Method
Econometric Model
Main Specification (Triple DiD)
GenAI_Adoption_{it} = beta_0
+ beta_1 * Post_ChatGPT_t
+ beta_2 * TechBoard_i
+ beta_3 * (Post x TechBoard)
+ beta_4 * (Post x TechBoard x Vintage_i)
+ gamma * Controls_{it}
+ alpha_i (firm fixed effects)
+ delta_t (time fixed effects)
+ epsilon_{it}Key coefficient: beta_4 captures whether education vintage moderates the effect of board technical competence on GenAI adoption in the post-ChatGPT period.
Hypothesis-Specific Models
H1 (Vintage -> Adoption Speed):
AI_Adoption_Speed_{it} = ... + beta * TKL_{it} + ...Where TKL = Tech Knowledge Latency (Current Year - STEM Degree Year)
H2 (Tech Committee Vintage):
GenAI_Adoption_{it} = ... + beta_1 * TechCommittee_i
+ beta_2 * (TechCommittee x OldVintage_i) + ...H3 (Vintage Faultlines):
GenAI_Adoption_{it} = ... + beta * Vintage_Variance_{it} + ...H4 (Reskilling Mitigation):
GenAI_Adoption_{it} = ... + beta_1 * OldVintage_i
+ beta_2 * (OldVintage x ReskillingIndex_i) + ...Supplementary Analyses
| Analysis | Method | Purpose |
|---|---|---|
| Event Study | Quarterly coefficient plot | Verify parallel trends assumption |
| Heterogeneity | Subgroup DiD | Industry (tech vs. non-tech), firm size effects |
| Instrumental Variable | 2SLS with birth year | Address endogeneity of vintage |
| Robustness | Varying cutoffs | Sensitivity to vintage classification (2010, 2012, 2015) |
| Mediation | Baron-Kenny / Causal Mediation | Board decision process as mediator |
| Placebo | Alternative tech DiD | Cloud/blockchain should show no vintage effect |
Software and Tools
- Statistical software: Stata / R
- Text analysis: Python (for 10-K keyword extraction)
- Visualization: R (ggplot2) / Stata