Convergent and Discriminant Validity

Convergent and discriminant validity are the two halves of construct validity evidence. A measure has convergent validity when it correlates with other measures of the same or a closely related construct. It has discriminant validity (also called divergent validity) when it does not correlate strongly with measures of conceptually distinct constructs it should be distinguishable from. Neither is sufficient alone — a scale that correlates with everything demonstrates convergence without discrimination, and a scale that correlates with nothing demonstrates discrimination without convergence. This guide covers how to define each, how they were originally tested together with the multitrait-multimethod (MTMM) matrix, and the modern quantified thresholds (AVE, Fornell-Larcker, HTMT) researchers report today.

Convergent Validity: Correlating With What It Should

Convergent validity is the evidence that a measure correlates with other measures of the same or a theoretically related construct, at a magnitude consistent with theory. If a new test-anxiety scale correlates strongly (say, r > 0.50) with an established general-anxiety inventory and with self-reported physical symptoms during exams, that pattern supports convergent validity. The correlation does not need to approach 1.0 — different instruments measuring the same construct through different methods (self-report, observer rating, physiological measure) are expected to agree substantially but not perfectly, since each also carries method-specific variance.

A common source of confusion: convergent validity is not the same claim as reliability. Two administrations of the same instrument correlating with each other is test-retest reliability. Convergent validity requires a genuinely different measure of the same construct, ideally using a different method.

Discriminant Validity: Not Correlating With What It Shouldn’t

Discriminant validity is the mirror-image evidence: the measure does not correlate strongly with measures of constructs it is theoretically supposed to be distinct from. A test-anxiety scale that correlates almost as strongly with general life satisfaction, or with an unrelated cognitive-ability measure, as it does with other anxiety measures raises a construct-validity problem — it may be capturing generalized negative affect, or a response-style artifact, rather than test anxiety specifically.

Discriminant validity is a statement about a pattern of correlations relative to other correlations, not an absolute threshold on its own. A moderate correlation (r = 0.30) between two theoretically related-but-distinct constructs (e.g., test anxiety and general anxiety) can still support discriminant validity if it is clearly lower than the convergent correlations for either construct with its own alternate measures.

Why Both Matter Together

Donald Campbell and Donald Fiske, in a widely cited 1959 Psychological Bulletin paper, argued that neither type of evidence means much in isolation, and that testing them together also exposes a third problem: method variance. Two measures can correlate not because they tap the same trait, but because they share the same method — two self-report scales might correlate partly because both are vulnerable to the same social-desirability or acquiescence bias, independent of what they claim to measure. Campbell and Fiske’s answer was a design that crosses multiple traits with multiple methods at once: the multitrait-multimethod matrix.

The Multitrait-Multimethod (MTMM) Matrix

The MTMM design measures several traits (constructs), each using several methods (e.g., self-report, peer-report, behavioral/observational task). Every measure-pair correlation in the resulting matrix falls into one of four types:

• Validity diagonal (monotrait-heteromethod) — same trait, different methods. This is the convergent-validity evidence: these correlations should be the highest in each row/column.
• Heterotrait-monomethod — different traits, same method. High values here suggest method variance is inflating agreement.
• Heterotrait-heteromethod — different traits, different methods. This is the strictest discriminant-validity comparison: these should be the lowest correlations in the matrix.
• Reliability diagonal (monotrait-monomethod) — the same measure with itself (test-retest or internal-consistency reliability), included as the reference ceiling every other value is judged against.

Campbell and Fiske’s original four criteria for a well-behaved matrix, restated in modern terms:

1. Validity-diagonal (convergent) correlations should be significantly different from zero and reasonably large.
2. A validity-diagonal correlation should be higher than the heterotrait-heteromethod correlations in the same row and column.
3. A validity-diagonal correlation should be higher than the heterotrait-monomethod correlations for the same trait pair.
4. The same pattern of trait interrelationships should hold across the heterotrait triangles, regardless of method.

A Worked Example

The table below is a simulated illustrative dataset (not real study data), generated for this guide from 200 simulated respondents: each observed score was built from an independent trait factor, an independent method factor, and random error, then correlated. It is included to show what a clean MTMM pattern looks like, not to report a real finding. Two traits are shown, each measured three ways:

TA = Test Anxiety, VR = Verbal Reasoning. Reading it against Campbell and Fiske’s criteria: the convergent (validity-diagonal) values — TA–TA across methods at 0.23, 0.22, 0.33, and VR–VR across methods at 0.37, 0.37, 0.33 — are all clearly higher than the heterotrait-heteromethod values in the same rows and columns (which cluster near zero, several even negative: -0.05, -0.09, -0.02). They are also higher than the heterotrait-monomethod values (TA–VR within the same method: 0.08, 0.16, 0.09). That combination — convergent correlations standing well above both heterotrait triangles — is what a construct-validity argument built on an MTMM matrix is actually looking for. If the heterotrait-monomethod values had instead been the largest correlations in the matrix, that would point to shared method variance driving the pattern rather than the traits themselves.

