What Is Discriminant Validity?
Discriminant validity is the degree to which a measure of one construct is empirically distinct from measures of other, theoretically different constructs. If your customer satisfaction scale correlates almost perfectly with your brand trust scale, one of two things is happening: either satisfaction and trust aren't really separate concepts, or your measures aren't capturing the distinction between them. Discriminant validity testing asks whether the constructs your research treats as separate are actually measured as separate by your instruments. It's the flip side of convergent validity, while convergent validity shows that your measure agrees with related measures, discriminant validity shows that it disagrees (or at least doesn't agree too much) with measures of different constructs. Together, they provide the core evidence for construct validity.
Why Discriminant Validity Matters in Research
If two supposedly distinct constructs are measured by scales that correlate at r = 0.95, they're functionally the same variable in your data, treating them as separate predictors, mediators, or outcomes in your analysis is misleading. Discriminant validity ensures that the conceptual distinctions in your theoretical framework actually show up in your measurement. Without it, you risk building complex models with redundant variables, or segmenting audiences on dimensions that aren't really different.
How Discriminant Validity Works
Several statistical approaches assess whether measures capture truly distinct constructs.
Fornell-Larcker Criterion
The most commonly cited test compares the average variance extracted (AVE) for each construct with the squared correlation between constructs. If a construct's AVE is greater than its squared correlation with every other construct, it shares more variance with its own indicators than with other constructs, evidence of discriminant validity. For example, if satisfaction has an AVE of 0.60 and its correlation with trust is 0.70 (squared = 0.49), discriminant validity holds because 0.60 > 0.49.
Heterotrait-Monotrait Ratio (HTMT)
The HTMT ratio, introduced by Henseler, Ringle, and Sarstedt in 2015, compares the average correlations between items of different constructs (heterotrait-heteromethod) to the average correlations between items within the same construct (monotrait-heteromethod). An HTMT value below 0.85 (conservative) or 0.90 (liberal) suggests discriminant validity. HTMT has been shown to outperform the Fornell-Larcker criterion in detecting discriminant validity problems, particularly when factor loadings are uniform.
Confidence Interval Approach
If you calculate the correlation between two constructs in a CFA model and the 95% confidence interval includes 1.0, discriminant validity is questionable, the two constructs may not be statistically distinguishable. This approach is straightforward and doesn't require additional computations beyond what a standard CFA provides.
Cross-Loadings
In factor analysis, items should load substantially on their intended factor and weakly on other factors. If items from your satisfaction scale load nearly as strongly on the trust factor as on the satisfaction factor, the two constructs aren't being measured distinctly. A rule of thumb: items should load at least 0.20 higher on their intended factor than on any other factor.
Multitrait-Multimethod Matrix
The MTMM approach compares same-trait correlations (convergent) with different-trait correlations (discriminant) across different measurement methods. Convergent coefficients should exceed discriminant coefficients systematically throughout the matrix. This method is thorough but requires measuring all constructs with at least two different methods.
When to Use Discriminant Validity Assessment
- Multi-construct survey development: whenever your instrument claims to measure two or more separate constructs (like satisfaction and loyalty, or engagement and commitment), discriminant validity should be tested
- Structural equation modeling: before interpreting paths between latent variables, ensure the variables are empirically distinct
- Brand perception research where you're measuring multiple brand attributes that might overlap (trust, quality, reliability)
- Employee experience research where engagement, satisfaction, and commitment are often treated as separate outcomes
- Evaluating existing scales before using them together in a study, published scales occasionally lack discriminant validity when combined in new contexts
Common Mistakes to Avoid
- Only checking the Fornell-Larcker criterion while ignoring newer, more sensitive tests like HTMT, the Fornell-Larcker criterion can pass even when discriminant validity is genuinely problematic, especially with similar factor loadings across items
- Assuming theoretical distinctness guarantees empirical distinctness: just because two constructs are defined differently in the literature doesn't mean your measures capture that difference; test, don't assume
- Treating minor discriminant validity failures as catastrophic: moderate correlations between related constructs (r = 0.70-0.80) may be theoretically expected; the concern is when correlations approach 1.0 or AVE/HTMT thresholds are clearly violated
How Quali-Fi Supports Discriminant Validity
Quali-Fi's survey platform supports multi-construct measurement designs with matrix questions, section randomization, and flexible scale configurations, letting you collect the multi-scale data that discriminant validity analysis requires. Export to SPSS or CSV for CFA-based discriminant validity testing, or use the platform's built-in cross-tabulation tools for preliminary correlation checks between constructs during data collection.
Frequently Asked Questions
What's the difference between discriminant validity and divergent validity?
They're the same concept, different fields and textbooks use different terms. Both refer to demonstrating that a measure is empirically distinct from measures of theoretically different constructs. "Discriminant validity" is the more widely used term in psychometrics and marketing research.
What do you do when discriminant validity fails?
First, examine whether the constructs are genuinely distinct or whether your theory needs revision, maybe satisfaction and loyalty really are one construct in your context. If the constructs should be distinct, look at the items: are some items cross-loading because they're ambiguously worded? Remove or revise problematic items and retest. In some cases, combining the two constructs into one and measuring it with a single scale is the most honest solution.
Can discriminant validity be assessed with exploratory factor analysis?
Yes, as a preliminary check. If items from two constructs load on the same factor in EFA, discriminant validity is likely poor. However, CFA provides a more rigorous test because it allows you to formally compare model fit with and without the constraint that the constructs are distinct.
Related Topics
- Convergent Validity
- Construct Validity
- Criterion Validity
- Face Validity
- Reliability in Research
- Predictive Validity
Measuring multiple constructs in one survey? See how Quali-Fi's multi-scale survey tools support the data collection that rigorous validity testing requires.