← Research
Data
Pooled RCT dataset provided by Diego Jannace. Cleaned and merged at baseline (P0). Moderator set covers demographics, experience, venture characteristics, illusion of control, uncertainty beliefs, idea scope, and dosage. Pending: cognitive battery (ambiguity aversion, risk aversion, learning orientation) — requires recovery from raw country-level files.
Framework — Three Categories
Person-level
SDM capability, ambiguity aversion, learning orientation, experience, illusion of control, uncertainty beliefs
Venture-level
BM development stage, sector, team composition, idea scope
Dosage
Attendance, sessions completed, mentor engagement
Method: GenericML (Chernozhukov et al., Econometrica 2025) run separately by category —
variable importance to select moderators → interaction OLS. Reference: McKenzie et al., VoxDevLit 2025.
Status — May 2026
Currently in the analysis phase. Two main analyses complete (learning HTE across sub-indices and periods, performance HTE via full outcome scanner). Writing has not started. Results are being consolidated before any presentation.
Analyses running / complete
- GenericML scanner — BLP, GATES, CLAN across 14 outcomes × 26 moderators × 2 comparisons (TE vs EB, TE vs C), 250 splits
- SI sub-index trajectories and mediation analysis (T → ΔSI → pivot)
- Translation HTE within treated arm (ΔSI → pivot heterogeneity)
- Control arm dynamics and attrition HTE
- OLS interaction robustness (TFP × moderators)
Pending
- Mechanism test: T → ΔSI → performance, heterogeneous by founder type
- Cognitive battery (Diego): risk aversion, uncertainty aversion, learning orientation — only untested moderator family
- P1 behavioral robustness: restrict pivoting outcomes to P1 before attrition diverges
- Targeting welfare analysis
Progress Log
2026-05-20
GenericML scanner re-run and confirmed stable. Figures and tables re-exported.
2026-05-19
Analysis 02 complete: full HTE scanner across all 14 outcomes, 26 moderators, 2 comparisons (TE vs EB, TE vs C). ITT fix applied. Onepager v10 and seminar slides v2 written.
2026-05-18
Methodology note written (RF vs OLS vs GenericML). SI sub-index deep exploration complete: trajectories, correlation structure, mediation (T → ΔSI → pivot), sub-index HTE. Onepager v9 written.
2026-05-18
Five heterogeneity angles complete: variance, trajectory maintenance, sub-indices, pivot count/quality, ΔSI × attrition. Onepager v8 written.
2026-05-16
Literature audit on pivots. Routes 1, 2, 4, 5 complete: mediation, translation HTE, control arm dynamics, attrition HTE. Onepager v7 written.
2026-05-16
HTE verdict on learning outcomes closed. D5/D6/D7 scans complete. Five new analysis routes identified.
2026-05-15
Diagnostics complete (attrition, multiple outcomes, TE vs EB). GenericML scanner expanded to 12 moderators. Onepager v5 and v6 written.
2026-05-14
GenericML (BLP/GATES/CLAN) first run complete. Implementation verified against Chernozhukov et al. (2025). DML analysis log published.
2026-05-13
RF exploratory analysis complete. Moderator shortlist produced for GenericML.
2026-05-11
Data received and cleaned. Full variable audit complete (199 variables). Clean dataset ready.
2026-04-27
Variable list finalized and sent to Diego.
2026-04-24
Framework agreed: three moderator categories (person, venture, dosage). Method: RF exploration then GenericML. Onepagers v1 and v2 written.
Next Steps
- Mechanism test: does ΔSI at P1 predict G2 membership for performance? T→ΔSI→performance, heterogeneous by founder type
- Cognitive battery (Diego):
risk_aversion, unc_aversion, learning_orientation_1–6 — only untested moderator family; could explain G1–G2 conversion gap
- P1 behavioral robustness: restrict pivoting analysis to P1 (before attrition diverges from Control)
- Targeting welfare analysis: if G1 is hurt by training, who should be enrolled to maximize program ROI?
- Present integrated findings (onepager v10 + seminar slides v2) to Alfonso