How Virtual Celebrity Characteristics Drive Purchase Intention: Testing the Stimulus-Organism-Response Framework with Structural Equation Modeling

This study analyzes how virtual celebrity attributes influence consumer satisfaction, royalty, purchase intention, and how personalization strengthens satisfaction-driven royalty. Existing studies identify alternative means and the credibility, but inadequately explain process outcomes. This study clarifies the mechanisms via SOR framework, mediation, personalization, moderation.

To begin, define the structural model for Partial Least Squares Structural Equation Modeling, or PLS-SEM. Specify a maximum of six predictors for the endogenous constructs, including three virtual celebrity characteristics and three control variables. Set the statistical parameters, such as medium effect size, significance level, and statistical power for the power analysis.

Now, run the power analysis to estimate the minimum required sample size. Recruit participants through an online research panel using purposive non-probability sampling. Restrict eligibility to individuals aged 18 years or older.

Insert screening questions and attention check items in the survey to verify eligibility and response quality. Review survey responses and remove entries that fail screening or attention check criteria. Stop data collection once the predetermined target sample size is achieved.

Start by presenting a definition describing virtual celebrities as digitally created audio-visual personas with stable visual identity, voice, and recurring narrative content. Instruct participants to answer survey questions based on their actual experiences with such virtual celebrities. Develop the questionnaire by including survey sections covering appearance, voice, content quality, satisfaction, loyalty, personalization, purchase intention, and demographics.

Measure all constructs using 7-point Likert scales where one indicates strongly disagree and seven indicates strongly agree. Prevent duplicate submissions through platform control settings. Export the collected responses in comma-separated values or CSV format and import the cleaned dataset into the PLS-SEM software.

For measurement and structural model evaluation, specify all constructs as reflective constructs. Evaluate reliability using Cronbach's alpha and composite reliability or CR greater than or equal to 0.7. Assess convergent validity using average variance extracted or AVE greater than or equal to 0.5.

Evaluate discriminant validity using the Fornell-Larcker criterion and Heterotrait-Monotrait or HTM ratio and ensure that it is less than 0.9. Specify the structural paths according to the conceptual model and assess multi-collinearity using the variance inflation factor or VIF values below five. Run bootstrapping with 5, 000 subsamples to estimate path coefficients, T-values, confidence intervals, and effect sizes.

Estimate indirect effects using bootstrapping with 5, 000 subsamples and 95%confidence intervals to test the sequential mediation from virtual celebrity attributes to satisfaction, satisfaction to loyalty, and loyalty to purchase intention. Create interaction terms to test the moderating effects of personalization. Estimate moderation coefficients and affect sizes using bootstrapping.

Generate interaction plots to visualize conditional effects. After predictive performance assessment, compare Q-squared predict values, root mean square error, and mean absolute error against a linear benchmark model. Confirm predictive relevance when Q-squared predict values exceed zero and PLS errors are lower than the benchmark.

Reliability and convergent validity tests showed that all constructs exceeded recommended thresholds, confirming the measures are both reliable and valid. Discriminant validity assessed using the Fornell-Larcker criterion and Heterotrait-Monotrait ratio confirms that all constructs are empirically distinct. Structural model evaluation indicates moderate explanatory power and acceptable predictive performance.

Predictive relevance is confirmed with Q-square predict values ranging from 0.280 to 0.485. The predictive performance of the PLS-SEM model is compared with a linear regression benchmark. Lower root mean square error and mean absolute error values indicate better prediction accuracy.

Statistical analysis identifies which hypotheses in the model are supported by the data. Appearance, voice, and content quality significantly influence customer satisfaction. Satisfaction increases loyalty, which in turn leads to higher purchase intention.

Mediation analysis indicates that satisfaction affects purchase intention indirectly through loyalty. Personalization strengthens the relationship between satisfaction and loyalty, but does not affect purchase intention. The protocol can help measure how virtual celebrity attributes affect satisfaction, loyalty, and purchase intention using surveys and structural modeling.

In this protocol, key challenges included recruiting participants virtual celebrity and clarifying inconsistent relationships between satisfaction and purchase intention reports in prior studies. This research facilitates future studies on emerging AI technologies across cultural differences and the long-term effect of personalization and virtual celebrity trait.