## wpiea2025078-print-pdf — section 3 contains our measurement exercises; section 4 presents the econometric method-

## Source details

**Canonical URL:** [wpiea2025078-print-pdf — section 3 contains our measurement exercises; section 4 presents the econometric method-](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025078-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025078-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025078-print-pdf.pdf.json)

---

### Methodological framing and literature
- The paper builds on Khwaja and Mian (2008), Chodorow-Reich (2013) and Chopra et al. (2020).
- Key theoretical and empirical context:
  - Real activity contracts following credit supply shocks; time to build can generate oscillatory fluctuations (Asea and Zak, 1999), and is central to debates following Kydland and Prescott (1982) vs. Rouwenhorst (1991).
  - Time-to-build across heterogeneous, complementary capital stocks has quantitative implications for business cycles and monetary transmission (Casares, 2006; Edge, 2007; Lucca, 2007).
  - Exogenous time to build is first-order equivalent to investment adjustment costs and relevant for matching investment dynamics (Christiano et al., 2005; Eberly et al., 2012; Bloesch and Weber, 2023).
  - Meier (2020) and Liu and Tsyvinski (2023) emphasize supply-chain/adjustment-cost channels that act similarly to exogenous time to build.
- Backlog ratio suitability and fiscal gestation lags:
  - Backlog ratio appropriate for inventories/intermediate inputs but not for large structures; prior backlog-based work is interpreted as informative about intermediate inputs.
  - Fiscal multiplier potency varies with gestation lags of public investment (Leeper et al., 2010; Ramey, 2020).

### Data sources and linkage
- Primary project-level dataset: CapexDx (Center for Monitoring of the Indian Economy, CMIE).
  - Project definition: construction enterprise with “reasonable clarity on the capacity being created”; classified as “structures” in national accounts (examples: steel plants, irrigation canals, airport terminals, office and residential buildings).
  - Records: project name, cost, geographical location, industry, event-level lifecycle histories (announcement, start, stalls, completion, abandonment).
  - Right-censoring: some project histories are right-censored; econometric methods account for right-censoring.
  - Appendix A.1 details project definition rules and data collection.
- Time coverage and size threshold:
  - CMIE collected projects since 1976; systematized from the mid-1990s; bulk of projects start in latter period.
  - Data covers projects above 10 million rupees (150-300 thousand USD, depending on foreign exchange over sample).
  - Focus: new capacity creation with some expansions/modernizations.
- Linking to firm and loan data:
  - Projects linked to sponsoring firms via CMIE’s ProwessDx; loan-level data from Ministry of Corporate Affairs (MCA) records secured loans (amount, date, borrower and lender names; not interest rates, maturity, or covenants).
  - Borrower matching via unique Company Identification Number; lender matching via fuzzy matching and manual checks; matches for over 90% of loans (by volume and number of charges).
  - Regulatory incentive: registration with MCA creates near–credit-registry-like dataset for secured loans.

### Deconstructing time to build: micro-data findings
- Sample scope and descriptive statistics:
  - Number of projects 106,626; Number of firms 32,456.
  - Average (earliest) project first observed in 2010; earliest projects enter mid-1990s.
  - Median (average) projects per sponsor: 11 (113).
  - Costs observed for 70% of projects; cost distribution right-skewed:
    - Median project cost 503 million rupees; projects above 90th percentile cost over 10 billion rupees.
  - Project types: new units 0.81, large expansions 0.15, modernizations 0.04.
  - Public sponsors account for almost 0.38 of projects; Private mean 0.57.
- Lifecycle overview (Panel B summary statistics reproduced exactly where reported):
  - Ever in Status (dummies; value 1 if project ever in that stage):
    - Under Imp 0.71
    - Stalled 0.06
    - Completed 0.40
    - Shelved or Abandoned 0.05
    - No Information 0.31
  - Time in Status (Qtrs) — total time spent in a stage over observable history (non-terminal stages only):
    - Announced 6.35
    - Under Imp 8.74
    - Stalled 0.69
  - Additional panel A cross-sectional exact table values (selected):
    - Projects per Sponsor mean 113.05 SD 257.90 10th 1.00 25th 2.00 50th 11.00 75th 77.00 90th 366.00
    - Year Started mean 2010 SD 8 10th 1997 25th 2005 50th 2011 75th 2017 90th 2020
    - Last Observed mean 2014 SD 8 10th 2001 25th 2009 50th 2016 75th 2021 90th 2022
    - Cost (Rs. Million) mean 6428.764 SD 40583.47 10th 4.00 25th 125.70 50th 503.30 75th 2500.00 90th 10000.00
    - Missing Cost 0.29 SD 0.45
    - Share: Section mean 0.19 SD 0.11 10th 0.02 25th 0.07 50th 0.23 75th 0.31 90th 0.31
    - Share: Industry mean 0.04 SD 0.04 10th 0.00 25th 0.01 50th 0.02 75th 0.04 90th 0.12
    - Share: State mean 0.06 SD 0.04 10th 0.02 25th 0.03 50th 0.05 75th 0.09 90th 0.14
    - New Unit mean 0.81 SD 0.39
    - Large Expansion mean 0.15 SD 0.35
    - Private mean 0.57 SD 0.50
    - Public mean 0.38 SD 0.49
- Transition and censoring patterns:
  - Most announced projects move to implementation; most under implementation are eventually completed.
  - Stallments and abandonments are rare: 6% ever stalled; 5% ever abandoned.
  - Most stalled projects do not remain stalled for over a quarter.
  - Right-censoring (“lost to follow-up”): 30% of projects have unknown outcomes; CMIE discontinues tracking if information not available for over 30 months (manufacturing) or 36 months (infrastructure).
  - Appendix tabulations: about 20% of announced projects, 6% of stalled projects, and 27% of projects under implementation are lost to follow-up.
- Distribution of gestation lags (completed projects):
  - Median project takes 2.5 years to be completed.
  - Aggregate distribution right-skewed: mean (median) duration reported for 42,691 projects as 14.37 (10.00) quarters; sizable share takes over 15 years.
- Cross-sectional heterogeneity drivers:
  - Sectoral differences: Transportation, construction, and other infrastructure sectors have the longest time to build; Retail, information technology, and manufacturing shortest.
  - Sponsor ownership: Public sponsors build projects with longer time to build; public sponsorship explains little variation once industry accounted for (no causal claim).
  - Project size relationship:
    - Bivariate elasticity: a 1% increase in costs associated with a 0.18% increase in completion time.
    - Including industry fixed effects: elasticity changes from 0.18% to 0.15%.
    - Elasticity robustly constant after residualizing by industry dummies.

### Measurement approach and caveats
- Time to build measured from announcement to eventual completion; analysis restricts to projects eventually completed; results conditional on completed projects possibly differing systematically from abandoned or censored projects.
- Right-censoring handled via duration models and econometric methods.
- Event-level diary data cleaned and reconstructed into consistent project lifecycles (Appendix A.1); “ever in status” variables can overlap; rare resuscitations treated as stalled for exposition and results insensitive to grouping.
- Limitations:
  - Cost as size proxy imperfect (cost overruns/accelerations can bias relationships).
  - Backlog-based measures better capture intermediate inputs/inventories, not large structures.

### Identification: the Asset Quality Review (AQR) as a credit supply shock
- Institutional shock:
  - RBI conducted Asset Quality Review between August and December 2015 (with annual follow-ups), uncovered under-provisioning, required additional provisions and mandated disclosure.
  - Forbearance period: automatic reclassification suspended from August 27, 2008 until April 1, 2015.
- Constructing bank shock:
  - Bank Shock_bt = (Provisions Required by the RBI Post-Audit − Provisions Pre-Audit) / Total Assets_bt
- Empirical approach to separate supply from demand:
  - Lender-borrower-year dataset using outstanding loans 2013–2015 to fix relationship composition pre-audit; balanced panel 2013–2019.
  - Estimate Loans_bft = exp(γ × Bank Shock_bt + λ_ft + ω_b + X_bft Γ) e_bft using PPML; λ_ft firm-by-year fixed effects absorb credit demand shocks; ω_b bank fixed effects; cluster standard errors at lender level.
- Table 3 key result:
  - Column (1): one-standard deviation increase in bank exposure reduces outstanding credit in average pair by 5.1% (Exposure coefficient = -0.0517, standard error 0.0106; ∗∗∗). Outc Mean = 2497.20; Obs = 268,858.
  - Restricted sample to borrowers with observed Capex projects: audit shock resulted in a 4.9% average contraction of credit supply (Exposure ≈ -0.0497 to -0.0491; standard errors 0.0122–0.0131; Outc Mean = 450.34; Obs = 109,929).

### From banks to firms to projects: exposure and hazard estimation
- Firm exposure mapping:
  - Firm Exposure_ft = Σ_{b ∈ B_{f,t−1}} w_{b,f,t−1} × Bank Shock_bt
  - w_{b,f,t−1} = Outstanding Loans_{b,f,t−1} / Σ_{b ∈ B_{f,t−1}} Outstanding Loans_{b,f,t−1}
- Project completion hazard estimation:
  - φ(Completed_pft) = β × Firm Exposure_ft + X_pft Γ + υ_pft with φ a link function; outcome is binary in discrete-time survival panel; standard errors clustered at firm level.
  - Discrete-time survival / duration models used to handle right-censoring and avoid selection bias.
- Link function and estimation choice (Appendix):
  - Discrete-time hazards coincide with logistic likelihood but authors use log link φ(x)=log(x) implying semi-parametric, piece-wise exponential model and estimation via Poisson pseudo-maximum likelihood (PPML).
  - Unconditional probability of completion equals "4%".

### Empirical effects on completion hazards (Table 4 / robustness A6)
- PPML semi-elasticities: percent effect on completion hazard for one-standard-deviation increase in firm exposure.
- Key point estimates (columns summarized; DepVar Mean = 0.0397; Observations = 122,123 unless noted):
  - Column (1): Exposure = 0.1094 ∗∗∗ (0.0270) — 10.94% increase in hazard.
  - Column (2): Exposure = 0.0667 ∗∗ (0.0272).
  - Column (3): Exposure = 0.0636 ∗∗ (0.0273).
  - Column (4): Exposure = 0.0605 ∗∗ (0.0264).
  - Column (5) (preferred): Exposure = 0.0573 ∗∗ (0.0266).
  - Column (6): Exposure = 0.0734 ∗∗∗ (0.0264); Observations = 122,100.
  - Column (7): Exposure = 0.0716 ∗∗∗ (0.0266); Observations = 122,100.
- Interpretation and robustness:
  - Preferred specification (column 5) yields a 5.7% increase in completion hazards for one-standard-deviation higher exposure.
  - Adding firm FE and state×year FE reduces effect from 11% to 6.6% (column 2), remaining statistically significant across specifications.
  - Robust in Appendix C: alternative weights, linear models, different fixed-effect combinations; effects largest at impact and decline over time.

### Empirical effects on project starts (Table 5 / A8, A9)
- PPML semi-elasticities: percent effect on firm-level new project starts for one-standard-deviation increase in firm exposure.
- Key point estimates (DepVar Mean = 0.8362; Observations = 89,948 unless noted):
  - Column (1): Exposure = -0.0775 ∗∗∗ (0.0286) — 7.75% reduction in new project starts.
  - Column (2): Exposure = -0.0698 ∗∗ (0.0325).
  - Column (3): Exposure = -0.0670 ∗∗ (0.0322).
  - Column (4): Exposure = -0.0806 ∗∗ (0.0316).
  - Column (5) (preferred): Exposure = -0.0763 ∗∗ (0.0314).
  - Column (6): Exposure = -0.0584 ∗ (0.0299); DepVar Mean = 0.8326; Observations = 88,968.
  - Column (7): Exposure = -0.0588 ∗∗ (0.0296); Observations = 88,968.
- Interpretation and robustness:
  - Preferred specification shows a 7.6% reduction in new project starts for one-standard-deviation higher exposure.
  - Estimates stable across fixed-effect sets and controls; Appendix robustness tables show similar semi-elasticities under alternative weighting.

### Heterogeneous effects by project stage (Section 6.2; Table 6)
- Model prediction: discount-rate shocks raise investment for sufficiently high-output projects or sufficiently late-stage projects; effect larger the later the stage.
- Empirical proxy for stage:
  - Stage_p = Initial Age_p / Predicted Time to Build_p; predicted time to build obtained from out-of-sample regression including industry dummies; proxy winsorized at 1st and 99th percentiles and standardized.
- Regression evidence (PPML; interaction specifications; standard errors clustered at firm level):
  - Exposure main effects (examples): 0.090 ∗∗∗, 0.086 ∗∗∗, 0.094 ∗∗∗, 0.090 ∗∗∗ (standard errors (0.025) each).
  - Stage × Exposure interaction estimates: 0.068 ∗∗∗, 0.069 ∗∗∗, 0.078 ∗∗∗, 0.079 ∗∗∗ (standard errors ≈ (0.021)–(0.023)).
- Interpretation:
  - For project at mean proxy stage, one-standard-deviation increase in exposure raises hazard by roughly 9%.
  - For project one standard deviation above mean stage, total effect ≈ 15.5%; one standard deviation below mean, total effect ≈ 2.5%.
- Model-data comparison:
  - Model implies a unit-increase in proxied stage raises hazard by 8% — within regression confidence bands and not targeted in calibration.

### Robustness, predictability, and selection diagnostics (Appendices D, A10–A11, A14–A18)
- Predictability tests:
  - Regressions of one-period lagged firm characteristics on contemporaneous exposure detect predictability in profitability and cash over assets; no detectable predictability among other covariates (Appendix Table A11).
- Differential censoring/backfilling concerns:
  - Two analyses: restrict to projects present in pre-audit vintage (Dec 31, 2014) and test for differential pre-trends in firm-level project stocks — no evidence that results driven by differential censoring (Tables A14–A16; Figure A13).
- Scope-reduction concerns:
  - Institutional and contractual context suggests large CAPEX projects are costly to rescope ex post; CapexDx records few drastic scope changes.
  - Robustness: exclude multi-unit projects, projects with downward cost revisions, projects with missing costs, ever-stalled projects, malleable industries, projects with regulatory events, projects under implementation — point estimates largely similar and remain significant in many subsamples (Appendix Tables A17–A18).
- Additional diagnostics:
  - Balance tests (Table A10) detect some cross-sectional predictability in log sales and equity issuance; panel balance (Table A11) shows predictability weakens with firm and time fixed effects.
  - Predictive models for time to build (Table A12) used to construct proxy stage; out-of-sample R-Squared chosen to select specification.

### Theory: investment model with endogenous time to build (continuous-time, frictional)
- Setup:
  - Projects require investing an amount ̄x to obtain z units of capital at completion; z drawn from f(z); capital price p taken as given; firm discount rate r>0.
  - Investment flow i(t) with constraint ∫_0^T i(t) dt = ̄x and frictional costs κ(i) (weakly increasing, weakly convex, deadweight).
  - Example κ(i): κ(i)=0 for i ≤ ̄i; κ(i)=κ/2 (i−̄i)^2 for i>̄i. As κ → ∞, model nests exogenous time-to-build with T = ̄x/̄i.
- HJB and FOC:
  - r v(x,z) = max_{i≥0} [ −i − κ(i) + i ∂v(x,z)/∂x ], terminal v(̄x,z)=p z.
  - First-order condition: 1 + κ′(i) = ∂v(x,z)/∂x (marginal cost equals marginal q).
- Entry and steady-state distribution:
  - Entry rule: start projects with z > ̄z where v(0,̄z)=0.
  - Steady-state distribution g(x,z) satisfies g(x,z) i(x,z) = f(z)·1{z ≥ ̄z}.
- Completion time:
  - T(x,z) = ∫_x^{̄x} 1/i(y,z) dy.
- Qualitative steady-state properties:
  - v(x,z) increasing in x; i(x,z) higher for more productive projects; lower-z projects have greater mass (longer durations).

### Response to a credit/provisioning shock (discount-rate formulation)
- Shock mapped to firm discount-rate increase:
  - r(t) = r + s e^{−ρ t} with persistence ρ and size s.
- First-order linearization (Lemma 1):
  - V(x,z,s,t) = v(x,z) + s e^{−ρ t} ˆv(x,z) + o(s)
  - I(x,z,s,t) = i(x,z) + s e^{−ρ t} ˆi(x,z) + o(s)
- Proposition 2 (value response) and Proposition 3 (investment response):
  - ˆv(x,z) expressed as weighted sum over remaining project life with duration-like weights; duration interpretation: greater sensitivity for projects with cash flows further in future.
  - ˆi(x,z) = 1/κ''( i(x,z) ) · ∂/∂x ˆv(x,z); near-completion corollary: for x sufficiently close to ̄x, ˆi(x,z) > 0.
- Implications:
  - Mature projects accelerated after discount-rate rise (intensive margin positive for late-stage projects).
  - Entry threshold ̄z rises (extensive margin contraction); Proposition 5 links ˆz to ˆv(0,̄z) and thereby predicts fewer new project starts.
- Aggregation:
  - Accounting investment I_acc(t) = ∫∫ g(x,z,t) · i(x,z,t) dz dx; fixed capital formation I_K(t) = ∫ z · g(x,z,t) · i(x,z,t) dz.
  - Aggregate responses reflect distributional shifts across project stages and entry effects.

