Model Testing Research — Structural Theory Testing Design
- Introduced by
- Karl G. Joreskog (SEM/LISREL framework); formalized through structural equation modeling tradition
- Year
- 1970s (Joreskog 1969–1973); widely adopted in social sciences by the 1980s–1990s
- Type
- Confirmatory quantitative research design
- Family
- Survey / observational design
What does it do?
Model testing research is a confirmatory quantitative design in which the researcher specifies a theoretical model — depicting hypothesized relationships among constructs — and then tests how well that model fits empirical data.
Drawing primarily on structural equation modeling (SEM) and confirmatory factor analysis (CFA), it evaluates whether the data-implied covariance structure is consistent with the theoretically derived one, yielding fit indices that indicate model-data correspondence.
The intuition behind the method
Think of model testing research as taking a blueprint and checking it against the actual building. You do not begin by asking 'what is going on in this data?' — you begin with a theoretically derived picture of how variables should relate to one another, then collect data and ask: does the world look like my blueprint? Fit indices tell you how close the building matches the plan, and modification indices help you identify where the blueprint needs revision. The entire logic is deductive and confirmatory: theory first, data second.
History
The statistical foundation for model testing research was laid by Karl Joreskog, whose work on confirmatory factor analysis (1969) and covariance structure analysis (1973) led to the LISREL software package. During the 1980s, SEM proliferated across psychology, education, management, and sociology, popularized by researchers such as Peter Bentler (EQS software, 1985) and Michael Browne. Rex Kline's textbook, first published in 1998 and now in its fourth edition (2015), remains the standard methodological reference for applied social science researchers.
When to use it
Use model testing research when you have a well-developed theory that specifies how multiple latent constructs relate, and you need to assess whether the theory is consistent with population-level covariance data. It is ideal for validating measurement instruments (CFA), testing mediation and moderation chains, and comparing competing theoretical frameworks. Require at least N = 200 and multiple indicators per construct. Do not use it as a first step when theory is nascent — exploratory factor analysis or grounded qualitative work should precede model specification. Avoid it when sample size is small, when constructs have only one indicator, or when the goal is purely descriptive rather than theory-driven.
Limitations and considerations
- Requires a large sample (typically N >= 200) that many researchers cannot feasibly obtain.
- Model fit can appear acceptable even when the theoretical interpretation is incorrect — fit indices confirm statistical consistency, not theoretical truth.
- Highly sensitive to model misspecification; omitting a theoretically relevant path inflates other coefficients and distorts fit.
- Assumes multivariate normality under ML estimation; violations require robust estimators (MLR, WLSMV) that are less familiar to many applied researchers.
- Post-hoc model modification without cross-validation converts a confirmatory study into an exploratory one, undermining its scientific value.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Tests entire theoretical frameworks simultaneously rather than isolated bivariate hypotheses.
- Separates measurement error from structural relationships, yielding more precise parameter estimates than regression alone.
- Allows formal comparison of competing models using fit indices and chi-square difference tests.
- Supports assessment of mediation, moderation, and indirect effects within a single integrated framework.
- Produces standardized and unstandardized coefficients that are directly interpretable in theoretical terms.
- Widely recognized and accepted in high-impact journals across social, behavioral, and health sciences.
Applications
- Validating psychometric scales and assessing convergent and discriminant validity through CFA.
- Testing structural mediation models — for example, whether the effect of job demands on burnout is mediated by emotional exhaustion.
- Comparing factor structures or path coefficients across demographic groups via multi-group SEM.
- Examining longitudinal growth trajectories and reciprocal causation through cross-lagged panel models.
- Testing theoretical frameworks in organizational behavior, health psychology, and educational research.
- Evaluating competing theoretical accounts by comparing their fit to the same dataset using AIC or BIC.
How it works
How it works
The researcher articulates a formal theoretical model — often depicted as a path diagram — that specifies which latent constructs exist, how observed variables (indicators) relate to those constructs (measurement model), and how the constructs relate to one another (structural model). Every path included or excluded reflects a substantive theoretical commitment, not a data-driven choice.
