Multitrait-Multimethod Matrix¶
Validity matrix — instantiates Shared-Source Variance Isolation
Crosses several traits with several measurement methods so that agreement which replicates across methods can be told apart from correlation manufactured by the shared method.
A Multitrait-Multimethod Matrix arranges the correlations you get when several distinct traits are each measured by several distinct methods, so that a shared measurement method shows up as a visible band rather than hiding inside apparent trait agreement. Its defining idea is replication across methods: if two traits correlate only because the same method measured both, that correlation collapses when each trait is re-measured through a different method. The matrix makes the collapse legible by placing same-trait / different-method cells (which should stay high) next to different-trait / same-method cells (which should stay low). It diagnoses method-as-source variance purely from the pattern of correlations — no model is fitted and nothing is residualized.
Example¶
A researcher wants to show that anxiety, conscientiousness, and sociability are three separable traits. She measures each by three methods: a self-report questionnaire, a peer rating, and a structured behavioral task, and assembles the full 9×9 correlation matrix. Convergent validity looks fine along one diagonal — self-reported anxiety agrees with peer-rated and task-measured anxiety. But the same-method blocks are alarming: within the self-report method, anxiety, conscientiousness, and sociability all correlate strongly, even though they should be nearly independent traits, and that inflation does not reappear across methods. The matrix's verdict is that a chunk of the apparent "trait structure" is really self-report style — a single respondent's framing and acquiescence — and that the cleaner trait signal is the weaker, cross-method agreement. She reports the three traits as separable but flags every self-report-only correlation as method-laden.
How it works¶
Define T traits × M methods, collect all T·M measures, and compute the correlation matrix. Then read four cell types: monotrait-monomethod (the reliability diagonal), monotrait-heteromethod (convergent validity — want high), heterotrait-monomethod (same method, different traits — the method-variance tell — want low), and heterotrait-heteromethod (want lowest). Shared-method variance announces itself whenever the heterotrait-monomethod cells sit well above the heterotrait-heteromethod cells. The distinctive move is that the diagnosis is structural and adjustment-free: you learn where the leakage is by contrasting method-matched against method-crossed correlations, not by fitting or subtracting anything.
Tuning parameters¶
- Number of methods — two is the minimum needed to contrast; three or more makes shared-method variance far easier to localize, at the cost of collecting every trait through every method.
- Method dissimilarity — how genuinely different the methods are (self-report versus another self-report is weak; self-report versus a behavioral task is strong). More dissimilar methods make a shared inflation more damning.
- Trait spread — including traits that should be nearly uncorrelated sharpens the discriminant contrast; near-synonym traits blur it.
- Reading rule — eyeballing the Campbell-Fiske comparisons versus a formal decomposition. Formalizing adds rigor but starts to slide into modeling, which is a sibling's job.
When it helps, and when it misleads¶
Its strength is that it needs no statistical model and exposes method variance a single-method study cannot see: it is the cleanest way to answer "would these constructs still look related if a different instrument had measured them?"[1] Its failure mode is the flip side of that power — it demands an expensive multi-method design up front, and with only one method it says nothing at all. It also misleads when the "different" methods secretly share a source (self- and peer-report both filtered through the same social context), which quietly reintroduces the leakage it was meant to expose. The classic misuse is declaring discriminant validity from a matrix whose methods were not actually independent. The guarding discipline is to verify method independence before trusting the contrast, and to treat the matrix as diagnosis, not correction — an informal read on structure, never a fix applied to it.
How it implements the components¶
dimension_claim_frame— the traits are the dimensions whose separateness is being claimed; the whole matrix is built to test exactly that claim.replication_source_contrast— its engine: each trait is re-measured through a contrasting method, and only agreement that replicates across methods is credited as trait signal.
It reads method variance straight off the correlation pattern and fits nothing, so it does not estimate or remove a common component (common_variance_adjustment_rule, precision_weight_update) — that's [Common Factor or Random-Effect Model] — and does not run the post-adjustment residual panel (independence_diagnostic_panel) — that's [Residual Correlation Diagnostic].
Related¶
- Instantiates: Shared-Source Variance Isolation — it supplies the convergent-vs-method-variance read the archetype's interpretation hinges on.
- Consumes: Batch, Rater, or Instrument Counterbalancing Protocol arranges the genuinely independent methods this matrix requires.
- Sibling mechanisms: Source Variance Audit Matrix · Common Factor or Random-Effect Model · Residual Correlation Diagnostic · Negative-Control Outcome Probe · Variance Partitioning Report · Leakage Sensitivity Grid
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Multitrait-Multimethod Matrix operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it crosses several traits with several measurement methods so that agreement which replicates across methods can be told apart from correlation manufactured by the shared method.
Independent corroboration: The frozen evidence defines Multitrait-Multimethod Matrix as 'Crosses several traits with several measurement methods so that agreement which replicates across methods can be told apart from correlation manufactured by the shared method', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Psychology
Origin pattern: Single lineage
Present-day reach: Specialized
Rationale: The multitrait-multimethod matrix was created in psychometrics to distinguish construct convergence from shared-method variance.
Related originating lineages:
- Statistics & Experimental Design — Correlation and latent-variable methods provide its statistical analysis.
Review outcome: Independent reviewer agreement; high confidence.
Notes¶
The matrix's power lives entirely in the quality of its "different methods." It is a design mechanism disguised as an analysis — which is why it consumes the counterbalancing protocol rather than standing alone.
References¶
[1] The multitrait-multimethod matrix is the classic validation device of Campbell and Fiske (1959), which formalized convergent validity (the same trait agreeing across methods) and discriminant validity (different traits staying apart within a method) as the two things a good measure must show at once. registry ↩