extradimen/llm_big5_ad_sem: Simulating Personality-Based Advertising Responses Using AI Agents and Structural Equation Modeling

This record contains the data and complete reproducibility package supporting the revised manuscript on persona-conditioned large-language-model simulations of advertising attitude and purchase intention. The prespecified primary analysis uses Chinese-culture, promotion-framed responses generated with DeepSeek-V2 236B across ten generation runs. The Big Five values were assigned before prompting and are therefore analyzed as observed experimental condition scores rather than reflective latent variables. The reproducibility package includes source and analysis-ready datasets, data-processing scripts, deterministic statistical-analysis scripts, measurement diagnostics, primary path estimates, run-stratified bootstrap indirect effects, a post-hoc measurement-screened sensitivity analysis, exploratory culture–model configuration comparisons, tables, and manuscript Figures 1–7. The primary analysis uses standardized observed-variable path regressions with generation-run fixed effects, HC3 heteroskedasticity-consistent standard errors and confidence intervals, and 2,000 run-stratified percentile bootstrap resamples with fixed seed 20251010. The primary source files contain 4,987 simulated response instances. A total of 4,695 complete cases were available for the purchase-intention equation. The deposited records are generated LLM responses and do not contain human-participant data. The corresponding version-controlled repository is available at:https://github.com/extradimen/llm_big5_ad_sem To reproduce the processed datasets, quality reports, measurement diagnostics, statistical results, and generated manuscript figures, run: python scripts/run_all.py