Each latent construct is operationalized by selecting or developing multi-item scales with established or hypothesized reliability and validity. Sample size requirements are estimated (a common benchmark is 10–20 cases per free parameter), and the sampling plan is aligned with the target population to support generalizability.
Data are collected — typically via self-report surveys or archival measures — from a sample large enough to meet estimation requirements (commonly N >= 200 for SEM). Multivariate normality, missing data patterns, and outlier profiles are examined before proceeding to model estimation.
The specified model is fitted to the observed covariance matrix using an estimation method such as maximum likelihood (ML). The estimation yields a reproduced covariance matrix; discrepancy between the observed and reproduced matrices is summarized by fit statistics: chi-square test, CFI, TLI, RMSEA, and SRMR.
Formula
\chi^2, \text{CFI}, \text{RMSEA}, \text{SRMR}Model fit is evaluated against conventional benchmarks (e.g., CFI >= 0.95, RMSEA <= 0.06, SRMR <= 0.08). Individual path coefficients, standard errors, and R-squared values are interpreted in light of theoretical predictions. Poor fit signals that the theoretical model does not adequately represent the data-generating process.
When fit is inadequate, modification indices may suggest theoretically justifiable amendments — adding paths, freeing correlated residuals, or reconsidering the measurement structure. Any post-hoc modifications must be re-validated on an independent sample to guard against capitalization on chance; otherwise the study shifts from confirmatory to exploratory.
Common pitfalls
- Specifying the model after inspecting the data — this transforms model testing into model building; the confirmatory logic is invalidated.
- Accepting mediocre fit as adequate by selectively reporting only favorable fit indices while omitting poor ones.
- Treating modification indices as a recipe for improvement without theoretical justification — each added path should reflect substantive reasoning, not mere statistical improvement.
- Using too few indicators per construct (single-indicator models cannot separate construct variance from measurement error, negating the key advantage of SEM).
- Ignoring invariance testing when comparing groups — factor loadings and intercepts must be shown to be equivalent across groups before latent mean differences are meaningful.
Frequently asked questions
How is model testing research different from confirmatory research in general?
Confirmatory research is a broad category that includes any study designed to test a predetermined hypothesis. Model testing research is a specific type of confirmatory design that evaluates an entire system of hypothesized relationships simultaneously — typically using SEM or CFA — rather than testing individual hypotheses one at a time. The defining feature is the formal specification of a structural or measurement model before data are collected.
What sample size do I need?
A common rule of thumb is 10–20 observations per freely estimated parameter, with an absolute minimum often cited as N = 200 for stable ML estimates. Complex models with many latent variables and paths require larger samples. Power analysis software such as WebPower or Monte Carlo simulation in Mplus can provide more precise sample size estimates for a given model structure.
What fit indices should I report and what are acceptable values?
Report at minimum: chi-square with degrees of freedom and p-value, CFI or TLI (acceptable >= 0.90, good >= 0.95), RMSEA with 90% confidence interval (acceptable <= 0.08, good <= 0.06), and SRMR (acceptable <= 0.10, good <= 0.08). No single index is definitive; convergent evidence across multiple indices is required. Chi-square alone is overly sensitive in large samples.
Can I modify my model after seeing the data and still call it model testing?
If modifications are made on the basis of modification indices or residuals from the same dataset, the study has shifted from confirmatory to exploratory. Post-hoc modifications are permissible only if cross-validated on an independent holdout sample. Otherwise, the study should be described as exploratory model development, not model testing.
Do I need to use SEM software, or can I use regression instead?
For simple models with observed variables only, multiple regression may suffice. However, regression cannot accommodate latent variables, separate measurement error, or test overall model fit. Dedicated SEM software — Mplus, lavaan (R), AMOS, or LISREL — is required whenever the model includes latent constructs, indirect effects, or multiple dependent variables measured simultaneously.
Related methods
Sources
Kline, R. B. (2015). Principles and Practice of Structural Equation Modeling (4th ed.). Guilford Press.
ISBN: 978-1462523344Copied.Joreskog, K. G., & Sorbom, D. (1993). LISREL 8: Structural Equation Modeling with the SIMPLIS Command Language. Scientific Software International.
Open publication referenceCopied.