### Calibration, identification and quantitative findings
- Steady-state illustrative parameters (median calibration for figures):
  - r = 0.01, p = 1, ̄x = 1, ̄i = 0.083, κ = 30.1.
- Shock illustration parameters (figure 5): ρ = 0.17, s = 0.02.
- Identification of (s, κ):
  - Use empirical semi-elasticities β_int (intensive) and β_ext (extensive) from Tables 4 and 5 (columns (5)).
  - β_ext informative about s; β_int informative about both s and κ; choose (s, κ) by minimum distance to match semi-elasticities.
- Reported calibrated estimates:
  - Shock size s = 0.02 (2%).
  - Acceleration cost κ = 3 (text reports κ=3 in main calibration narrative; Appendix A19 reports κ = 3.09 in calibration table).
- Calibration fit:
  - Appendix Table A20 shows model matches targeted moments: Hazard Regression β = 0.06 vs 0.06 (model vs data); Starts Regression β = -0.08 vs -0.08.
  - Table A19 reports parameter estimates including ρ = 0.17; σ2 ̄x = 2.12; β ̄i1 = 0.82; σ2 ̄i = 0.88; α3 = Pareto tail parameter; others as listed.

### Model implications and counterfactual dynamics
- Endogenous vs exogenous time to build:
  - Exogenous TTB (κ→∞): investment falls, trough 12–16 quarters later, hump-shaped IRF.
  - Endogenous TTB: initial rise in investment (impact spike) due to acceleration of mature projects; followed by sharp collapse as distribution skews to early-stage projects (boom-bust dynamics).
  - Endogenous TTB raises investment volatility by 30% relative to exogenous TTB in calibrated exercises.
- State-dependence:
  - Aggregate impact of discount-rate/monetary shocks depends on cross-sectional project-stage distribution.
  - If economy skewed toward mature projects, impact spike larger; if skewed to younger projects, impact can be negative.
- Fiscal policy (investment tax credit) implications:
  - Temporary investment tax credits mainly reshuffle investment intertemporally; endogenous TTB amplifies front-loading and subsequent collapse when credit expires.
  - Takeaway: temporary tax credits’ cumulative response comes from project starts; intensive margin drives impact effects; effects state-dependent.
- General equilibrium:
  - Embedding partial equilibrium mechanisms into GE (capital-price feedback) preserves boom-bust and persistence features; endogenous TTB can produce sharper, more persistent slumps in productive capital than models with convex adjustment costs only.
  - Proposition 7 gives linearized response of capital prices p̂(t)/p to announced paths Â(t) and r̂(t).

### Conclusion and policy-relevant takeaways
- Empirical + theoretical summary:
  - Time to build pervasive across industries, sponsor types, time periods, and project sizes.
  - Firms exercise causal control over gestation lags; endogenous TTB rationalizes empirical findings: contraction in new project starts (extensive margin) and acceleration of near-complete projects (intensive margin).
  - Endogenous TTB implies:
    - Investment more responsive and volatile to shocks.
    - Mature projects react more strongly than fledgling projects.
    - Distribution of projects across stages makes monetary policy state-dependent.
    - Investment in ongoing projects highly responsive to cash flow variations; investment subsidies can reshuffle aggregate demand and aid structural transitions.
- Open questions and extensions:
  - Designing project financing internalizing gestation-lag incentives.
  - Competitive and political-economy implications of gestation-lag strategies.
  - Quantifying targeted subsidies for infrastructure or green transition; generalizing to breakthrough or intangible capital.

*Source: wpiea2025078-print-pdf - section 3 contains our measurement exercises; section 4 presents the econometric method- (CapexDx Mar 2022 vintage; excerpts provided).*

### section 3 contains our measurement exercises; section 4 presents the econometric method-

### wpiea2025078-print-pdf - section 3 contains our measurement exercises; section 4 presents the econometric method-

### Methodological framing and literature
- Methodologically, the paper borrows from Khwaja and Mian (2008), Chodorow-Reich (2013) and Chopra et al. (2020).
- Thematic context:
  - A large body of evidence documents that real activity contracts following credit supply shocks.
  - Asea and Zak (1999) showed the neoclassical growth model with exogenous time to build predicts oscillatory fluctuations around a steady state.
  - The debate on whether time to build matters for business cycles is linked to Kydland and Prescott (1982) vs. Rouwenhorst (1991).
  - Casares (2006), Edge (2007) and Lucca (2007) find time to build across heterogeneous, complementary capital stocks has strong quantitative implications for business cycles and monetary transmission.
  - Lucca (2007) proves exogenous time to build is first-order equivalent to investment adjustment costs, important for:
    - (i) matching effects of monetary shocks on aggregate investment (Christiano et al., 2005),
    - (ii) explaining pervasive lagged investment effects (Eberly et al., 2012),
    - (iii) matching R&D investment responses to monetary shocks (Bloesch and Weber, 2023).
  - Meier (2020) argues time-series variation in time to build due to supply chain disruptions yields GDP and TFP losses.
  - Liu and Tsyvinski (2023) show adjustment costs for intermediate inputs in a dynamic input-output network model are first-order equivalent to exogenous time to build.
- The backlog ratio is an appropriate measure of time to build for inventories, intermediate inputs, or smaller equipment, but inadequate for large structures; the authors interpret prior backlog-based work as informative about backlogs in intermediate inputs rather than time to build in capital formation.
- Leeper et al. (2010) and Ramey (2020) show fiscal multiplier potency varies with gestation lags of public investment.

### Data sources and linkage
- Primary project-level dataset: CapexDx (Center for Monitoring of the Indian Economy, CMIE).
  - Project definition: construction enterprise with “reasonable clarity on the capacity being created”; projects are classified as “structures” in national accounts (examples: steel plants, irrigation canals, airport terminals, office and residential buildings).
  - CapexDx records project name, cost, geographical location, industry, and rich event-level lifecycle histories (announcement, start, stalls, completion, abandonment).
  - Censoring: some project histories are right-censored; authors leverage econometric methods that account for right-censoring.
  - CMIE documents monitoring efforts and follows transparent rules when discontinuing project coverage; appendix A.1 elaborates on project definition rules and data collection.
- Time coverage and size threshold:
  - CMIE collected projects since 1976, systematized from the mid-1990s; bulk of projects start in the latter period.
  - Data covers projects above 10 million rupees (150-300 thousand USD, depending on the foreign exchange rate over the sample period).
  - Focuses on new capacity creation, with some expansions or modernizations.
- Linking to firm and loan data:
  - Link projects to sponsoring firms via CMIE’s ProwessDx for firm and bank financial statements; data sourced from regulatory filings and not restricted to public firms.
  - Loan-level data from the Indian Ministry of Corporate Affairs (MCA): secured loans registered by lenders with the MCA; observe amount borrowed, date, borrower and lender name; do not observe interest rates, contractual maturity, or covenants.
  - Borrower matching uses unique Company Identification Number present in both datasets.
  - Lender matching uses fuzzy matching and manual checks; achieved matches for over 90% of loans (both in loan volume and number of charges) to a Prowess company.
- Regulatory incentive: failing to register secured charges with the MCA has material adverse consequences; registration creates a near–credit-registry-like dataset for secured loans.

### Deconstructing time to build: novel facts (micro-data findings)
- Sample scope:
  - Over 100,000 projects sponsored by over 30,000 entities.
  - Average (earliest) project was first observed in 2010, with earliest projects entering in the mid-1990s.
  - The median (average) project is sponsored by a firm with 11 (113) projects.
  - Costs observed for 70% of projects.
  - Distribution of costs is right-skewed: median project cost is 503 million rupees; projects above the 90th percentile cost over 10 billion rupees.
  - Projects are either new units (0.81), large expansions (0.15) or modernizations (0.04).
  - Public sponsors account for almost 0.38 of projects.
- Lifecycle overview (Panel B summary statistics reproduced exactly where reported):
  - Ever in Status (dummies; take value 1 if project ever in that stage):
    - Under Imp 0.71
    - Stalled 0.06
    - Completed 0.40
    - Shelved or Abandoned 0.05
    - No Information 0.31
  - Time in Status (Qtrs) — total time spent in a particular stage over the project’s entire observable history (non-terminal stages only):
    - Announced 6.35
    - Under Imp 8.74
    - Stalled 0.69
  - Sample counts:
    - Number of projects 106,626
    - Number of firms 32,456
  - Additional panel A cross-sectional entries (selected exact table values):
    - Projects per Sponsor mean 113.05 SD 257.90 10th 1.00 25th 2.00 50th 11.00 75th 77.00 90th 366.00
    - Year Started mean 2010 SD 8 10th 1997 25th 2005 50th 2011 75th 2017 90th 2020
    - Last Observed mean 2014 SD 8 10th 2001 25th 2009 50th 2016 75th 2021 90th 2022
    - Cost (Rs. Million) mean 6428.764 SD 40583.47 10th 4.00 25th 125.70 50th 503.30 75th 2500.00 90th 10000.00
    - Missing Cost 0.29 SD 0.45
    - Share: Section mean 0.19 SD 0.11 10th 0.02 25th 0.07 50th 0.23 75th 0.31 90th 0.31
    - Share: Industry mean 0.04 SD 0.04 10th 0.00 25th 0.01 50th 0.02 75th 0.04 90th 0.12
    - Share: State mean 0.06 SD 0.04 10th 0.02 25th 0.03 50th 0.05 75th 0.09 90th 0.14
    - New Unit mean 0.81 SD 0.39
    - Large Expansion mean 0.15 SD 0.35
    - Private mean 0.57 SD 0.50
    - Public mean 0.38 SD 0.49
- Transition and censoring patterns:
  - Most announced projects move to implementation; most projects under implementation are eventually completed.
  - Stallments and abandonments are rare: 6% of projects ever stalled; 5% ever abandoned.
  - Most stalled projects do not remain stalled for over a quarter.
  - Right-censoring (“lost to follow-up”): 30% of projects have unknown outcomes; CMIE discontinues tracking if projects are stalled/shelved/abandoned or information unavailable for over 30 months (manufacturing) or over 36 months (infrastructure).
  - Appendix tabulations (described) show about 20% of announced projects, 6% of stalled projects, and 27% of projects under implementation are lost to follow-up; quarterly transition probabilities show persistence in intermediate stages.
- Distribution of gestation lags (completed projects):
  - The median project takes 2.5 years to be completed.
  - Aggregate distribution is right-skewed: most projects complete in a few years, but a sizable share takes over 15 years.
  - Figure 1 note: 42691 projects with mean (median) duration of 14.37 (10.00) quarters; histogram of gestation lags measured from announcement to completion; distribution winsorized at 1st and 99th percentile for visualization (mean and median reported are of raw distribution).
- Cross-sectional heterogeneity drivers:
  - Sectoral differences:
    - Transportation, construction, and other infrastructure-related sectors have the longest time to build.
    - Retail, information technology, and manufacturing have the shortest time to build.
    - Panel (a) of Figure 2 plots medians and interquartile ranges by sector (sectors based on aggregated NIC-defined sectors; water and sanitation excluded as an outlier).
  - Sponsor ownership:
    - Public sponsors build projects with longer time to build; public sector frequently sponsors heavy infrastructure projects (e.g., Indian Railways).
    - Panel (b) of Figure 2 reports distributions by private, foreign, and public ownership (distributions winsorized at 1st and 99th percentile for visualization).
    - Public sponsorship explains little variation in gestation lags once industry is accounted for; no causal claim of public sponsorship causing longer gestation lags is made.
  - Project size (cost) relationship:
    - Cost is used as a common metric for project “size” across heterogeneous structures.
    - Figure 3 binned scatterplots (log time to completion vs. log costs):
      - Bivariate elasticity: a 1% increase in costs is associated with a 0.18% increase in completion time.
      - Including industry fixed effects: elasticity changes from 0.18% to 0.15%.
      - Panel (b) residualizes by industry dummies; elasticity is robustly constant.

### Measurement approach and caveats
- Measuring time to build:
  - Time to build is measured from announcement to eventual completion.
  - Analysis restricting to projects that are eventually completed; results conditional on the possibility that completed projects differ systematically from abandoned or censored projects.
  - Right-censoring handled via econometric methods (described elsewhere in the paper).
- Data cleaning and lifecycle reconstruction:
  - Event-level diary data can record multiple events on the same date; authors reconstruct consistent project lifecycles and conduct extensive data cleaning (detailed in Appendix A.1).
  - “Ever in status” variables can overlap because projects transit multiple stages.
  - Some projects labeled “shelved or abandoned” can later be resuscitated; authors group rare resuscitations with stalled projects for expositional purposes and show results are insensitive to this grouping.
- Limitations noted by authors:
  - Cost as size proxy is imperfect (cost overruns or accelerations can bias relationships).
  - Backlog-based measures from other literature better capture intermediate inputs or inventories, not large structures.

*Source: wpiea2025078-print-pdf - section 3 contains our measurement exercises; section 4 presents the econometric method- (CapexDx Mar 2022 vintage; excerpts provided).*

### Appendix Table A4 collects estimates from the regression specifications.

### wpiea2025078-print-pdf - Appendix Table A4 collects estimates from the regression specifications.

### Decomposing Variation in Time to Build
- Measurement shows sizable variation in time to build.
- Strategy: decompose variance using fixed effects, including only “exogenous” predictors to assess technological drivers.
- Table 2 key findings:
  - Column (1): project sector accounts for 24% of the variation (R2 = 0.239).
  - Column (2): moving from a dozen sector bins to over a hundred industry bins marginally increases explained variance to 30% (R2 = 0.302).
  - Column (3): year dummies (year the project was started) account for 28% of the variance (R2 = 0.285).
  - Column (4): sector and year started fixed effects together explain about 40% of the variation in time to build (R2 = 0.395).
  - Columns (5)–(8): keep sector and year started fixed effects; add other fixed effects sequentially:
    - Column (5) adds state fixed effects (R2 = 0.409).
    - Column (6) adds ownership dummies (public/private/foreign) (R2 = 0.410).
    - Column (7) adds dummies for whether the sponsor is a new unit (R2 = 0.396).
    - Column (8) includes all preceding variables (R2 = 0.441).
    - Geography, ownership and unit type add little explanatory power relative to sector and year.
  - Column (9): include sponsor and year started fixed effects (abandon dimensionality reduction) and account for 58% of the variation (R2 = 0.579).
- Interpretation:
  - Considerable within-sector (and within-firm) variation in gestation lags remains.
  - Parsimonious, low-dimensional specifications learn at most about 30% of the variation in construction times; abandoning parsimony raises explanatory power but reduces interpretability.
- Table 2 numeric summary (as presented):
  - R2 row: 0.239 0.302 0.285 0.395 0.409 0.410 0.396 0.441 0.579
  - Unc. Var: 0.014 0.015 0.001 0.046
  - Partial R2: 0.023 0.024 0.002 0.077
  - Fixed effects indicated in table: Ownership FEX, State FEX, Sector FEX, Industry FEX, Year Started FEX, Type of unit FEX, Firm FEX (with the specific pattern of inclusion marked in the source table).

### Testing for Endogenous Time to Build
- Research question: Are gestation lags purely exogenous constraints, or can firms influence completion speed (“endogenous time to build”)?
- Conceptual contrast:
  - Null (purely exogenous): project completion determined by natural progression; firm-level shocks (weak sales, cash shortages, credit cuts) do not affect completion; observed completion variation driven by project-level disturbances unrelated to firm decisions.
  - Alternative (endogenous): firms can accelerate completion in favorable conditions (exceptional sales) or stall/shelve projects when performance is weak; firm-level shocks would then alter project completion speeds.
- Identification challenge: need a shock that affects firms but is plausibly unrelated to day-to-day project construction; must rule out confounding where “bad firms” choose worse projects.
- Identification strategy: use a contraction in credit supply originating in the banking sector (the Indian Asset Quality Review) and compare projects of more versus less exposed firms, leveraging pre-shock observational similarity.

