Bayesian Discriminant Validity Assessment
Also known as: Bayesian HTMT, Bayesian HTMTb, Bayesian discriminant evidence, Bayesian CFA discriminant validity
Bayesian discriminant validity assessment evaluates whether two theoretically distinct latent constructs are empirically separable, using posterior distributions and credible intervals rather than single-point null-hypothesis tests. It is applied within Bayesian confirmatory factor analysis or via the Bayesian heterotrait-monotrait ratio (HTMTb) to determine whether constructs measuring different traits are sufficiently differentiated.
Key highlights
- Yields a full posterior distribution over the discriminant validity index, providing richer evidence than a single point estimate or p-value.
- Credible intervals directly communicate the probability that discriminant separation meets or falls short of a threshold, aiding transparent reporting.
- Performs well in small-to-moderate samples where frequentist HTMT confidence intervals are unstable, provided reasonable priors are used.
- Accommodates ordinal item data naturally through appropriate likelihood models (e.g., ordered-probit CFA), unlike standard ML-CFA which assumes continuous indicators.
- Can incorporate prior knowledge from earlier studies or meta-analyses to improve precision of the inter-factor correlation estimates.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Bayesian discriminant validity when you have a multi-factor scale and want to verify that the latent constructs are empirically distinct, especially when sample sizes are small or moderate (under 200), where frequentist tests are unstable or underpowered. It is particularly valuable when prior research provides informative expectations about inter-factor correlations, allowing those priors to stabilize estimates. Do not use it as a substitute for theory: if two constructs are expected to be highly correlated on conceptual grounds, a high inter-factor correlation is not evidence of poor discriminant validity but of theoretical convergence. Avoid this method when you lack the software expertise or computing resources for MCMC estimation; in those cases, frequentist CFA with HTMT and bootstrap confidence intervals is a pragmatic alternative.
Strengths & limitations
- Yields a full posterior distribution over the discriminant validity index, providing richer evidence than a single point estimate or p-value.
- Credible intervals directly communicate the probability that discriminant separation meets or falls short of a threshold, aiding transparent reporting.
- Performs well in small-to-moderate samples where frequentist HTMT confidence intervals are unstable, provided reasonable priors are used.
- Accommodates ordinal item data naturally through appropriate likelihood models (e.g., ordered-probit CFA), unlike standard ML-CFA which assumes continuous indicators.
- Can incorporate prior knowledge from earlier studies or meta-analyses to improve precision of the inter-factor correlation estimates.
- Requires specification of prior distributions for loadings and inter-factor correlations; poorly chosen priors can bias results, especially in small samples.
- MCMC estimation is computationally intensive and requires more time and technical skill than standard maximum-likelihood CFA.
- The choice of HTMT threshold (0.85 vs. 0.90) remains a matter of convention rather than formal theory, and Bayesian framing does not resolve this ambiguity.
- Software options (blavaan, Stan, JAGS) have steeper learning curves than frequentist SEM packages, limiting accessibility.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is the HTMT threshold for discriminant validity?
Henseler et al. (2015) proposed 0.85 as a conservative threshold and 0.90 as a more lenient one. In the Bayesian framework, the question becomes: what is the posterior probability that HTMTb exceeds 0.85 (or 0.90)? A probability below about 5% is typically taken as support for discriminant validity. The choice of threshold remains a disciplinary convention, not a statistical law.
How is Bayesian discriminant validity different from the standard CFA approach?
Standard (frequentist) CFA tests discriminant validity using a point estimate of the inter-factor correlation with a 95% confidence interval, or HTMT with a bootstrap CI. Bayesian discriminant validity produces a posterior distribution for the same quantities, yielding probability statements such as 'there is a 97% posterior probability that HTMTb < 0.85', which is more directly interpretable than a frequentist confidence interval.
Do I need a large sample to use Bayesian discriminant validity?
No — this is one of its main advantages. With informative or weakly informative priors, Bayesian estimation can yield stable inter-factor correlation estimates in samples as small as 50–100, where frequentist bootstrap CIs for HTMT are unstable. Larger samples (200+) are still preferable for overall model stability, but Bayesian methods tolerate smaller samples better than maximum-likelihood CFA.
Which software can I use for Bayesian discriminant validity?
The blavaan package in R is the most accessible entry point, as it mirrors the lavaan syntax but adds Bayesian estimation via Stan or JAGS. Alternatively, Stan or PyMC can be used for fully custom CFA models. For the HTMT index, manual computation from posterior draws is straightforward once the factor correlation posterior is obtained.
What if the credible interval for the inter-factor correlation overlaps 0.85?
This is an ambiguous result: the data are consistent with both adequate and inadequate discriminant validity. Options include collecting more data, reconsidering whether the two constructs are truly conceptually distinct, revising or dropping items that cross-load heavily between factors, or reporting the uncertainty transparently and letting substantive theory guide the interpretation.
Sources
- 1.Garnier-Villarreal, M. & Jorgensen, T. D. (2020). Adapting fit indices for Bayesian structural equation modeling: Comparison to maximum likelihood. Psychological Methods, 25(1), 46–70.
- 2.Campbell, D. T. & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 56(2), 81–105.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Discriminant Validity. ScholarGate. https://scholargate.app/psychometrics/bayesian-discriminant-validity