Analyzing scenario-based questionnaires—also known as vignette studies—with Structural Equation Modeling (SEM) is a highly effective way to measure how specific, situational variables trigger psychological, behavioral, or decision-making responses, while simultaneously accounting for measurement error and unobserved latent traits.
The recommended analytical process follows a step-by-step approach tailored to the unique structure of vignette designs.
1. Structure of Vignette Data
In a vignette study, respondents are exposed to hypothetical scenarios where specific attributes (cues) are experimentally manipulated.
- Within-Subjects Design: Each respondent evaluates multiple scenarios.
- Level 1 Data: The vignette-level variables (e.g., specific scenarios, manipulated cues).
- Level 2 Data: The respondent-level variables (e.g., demographics, baseline attitudes).
2. The Two-Step SEM Approach
To analyze this type of survey data, the analysis follows the standard two-step SEM process: the measurement model and the structural model.
Step A: Validate the Measurement Model (CFA)
Before testing structural paths, you must confirm that the questionnaire items accurately measure the unobserved constructs (latent variables) related to the scenarios (e.g., Perceived Risk, Empathy, Intent to Purchase).
- Confirmatory Factor Analysis (CFA): Load the scenario responses into your latent constructs.
- Validity & Reliability: Check the Composite Reliability (CR) (should be >0.7) and Average Variance Extracted (AVE) (should be >0.5).
- Discriminant Validity: Ensure that the constructs are distinct from one another, usually by confirming that the square root of the AVE for each construct is greater than its correlations with other constructs.
Step B: Estimate the Structural Model
Once the measurement model fits properly, you connect the latent variables with directional pathways to test your theoretical hypotheses.
- Direct Effects: How specific scenario conditions influence a participant's reaction.
- Mediation/Moderation: Assessing whether the effect of a scenario cue on behavior is transmitted through an emotional or cognitive variable.
3. Key Considerations for Vignette-SEM
Handling scenario data requires specialized statistical considerations to avoid violating assumptions of independence:
- Multilevel SEM (MLSEM): Because respondents evaluate multiple vignettes, observations within the same person are correlated. Multilevel SEM accounts for both within-person (Level 1) and between-person (Level 2) variance.
- SEM for Within-Subject Experiments: Often used in factorial surveys, this isolates the effects of specific manipulated vignette attributes from individual baseline traits by treating the vignette situations as experimental conditions.
You must evaluate how well your hypothesized model reproduces the observed covariance matrix. Standard thresholds for good model fit include:
- Chi-Square (chi ^2): Should ideally be non-significant, though it is heavily impacted by large sample sizes.
- RMSEA (Root Mean Square Error of Approximation): Values should be <0.05 (acceptable up to 0.08).
- CFI (Comparative Fit Index): Values should be (> 0.95).
- SRMR (Standardized Root Mean Residual): Values should be <0.08.
Depending on your exact modeling needs (e.g., Covariance-based vs. Variance-based SEM), you can utilize several tools:
- Mplus: The gold standard for handling complex multilevel SEM and categorical outcomes typical of scenario-based choices.
- lavaan: A robust, free, and open-source R package explicitly designed for latent variable analysis.
- SmartPLS: Ideal if your vignette data is non-normal, highly complex, or you are using formative (rather than reflective) indicators for your scenario constructs.
I've explained the working of