### A Credit Supply Shock: the Indian Asset Quality Review
- Quasi-experimental setting: Indian Asset Quality Review (AQR), an RBI-conducted audit of scheduled commercial banks between August and December 2015, with annual follow-ups under same guidelines.
- Key institutional points:
  - Prior to the Global Financial Crisis, banks had to automatically reclassify restructured loans into non-performing assets and set aside provisions. On August 27, 2008, RBI suspended automatic reclassification; the forbearance lasted until April 1, 2015.
  - The AQR audited banks, estimated non-performing assets, required additional provisions, and mandated disclosure of audit results in bank financial statements.
  - Audits uncovered sizable under-provisioning across systemic banks and thus could reduce bank lending.
- Constructing the bank shock:
  - Bank Shock_bt = (Provisions Required by the RBI Post-Audit − Provisions Pre-Audit) / Total Assets_bt
  - Interpretation: banks requiring larger additional provisions experience larger reductions in lending capital; normalization by total assets avoids spurious comparisons across bank sizes.
- Empirical approach to separate credit supply from demand:
  - Construct lender-borrower-year dataset using outstanding loans between 2013 and 2015 to fix relationship composition pre-audit; create a balanced panel for years 2013–2019.
  - Estimate Loans_bft = exp(γ × Bank Shock_bt + λ_ft + ω_b + X_bft Γ) e_bft using PPML, where λ_ft are firm-by-year fixed effects (absorb credit demand shocks), ω_b are bank fixed effects, and X_bft are controls. Cluster standard errors at the lender level.
- Table 3 key empirical results:
  - Column (1): a one-standard deviation increase in bank exposure reduces outstanding credit in the average pair by 5.1% (Exposure coefficient = -0.0517, standard error 0.0106; significance denoted ∗∗∗). Outc Mean = 2497.20; Obs = 268,858.
  - Columns (2)–(3): point estimates are economically indistinguishable when relaxing controls and adding firm-level time-varying controls (standard errors and coefficients shown in table).
  - Columns (4)–(6): restricted sample to borrowers with observed Capex projects:
    - Audit shock resulted in a 4.9% average contraction of credit supply among this subsample (Exposure coefficients around -0.0497 to -0.0491 with standard errors 0.0122–0.0131; significance ∗∗∗). Outc Mean = 450.34; Obs = 109,929.
  - Notes on specifications:
    - Coefficients are read as percent effect on outcome for a one standard-deviation increase in lender exposure.
    - Columns (1) and (4) include lender and borrower-by-year fixed effects.
    - Columns (2) and (5) include lender, borrower, and year fixed effects.
    - Columns (3) and (6) add firm controls (sales, sales growth, profitability, debt over equity, book leverage, equity issuances, cash over liquid assets, debt over EBITDA), winsorized at 1st and 99th percentiles, standardized, and lagged one period.
    - Standard errors clustered at lender level; ∗, ∗∗, ∗∗∗ denote 10%, 5%, 1% significance.

### From Banks to Projects: Estimation and Identification
- Mapping bank shock to firm exposure:
  - Firm Exposure_ft = Σ_{b ∈ B_{f,t−1}} w_{b,f,t−1} × Bank Shock_bt
  - w_{b,f,t−1} = Outstanding Loans_{b,f,t−1} / Σ_{b ∈ B_{f,t−1}} Outstanding Loans_{b,f,t−1}
  - Exposure aggregates bank-level shocks weighted by a bank’s lagged share of the firm’s total borrowing.
- Estimating effect of firm exposure on project completion hazards:
  - Generalized linear model: φ(Completed_pft) = β × Firm Exposure_ft + X_pft Γ + υ_pft
    - p indexes project, f sponsor (firm), t calendar time.
    - φ(·) is a link function; X_pft collects controls; υ_pft residuals.
    - Effect of interest: β. Standard errors clustered at firm level.
- Methodological considerations:
  - Use discrete-time survival analysis / duration models to handle right-censoring and avoid selection bias from “complete case analysis.”
  - Data organized into a quarterly panel of projects.
  - Outcome Completed_pft equals zero in all periods before completion and one in the completion period; censoring handled by survival framework.

_Italic source attribution: Content unit: wpiea2025078-print-pdf - Appendix Table A4 collects estimates from the regression specifications._

### Appendix  Table  A5  reports  results  from  such  specification.   Although  our  results  are  slightly  weaker  in  t

### Appendix — Robustness and Identification for Project Completion and Starts

### Specification and Link Function
- Appendix Table A5 reports results from an alternative specification; results are "slightly weaker in the broader sample" but still show evidence of a credit supply contraction.
- Semi-elasticities obtained by dividing the coefficients by the mean of the outcome are similar in magnitude to those estimated in Table 3.
- Quarterly panel preferred despite exposure varying at the yearly level: "Since completion times are small in yearly scale, we would not account well for age effects in an yearly panel; and would not disentangle well the role of project age from shocks to firms, except for firms with the largest project portfolios."
- Right-censoring: if a project is right-censored, either because CapexDx loses track or because the project is ongoing, Completed_pft takes value zero in all time periods.
- Link function choice:
  - Discrete-time hazards coincide with logistic likelihood (Allison, 1982) but logistic regressions may face incidental parameter problems (Neyman and Scott, 1948) when including high-dimensional fixed effects.
  - Authors opt for log link φ(x)=log(x), implying a semi-parametric, piece-wise exponential model for hazards (log-linear hazard, Poisson rate conditional on surviving up to period t).
  - Outcome is binary with unconditional probability of completion "4%".
  - Interpretation can be as Poisson pseudo-maximum likelihood (Silva and Tenreyro, 2006; Gourieroux et al., 1984).

### Identification and Controls
- Identification requires satisfying non-linear GMM moment condition E[υ_pft Exposure_ft | X_pft] = 0 (equation (2)).
- Sufficient that bank shocks are (conditionally) uncorrelated with (loan-weighted) project-level unobservables; identification leverages exogenous variation in the shocks.
- Authors consider two main threats: (1) other factors influence project completion and correlate with the credit supply shock; (2) failure to control for such confounders.
- Controls and strategies used:
  - Firm fixed effects to address persistent firm sorting (e.g., firms with systematically slow build times or relationships with unhealthy banks).
  - State-by-year fixed effects to absorb time-varying regional confounders (political instability, local unemployment, construction labor supply or wages).
  - Year fixed effects or sector-by-year fixed effects to mitigate aggregate or industry-level shocks.
  - Cubic spline on project age to remove baseline hazards and compare projects of similar ages.
  - In the most conservative specifications, controls related to recent firm performance (e.g., sales growth) and firm financial solvency (e.g., leverage, cash-to-assets, debt-to-gross profits). These firm controls are winsorized at the 1st and 99th percentile, standardized, and lagged by one period.
- Practical considerations:
  - Including firm-year fixed effects is infeasible in project-level regressions because exposure varies at the firm-year level and such fixed effects would absorb all variation.
  - For firms with a single project, firm fixed effects compare project-level outcomes before and after exposure; for firms with multiple projects, identification relies on within-firm variation.
  - The authors argue firm and year fixed effects approximate well the information set spanned by firm-year effects based on comparisons in Table 3.

### Table 4: Effects of Firm-Level Credit Supply Shocks on Project Completion Hazards
- Estimation method: PPML estimates of the semi-elasticity of project-level hazards to firm exposure to the Asset Quality Review (model (1)).
- Sample: projects observed quarterly between 2013-2019; firm exposure is weighted-average of lender exposure with weights = outstanding loans in the previous year.
- Coefficients interpreted as percent effect on outcome for a one standard-deviation increase in firm exposure.
- Key point estimates and statistics (columns correspond to progressively richer control sets):
  - Column (1): Exposure = 0.1094 ∗∗∗ (standard error (0.0270)); DepVar Mean = 0.0397; Observations = 122123.
    - Interpretation: moving from average exposure to one-standard deviation higher increases hazard by 10.94%.
  - Column (2): Exposure = 0.0667 ∗∗ (0.0272); DepVar Mean = 0.0397; Observations = 122123.
  - Column (3): Exposure = 0.0636 ∗∗ (0.0273); DepVar Mean = 0.0397; Observations = 122123.
  - Column (4): Exposure = 0.0605 ∗∗ (0.0264); DepVar Mean = 0.0397; Observations = 122123.
  - Column (5) (preferred specification): Exposure = 0.0573 ∗∗ (0.0266); DepVar Mean = 0.0397; Observations = 122123.
  - Column (6): Exposure = 0.0734 ∗∗∗ (0.0264); DepVar Mean = 0.0397; Observations = 122100.
  - Column (7): Exposure = 0.0716 ∗∗∗ (0.0266); DepVar Mean = 0.0397; Observations = 122100.
- Fixed effects and controls across columns include age splines (cubic splines on project age measured in quarters), firm FE, state x year FE, sector x year FE, industry x year FE, and firm controls (sales, sales growth, profitability, debt over equity, book leverage, equity issuances, cash over liquid assets, debt over EBITDA).
- Standard errors clustered at the firm level. Significance: ∗, ∗∗, ∗∗∗ denote 10%, 5% and 1% levels respectively.
- Interpretation and robustness:
  - Column (1) suggests an 11% contemporaneous increase in hazard for one-standard deviation higher exposure.
  - Adding firm FE and state-by-year FE (column 2) reduces effect to 6.6%.
  - Firm-year controls and sector-by-year FE (columns 3 and 4) leave effects largely unchanged (6.36% and 6.05%).
  - Preferred specification (column 5) yields a 5.7% increase in hazards.
  - Tighter industry-by-year specifications (columns 6–7) yield larger completion effects.
  - Authors report robustness in Appendix C: Table A6 (choice of weights conservative; pre-period weights strengthen results), Table A7 (linear models yield similar semi-elasticities), other fixed-effect combinations and windows, and effects largest on impact and decline over time.

### Table 5: Effects of Firm-Level Credit Supply Shocks on Project Starts
- Estimation method: PPML estimates of the semi-elasticity of firm-level project starts to firm exposure to the Asset Quality Review (model (3)).
- Sample: Capex-Prowess matched sample of firms observed yearly between 2013-2019; firm exposure defined analogously to Table 4.
- Coefficients interpreted as percent effect on outcome for a one standard-deviation increase in firm exposure.
- Key point estimates and statistics:
  - Column (1): Exposure = -0.0775 ∗∗∗ (0.0286); DepVar Mean = 0.8362; Observations = 89948.
    - Interpretation: one-standard deviation increase in exposure reduces new project starts by 7.75%.
  - Column (2): Exposure = -0.0698 ∗∗ (0.0325); DepVar Mean = 0.8362; Observations = 89948.
  - Column (3): Exposure = -0.0670 ∗∗ (0.0322); DepVar Mean = 0.8362; Observations = 89948.
  - Column (4): Exposure = -0.0806 ∗∗ (0.0316); DepVar Mean = 0.8362; Observations = 89948.
  - Column (5) (preferred specification): Exposure = -0.0763 ∗∗ (0.0314); DepVar Mean = 0.8362; Observations = 89948.
  - Column (6): Exposure = -0.0584 ∗ (0.0299); DepVar Mean = 0.8326; Observations = 88968.
  - Column (7): Exposure = -0.0588 ∗∗ (0.0296); DepVar Mean = 0.8326; Observations = 88968.
- Fixed effects and controls across columns include firm FE, year FE, NIC section x year FE, division x year FE, and firm controls (same list as Table 4).
- Standard errors clustered at the firm level. Significance: ∗, ∗∗, ∗∗∗ denote 10%, 5% and 1% levels respectively.
- Interpretation and robustness:
  - Bivariate specification (column 1) shows a 7.7% reduction in new project starts for one-standard deviation higher exposure.
  - Adding firm and year FE reduces point estimate to 7.0% (column 2); adding time-varying firm controls barely changes estimates.
  - Preferred specification (column 5) shows a 7.6% reduction; replacing sector-by-year with division-by-year reduces to 5.8% (columns 6–7).
  - Effects are significant and magnitudes are stable across specifications.
  - Appendix C robustness: Table A8 (choice of weights conservative; pre-period weights strengthen results), Table A7 (linear models produce similar semi-elasticities).

### Extensive Margin and Interpretation
- Firms discount future cash flows; accelerating completion increases project NPV but speeding up is costly.
- Evidence consistent with decline in firm investment and reduction in new project starts as an extensive margin response.
- Authors estimate New Project Starts_ft = exp(β × Firm Exposure_ft + X_ft Γ) ν_ft (model (3)), clustering standard errors at the firm level.

### Other Threats to Identification and Diagnostics
- Main threat: absence of firm-time effects (unobserved characteristics that banks internalized when reducing credit and that simultaneously affect firms' willingness to start or complete projects).
- Authors present balance tables and use prior credit regressions (Table 3) as suggestive evidence that borrower-year unobservables are unlikely to drive results: moving from borrower-year to borrower and year fixed effects barely changed credit results.
- Authors argue it is hard to concoct plausible stories where time-varying idiosyncratic firm characteristics generate sizable distortions in banks' propensities to hide losses, absent extreme granular firms or a dispersed lending distribution through many small banks (which they state is not the case).

*Source: Appendix text (Tables 4 and 5 and accompanying discussion) from the provided IMF PDF content unit.*

### Appendix Table A11 ameliorates these concerns. We regress one-period lagged firm char-

### Appendix Table A11 ameliorates these concerns.

### Robustness to predictability and selection concerns
- Regressions of one-period lagged firm characteristics on contemporaneous exposure (including firm and year effects) test for predictability in trends; the design relies on within-firm time-series variation.
- Detectable predictability:
  - profitability
  - cash over assets
- No detectable predictability among other covariates.
- A second concern: the shock could change the data provider’s ability to collect data on projects, creating differential censoring (post-shock sample of projects differing from pre-shock sample), which would invalidate causal statements if present.
- Two analyses in Appendix D to address differential censoring:
  - Restrict completion analyses to projects that were in data vintages prior to the first audits; such projects would be immune to post-shock changes in data collection.
  - Test for differential pre-trends in firm-level project stocks as a direct implication of differential backfilling: if CapexDx became less able to backfill data on treated relative to untreated projects, a downward pre-trend in treated relative to untreated projects should be observed. Across both tests, no evidence was found that results are driven by differential censoring.
- Final concern: projects could have had their scopes reduced after the shock, so a positive effect on the completion hazard might reflect a reduction in exogenous time to build.
  - Appendix E provides:
    - Qualitative evidence on the construction industry arguing common contracts and technical considerations hinder changes in scope of ongoing projects; frictions are more pronounced for large capital expenditure projects undertaken by sizable sponsors (as in the sample).
    - Quantitative robustness: repeat analyses discarding projects proxied to be most susceptible to potential reductions in scope, where classification relies on (a) how the data provider tracks changes in project scopes in the raw data; or (b) technological or regulatory aspects of construction that preclude large within-project variations. Results are similar to baseline, supporting the preferred interpretation.
- Cross-sectional balance tests (Table A10, column (3)):
  - Regress pre-audit firm characteristics on average exposure ex post.
  - Detect cross-sectional predictability in sales and equity issuance, but no predictability otherwise.
  - Note: this test is more stringent than required by the identification assumption, but overall encouraging.

### A Theory of Investment with Endogenous Time to Build — purpose and intuition
- Empirical puzzle: firms facing an adverse shock cut back on new projects (extensive margin) but speed up existing projects (intensive margin), implying speeding up is costly by revealed preference.
- The model rationalizes these empirical findings by letting firms choose time to build endogenously (time to build is a choice, not an exogenous primitive).
- The model is minimalist to allow analytical statements and to lend itself to quantitative applications; richer quantitative exercises are deferred to section 6.

### Setup: projects, payoffs, and frictional investment
- Time is continuous, indexed by t.
- A firm manages a collection of capital investment projects. Each period, a continuum of projects arrive.
- Completing a project requires investing an amount ̄x (bar x). After completion at time T, the firm obtains z units of capital; z is drawn at project inception from a p.d.f. f(z).
- Capital is valued at a common price p; firms take p as given.
- Project returns accrue only upon completion; firm discount rate is r>0.
- Investment flow for a project at time t is i(t). Normalize start time to t=0. Project completion satisfies:
  - ∫_0^T i(t) dt = ̄x
- In absence of frictions, optimal behavior:
  - Invest ̄x immediately (since r>0), so discounted profits are e^{−rT} (p z − ̄x), maximized at T=0 if p z > ̄x; otherwise do not build.
- Introduce frictional costs κ(i), weakly increasing and weakly convex, which do not contribute to installed capacity (deadweight losses). The firm problem becomes:
  - max_{i(t),T} e^{−rT} p z − ∫_0^T e^{−r t} [i(t) + κ(i(t))] dt subject to ∫_0^T i(t) dt = ̄x
- Example functional form for κ(i) (analytical results do not rely on this exact form):
  - κ(i) = 0, i ≤ ̄i
  - κ(i) = κ/2 (i − ̄i)^2, i > ̄i
- Interpretation:
  - ̄i is a natural rate of project completion; accelerating beyond ̄i incurs quadratic cost parametrized by κ.
  - As κ → ∞, model nests exogenous time-to-build with T = ̄x / ̄i.
  - κ captures technological constraints, congestion costs, moral hazard, overtime pay, renegotiation costs, convex labor supply, and other construction frictions.

### Optimal investment: HJB characterization and first-order condition
- Define project state x(t) = ∫_0^t i(s) ds.
- Projects are heterogeneous by stage x and return z. The firm’s HJB with state x is:
  - r v(x,z) = max_{i≥0} [ −i − κ(i) + i ∂v(x,z)/∂x ]
- Terminal condition: v(̄x, z) = p z.
- First-order (necessary) condition (since κ differentiable):
  - 1 + κ′(i) = ∂v(x,z)/∂x
- Interpretation: invest up to the point where marginal cost of investment inclusive of acceleration costs equals the marginal benefit of moving the project forward (project-level marginal q).
- Denote optimal policy by i(x,z); HJB is solved numerically using finite-difference methods.

### Optimal entry and distribution of projects
- Entry rule: start all projects with high enough return z > ̄z, where ̄z solves v(0, ̄z) = 0.
- Firms manage many projects; the distribution over stages evolves. Let g(x,z) denote steady state distribution satisfying the Kolmogorov Forward Equation (KFE):
  - ∀(x,z), ∂/∂x [ i(x,z) g(x,z) ] = 0
- With a constant flow of entrants f(z)·1{z ≥ ̄z}, steady state implies:
  - g(x,z) i(x,z) = f(z)·1{z ≥ ̄z}
  - Hence g(x,z) = f(z)·1{z ≥ ̄z} / i(x,z)

### Completion times
- Time remaining until project completion T(x,z) is given by solving the Kolmogorov Backward Equation with T(̄x,z) = 0, yielding:
  - T(x,z) = ∫_x^{̄x} 1 / i(y,z) dy
- Interpretation: in small dt, dx = i(x,z) dt, so time to move dx is dx / i(x,z); integrate until ̄x.