The Modern, Quantified Approach: AVE, Fornell-Larcker, and HTMT

The classic MTMM matrix requires multiple methods per trait, which is often impractical outside dedicated validation studies. In confirmatory factor analysis and structural equation modeling, three statistics are now more commonly reported as quantified convergent/discriminant evidence from a single-method, multi-item dataset:

• Average variance extracted (AVE) — the average proportion of variance in a construct’s indicators explained by the construct itself rather than by error. An AVE of 0.50 or higher is conventionally treated as adequate convergent validity.
• Fornell-Larcker criterion — discriminant validity is supported when a construct’s AVE (or its square root) exceeds its correlations with every other construct in the model.
• Heterotrait-monotrait ratio (HTMT) — a diagnostic from the partial least squares SEM literature, shown to be more sensitive than Fornell-Larcker at detecting discriminant-validity failures, particularly with reflective indicators.

CASRAI’s guide to calculating AVE and applying Fornell-Larcker and HTMT covers the formulas, thresholds, and a worked numeric example in depth — use that page when you need the quantified SEM-based approach; this page is the conceptual and MTMM-design starting point both build on.

Where This Fits in the Broader Construct-Validity Argument

Convergent and discriminant validity are two of several evidence types that together make up a construct-validity argument — alongside nomological-network testing, known-groups comparisons, and factorial validity. No single correlation or coefficient “proves” construct validity; it accumulates across studies. CASRAI’s construct validity guide covers the full evidence-type framework and the two practical threats (construct underrepresentation and construct-irrelevant variance) that undermine it.

Construct validity is also distinct from, though related to, content validity (does the instrument cover the whole domain of the construct) and criterion validity (does it predict a real-world outcome). CASRAI’s overview of validity types maps how all of these relate to each other and to design-level validity (internal, external, statistical-conclusion).

Practical Guidance for Reporting

• Name the specific evidence, not just “the scale was validated.” Report which correlations or coefficients support convergence and which support discrimination, with the actual values.
• Choose measures deliberately when planning a validation study. A convergent measure should be a genuinely different operationalization of the same construct, not a near-duplicate of your own instrument. A discriminant comparator should be theoretically related enough that low correlation is informative — comparing against something obviously unrelated (e.g., shoe size) proves nothing.
• If you only have single-method, multi-item data, report AVE, Fornell-Larcker, and ideally HTMT rather than claiming an MTMM analysis you didn’t run.
• If you’re running a true MTMM study, report the full matrix (or at minimum the validity-diagonal and heterotrait values), not just a summary claim that “convergent and discriminant validity were established.”

Frequently Asked Questions

What is the difference between convergent and discriminant validity?

Convergent validity is evidence that a measure correlates with other measures of the same construct. Discriminant validity is evidence that it does not correlate strongly with measures of conceptually distinct constructs. Both are needed to support that a measure captures its intended construct specifically, rather than something broader or something shared with unrelated measures.

What counts as a “high enough” convergent correlation?

There is no single universal cutoff — it depends on the constructs and methods involved. What matters more than any absolute value is the relative pattern: convergent correlations should clearly exceed the discriminant correlations for the same measures, per Campbell and Fiske’s criteria. When quantified via CFA, an AVE of 0.50 or higher is the conventional benchmark.

Can a measure have convergent validity but not discriminant validity?

Yes, and it is a common failure pattern. A scale that correlates strongly with everything — both the constructs it should relate to and the ones it shouldn’t — has convergent evidence without discriminant evidence, which usually signals the scale is capturing something too broad (e.g., general negative affect instead of a specific construct) or that shared method variance is inflating every correlation.

Do I need a full MTMM study to establish discriminant validity?

No. A full multi-method MTMM design is the strongest evidence but is often impractical. Reporting AVE alongside the Fornell-Larcker criterion or HTMT from a single-method CFA/SEM model is the standard, widely accepted substitute in most published validation studies today.

Who introduced the multitrait-multimethod matrix?

Donald T. Campbell and Donald W. Fiske, in “Convergent and discriminant validation by the multitrait-multimethod matrix,” Psychological Bulletin, 1959.