# Hypothesis Test: β_VF > 0 **Theory**: Flexibility (F) amplifies the effect of Vagueness (V) on Later success (L) ## Research Question **H***: Does flexibility make it easier for vague companies to succeed? **Formal hypothesis**: ``` Pr(L=1) = logit^-1(α + β_V·V + β_F·F + β_VF·(V×F) + controls) H0: β_VF = 0 H1: β_VF > 0 (Flexibility amplifies Vagueness effect) ``` **Key principle**: E (Early funding) is a **mediator**, NOT a confounder - Do NOT include E as control in L regression - Causal path: V → E → L (E mediates V's effect on L) ## Minimal Specification **Model**: Non-Bayesian Logit with V×F interaction - Response: L (Later Stage VC at 2025.11) - Main effects: z_V (Vagueness), F_flexibility (1=SW, 0=HW) - **Interaction**: z_V × F_flexibility (this is β_VF!) - Controls: founding_cohort, region - **NO E control** (E is mediator) **Statistical test**: One-tailed Wald test - If β_VF > 0 and p < 0.05 → Reject H0 - Conclusion: SW companies benefit more from vagueness than HW ## Usage ### All companies ```bash python scripts/test_hypothesis_VxF.py ``` ### Quantum computing only ```bash python scripts/test_hypothesis_VxF.py --industry quantum ``` ### Transportation only ```bash python scripts/test_hypothesis_VxF.py --industry transportation ``` ## Output ### 1. Terminal ``` ================================================================================ HYPOTHESIS TEST: β_VF > 0 (Quantum Companies) ================================================================================ H*: Flexibility amplifies the effect of Vagueness on Later success H0: β_VF = 0 H1: β_VF > 0 LOGIT REGRESSION Formula: L ~ z_V * F_flexibility + C(founding_cohort) + C(region) Coefficients: Coef Std.Err z P>|z| z_V -0.234 0.089 -2.63 0.0085 F_flexibility 0.456 0.123 3.71 0.0002 z_V:F_flexibility 0.312 0.145 2.15 0.0316 ← β_VF HYPOTHESIS TEST Interaction term: z_V:F_flexibility Coefficient: β_VF = 0.3120 Std Error: SE = 0.1450 z-statistic: z = 2.152 p-value (one-tailed): p = 0.0158 ================================================================================ VERDICT: ✓ REJECT H0 ================================================================================ Flexibility AMPLIFIES Vagueness effect (β_VF = 0.3120 > 0, p = 0.0158) Significance: * ``` ### 2. Files **Coefficient table:** - `outputs/hypothesis_VxF/coefficients_quantum.csv` **Summary stats:** - `outputs/hypothesis_VxF/summary_quantum.csv` **Interaction plot:** - `outputs/hypothesis_VxF/interaction_VxF_quantum.png` - `outputs/hypothesis_VxF/interaction_VxF_quantum.pdf` ### 3. Interaction Plot ![Interaction plot showing two curves: - Blue solid line (F=1, SW/Flexible, skyblue) - Gray dashed line (F=0, HW/Rigid, gray) X-axis: Vagueness (green) Y-axis: Pr(L=1|V,F) (blue) Annotation: β_VF = 0.312* ] **Interpretation:** - If lines diverge (positive interaction) → SW benefits more from high V - If lines parallel (β_VF ≈ 0) → No differential effect - If lines converge (negative interaction) → HW benefits more ## Color Code (W2 Standard) | Variable | Color | Usage | |----------|-------|-------| | L (Later success) | #0000FF (blue) | Y-axis label | | V (Vagueness) | green | X-axis label | | F=1 (Flexible/SW) | skyblue | Solid line | | F=0 (Rigid/HW) | gray | Dashed line | ## Current Limitations ### ⚠️ Mock Vagueness Data Current script uses **random V** for demonstration. **To use real vagueness**: 1. Ensure consolidated data includes `Description` and `Keywords` columns 2. Run `compute_vagueness_vectorized()` from `modules/features.py` 3. Or re-consolidate from original .dat files with Description/Keywords **To fix consolidation script:** ```python # In consolidate_2021_cohort.py, change: baseline_cols = ['CompanyID', 'CompanyName', 'LastFinancingDealType', 'Description', 'Keywords'] # ← Add these ``` ### Mock Controls - `founding_cohort`: Currently all 'cohort_1' - `region`: Currently all 'US' **Real controls** should come from original .dat files. ## Relationship to W2 Slides - **H* = W2 H₂** (p.29-31): "Flexibility amplifies the effect of Vagueness on later survival" - **E as mediator** (p.49): Why we exclude E from L regression - **Color palette** (p.XX): Standard colors for all figures ## Comparison with Previous Analysis ### ❌ Old approach (`analyze_2021_cohort.py`): - Chi-square test comparing HW rate vs SW rate - Ignores Vagueness (V) - No interaction term - Bar chart (rates only) ### ✓ New approach (`test_hypothesis_VxF.py`): - Logit regression with V×F interaction - Directly tests β_VF > 0 - Accounts for continuous V - Interaction curve showing mechanism ## Next Steps 1. **Add real Vagueness measure** - Include Description/Keywords in consolidation - Run `compute_vagueness_vectorized()` 2. **Add real controls** - founding_cohort from YearFounded - region from HQCountry 3. **Multiple time points** - Test β_VF at 2yr, 3yr, 4yr separately - Check if interaction grows over time 4. **Robustness checks** - Probit instead of Logit - Different V measures (alternative scorers) - Subsample analysis (by cohort, region)