### Steady state qualitative properties (referenced)
- Value function v(x,z) is monotonically increasing in x and reaches terminal value p z at x = ̄x.
- Optimal investment i(x,z) is endogenously higher for more productive projects (higher z).
- Stationary distribution normalized by entrant flow, g(x,z)/f(z) = 1 / i(x,z), implies greater mass of active projects for lower z because lower-z projects take longer to complete.
- Time to completion T(x,z) is higher for low-productivity projects until reaching zero at x = ̄x.

### Dynamics: response to a credit (provisioning) shock
- The provisioning shock is modeled as an increase in the firm’s discount rate r; Appendix G.5 discusses theoretical channels and empirical evidence supportive of this mapping.
- Intuition: as credit dries up, debt becomes costlier and firms raise internal discount rates; the resulting patterns are consistent with endogenous time to build within a neoclassical framework.
- Specify a perfect-foresight AR(1) path for discount rates:
  - r(t) = r + s e^{−ρ t}
- Let V(x,z,s,t) and I(x,z,s,t) be the value function and investment policy following the shock; they depend on shock size s and calendar time t.
- Lemma 1 (first-order responses):
  - V(x,z,s,t) = v(x,z) + s e^{−ρ t} ˆv(x,z) + o(s)
  - I(x,z,s,t) = i(x,z) + s e^{−ρ t} ˆi(x,z) + o(s)

*Source: wpiea2025078-print-pdf - Appendix Table A11 ameliorates these concerns. We regress one-period lagged firm char-*

### Appendix G.6, we assess whether a model with borrowing constraints can rationalize our empirical findings, and

### Appendix G.6 — Model with Borrowing Constraints and Discount-Rate Shock Formulation

### Model formulation and computational tractability
- The authors prefer the discount rate shock formulation because it is computationally tractable and allows meaningful estimation and counterfactual exercises focused on:
  - Theoretical modeling of endogenous time to build.
  - Rich disaggregated project-level data.
- Steady state parameters used in figure illustrations (median calibration in section 6.1): r=0.01, p=1, ̄x=1, ̄i=0.083, κ=30.1.
- Shock specification used for response illustrations (figure 5): shock persistence ρ=0.17 and shock size s=0.02.

### Analytical characterization: value, investment, and duration
- First-order deviations from steady state are defined by
  - ˆv(x,z) = ∂V/∂s |t,s=0 (steady state value response)
  - ˆi(x,z) = ∂I/∂s |t,s=0 (steady state investment policy response)
- Proposition 2 — First-order response of project value:
  - ˆv(x,z) = −(1−e^{−ρT(x)}/ρ) e^{−rT(x)} p z + ∫_{0}^{T(x)} (1−e^{−ρτ}/ρ) e^{−rτ} [ i(y(τ),z) + κ( i(y(τ),z) ) ] dτ.
  - Limiting case as ρ→0:
    - lim_{ρ→0} ˆv(x,z) = −T(x) e^{−rT(x)} p z + ∫_{0}^{T(x)} τ e^{−rτ} [ i(y(τ),z) + κ( i(y(τ),z) ) ] dτ.
  - Interpretation: the change in project value equals its duration — a weighted sum of NPVs over the remaining life of the project. Duration captures sensitivity to discount rate changes: projects with cash flows further in future fall more in value.
- Proposition 3 — First-order response of optimal investment:
  - ˆi(x,z) = 1/κ''( i(x,z) ) · ∂/∂x ˆv(x,z),
  - with ∂/∂x ˆv(x,z) = 1/i(x,z) [ v(x,z) + (r+ρ) ˆv(x,z) ].
  - Interpretation: re-optimization equates marginal cost of accelerating completion to marginal value of reducing duration. Two opposing effects in ∂/∂x ˆv: (i) impatience increases value of earlier cash flows (positive, via v(x,z)), (ii) higher discounting lowers future cash flows (negative, via ˆv).
- Corollary 4 — Near-completion effect:
  - For x sufficiently close to ̄x, ˆi(x,z) > 0.
  - Intuition: as x→̄x, ˆv(x,z)→0 while v(x,z)→p z > 0, so marginal incentive to accelerate is positive. Thus mature projects are accelerated after an increase in discount rates.
- Qualitative implications:
  - Projects close to completion may increase investment following a discount rate rise (intensive margin).
  - Projects with long time to build may optimally slow down if negative duration effects dominate.
  - Aggregate (firm-level) investment responses require accounting for distribution of projects across stages g(x,z) and entry effects (extensive margin).

### Entry threshold and extensive margin
- Proposition 5 — First-order response of the entry threshold ̄z:
  - Let ̄Z(t,s) = ̄z + s e^{−ρt} ˆz + o(s), where ˆz = ∂ ̄Z/∂s |t,s=0.
  - The first-order response satisfies: e^{−r T(0,̄z)} × p ˆz = − ˆv(0,̄z).
  - Interpretation: rise in discount rates lowers value of unbegun projects by ˆv(0,̄z); to restore positive NPV at the margin, the entry hurdle ̄z must rise by ˆz > 0. Hence the flow of newly started projects decreases after the increase in discount rates (extensive margin), consistent with empirical results.

### Aggregation to firm-level investment
- Law of motion for capital stock K:
  - ̇K(t) = I_K(t) − δ K(t).
- Fixed capital formation equals aggregate flow of finished projects weighted by their output z:
  - I_K(t) = ∫ z · g(x,z,t) · i(x,z,t) dz.
- Accounting investment aggregates ongoing project investment flows:
  - I_acc(t) = ∫∫ g(x,z,t) · i(x,z,t) dz dx.
- In steady state (footnote): accounting investment is independent of time to build; fixed capital formation depends on time to build via its effect on the entry threshold.

### Calibration and identification strategy
- Heterogeneity:
  - Project sizes ̄x drawn from lognormal: log ̄x ∼ N(μ_{̄x}, σ^2_{̄x}). Parameters set to match the 10th, 25th, 50th, 75th, and 90th percentiles of empirical log costs.
  - Costless rate of completion ̄i drawn from lognormal with mean dependent on log project size:
    - log ̄i ∼ N(β_{̄i 0} + β_{̄i 1} log ̄x, σ^2_{̄i}).
  - Pareto distribution for project returns f(z) with minimum draw z_min and tail exponent α; α estimated from firm revenue rank regressions after trimming 1% outliers.
- Mapping banking shocks:
  - Banking shocks are mapped to increases in firms’ discount rates.
  - Shock persistence ρ is estimated from panel autoregressions of Exposure_{f t} on its first lag with firm and year fixed effects, using various dynamic panel estimators (OLS without fixed effects, Han and Phillips (2010) GMM, Arellano and Bover (1995) and Blundell and Bond (1998) system GMM). Quarterly figures are constructed from estimates of ρ.
- Identification of shock size s and endogeneity parameter κ:
  - Use empirical semi-elasticities β_int (intensive) and β_ext (extensive) from sections 4.2 and 4.3 (point estimates in columns (5) of Tables 4 and 5).
  - β_ext informative about s but uninformative about κ.
  - β_int informative about both s and κ.
  - Parameters (s, κ) chosen by minimum distance to match empirical semi-elasticities.

### Quantitative findings and calibrated values
- Calibration/estimation outcomes reported in the text:
  - Shock size estimated s=2%.
  - Acceleration cost parameter estimated κ=3.
- Notes on figures and fitted responses:
  - Figure 4 (steady state objects as functions of project stage) uses median calibration values r=0.01, p=1, ̄x=1, ̄i=0.083, κ=30.1.
  - Figure 5 (first-order responses to discount rate shock) plots percent-deviations from steady state for ˆv(x,z) and ˆi(x,z); shock persistence ρ=0.17 and shock size s=0.02 used for illustration. Solid green lines represent low output projects (z=2); black dashed lines represent high output projects (z=8).
- Model fit:
  - The authors state that Appendix Table A20 reports the calibrated model fits well the targeted moments.
  - Appendix Table A19 collects calibrated parameters (the text reports the calibrated shock size s=2% and acceleration cost κ=3).

*Source: wpiea2025078-print-pdf — Appendix G.6*

### 6.2    A Sharper Test: Heterogeneous Effects

### 6.2    A Sharper Test: Heterogeneous Effects

### Model prediction: Heterogeneous effects by project stage
- The model predicts that the effect of discount rate shocks on investment is heterogeneous depending on the stage of the project when the shock hits:
  - Investment rises when discount rates rise for either (i) sufficiently high output projects, or (ii) sufficiently late-stage projects.
  - The effect of the shock would have been greater had the project been at a later stage relative to an earlier stage. This holds for any z and provides a sharper test of the mechanism.

### Empirical challenges and proxy construction
- Two empirical challenges for testing heterogeneous treatment effects:
  - The project’s value put in place (the empirical counterpart of model stage x) is not observed.
  - A project cannot simultaneously be in two distinct stages when the shock hits (no counterfactual stage observed).
- Proxy stage variable constructed as:
  - Stage_p = Initial Age_p / Predicted Time to Build_p
  - Intuition: fixing expected time to build, older projects are more advanced; fixing age, projects with shorter average time to build are more advanced.
- Prediction step for time to build:
  - Use completed projects in CapexDx not in regression sample; regress time to build on exogenous predictors.
  - Compare specifications by out-of-sample R^2 in the regression sample; choose model with highest out-of-sample R^2 (model including industry dummies).
  - Proxy measure winsorized at the 1st and 99th percentile and then standardized before inclusion.
- Preference for using age when entering the hazard regression sample rather than age at time of shock:
  - Some projects enter and leave sample before/after asset quality review audits; age at entry is always definable.
  - Initial age and year fixed effects help partly reconstruct age at shock.

### Regression evidence (Table 6) — effects by proxied stage
- Estimation: PPML estimates of semi-elasticity of project-level hazards to firm exposure to the Asset Quality Review; sample: projects observed quarterly between 2013-2019.
- Firm exposure: weighted-average of lender exposure, weights correspond to outstanding loans in the previous year.
- Key coefficient estimates (reported in Table 6 across specifications):
  - Exposure: 0.090 ∗∗∗, 0.086 ∗∗∗, 0.094 ∗∗∗, 0.090 ∗∗∗ (standard errors: (0.025) each)
  - Stage x Exposure: 0.068 ∗∗∗, 0.069 ∗∗∗, 0.078 ∗∗∗, 0.079 ∗∗∗ (standard errors: (0.021), (0.021), (0.023), (0.023))
- Notes on interpretation and controls:
  - Coefficient on exposure interpreted as percent effect on outcome for a one standard-deviation increase in firm exposure for the project with mean proxy stage.
  - Coefficient on interaction is change in main effect resulting from a one-standard deviation level change in proxy stage.
  - Example from Table 6, column (1): moving from average exposure to one-standard-deviation higher exposure increases hazard by 9% for project at mean proxy stage; if project is one standard deviation above (below) mean stage, effect would increase (decline) by 6.8%, for total effect of 15.8% (2.2%) as reported in the table notes.
  - Regression controls include initial age fixed effects, firm fixed effects, state x year fixed effects, sector x year or industry x year fixed effects (depending on column), firm controls (sales, sales growth, profitability, debt over equity, book leverage, equity issuances, cash over liquid assets, debt over EBITDA). Firm controls winsorized at 1st and 99th percentile and standardized; lagged by one period.
  - Standard errors clustered at firm level. ∗, ∗∗, ∗∗∗ denote significance at 10%, 5% and 1% levels.
- Text summary of regression magnitudes:
  - Across specifications, for a project at the mean proxy stage, a one-standard deviation increase in exposure raises the hazard by roughly 9%.
  - If the project stage had been one standard deviation above (below) the mean, the effect would have been increased (reduced) by roughly 7.5%, so the total effect of the shock would have been 15.5% (2.5%).
  - Appendix Table A13: similar results if initial age is not scaled by predicted time to build (robustness to prediction step).

### Model-data comparison (Figure 6)
- Model plots:
  - Red dashed line: cumulative distribution of the proxy stage variable in the model (use unconditional expected time to build in the model).
  - Blue solid line: heterogeneous effects of a discount rate shock on hazards by the proxy stage variable.
- Quantitative match:
  - A unit-increase in proxied stage increases the hazard by 8% in the model.
  - This model-based estimate lies within the regression’s confidence bands and is untargeted in calibration — interpreted as encouraging agreement between model and data.

### Investment dynamics: endogenous versus exogenous time to build (Figures 7–9)
- Partial equilibrium setup: responses of firm-level aggregate investment to (i) discount rate shock (holding p fixed) and (ii) price/payoff shock (holding discount rates fixed); shock processes are AR(1) perfect-foresight with parameters as in appendix table A19.
- Exogenous time to build (κ→∞, completely exogenous TTB):
  - Investment falls, reaches a trough 12-16 quarters later, and then returns to steady state (hump-shaped IRF).
  - Mechanism: higher discount rates lower value of projects, raise threshold for positive NPV, fewer projects started today → lower future investment on ongoing projects.
- Endogenous time to build:
  - Investment initially rises following a discount rate increase (short-lived spike at impact).
  - Mechanism: mass of projects close to completion are sped up (acceleration effect) → immediate surge in investment flows.
  - Shortly after impact, investment collapses as intensive margin effects become negative; completion of late-stage projects skews remaining distribution toward early-stage projects with lower optimal investment rates → drives drop in investment.
  - Decomposition: intensive margin (changes in investment rates of ongoing projects) and extensive margin (changes due to entry of new projects) add together to generate steeper fall.
  - Endogenous time to build generates boom-bust dynamics analogous to investment accelerator theories, microfounded by optimal investment timing.
- Price/payoff shocks (Figure 8):
  - Exogenous TTB: positive price shock raises entry and investment analogously.
  - Endogenous TTB: short-run spike in investment followed by a long-lived mild slump due to distribution skewing toward earlier-stage projects.
- Volatility and amplification:
  - Moving from exogenous to endogenous time to build increases investment volatility by 30%.
  - Endogenous TTB produces endogenous amplification of structural shocks rather than needing larger or more volatile shocks.
- Proposition 6 (brief statement):
  - Fix any z > z̄. For a sufficiently small and temporary shock, the cumulative effect on the intensive margin of investment is zero.
  - Interpretation: intensive margin IRF has no cumulative effects for shocks that allow the distribution to return to steady state; endogenous TTB reshuffles investment flows over time producing time-series variation with null cumulative effect and hence higher volatility.

### State-dependence of monetary policy effects (Figure 9)
- Heterogeneous effects imply the cross-sectional distribution of projects at shock time determines aggregate response.
- Scenario analysis (three distributions over project stages):
  - Steady state distribution (baseline).
  - Distribution skewed towards younger projects (economy recently had influx of new projects).
  - Distribution skewed towards older projects (economy experienced prolonged period of no starts).
- Impulse responses to a discount rate shock starting from each distribution:
  - If economy starts filled with mature projects, the initial investment spike is larger.
  - If economy starts filled with fledgling projects, the impact response can turn negative.
  - The distributional dependence is strong enough that impact effect can flip sign depending on initial conditions.
- Policy implication:
  - Interest rate changes’ effects on investment are state-dependent when time to build is endogenous.
  - Interest rate hikes (cuts) will be less contractionary (expansionary) on impact when the project distribution is skewed towards more mature projects.
  - The investment channel of monetary policy may face headwinds absent in models with exogenous time to build.

*Source: IMF working paper chapter "6.2    A Sharper Test: Heterogeneous Effects" (wpiea2025078-print-pdf).*

### appendix table A19.

### appendix table A19.

### 7.3 Tax Policy: Reshuffling Investment over Time
- Setup:
  - Study of an unanticipated investment tax credit announced at t=0 lasting for 10 quarters.
  - Firms know the policy duration and respond in general equilibrium (allowing p(t) to vary).
  - Remaining parameters are as in appendix table A19.

- Empirical/model dynamics (IRFs shown in Figure 10):
  - Exogenous time to build (TTB) case:
    - A positive cash flow shock induces project entry.
    - Accounting investment slowly rises as firms build nascent projects.
    - Investment starts to return to steady state after the credit expires.
    - Long-lived effects arise from distributional dynamics: since projects were started, the IRF accounts for cash flows that will be spent on these projects even after the tax credit is gone; investment returns to steady state only as projects are completed.
  - Endogenous TTB case:
    - Investment rises on impact, keeps rising, then collapses at the time the credit expires.
    - Investment continues to decline back to steady state and eventually runs into negative territory — behaving “as if” the tax credit had reallocated investment flows across time, front-loading them to periods when the credit was in place.
  - Decomposition (panels c and d):
    - Extensive margin: accounts for the exogenous case (fix investment policies to steady state, allow entry to respond).
    - Intensive margin: explains reshuffling of investment flows over time (shut down entry, allow investment rates to respond).
    - Intensive margin is entirely responsible for the impact effect: investment flows rise for all ongoing projects; when the credit expires, investment in ongoing projects collapses.
    - Distributional dynamics: more mature projects respond more strongly; as projects are completed, the distribution skews to younger projects, producing a bust when the credit expires; as these projects age, investment flows pick up and aggregate impulse returns to steady state.
  - Productive capital:
    - Accumulates more quickly in the endogenous case, reflecting front-loading of investment flows, particularly for projects closer to completion.

