross-feature interaction or segment combination must satisfy before entering.validation_regime_boundary— it requires that an admitted interaction have the held-out data and power to be tested within the validated regime, refusing those whose cells fall outside it.
It does not set the model's overall standing complexity budget (dimension_budget, computational_scaling_breakpoint, effective_dimension_statement) — that is Dimension Budget Review, its nearest governance twin, which this gate consumes: the review sets the total budget, while the gate enforces a per-interaction admission policy against it. It also does not rank or select main-effect variables (feature_relevance_boundary, that is Feature Selection Pass).
Related¶
- Instantiates: High-Dimensional Tractability Control — supplies the admission control that stops combinatorial interaction growth from outrunning evidence.
- Consumes: Dimension Budget Review supplies the standing complexity budget the gate charges admitted interactions against.
- Sibling mechanisms: Dimension Budget Review · Feature Selection Pass · Regularized Model Selection · Sparse / Low-Rank Prior · Sample Density Stress Test · Distance Metric Audit · Dimensionality Reduction Probe · Manifold / Embedding Validation · Cross-Validation Under Dimensional Stress
Editorial Notes¶
Form Classification¶
Form family: Decision, Gate & Allocation
Rationale: Interaction Term Gate operates as a case-specific gate, selection, routing, prioritization, or resource disposition because it requires evidence, rationale, and validation capacity before adding cross-feature interactions or segment combinations
Independent corroboration: The frozen evidence defines Interaction Term Gate as 'Requires evidence, rationale, and validation capacity before adding cross-feature interactions or segment combinations', so its operative form is Decision, Gate & Allocation.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Pre-specification, multiplicity control, and evidence requirements for interactions arise from statistical safeguards against false discovery.
Related originating lineages:
- Data Science & Analytics — High-dimensional feature governance and validation-capacity limits materially shape the operational admission gate.
Review resolution: Both independent reviews place the primary lineage in statistics_experimental_design. The queued differences (origin_mode_disagreement, domain_reach_disagreement) concern secondary metadata rather than primary provenance. The final retains data_science only where a reviewer supplied a formative-lineage rationale; this does not convert downstream applicability into origin. origin_mode=cross_disciplinary_synthesis because the entry's present form deliberately composes methods from the documented lineages. domain_reach=multi_domain records application breadth separately from provenance.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] The garden of forking paths (Gelman and Loken): even without deliberate p-hacking, the many analysis choices a researcher could have made in response to the data — which subgroups, which interactions — inflate the false-positive rate, because some comparison will look significant by chance. It is the core reason interactions and subgroups must be pre-specified and multiplicity-adjusted rather than admitted on discovery. ↩a ↩b