# Diagnostic Tests for E/L/S Analysis This folder contains critical diagnostic tests for validating data quality and methodology assumptions in the entrepreneurship strategy analysis. ## Test Scripts ### 1. **verify_EL_definition.py** - E/L Definition Validation **Purpose:** Checks whether the current E (Early funding) and L (Later success) definitions correctly capture the intended populations. **Research Question:** Is the "narrow" definition (E=1 only if LastFinancing = Series A at baseline) excluding successful "fast-growing" companies who progressed to Series B+ before baseline? **What it Tests:** - Compares **narrow** vs **broad** E/L definitions - Narrow: E=1 if LastFinancing = "Series A" at baseline (2022.01) - Broad: E=1 if LastFinancing ≥ "Series A" at baseline (includes B+) - Identifies companies excluded by narrow definition - Calculates L (later success) rates for each group - Detects logically impossible cases (E=0 & L=1) **Key Outputs:** ``` E (narrow): N companies E (broad): N+X companies Excluded: X companies (these got A→B+ before baseline) L rate for narrow E: Y% L rate for excluded: Z% ``` **Decision Criteria:** - If X > 25,000 AND Z% > Y% → Switch to broad definition (features.py:466) - If X < 10,000 OR Z% < Y% → Keep narrow definition, document limitation - Critical finding: Excluded companies are more successful, creating selection bias **Usage:** ```bash python tests/verify_EL_definition.py ``` **Data Requirements:** - `data/raw/Company20220101.dat` (baseline snapshot) - `data/raw/Company20231201.dat` (endpoint snapshot) - Pipe-delimited PitchBook format with `LastFinancingDealType` column --- ### 2. **test_consecutive_rounds.py** - Bridge Rounds Detection **Purpose:** Tests whether later VC funding (L) was the VERY NEXT round after early funding (E), or if there were bridge rounds in between. **Research Question:** Does the step-up calculation (S = PreMoney_t2 / PostMoney_t1) accurately measure A→B growth, or is it confounded by intermediate bridge rounds? **What it Tests:** - Identifies E→L transition companies (Series A at baseline → B+ at endpoint) - Checks for round count columns (TotalFundingRounds, etc.) - Calculates round_diff = rounds_t2 - rounds_t1 - Analyzes deal type transition patterns (A→B vs A→C vs A→B via bridge) - Date-based timing analysis as fallback proxy **Key Outputs:** ``` E→L transitions: N companies Round count data available: Yes/No IF round data available: Exactly 1 round difference (consecutive): X% >1 round difference (bridge rounds): Y% Deal type transitions: A → B: N1 companies A → C: N2 companies A → A: N3 companies (concerning - stalled?) ``` **Decision Criteria:** - If Y% > 40% (bridge rounds common) → Options: - **Censor**: Exclude non-consecutive from H3/HSF analysis - **Request data**: Get full deal-level history to properly code rounds - **Document**: Note as major limitation - If Y% < 20% (mostly consecutive) → Use full E→L sample **Usage:** ```bash python tests/test_consecutive_rounds.py ``` **Data Requirements:** - Same as verify_EL_definition.py - Ideally also has round count columns (e.g., `TotalFundingRounds`) --- ## Classification Logic Both scripts use **identical regex patterns** from `modules/features.py` to ensure consistency: ```python PAT_A = r"(?:\bEarly\s*Stage\s*VC\b|\bSeries\s*A(?:\b|[\s\-]?\d*)\b)" PAT_Bp = r"(?:\bLater\s*Stage\s*VC\b|\bSeries\s*[B-G](?:\b|[\s\-]?\d*)\b)" ``` **Matches for Series A (E):** - "Early Stage VC" - "Series A" - "Series A-II" (second closing) - "Series A - 1" (tranche) **Matches for Series B+ (L):** - "Later Stage VC" - "Series B", "Series C", "Series D", "Series E", "Series F", "Series G" - With variants: "Series B-II", "Series C - 1", etc. --- ## Relationship to Research Design ### Theoretical Model (W2 Slides) ``` V (Vagueness) → E (Early funding) → L (Later success) → S (Step-up) ↓ ↗ └─────(mediator)────┘ ``` **Why These Tests Matter:** 1. **verify_EL_definition.py:** Tests whether E correctly captures "received early VC" - Affects HEV (H1): E ~ V + controls - Affects HLVF (H2): L ~ V × F (NO E control - mediator principle) - Wrong E definition → biased coefficient estimates 2. **test_consecutive_rounds.py:** Tests whether S measures true A→B growth - Affects HSF (H3): S ~ V × F + controls (survivors only) - Bridge rounds → S confounded by multiple funding events - Wrong S measurement → spurious interaction effects --- ## Expected Findings (Pre-Analysis Hypotheses) ### verify_EL_definition.py: - **Predicted:** 30,000-40,000 fast-growing companies excluded - **Predicted:** Excluded companies have HIGHER L rates (15-25% vs 3-5%) - **Implication:** Current narrow definition creates downward bias in E→L effects ### test_consecutive_rounds.py: - **Predicted:** 40-60% of E→L companies had bridge rounds - **Predicted:** Round count data available (60% probability) - **Implication:** H3 (step-up) analysis needs censoring or additional controls --- ## Reporting to Advisors When presenting these diagnostic findings, include: 1. **Tables:** - E/L classification comparison (narrow vs broad) - Transition matrices (baseline → endpoint) - Round count distributions 2. **Key Metrics:** - N excluded by narrow definition - L rates by group - % with consecutive rounds - Median days between funding 3. **Methodological Decisions:** - Which E/L definition chosen and why - How bridge rounds handled in H3 analysis - Robustness checks planned 4. **Limitations Acknowledged:** - Companies may have had funding before entering PitchBook - "LastFinancing" only shows most recent, not full history - Timing based on deal dates, not actual cash-in-bank dates --- ## Notes - Both scripts auto-detect pipe (`|`) or tab (`\t`) delimiters - Runtime: ~30-60 seconds per script on 400k+ companies - Memory usage: ~2GB for full dataset - Compatible with Python 3.8+ - No external dependencies beyond pandas, numpy --- **Last Updated:** 2025-11-12 **Author:** Claude (with human validation) **Related:** See `modules/features.py` for production classification logic