- Policy takeaways:
  - Takeaway 1: The cumulative response of investment to tax credits comes from project starts. To first order, temporary tax credits imply intertemporal reshuffling of expenditures rather than net permanent increases. Duration and size of the tax credit are two levers to intertemporally reshuffle aggregate demand.
  - Takeaway 2: Investment tax credits are a particularly flexible tool to spur investment relative to monetary policy because tax policy can more flexibly target the “short-end” of the cash flow curve relative to the “long-end”.
  - Takeaway 3 (state-dependence): Effects are state-dependent. If the distribution of projects is skewed to later-stage projects when the credit is introduced, stronger intertemporal reshuffling and a stronger short-term investment boom will occur; if skewed to earlier-stage projects, aggregate effects are muted.
  - Additional note: Investment subsidies can speed up structural change (e.g., green transition or reshoring). The model clarifies trade-offs between speed gains and acceleration costs and provides a framework to quantify such costs and benefits.

### 7.4 General Equilibrium
- Partial vs general equilibrium:
  - Estimated causal elasticities inform partial equilibrium responses (state-by-year and sector-by-year fixed effects absorb local and sectoral price changes).
  - Partial equilibrium isolates mechanisms but is not informative about general equilibrium responses (cannot escape the “missing intercept” problem).
  - The partial equilibrium model is embedded into a minimalist general equilibrium setting to endogenize capital prices while retaining exogenous discount rates r(t).

- Production and capital price setup:
  - Final goods production: Y(t) = A(t) K(t)^α where α ∈ (0, 1] and A(t) is stochastic productivity; price of final good normalized to one.
  - Marginal product of capital: MPK(t) = α A(t) K(t)^{α−1}.
  - Capital prices reflect future marginal products of capital:
    - p(t) = ∫_{0}^{∞} e^{−(r+δ) s} MPK(t+s) ds.
  - When α = 1, constant returns to scale; partial equilibrium with given price sequence can be recast as GE with particular A(t). When α < 1, decreasing returns imply GE forces dampen partial equilibrium responses as capital accumulation lowers MPK and capital prices.

- Proposition 7 (linearized response of capital prices):
  - Let hatted variables denote linearized deviations from steady state. Suppose announced paths Â(t) and r̂(t) for aggregate productivity and real interest rates. Then, to first-order, the response of capital prices is given by:
    - p̂(t)/p = (r+δ) ∫_{0}^{∞} e^{−(r+δ)s} [ Â(t+s)/A − (1−α) K̂(t+s)/K − r̂(t+s)/(r+δ) ] ds (18)
  - Interpretation: three channels change capital prices — direct productivity effect on MPK, direct effect of interest rate shocks on discounting of future MPK, and feedback from changes in future capital stock K̂(t+s)/K.

- GE impulse responses to discount rate shocks (Figures 11 and 12):
  - Setup: discount rate shock magnitude and persistence same as partial equilibrium exercise; productivity shocks shut down; shock responses plotted for accounting investment and fixed capital.
  - Exogenous TTB:
    - Response of accounting investment similar to partial equilibrium but more pronounced due to GE lowering capital prices and discouraging project entry; productive capital falls.
    - End results: productive capital falls; persistent falls arise despite short-lived shocks.
  - Endogenous TTB:
    - Predicts a sharp contraction in investment on impact (a bust rather than a boom compared to exogenous TTB).
    - Two reasons:
      1. Direct impact of discount rate shock on capital prices lowers project returns and discourages investment.
      2. Projects close to completion are accelerated, generating an upward bump in productive capital and downward pressure on MPK, further depressing incentives.
    - Slower project completion causes more persistent fall in productive capital relative to exogenous case.
  - Decomposition in endogenous TTB (Figure 12):
    - Intensive margin effects: large and short-lived; drive short-run response of investment.
    - Extensive margin effects: sizable and delayed; drive long-lived dynamics.
    - The boom-bust dynamics from intensive margin effects survive in general equilibrium because distributional shifts toward later-stage projects produce temporary elevated investment in ongoing projects as shocks fade; GE changes in capital prices do not undo these mechanisms.

- Comment on counterfactual policy analysis:
  - Comparison with a model featuring convex adjustment costs of installation (no TTB):
    - Convex installation costs predict a sharp, short-lived contraction in investment with no hump-shaped dynamics; productive capital falls more quickly but converges faster to the original steady state.
    - Time-to-build (especially endogenous) can generate hump-shaped investment dynamics and more persistent slumps in capital supply.
  - Adjustment costs at aggregate or investment level are reduced-form devices for gestation lags; with endogenous TTB, effects of shocks vary with the underlying distribution of projects (which depends on steady-state policy stances). Thus, adjustment costs become state-dependent and ad hoc adjustment-cost models may yield inappropriate counterfactuals.
  - Illustrations:
    - Monetary policy: impact sign and strength of investment response depend on distribution of project completion stages.
    - Fiscal policy: short-run effects of investment tax credits are stronger and state-dependent.
  - Conclusion about embedding: Endogenous time to build provides empirically plausible microfoundation for adjustment costs and can be embedded in richer models of firm, industry, or aggregate dynamics.

### 8 Conclusion (selected points)
- Empirical and theoretical summary:
  - Time to build is pervasive across industries, sponsor types, time periods, and project sizes.
  - Firms have causal control over gestation lags; endogenous TTB rationalizes empirical findings.
  - Endogenous TTB implies:
    - Investment is more responsive and more volatile to shocks.
    - More mature projects react more strongly than fledgling projects (empirically borne out).
    - Distribution of projects across completion stages shapes aggregate dynamics and makes monetary policy state-dependent.
    - Investment in ongoing projects is highly responsive to cash flow variations, enabling time-varying investment credits to reshuffle aggregate demand across cycles and offering roles for investment subsidies in structural transitions (e.g., green transition).

- Open questions and extensions for future research:
  - Designing project financing that internalizes gestation lag incentives.
  - Whether firms internalize competitors' gestation lags and implications for industry dynamics.
  - Political economy: do short-termist politicians accelerate projects and overruns to claim credit before tenure ends?
  - Quantifying how targeted subsidies can speed infrastructure development or a global green transition.
  - Generalizing insights to breakthrough innovations or intangible capital.

*Source: appendix table A19.*

### References

### References

### Key citations
- Abadie, A., Athey, S., Imbens, G. W., and Wooldridge, J. M. (2022). When Should You Adjust Standard Errors for Clustering? The Quarterly Journal of Economics, 138(1):1–35.
- Abel, A. B. (1981). A dynamic model of investment and capacity utilization*. The Quarterly Journal of Economics, 96(3):379–403.
- Abel, A. B. and Eberly, J. C. (1994). A unified model of investment under uncertainty. The American Economic Review, 84(5):1369–1384.
- Achdou, Y., Han, J., Lasry, J.-M., Lions, P.-L., and Moll, B. (2021). Income and Wealth Distribution in Macroeconomics: A Continuous-Time Approach. The Review of Economic Studies, 89(1):45–86.
- Aftalion, A. (1927). The theory of economic cycles based on the capitalistic technique of production. The Review of Economics and Statistics, 9(4):165–170.
- Aïd, R., Federico, S., Pham, H., and Villeneuve, B. (2015). Explicit investment rules with time-to-build and uncertainty. Journal of Economic Dynamics and Control, 51:240–256.
- Alfaro, L. and Chari, A. (2014). Deregulation, Misallocation, and Size: Evidence from India. Journal of Law and Economics, 57(4):897–936.
- Allison, P. D. (1982). Discrete-time methods for the analysis of event histories. Sociological Methodology, 13:61–98.
- Antras, P. (2023). An ’Austrian’ Model of Global Value Chains. NBER Working Papers 30901, National Bureau of Economic Research, Inc.
- Arellano, M. and Bover, O. (1995). Another look at the instrumental variable estimation of error-components models. Journal of Econometrics, 68(1):29–51.
- Asea, P. K. and Zak, P. (1999). Time-to-build and cycles. Journal of Economic Dynamics and Control, 23(8):1155–1175.
- Bajari, P. and Tadelis, S. (2001). Incentives versus transaction costs: A theory of procurement contracts. The RAND Journal of Economics, 32(3):387–407.
- BBC (2020). Coronavirus: The hospital built in a matter of days. https://www.bbc.com/news/in-pictures-51280586. Last accessed Oct 2024.
- Berger, D., Milbradt, K., Tourre, F., and Vavra, J. (2021). Mortgage prepayment and path-dependent effects of monetary policy. American Economic Review, 111(9):2829–78.
- Bloesch, J. and Weber, J. (2023). Congestion in onboarding workers and sticky r&d. tech report.
- Blundell, R. and Bond, S. (1998). Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics, 87(1):115–143.
- Bok, K., Sitar, S., Graham, B. S., and Mascola, J. R. (2021). Accelerated covid-19 vaccine development: milestones, lessons, and prospects. Immunity.
- Borusyak, K., Hull, P., and Jaravel, X. (2021). Quasi-Experimental Shift-Share Research Designs. The Review of Economic Studies, 89(1):181–213.
- Brooks, P. K. (2000). The facts about time-to-build. IMF Working Papers 2000/138, International Monetary Fund.
- BÃ¶hm-Bawerk, E. (1889). The Positive Theory of Capital. Macmillan.
- Caballero, R. J. and Engel, E. M. R. A. (1999). Explaining investment dynamics in u.s. manufacturing: A generalized (s,s) approach. Econometrica, 67(4):783–826.
- Casares, M. (2006). Time-to-build, monetary shocks, and aggregate fluctuations. Journal of Monetary Economics, 53(6):1161–1176.
- Chari, A., Jain, L., and Kulkarni, N. (2021). The unholy trinity: Regulatory forbearance, stressed banks and zombie firms. Working Paper 28435, National Bureau of Economic Research.
- Chen, J. and Roth, J. (2022). Logs with zeros? Some problems and solutions. Papers 2212.06080, arXiv.org.
- Chodorow-Reich, G. (2013). The Employment Effects of Credit Market Disruptions: Firm-level Evidence from the 2008-9 Financial Crisis *. The Quarterly Journal of Economics, 129(1):1–59.
- Chodorow-Reich, G., Nenov, P. T., and Simsek, A. (2021). Stock market wealth and the real economy: A local labor market approach. American Economic Review, 111(5):1613–57.
- Chopra, Y., Subramanian, K., and Tantri, P. L. (2020). Bank Cleanups, Capitalization, and Lending: Evidence from India. The Review of Financial Studies, 34(9):4132–4176.
- Christiano, L., Eichenbaum, M., and Evans, C. (2005). Nominal rigidities and the dynamic effects of a shock to monetary policy. Journal of Political Economy, 113(1):1–45.
- CMIE (2023a). Capexdx: Database of the life cycle of new investments. Last Accessed: 2023-09-01.
- CMIE (2023b). Prowessdx: Database of financial performance of indian companies. Last Accessed: 2023-09-01.
- Cohn, J. B., Liu, Z., and Wardlaw, M. I. (2022). Count (and count-like) data in finance. Journal of Financial Economics, 146(2):529–551.
- Correia, S., Guimarães, P., and Zylkin, T. (2020). Fast poisson estimation with high-dimensional fixed effects. The Stata Journal, 20(1):95–115.
- Correia, S., Guimarães, P., and Zylkin, T. (2021). Verifying the existence of maximum likelihood estimates for generalized linear models.
- Dalvi, A. (2024). Vinfast plant in tamil nadu to open six months ahead of schedule in first-half 2025. Technical report, Autocar Professional, https://www.autocarpro.in/news-national/vinfast-plant-in-tamil-nadu-to-open-six-months-ahead-of-schedule-in-first-half-2025-121239. Last acccessed Oct 2024.
- Eberly, J., Rebelo, S., and Vincent, N. (2012). What explains the lagged-investment effect? Journal of Monetary Economics, 59(4):370–380.
- Edge, R. M. (2007). Time-to-build, time-to-plan, habit-persistence, and the liquidity effect. Journal of Monetary Economics, 54(6):1644–1669.
- Eichenbaum, M., Rebelo, S., and Wong, A. (2022). State-dependent effects of monetary policy: The refinancing channel. American Economic Review, 112(3):721–61.
- Flanagan, T. and Purnanandam, A. (2023). Why do banks hide losses? Working paper.
- Freixas, X., Peydro, J.-L., and Laeven, L. (2015). "Systemic Risk, Crises, and Macroprudential Regulation". The MIT Press.
- Freixas, X. and Rochet, J.-C. (2008). Microeconomics of Banking, 2nd Edition, volume 1 of MIT Press Books. The MIT Press.
- Gabaix, X. and Ibragimov, R. (2011). Rank1/2: A simple way to improve the ols estimation of tail exponents. Journal of Business & Economic Statistics, 29(1):24–39.
- Goldberg, A. (1992). The Pentagon: The First Fifty Years. Government Printing Office.
- Gomtsyan, D. (2024). Building the city under financial frictions. Review of Economic Dynamics, 52:70–83.
- Gourieroux, C., Monfort, A., and Trognon, A. (1984). Pseudo maximum likelihood methods: Theory. Econometrica, 52(3):681–700.
- Greenwood, R. and Hanson, S. G. (2014). Waves in Ship Prices and Investment *. The Quarterly Journal of Economics, 130(1):55–109.
- Hall, R. (1977). Investment, interest rates, and the effects of stabilization policies. Brookings Papers on Economic Activity, 8(1):61–122.
- Han, C. and Phillips, P. C. B. (2010). Gmm estimation for dynamic panels with fixed effects and strong instruments at unity. Econometric Theory, 26(1):119–151.
- Hayashi, F. (1982). Tobin’s marginal q and average q: A neoclassical interpretation. Econometrica, 50(1):213–224.
- Jampol, N. (2024). Tesla shares first render of its new megafactory located outside of the u.s.:’speeding up construction’. Technical report, The Cool Down, https://www.thecooldown.com/green-business/tesla-megapacks-shanghai-factory-update/.
- Judy, S. (2020). In atlanta, miami beach, contractors speed delivery of covid-19 hospital facilities. Technical report, Engineering News-Record. Last accessed Oct 2024.
- Kalecki, M. (1935). A macrodynamic theory of business cycles. Econometrica, 3(3):327–344.
- Kalouptsidi, M. (2014). Time to build and fluctuations in bulk shipping. American Economic Review, 104(2):564–608.
- Khwaja, A. I. and Mian, A. (2008). Tracing the impact of bank liquidity shocks: Evidence from an emerging market. American Economic Review, 98(4):1413–42.
- Kulkarni, N., Ritadhi, S., Vij, S., and Waldock, K. (2021). Unearthing zombies. Working paper, Georgetown McDonough School of Business.
- Kydland, F. E. and Prescott, E. C. (1982). Time to build and aggregate fluctuations. Econometrica, 50(6):1345–1370.
- Leeper, E. M., Walker, T. B., and Yang, S.-C. S. (2010). Government investment and fiscal stimulus. Journal of Monetary Economics, 57(8):1000–1012.
- Lewis, G. and Bajari, P. (2011). Procurement contracting with time incentives: Theory and evidence. The Quarterly Journal of Economics, 126(3):1173–1211.
- Lewis, G. and Bajari, P. (2014). Moral Hazard, Incentive Contracts, and Risk: Evidence from Procurement. The Review of Economic Studies, 81(3):1201–1228.
- Liu, E. and Tsyvinski, A. (2023). A dynamic model of input-output networks. Review of Economic Studies.
- Lucca, D. O. (2007). Resuscitating Time-to-Build. 2007 Meeting Papers 909, Society for Economic Dynamics.
- Majd, S. and Pindyck, R. S. (1987). Time to build, option value, and investment decisions. Journal of Financial Economics, 18(1):7–27.
- Mayer, T. (1960). Plant and equiptment lead times. The Journal of Business, 33(2):127–132.
- McDonald, R. and Siegel, D. (1986). The value of waiting to invest. The Quarterly Journal of Economics, 101(4):707–727.
- McKay, A. and Wieland, J. F. (2021). Lumpy durable consumption demand and the limited ammunition of monetary policy. Econometrica, 89(6):2717–2749.
- Meier, M. (2020). Supply Chain Disruptions, Time to Build, and the Business Cycle. Crc tr 224 discussion paper series.
- Montgomery, M. R. (1995). Time-to-build completion patterns for nonresidential structures, 1961-1991. Economics Letters, 48(2):155–163.
- Neyman, J. and Scott, E. L. (1948). Consistent estimates based on partially consistent observations. Econometrica, 16(1):1–32.
- Nickell, S. (1981). Biases in dynamic models with fixed effects. Econometrica, 49(6):1417–1426.
- Oh, H., Yang, C., and Yoon, C. (2024). Land development and frictions to housing supply over the business cycle. Finance and Economics Discussion Series.
- Oh, H. and Yoon, C. (2020). Time to build and the real-options channel of residential investment. Journal of Financial Economics, 135(1):255–269.
- Pigou, A. (1920). The Economics of Welfare. Macmillan.
- Ramey, V. (2020). The macroeconomic consequences of infrastructure investment. NBER Working Papers 27625, National Bureau of Economic Research, Inc.
- Ramsey, M. (2016). Tesla races to finish âgigafactoryâ in time for model 3 rollout. Technical report, WSJ, https://www.wsj.com/articles/tesla-races-to-finish-gigafactory-in-time-for-model-3-rollout-1469404142.
- Rollo, C. (2024). 10 things you probably didn’t know about the pentagon. https://www.defense.gov/News/Feature-Stories/story/Article/1650913/10-things-you-probably-didnt-know-about-the-pentagon/.
- Rouwenhorst, K. (1991). Time to build and aggregate fluctuations: A reconsideration. Journal of Monetary Economics, 27(2):241–254.
- Samuelson, P. A. (1939). Interactions between the multiplier analysis and the principle of acceleration. The Review of Economics and Statistics, 21(2):75–78.
- Silva, J. M. C. S. and Tenreyro, S. (2006). The Log of Gravity. The Review of Economics and Statistics, 88(4):641–658.
- Singer, J. D. and Willett, J. B. (1993). It’s about time: Using discrete-time survival analysis to study duration and the timing of events. Journal of Educational Statistics, 18(2):155–195.
- Taylor, J. B. (1979). Staggered wage setting in a macro model. The American Economic Review, 69(2):108–113.
- Taylor, J. B., Baily, M. N., and Fischer, S. (1982). The swedish investment funds system as a stabilization policy rule. Brookings Papers on Economic Activity, 1982(1):57–106.
- Van Den Berg, G. J. (2001). Chapter 55 - duration models: Specification, identification and multiple durations. volume 5 of Handbook of Econometrics, pages 3381–3460. Elsevier.
- Vogel, S. (2007). The Pentagon â A History: The Untold Story of the Wartime Race to Build the Pentagon and to Restore it Sixty Years Later. Random House.
- Winberry, T. (2021). Lumpy investment, business cycles, and stimulus policy. American Economic Review, 111(1):364–96.
- Xuequan, M. (2020). Commentary: 2 in 2 weeks – hospital building speed matters in china’s virus battle. http://www.xinhuanet.com/english/2020-02/06/c138761691.htm. Last accessed October 2024.

### Data Appendix — A: Overview
- Sections included: A.1 CapexDx; A.1.1 Defining Projects; A.1.2 Data Collection; A.1.3 Raw Data Structure; A.2 Constructing Event Histories.
- Data vintage used for analyses: March 31st 2023 vintage.

### A.1 CapexDx — Definitions and scope
- CMIE defines a project when it can assess the capacity that will be created.
- Illustrative example: a sponsor investing "a hundred million rupees" does not constitute a project unless capacity is specified (factory, road, school, etc.).
- If sponsor specifies a "glass manufacturing plant in Chennai" for "a hundred million rupees", CMIE tracks the project.
- CMIE would not track acquiring an already existing plant; project must be development of capacity.
- Multi-faceted projects: CMIE recognizes smaller logical components as independent projects if:
  - "the cost for each of them is available separately" and
  - "the implementation is clearly separate".
- Example: four plants each costing "25 million rupees" would be tracked individually if specified; exception if sponsor treats them as a single integrated structure.
- Resulting observed label example: "Glass Manufacturing Project" with a total cost of "100 rupees" located in Chennai (as described in documentation).

### A.1.2 Data Collection — procedures and monitoring
- CMIE builds Capex by combing publicly available sources and by actively contacting sponsors via emails and phone calls.
- Monitoring frequency: "all projects 'are touched for updates once in a quarter.'"
- CMIE validates sponsor information against "Media, government websites" and other public disclosures when available.
- CMIE checks and updates outdated implementation dates when corroborating evidence exists in the public domain.
- Dating convention: announcement dates correspond to the first time the project is mentioned in the public domain; implementation start dates mark the first time CMIE learns an implementation-related event has taken place.
- Right-censoring policy: CapexDx discloses it stops tracking projects "if they are stalled or shelved or abandoned or information regarding them is not available for over 30 months in the case of manufacturing projects and for over 36 months in the case of infrastructure projects."
- CMIE may resume tracking previously censored projects if it obtains reliable information on project status from the public domain.

### A.1.3 Raw Data Structure — files used
- Principal files (obtainable with a CapexDx subscription):
  - Project Detail file: cross-sectional data on project characteristics, such as project name and id, sponsor name and id, state and district of project, most recent project status, etc.
  - Project Costs file: data at the project-cost estimate level, including date of cost estimate. Projects may have multiple cost estimates over their lifetime, referred to as "restatements".
  - Project Completion file: cross-sectional data for completed projects, specifying dates of completion and previous estimates of completion dates.
  - Project Events file: data at the project-event-event date level, tracks project lifecycle.

### A.2 Constructing Event Histories — methodology for panel creation
- Raw event-level data contain one "event" per row; multiple events can occur on the same date.
- For hazard models, the authors construct a project-time panel tracking whether the project is in one of these stages: announced, shelved, stalled, under implementation, completed, abandoned, or lost to follow-up. These correspond to labels in the Project Detail file.
- Terminal (absorbing) stages: Completion, abandonment/shelving, lost to follow-up.
- Intermediate stages: announced, stalled, under implementation — projects may transition in and out.
- Pre-processing steps (examples and rules):
  - Drop non-informative monitoring events such as "Company contacted successfully" or "Company contacted but no response" under specific conditions:
    - if another event on the same date is not a monitoring marker;
    - or if they occur alone but are neither the first nor last event in the project’s history;
    - or if they are the last observations but the preceding event is completion, abandonment, or shelving.
  - Remove events tied to particular phases (substring "for pr/s/u/ph") when CapexDx also creates event markers applying to the entire project.
- Terminal-stage assignment rules:
  - Completed: any event date at or after events "Completed" or "Commercial Production"; if both appear with different dates, consider the last event as marking completion; project marked as "Completed" in the Detail file.
  - Abandoned (Shelved): any event date at or after the events "Abandoned on" ("Shelved on"); project marked as "Abandoned" in the Detail file.
  - Lost to Follow-Up:
    - If only one event "No information as on": any event date at or after that event; project must be marked as "No Information" in the Detail file.
    - If more than one "No information as on": identify temporary loss-to-follow-up spells by initial "No Information as on" and terminal "No information as up to"; for completion dynamics, take the last-occurring "No Information as on" marker as when the project becomes censored and assume the pre-no-information state persisted during the no-information period.
- Intermediate-stage assignment rules and spell identification:
  - Stalled: marked during
    - event dates between "Implementation stalled on" and exit "Stalled up to" or any terminal-stage exit; or
    - event dates between "Announced and stalled on" and exit "Stalled up to" or any terminal-stage exit; or
    - event dates starting at "Implementation stalled on" or "Announced and stalled on" without a subsequent terminal transition — these are manually checked and often feature many monitoring markers or are recent active stallment cases.
  - Under Implementation: marked during
    - event dates between "Implementation started" and exit "Implementation stalled on" or any terminal-stage exit; or
    - event dates starting at "Implementation stalled on" or "Stalled up to" without a subsequent terminal transition — manually checked and often recent active implementation spells.
    - Note: a project may have more than one "Implementation started" marker; multiple spells are marked separately to capture multiple implementation spells.
  - Announced: marked during
    - event dates between "Date of Announcement" and exit events "Announced and stalled on", "Shelved on" for projects never started;
    - event dates between "Date of Announcement" and exit events "Announced and stalled on", "Implementation started", or any terminal-stage exit for projects eventually started;
    - event dates after "Date of Announcement" without a transition for very recent projects — these are manually checked and often reflect recent activity or monitoring markers.
- Manual checks and assertion tests are used to ensure resulting histories are sensible and to verify transitions and spell boundaries.
- Notes on edge cases:
  - There are "16 projects in the entire dataset that move from stallment from completion" — treated as cases where interpretation is ambiguous, and results are not sensitive to including or excluding such projects.
  - The marker "Announced and stalled on" is used when the first contact occurs at a point where the project is already stalled.
  - In a minuscule number of cases, projects flow in and out of abandonment or shelving; CMIE indicated some sponsors reversed abandonment decisions — such cases are treated as intermediate rather than terminal in robustness checks.

*Content derived from the "References" and Data Appendix sections of the source PDF.*

### 1.  At any point in time, a project is assigned to one of the above stages.

### wpiea2025078-print-pdf - 1.  At any point in time, a project is assigned to one of the above stages.

### Data construction and stage-assignment rules
- At any point in time, a project is assigned to a single stage (a project cannot be simultaneously under implementation and stalled, or completed and lost to follow-up).
- Event histories are constructed from an event date–event–project level dataset:
  - Convert to a project-day level dataset, then aggregate to quarterly frequency.
  - When more than two stages occur within a quarter, take the last stage for the entire quarter (standard survival-analysis convention). Example: if a project moves from announcement to implementation mid-2011:2, the project is considered “under implementation” for the entirety of 2011:2.
  - The last row in the constructed series should capture the transition into the terminal stage; after this the final quarterly panel on project histories is produced.
- Manual interventions and checks:
  - Cases where an entry event is “Stalled up to” without a contemporaneous “Implementation started” event (but with a preceding “Implementation started” earlier) are manually checked.
  - Multiple announcement or “announce and hold” spells are recorded as separate spells and manually checked in rare ambiguous cases.
  - Overlapping stages produced by automated procedures are manually disambiguated by examining the full project history.

### Illustrative examples (event-time → project-time conversion)
- Example 1 (Project ID 1) — straightforward lifecycle:
  - Event-Time Panel (selected events):
    - Date of announcement 05/20/2006
    - Implementation Started 09/02/2007
    - Completed 07/13/2012
  - Project-Time Panel (excerpted):
    - 1 2006:2 Announced 1 2006:2 Under Imp 0 Completed 0
    - 1 2007:2 Announced 1 2007:2 Under Imp 0 Completed 0
    - 1 2007:3 Announced 1 2007:3 Under Imp 0 Completed 0
    - 1 2012:2 ... 1 2012:2 Under Imp 0 Completed 0
    - 1 2012:3 ... 1 2012:3 Completed 1
- Example 2 (Project ID 2) — multiple stalls, monitoring markers, and loss to follow-up:
  - Event-Time Panel (selected events with monitoring markers):
    - Date of announcement 11/05/2005
    - Contacted Successfully 08/01/2006
    - Land acquired 11/11/2008
    - Implementation started 11/11/2008
    - Implementation stalled on 04/20/2009
    - Stalled up to 04/01/2010
    - Construction commenced 04/01/2010
    - Implementation stalled on 08/21/2012
    - Stalled up to 10/09/2012
    - No response 02/09/2013; 07/10/2013; 12/01/2013; 04/13/2014
    - No information as on 06/27/2014
  - Project-Time Panel (selected quarterly snapshots):
    - 2 2005:4 Announced 1 Under Imp 0 Stalled 0 No Info 0
    - 2 2008:3 Announced 1 Under Imp 0 Stalled 0 No Info 0
    - 2 2008:4 Announced 1 Under Imp 1 Stalled 0 No Info 0
    - 2 2009:1 Announced 1 Under Imp 0 Stalled 0 No Info 0
    - 2 2009:2 Announced 1 Under Imp 0 Stalled 1 No Info 0
    - 2 2010:1 Announced 1 Under Imp 0 Stalled 0 No Info 0
    - 2 2012:2 Announced 1 Under Imp 0 Stalled 0 No Info 0
    - 2 2012:3 Announced 1 Under Imp 0 Stalled 1 No Info 0
    - 2 2014:1 Announced 1 Under Imp 0 Stalled 0 No Info 1

### Merging with firm- and charge-level data (ProwessDx and MCA charges)
- Annual and quarterly accounting data are obtained from ProwessDx and merged at the firm-time level with project-level data.
- Charge-level data are obtained from the Ministry of Corporate Affairs (MCA) for entities whose CIN codes are in ProwessDx.
  - A cross-walk between lending-entity names in MCA data and ProwessDx names is constructed using fuzzy matching algorithms plus extensive manual checks.
  - Borrower matching is straightforward when CIN codes exist in both datasets.
- Construction of a bank-firm-time panel:
  - Expand data to loan-time level, compute relevant variables (e.g., total amount outstanding in a year) at bank-firm-time level, then collapse/aggregate to firm-time or bank-time as needed.

### Measurement and descriptive statistics (selected tables and figures)
- Concentration statistics (Table A1, selected entries):
  - Projects per Company: N 324563 Mean 3.29 SD 18.99 10th 1.00 25th 1.00 50th 1.00 75th 2.00 90th 4.00
  - Share: Section 1 Mean 0.06 SD 0.09 10th 0.01 25th 0.01 50th 0.02 75th 0.05 90th 0.23
  - HHI: Section 1 0.19
  - Share: Industry 1 N 340 Mean 0.01 SD 0.02 10th 0.00 25th 0.00 50th 0.00 75th 0.01 90th 0.02
  - HHI: Industry 1 0.04
  - Share: State 4 N 20 Mean 0.02 SD 0.03 10th 0.00 25th 0.00 50th 0.01 75th 0.04 90th 0.05
  - HHI: State 4 0.06
- Transition patterns across project stages (Table A2 — fractions ×100, departing rows → arriving columns):
  - From Announced: Announced 0.00; Completed 6.52; No Information 20.49; Abandoned 2.47; Stalled 2.45; Under Imp 68.07
  - From Stalled: Announced 10.63; Completed 1.31; No Information 6.35; Abandoned 22.85; Stalled 0.00; Under Imp 58.86
  - From Under Imp: Announced 0.00; Completed 62.15; No Information 27.55; Abandoned 3.26; Stalled 7.04; Under Imp 0.00
- Quarterly transition probabilities (Table A3 — fractions ×100, departing rows → arriving columns):
  - From Announced: Announced 87.92; Completed 0.79; No Information 2.48; Abandoned 0.30; Stalled 0.30; Under Imp 8.22
  - From Abandoned: Announced 0.66; Completed 0.17; No Information 0.00; Abandoned 97.98; Stalled 0.00; Under Imp 1.20
  - From Stalled: Announced 0.63; Completed 0.08; No Information 0.38; Abandoned 1.36; Stalled 94.05; Under Imp 3.50
  - From Under Imp: Announced 0.00; Completed 4.02; No Information 1.78; Abandoned 0.21; Stalled 0.46; Under Imp 93.53
- Cost–time correlations (Table A4):
  - Log Cost coefficients in regressions of log time to build: 0.178; 0.163; 0.162; 0.152; 0.155 (standard errors in parentheses reported in source).

### Key empirical estimates (selected results)
- Effects of Asset Quality Review on credit supply (Table A5):
  - Exposure coefficients (various specifications):
    - Column (1): Exposure -0.0021 (std. err. 0.0015)
    - Column (2): Exposure -0.0069 ∗∗∗ (std. err. 0.0018)
    - Column (3): Exposure -0.0027 ∗ (std. err. 0.0015)
    - Column (4): Exposure -0.0046 ∗ (std. err. 0.0024)
    - Column (5): Exposure -0.0085 ∗∗∗ (std. err. 0.0028)
    - Column (6): Exposure -0.0051 ∗∗ (std. err. 0.0024)
  - Outcome mean reported as -0.08 (columns 1–3), -0.10 (columns 4–6).
  - Interpretation example from source: column (1) — moving from average lender to one with one-standard deviation higher exposure decreases outstanding loans by -0.21 percentage points, relative to mean growth rate of 8 percentage points (textual interpretation preserved from source).
- Effects of firm-level credit supply shocks on project completion hazards (PPML, Table A6 — semi-elasticities interpreted as percent changes for one standard-deviation increase in exposure):
  - Exposure (Pre) coefficients:
    - Column (1): 0.1057 ∗∗∗ (std. err. 0.0277)
    - Column (2): 0.0772 ∗∗∗ (std. err. 0.0273)
    - Column (3): 0.0766 ∗∗∗ (std. err. 0.0273)
    - Column (4): 0.0754 ∗∗∗ (std. err. 0.0267)
    - Column (5): 0.0749 ∗∗∗ (std. err. 0.0267)
    - Column (6): 0.0844 ∗∗∗ (std. err. 0.0265)
    - Column (7): 0.0845 ∗∗∗ (std. err. 0.0266)
  - Exposure (2015) coefficients similar in magnitude (e.g., 0.1067 ∗∗∗ in column (1)).
  - Dependent-variable mean ranges: 0.0397–0.0414 across columns.
  - Example interpretation from source: column (1) — moving from a firm with average exposure to a firm with one-standard deviation higher exposure increases the hazard by 10.57%.
- Effects of firm-level credit supply shocks on project completion hazards (linear regressions, Table A7 — level effects per one standard-deviation increase in exposure):
  - Exposure coefficients:
    - Column (1): 0.0057 ∗∗∗ (std. err. 0.0017)
    - Column (2): 0.0030 ∗∗ (std. err. 0.0012)
    - Column (3): 0.0029 ∗∗ (std. err. 0.0012)
    - Column (4): 0.0028 ∗∗ (std. err. 0.0012)
    - Column (5): 0.0027 ∗∗ (std. err. 0.0012)
    - Column (6): 0.0029 ∗∗∗ (std. err. 0.0011)
    - Column (7): 0.0028 ∗∗ (std. err. 0.0011)
  - Dependent-variable mean 0.0397 across columns.
  - Example interpretation from source: column (1) — moving to a firm with one-standard deviation higher exposure increases the hazard by 0.57 percentage points, relative to an unconditional completion probability of 3.97 percentage points.
- Effects of firm-level credit supply shocks on project starts (PPML, Table A8 — semi-elasticities interpreted as percent change for one standard-deviation increase in exposure):
  - Exposure (Pre) coefficients:
    - Column (1): -0.0740 ∗∗∗ (std. err. 0.0278)
    - Column (2): -0.0708 ∗∗ (std. err. 0.0324)
    - Column (3): -0.0673 ∗∗ (std. err. 0.0321)
    - Column (4): -0.0808 ∗∗ (std. err. 0.0321)
    - Column (5): -0.0753 ∗∗ (std. err. 0.0319)
    - Column (6): -0.0562 ∗ (std. err. 0.0305)
    - Column (7): -0.0556 ∗ (std. err. 0.0303)
  - Exposure (2015) coefficients similar in magnitude (e.g., -0.0755 ∗∗∗ in column (1)).
  - Dependent-variable mean 0.8362 (columns 1–5) and 0.8326 (columns 6–7).
  - Interpretation from source: the estimates reflect semi-elasticities of firm-level project starts to firm exposure to the Asset Quality Review (model (3) in source).

### Data quality and robustness notes preserved from source
- Aggregation checks:
  - Authors compute number of events per month and per quarter before aggregating and do not find considerable discrepancies, alleviating concerns of information loss from aggregation.
- Treatment of rare events:
  - Abandonment is treated as terminal in most results because returns from abandonment to activity are exceedingly rare.
- Controls and estimation details (preserved terminology):
  - Firm controls include sales, sales growth, profitability, debt over equity, book leverage, equity issuances, cash over liquid assets, and debt over EBITDA.
  - Firm controls are winsorized at the 1st and 99th percentile, then standardized, and lagged by one period.
  - Standard errors clustered at lender or firm level as specified in tables.
  - Fixed effects applied vary by specification (e.g., borrower-by-year, firm, state×year, sector×year, industry×year, age splines).

_Italic: Source — wpiea2025078-print-pdf (canonical URL provided in the pipeline)_

### 2019. Each row corresponds to a measure of firm exposure. Firm exposure is defined as the weighted-average of lender exp

### Table A9: Effects of Firm-Level Credit Supply Shocks on Project Starts

### Key findings
- Coefficient estimates (Exposure) across columns (1)–(7):
  - (1) Exposure = -0.0560 ∗∗∗ (standard error (0.0195))
  - (2) Exposure = -0.0534 ∗∗ (standard error (0.0226))
  - (3) Exposure = -0.0520 ∗∗ (standard error (0.0223))
  - (4) Exposure = -0.0574 ∗∗ (standard error (0.0226))
  - (5) Exposure = -0.0562 ∗∗ (standard error (0.0223))
  - (6) Exposure = -0.0471 ∗∗ (standard error (0.0218))
  - (7) Exposure = -0.0462 ∗∗ (standard error (0.0215))
- Interpretation: coefficient estimates are interpreted as the percent effect on the outcome for a one standard-deviation increase in firm exposure. Example given: moving from a firm with average exposure to a firm with one-standard deviation higher exposure reduces project starts by 7.4% (illustrative example in the text).

### Estimation details and controls
- Dependent variable mean (DepVar Mean) reported by column:
  - Columns (1)–(5): 0.8362
  - Columns (6)–(7): 0.8281
- Observations by column:
  - Columns (1)–(5): 8,994
  - Columns (6)–(7): 8,941
- Fixed effects included by column (indicated as "X"):
  - Firm fixed effects: included in columns (1)–(3)
  - Year fixed effects: included in columns (1)–(2)
  - NIC Section x Year fixed effects: included in column (3)
  - NIC Division x Year fixed effects: included in column (4)
- Firm controls: included in columns (1)–(3) as indicated; full list of firm controls:
  - sales
  - sales growth
  - profitability
  - debt over equity
  - book leverage
  - equity issuances
  - cash over liquid assets
  - debt over EBITDA
- Firm controls are winsorized at the 1st and 99th percentile, then standardized, and lagged by one period.

### Sample and exposure definition
- Sample: the Capex-Prowess matched sample of firms observed yearly between 2013-2019.
- Firm exposure definitions:
  - For the main table text: “Firm exposure is defined as the weighted-average of lender exposure, where weights correspond to outstanding loans in the previous year.”
  - In other descriptive rows: “(Pre)” indicates weights are constructed as average outstanding loans between 2013-2015 for a borrower-lender pair relative to average outstanding loans between 2013-2015 for a given borrower.
  - “(2015)” indicates weights are constructed as outstanding loans in 2015 between a borrower-lender pair relative to outstanding loans in 2015 for a given borrower.
- Coefficients in other specification discussion are interpreted as the level effect on the outcome for a one standard-deviation increase in firm exposure.

### Inference and sample size note
- Standard errors are clustered at the firm level and reported in parentheses.
- Statistical significance notation: ∗, ∗∗, and ∗∗∗ denote significance at the 10%, 5% and 1% levels, respectively.
- The reduction in sample size from columns (5) to columns (6) and (7) results from separation of additional observations by the more granular division-by-year fixed effects relative to the section-by-year fixed effects.

*Source: Table A9, wpiea2025078-print-pdf - 2019.*

### 0.056 projects relative to an unconditional mean of 0.8362 projects.  NIC Section and Division classifications are const

### wpiea2025078-print-pdf - 0.056 projects relative to an unconditional mean of 0.8362 projects.  NIC Section and Division classifications are const

### Data construction, firm controls, and estimation details
- NIC Section and Division classifications are constructed from NIC codes and categorizations.
- Firm controls include sales, sales growth, profitability, debt over equity, book leverage, equity issuances, cash over liquid assets, and debt over EBITDA.
- All firm controls are winsorized at the 1st and 99th percentile, and then standardized.
- All firm controls are lagged by one-period.
- Standard errors are clustered at the firm level and reported in parentheses.
- Statistical significance annotations: ∗, ∗∗, and ∗∗∗ denote significance at the 10%, 5% and 1% levels, respectively.

### Balance tests — cross-section and panel (Tables A10–A11)
- Table A10 (Cross-Section): regressions of firm-level characteristics averaged over 2013-2015 on average firm exposure over 2016-2019. Outcomes are winsorized at the 2.5th and 97.5th percentile, then standardized. Heteroscedasticity-robust standard errors.
  - Selected coefficients (Column (3)):
    - Log Sales: -0.1998 (0.0826) ∗∗
    - Sales Growth: -0.0002 (0.0525)
    - Profitability (EBIT): -0.0149 (0.0325)
    - EBIT<0: 0.0419 (0.0402)
    - Debt/Equity: -1.9474 (4.5947)
    - Book Leverage: 1.6844 (3.1223)
    - Equity Issuance: -0.2403 (0.1041) ∗∗
    - Cash/Liquid Assets: 0.0150 (0.0116)
    - Debt/EBIT: -3.6173 (2.8500)
- Table A11 (Panel): sample is Capex-Prowess matched firms observed yearly 2013–2019. Regress lagged firm characteristics Xf,t−1 on contemporaneous firm exposure Expof,t. Standard errors clustered at firm level.
  - Selected coefficients (Column (3)):
    - Log Sales: -0.3798 (0.4663)
    - Sales Growth: -0.1046 (0.0782)
    - Profitability (EBIT): -0.2861 (0.1177) ∗∗
    - EBIT<0: 0.0906 (0.0644)
    - Debt/Equity: -23.0500 (46.1865)
    - Book Leverage: -17.9970 (24.3064)
    - Cash/Liquid Assets: -0.0710 (0.0377) ∗
    - Debt/EBIT: 0.2829 (5.7189)

### Predictive models of Time to Build (Table A12, Figure A12)
- Table A12 compares predictive models for Time to Build and Log Time to Build.
  - Panel A: RMSE
    - Out-of-Sample (levels): 13.718, 13.382, 13.552, 13.525
    - Out-of-Sample (logs): 0.860, 0.822, 0.848, 0.854
    - In-Sample (levels): 11.011, 10.562, 11.639, 10.436
    - In-Sample (logs): 0.819, 0.786, 0.791, 0.778
  - Panel B: R-Squared
    - Out-of-Sample (levels): 0.123, 0.166, 0.144, 0.148
    - Out-of-Sample (logs): 0.110, 0.187, 0.134, 0.122
    - In-Sample (levels): 0.386, 0.435, 0.314, 0.448
    - In-Sample (logs): 0.138, 0.206, 0.196, 0.223
  - Panel C: OLS regressions of observed gestation lags on predicted gestation lags (hazard regression sample)
    - Selected slope estimates and standard errors (columns shown):
      - 1.076 (0.057)
      - 1.093 (0.046) ∗∗
      - 1.208 (0.033) ∗∗∗
      - 0.945 (0.040)
      - 1.312 (0.023) ∗∗∗
      - 1.196 (0.017) ∗∗∗
      - 0.989 (0.015)
      - 0.945 (0.015) ∗∗∗
    - Constants and standard errors likewise reported; see table for exact values.
- Notes:
  - Columns correspond to models including covariates marked by X in bottom panel.
  - Sector and Industry fixed effects and additional covariates are flexibly included; the model of column (2) is used for predicting project stages in the paper.
- Figure A12:
  - Binned scatter plots and lines of best fit for different models. First row predicts gestation lags in levels; second row predicts logs. Columns: Sector, Industry, Sector Plus, Industry Plus. In-sample and out-of-sample samples plotted separately.

### Heterogeneous effects by Initial Age (Table A13)
- Sample: projects observed quarterly 2013–2019. PPML estimates of semi-elasticity of project-level hazards to sponsor exposure to the Asset Quality Review.
- Initial Age is proxy for stage (age when first observed), winsorized at the 1st and 99th percentile and standardized.
- Interpretation:
  - Coefficient on Exposure is percent effect on outcome for a one standard-deviation increase in firm exposure for project at mean proxy stage.
  - Coefficient on Age x Exposure is change in main effect resulting from a one-standard-deviation change in proxy stage.
- Selected results (Columns (1)–(4)):
  - Exposure: 0.094 ∗∗∗ ; 0.090 ∗∗∗ ; 0.099 ∗∗∗ ; 0.096 ∗∗∗ (standard errors: (0.025), (0.026), (0.026), (0.026))
  - Age x Exposure: 0.060 ∗∗∗ ; 0.062 ∗∗∗ ; 0.066 ∗∗∗ ; 0.068 ∗∗∗ (standard errors: (0.023), (0.023), (0.025), (0.025))
  - Example interpretation from column (1): moving from average sponsor exposure to one-standard-deviation higher increases hazard by 9.4% at mean stage; a project one standard deviation above mean stage would have total effect 15.4% (9.4% + 6.0%), one below mean would have 3.4%.
- Notes: age spline omitted to avoid collinearity; firm controls as described above; sample reductions from additional fixed effects explained.

### Backfilling, censoring, and selection (Appendix D, Figures A13, Tables A14–A16)
- Data collection: CapexDx backfills project histories when they obtain information on ongoing projects via public-domain research and sponsor contact.
- Selection concern articulated:
  - RBI audits and resulting firm disruptions could alter data collection process (e.g., CapexDx less likely to find data on treated projects), producing differential censoring and selection bias.
  - If treated group has fewer early projects observed post-shock, observed association between treatment and completion could be driven by selection rather than causal effect.
- Two analyses to address concerns:
  1. Restrict completion analyses to projects present in data vintages prior to the first audits (first vintage: Dec 31, 2014).
     - These projects could not be affected by post-shock changes in data collection.
     - Reestimation of completion regressions on this subsample reported in Table A14.
  2. Test implication of differential censoring: check whether exposure predicts differential pre-trends in firm-level project stocks.
     - Use vintages and counts of projects by vintage (Figure A13) and balance tests on project stocks (Tables A15–A16).
- Figure A13:
  - Plots total and incremental counts of projects by vintage (vintages: 2014-12-31, 2016-03-31, 2017-03-31, 2018-03-31, 2019-03-31, 2020-03-31, 2021-03-31, 2022-03-31).
  - Bars for: All Projects, MCA-Prowess-Capex matched sample, projects in 2013–2019 window used for completion regressions.
- Table A14: Effect on Completion Hazards — subset of first vintage projects (panel)
  - Exposure (LTV) coefficients across columns (1)–(7): 0.0213, 0.0637, 0.0627, 0.0580, 0.0578, 0.0469, 0.0449 (standard errors: (0.0302),(0.0422),(0.0422),(0.0425),(0.0425),(0.0453),(0.0457))
  - Exposure (PreW) coefficients: 0.0260, 0.0802 ∗∗, 0.0817 ∗∗, 0.0782 ∗, 0.0804 ∗, 0.0717 ∗, 0.0743 ∗ (standard errors reported)
  - Exposure (2015W) coefficients: 0.0241, 0.0769 ∗, 0.0789 ∗∗, 0.0756 ∗, 0.0783 ∗, 0.0697 ∗, 0.0726 ∗
  - DepVar Mean: 0.0385 (columns 1–5), 0.0408 (columns 6–7)
  - Observations vary across columns: 53750, 53741, 53741, 53741, 53741, 27750, 277
  - Notes: LTV indicates lagged, time-varying weights; PreW averages outstanding loans 2013–2015; 2015W uses 2015 outstanding loans.
  - Example interpretation: row 1 col (1) — moving from average exposure to one-standard-deviation higher increases hazard by 2.13%.
- Table A15: Cross-Sectional Balance Tests — Project Stocks (mean comparisons 2013–2015)
  - Selected means and cross-sectional exposure coefficients (column (3)):
    - Active Projects: Mean|Exp=0 = 2.441, Mean|Exp>0 = 4.185, Mean|Exp = -3.3262 ∗∗∗ (1.1054)
    - Projects Started: 0.504, 0.683, -0.5060 ∗∗ (0.2276)
    - Projects Completed: 0.233, 0.342, -0.4336 ∗∗∗ (0.0998)
    - Projects Abandoned: 0.043, 0.075, -0.0377 (0.0415)
    - Projects Censored: 0.100, 0.192, -0.0240 (0.0779)
    - Projects Under Implementation: 1.428, 2.415, -1.9943 ∗∗∗ (0.6739)
    - Projects Announced: 0.976, 1.599, -1.3409 ∗∗∗ (0.4452)
    - Projects Stalled: 0.176, 0.355, 0.0812 (0.1501)
  - Notes: regressions of firm-level outcomes averaged over 2013–2015 on average firm exposure over 2016–2019; exposure defined as weighted-average of lender exposure with lagged loan weights.
- Table A16: Panel Balance Tests — Project Stocks (lagged outcomes)
  - Columns report means conditional on Exp=0 and Exp>0, and coefficients for Exposure in panel regressions with and without firm+year fixed effects.
  - Selected statistics (columns (3) and (4) coefficients):
    - L1 Active Projects: Mean|Exp=0 = 2.987, Mean|Exp>0 = 3.408, Mean|Exp = -2.0376 ∗∗∗ (0.5203), Mean|Exp (with FE) = -0.1223 (0.2349)
    - L1 Projects Started: 0.473, 0.649, 0.0162 (0.1167), -0.1163 (0.1217)
    - L1 Projects Completed: 0.255, 0.332, -0.0630 (0.0730), 0.0410 (0.0749)
    - L1 Projects Abandoned: 0.042, 0.032, -0.0510 ∗∗∗ (0.0174), 0.0292 (0.0242)
    - L1 Projects Under Implementation: 1.768, 2.038, -1.3085 ∗∗∗ (0.3334), -0.1103 (0.1944)
    - L1 Projects Announced: 1.104, 1.241, -0.6835 ∗∗∗ (0.2054), -0.0655 (0.1568)
    - L1 Projects Stalled: 0.228, 0.259, -0.1360 ∗∗ (0.0572), -0.0042 (0.0481)
  - Notes: regress lagged firm outcomes Yf,t−1 on contemporaneous firm exposure Expof,t. Column (4) includes firm and time fixed effects. Standard errors clustered at firm level.
- Interpretation and conclusion from these analyses:
  - Restricting to projects present in 2014 vintage yields point estimates similar to baseline (though with larger standard errors), reducing concern that post-shock changes in data collection drive results.
  - Panel balance tests with firm and time fixed effects show predictability of lagged project counts by contemporaneous exposure vanishes; this weakens the differential backfilling/differential censoring threat.
  - Overall evidence is suggestive but not definitive; authors are reassured that main results are unlikely driven by sample selection from differential backfilling.

### Changes in scope and robustness checks (Appendix E)
- Institutional context on construction contracts in India:
  - Contracts governed by Indian Contract Act of 1872; construction contracts defined per Indian Accounting Standard 7.
  - Common contractual documentation: drawings, specification (“specs”), bill of quantities, Conditions allocating responsibilities, sponsor requirements, performance requirements, dispute adjudication clauses.
  - Common practice to follow FIDIC or Indian Institute of Architects contract forms; design-build (and design-bid-build) forms common.
  - Separation of phases: design phase (determines (x̄, ī)), bid phase, build phase.
  - Variations in works:
    - Small variations: recorded and compensated; captured by adjustment costs κ(·) or marginal costs κ′(·).
    - Large or drastic changes in scope: may be denied, trigger contractual breach, or lead to partial/full termination and indemnification.
    - Sponsor cannot unilaterally and substantially change contract obligations without contractor consent; courts deem such unilateral changes unenforceable.
  - Dispute resolution: arbitration is usual; judicial congestion raises negotiation incentives.
- Communications with data provider (Capex) about scope changes:
  - Capex reports few drastic changes in scope across many years.
  - Recording logic:
    - Small scope changes: revised directly in project profile (e.g., capacity X to 0.9X, cost C to 0.95C) with new cost row indicating revision date and amount.
    - Large changes (e.g., change in type of use): Capex records a new project under same sponsor and stalls the original project.
    - Example of partial completion (three factories, two completed, third killed): Capex revises capacity/cost of original project, pushes incomplete plant into a separate stalled project; a project is considered completed only if all subunits have been completed.
- Robustness checks motivated by Capex communications:
  - Exclude projects with more than one phase or unit (Capex has less granular information on such projects).
  - Exclude projects that experience downward cost restatements (proxy for negative variations).
  - Exclude firms that had stalled projects over the sample period (stalled project coinciding with completion of others could indicate negative variation).
- Conclusion: qualitative and quantitative evidence, combined with Capex recording practices and robustness checks, alleviate concerns that observed increases in completion hazards are driven by ex-post reductions in project scope induced by the credit supply shock.

*Italic source: wpiea2025078-print-pdf - 0.056 projects relative to an unconditional mean of 0.8362 projects. NIC Section and Division classifications are const*

### Appendix Table A17 presents the results.  Panel A removes multi-unit projects.  The re-

### wpiea2025078-print-pdf - Appendix Table A17 presents the results.  Panel A removes multi-unit projects.  The re-

### Robustness checks on project-level hazards (Appendix Table A17)
- Purpose: assess robustness of PPML semi-elasticity estimates of project-level completion hazards to alternative subsamples that could have experienced reductions in scope following the credit shock.
- Panels and salient coefficient estimates (Exposure; standard errors in parentheses; N):
  - Panel A: Remove Multi-Unit Projects
    - Exposure: 0.1112 ∗∗∗ (0.0285), 0.0901 ∗∗∗ (0.0290), 0.0852 ∗∗∗ (0.0291), 0.0822 ∗∗∗ (0.0277), 0.0775 ∗∗∗ (0.0279), 0.0947 ∗∗∗ (0.0297), 0.0928 ∗∗∗ (0.0298)
    - N: 90289 90289 02899 02899 02899 02898 58128 5812 (as reported in table)
  - Panel B.1: Remove Downward Cost Revisions
    - Exposure: 0.0872 ∗∗∗ (0.0255), 0.1075 ∗∗∗ (0.0410), 0.1030 ∗∗ (0.0415), 0.1023 ∗∗ (0.0400), 0.0981 ∗∗ (0.0403), 0.1068 ∗∗ (0.0462), 0.1065 ∗∗ (0.0462)
    - N: 51589 51589 51589 51589 51589 51589 47911 47911
  - Panel B.2: Remove Projects with Missing Costs
    - Exposure: 0.0970 ∗∗∗ (0.0261), 0.0997 ∗∗∗ (0.0345), 0.0941 ∗∗∗ (0.0349), 0.0941 ∗∗∗ (0.0329), 0.0885 ∗∗∗ (0.0332), 0.1247 ∗∗∗ (0.0375), 0.1198 ∗∗∗ (0.0376)
    - N: 72958 72336 72336 72958 72958 72958 67968 7968 (as reported)
  - Panel C: Remove Ever Stalled Projects
    - Exposure: 0.0982 ∗∗∗ (0.0267), 0.0570 ∗∗ (0.0279), 0.0542 ∗ (0.0280), 0.0516 ∗ (0.0270), 0.0487 ∗ (0.0272), 0.0651 ∗∗ (0.0276), 0.0639 ∗∗ (0.0278)
    - N: 104186 104155 104155 104155 104155 99623 99623
  - Panel D: Remove Any Companies with Projects Stalled
    - Exposure: 0.0967 ∗∗∗ (0.0279), 0.0667 ∗ (0.0370), 0.0648 ∗ (0.0369), 0.0469 (0.0341), 0.0451 (0.0341), 0.0737 ∗∗ (0.0375), 0.0740 ∗∗ (0.0373)
    - N: 52493 52493 52493 52493 52493 11749 117
- Notes on interpretation:
  - Estimates are PPML semi-elasticities: percent effect on completion hazard for a one standard-deviation increase in firm exposure.
  - Example: column (1) in the main table interpretation: moving from average exposure to one-standard deviation higher exposure increases the hazard by 11.12%.
  - Age splines = cubic splines on project age (quarters). Firm controls include sales, sales growth, profitability, debt over equity, book leverage, equity issuances, cash over liquid assets, debt over EBITDA; winsorized at the 1st and 99th percentile, standardized, lagged by one period.
  - Standard errors clustered at firm level. ∗, ∗∗, ∗∗∗ denote significance at the 10%, 5% and 1% levels.

### Robustness based on ex-post malleability classifications (Appendix Table A18)
- Objective: remove projects more amenable to ex-post scope changes using three metrics (industry-based, regulatory-based, implementation-based) and reestimate hazards.
- Panel A: Remove “malleable” industries (examples removed: housing construction, computer software, ITES, business services and consultancy, retail trading, wholesale trading, commercial complexes, hotels and restaurants, education, health services, tourism)
  - Exposure: 0.0851 ∗∗∗ (0.0269), 0.0774 ∗∗ (0.0320), 0.0768 ∗∗ (0.0324), 0.0757 ∗∗ (0.0322), 0.0748 ∗∗ (0.0324), 0.0976 ∗∗∗ (0.0338), 0.0981 ∗∗∗ (0.0340)
  - N: 74483 74439 74439 74439 74439 96699 69960
- Panel B: Some Regulatory Event (projects with events related to regulatory approvals)
  - Exposure: 0.0789 ∗∗∗ (0.0258), 0.2180 ∗∗∗ (0.0805), 0.2193 ∗∗∗ (0.0789), 0.2186 ∗∗∗ (0.0799), 0.2182 ∗∗∗ (0.0779), 0.1650 (0.1009), 0.1634 (0.1006)
  - N: 29966 29966 29966 29966 29966 24765 24765
  - Interpretation: point estimates larger and remain significant except where industry-time fixed effects reduce power.
- Panel C: Under implementation before shock (implementation-based)
  - Exposure: 0.1180 ∗∗∗ (0.0351), 0.1402 ∗∗∗ (0.0388), 0.1348 ∗∗∗ (0.0390), 0.1269 ∗∗∗ (0.0388), 0.1220 ∗∗∗ (0.0391), 0.0665 ∗ (0.0398), 0.0614 (0.0400)
  - N: 65046 65046 65046 65046 65046 16116 16111
- Takeaway: across malleability definitions, point estimates largely similar or larger than baseline; significance can be lost in specifications with high-granularity fixed effects due to sample size/power reductions.

### Takeaways from robustness exercises (qualitative)
- The appendix addresses concern that results reflect premature project completion via scope reductions by:
  - Contractual/technical frictions that make large ex-post reductions costly, particularly for large CAPEX projects.
  - Dropping projects most susceptible to scope reductions (multi-unit, downward cost revisions, stalled projects, malleable industries, regulatory-event projects, projects under implementation) — point estimates remain broadly consistent with baseline, lending support to the interpretation that results are not driven by scope-reduction artifacts.
- Data limitations: direct observation of contractual/specification changes was not available in CapexDx; combination of qualitative evidence and robustness checks used instead.

### Calibration results (Figure A18; Tables A19–A20)
- Calibration moments and parameter estimates (Table A19):
  - r = 0.01 (Real Rate; Post-1990 Time Series Average)
  - p1 = Price of K (Normalization)
  - μ ̄x = 0 (Avg Project Size)
  - σ2 ̄x = 2.12 (SD Project Size; Cost-TTB Distribution)
  - β ̄i0 = −2.47 (Mean ̄i: intercept)
  - β ̄i1 = 0.82 (Mean ̄i: slope)
  - σ2 ̄i = 0.88 (SD ̄i)
  - ρ = 0.17 (Shock Persistence; Panel Autoregressions)
  - α3 = Pareto Tail (Log-Rank Regressions)
  - s = 0.02 (Shock Size; Reduced Form Regressions)
  - κ = 3.09 (Accel. Cost)
- Targeted moments (Table A20; Model vs Data):
  - Avg Log Size (Q1): −2.68 vs −2.58
  - Avg Log Size (Q2): −1.42 vs −1.57
  - Avg Log Size (Q3): 0.01 vs −0.18
  - Avg Log Size (Q4): 1.45 vs 1.42
  - Avg Log Size (Q5): 2.77 vs 2.80
  - SD Log Size: 0.91 vs 0.90
  - Avg Log TTB (Q1): 1.96 vs 2.02
  - Avg Log TTB (Q2): 2.30 vs 2.21
  - Avg Log TTB (Q3): 2.53 vs 2.48
  - Avg Log TTB (Q4): 2.88 vs 2.95
  - Hazard Regression β: 0.06 vs 0.06
  - Starts Regression β: −0.08 vs −0.08
- Figure A18 panels (calibration diagnostics):
  - A18a: joint distribution of costs and time to build compared between model and data; used to calibrate distributions of ̄x and ̄i.
  - A18b: log-rank regressions to estimate Pareto exponent; midpoint chosen as 3.
  - A18c: multiple panel autoregressive estimators for shock persistence ρ (OLS no FE; Han-Phillips; various system GMMs); blue dots for all firm-years in MCA-Prowess matched data; red dots restrict to firms in hazard regressions.

### Theoretical results and propositions (Appendix G)
- Kolmogorov Backward Equation and time-to-completion:
  - For remaining time to completion T(x,t): 0 = 1 + i(x,t) ∂/∂x T(x,t) + ∂/∂t T(x,t) (equation (19)) with terminal condition T( ̄x , t) = 0.
  - In steady state ∂/∂t T = 0 so T′(x) = −1 / i(x).
- Proposition 8: In steady state, completion times are uniformly distributed (derivation using g(x) and inverse transforms).
- First-order dynamics for terminal payoff shocks (Proposition 9):
  - For terminal payoff shock p(t) = p(1 + s e−ρt):
    - First-order response of project value: ˆv(x,z) = p z e−(r+ρ) T(x,z) (equation (21)).
    - First-order response of investment flows: ˆi(x,z) = (1 / κ′′(i(x,z))) ∂ˆv(x,z)/∂x (equation (22)).
  - Interpretation: persistent terminal-payoff shocks discount with rate r + ρ; mature projects react more to terminal-payoff shocks (shorter T(x,z) → larger impact).
- Extensive margin (Proposition 10):
  - Response of entry threshold ̄z to AR(1) terminal-payoff shock: ˆz = − ̄z e−ρ T(0, ̄z) (equation (23)).
- Endogenous p (Proposition 12, Corollary 13):
  - If terminal operating value p depends on r through p = c/(r+δ) and r experiences AR(1) shock, then investment response:
    - ˆi(x,z) = [1 / (i(x,z)·κ′′(i(x,z)))] · [ v(x,z) + (r+ρ)ˆv(x,z) − (r+ρ)/(r+δ+ρ) e−(r+ρ)T(x) p z ] (equation (25)).
  - Corollary 13: For x sufficiently close to ̄x, lim x→ ̄x ˆi(x,z) = (1 / (i(̄x,z)·κ′′(i(̄x,z)))) · δ/(r+δ+ρ) p z > 0 (equation (26)).
  - Interpretation: even when p falls with higher r, projects sufficiently close to completion optimally accelerate completion; acceleration incentives shrink by factor δ/(r+ρ+δ) ∈ (0,1).
- Analytical proofs provided for linearization of HJB, solution for ˆv(x), FOC for investment responses, and entry threshold dynamics (G.4 subsections).

### Discount rates, microfoundations, and interpretation of empirical shock (Appendix G.5)
- Rationale for mapping credit shock into higher firm discount rates:
  - Banking-sector turbulence increases funding costs; higher lending rates raise firms’ cost of debt and thus discount rates.
  - Literature evidence: higher bank distress or funding costs lead to higher lending rates (cited empirical studies summarized).
  - Morais et al (2020) causal example: regulatory provisioning reform increased provisions by 4% for affected borrowers, led to 50bps relative lending rate increase, 10% loan volume contraction, 5% decline in average maturity (effects reached within one year).
- Microfoundation sketch:
  - Banks set lending rates reflecting funding costs and provisioning opportunity costs; monitoring shocks (Asset Quality Review) force higher provisions → higher lending rates.
- Alternative channels captured by discount-rate formulation:
  - Higher value of cash within firm when credit supply tightens; managerial behavioral changes; internal capital market disciplining.
- Preference for discount-rate shock modeling:
  - Computational tractability for estimation and counterfactuals with rich project heterogeneity.
  - A plausible economic mapping from provisioning/lending-rate channel to firm-level discount rates.
  - Acknowledgement that borrowing-constraint formulations are also relevant (Appendix G.6) but more complex.

### Spending constraints vs. discount-rate shock (Appendix G.6)
- Alternative model: firm with N projects faces aggregate spending constraint E_N on total investment/spending; optimal investment per project can be constrained (i + κ(i) ≤ E).
- Key comparative implications:
  - With spending constraints, a tightening scale-down of investment across projects is mechanical; average project investment falls.
  - Spending-constraint mechanism explains heterogeneous prioritization across projects (closer-to-completion projects prioritized), but cannot by itself generate the empirically robust positive hazard effects for the average investment project observed.
- Numerical illustration:
  - Two-project example: when spending constraint binds, investment in the more advanced project increases in its stage; investment in the other project decreases with first project’s stage.
- Modeling complexity argument:
  - Jointly modeling (i) investment across multiple projects, (ii) endogenous firm capital structure, and (iii) occasionally binding constraints with full project-distribution dynamics is computationally burdensome.
- Authors’ preference:
  - Map credit disturbance to discount-rate shock for tractability, ability to exploit project heterogeneity, and because microfoundations linking provisioning shocks to lending rates/discount rates are plausible.

### Bond-pricing analogy for endogenous time to build (Appendix H)
- Mapping:
  - Bond with coupons c(t), face F, remaining maturity T: V = ˆT 0 e−rt c(t) dt + e−rT F.
  - ∂V/∂r = −Dur(V) parallels project value sensitivity ˆv(x,z) to discount-rate changes.
- “Accelerating” cash flows experiment:
  - Shift all bond cash flows backward by Δ; first-order optimality condition for paying κ(Δ) to accelerate cash flows:
    - κ′(Δ∗) = ∂V(Δ∗)/∂Δ (equation (36)), analogous to FOC for investment acceleration (equation (9) in main text).
  - ∂V(r,Δ)/∂r∂Δ = V(r,Δ) − r·Dur(V(r,Δ)) (equation (38)):
    - Two forces: “Impatience” (V positive) and “Discounting” (− r·Duration). Whether acceleration increases with r depends on net effect.
  - Close-to-maturity assets have small duration → acceleration is optimal (parallels Corollary 4 in main text).
- Conclusion: bond-pricing intuition reinforces mechanism that higher discount rates can make acceleration of near-complete projects optimal while potentially delaying less-mature projects.

### Convex adjustment cost model without time to build (Appendix I)
- Purpose: compare general equilibrium investment dynamics across models with instantaneous completion (no time to build) and models with endogenous time to build.
- Setup highlights:
  - Projects instantaneously built; entry threshold ̄z(t) = ̄x / p(t).
  - Final goods producer chooses I(t) with quadratic adjustment cost p(t) φ / 2 (I(t) − ̄I)^2; law of motion K̇ = I − δK; price p(t) relates to discounted MPK.
  - Pareto distribution for z: f(z) = α zαmin / zα+1 (equation (43)).
  - Fixed capital formation supply closed form: FCF(t) = α/(α−1) zαmin xα−1 p(t)α−1 (equation (44)).
- Equilibrium and first-order dynamics:
  - Investment equilibrium condition (after substitutions) leads to implicit equation (46) relating I(t), φ, and the discounted future MPK path.
  - First-order response of investment ˆI/I to a small AR(1) shock in r: ˆI(t)/I = (r+δ) / (1/(α−1) + φ I) × ˆλ-term (equation (48)), where ˆλ captures discounted MPK deviations net of ˆr/r+δ.
  - Accounting investment Iacc relates to I via Iacĉ/Iacc = α/(α−1) ˆI/I (equation (49)).
  - Capital stock dynamics K̂/K = δ ˆ0 t eδ(s−t) ˆI(s)/I ds (equation (50)).
- Comparative dynamics intuition:
  - In convex-adjustment model without time to build, higher discount rates lower capital prices and accounting investment rapidly via higher entry thresholds; as MPK rises, investment rebounds over time.
  - Time-to-build models produce richer heterogeneity across projects by stage (mature vs nascent) that the instantaneous model cannot capture.

*Italic: Source — Excerpt from “Wasn’t Built in a Day: Investment with Endogenous Time to Build”, Working Paper No. WP/2025/078 (appendices and tables as provided).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025078-print-pdf.pdf_
