## 2.1 Firm balance-sheets: The CADS dataset (wpiea2019107) ## Source details **Canonical URL:** [2.1 Firm balance-sheets: The CADS dataset (wpiea2019107)](https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019107.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2019/wpiea2019107.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2019/wpiea2019107.pdf.json) --- ### CADS database: coverage, construction, and sample - CADS administered by CERVED-Group Ltd.; detailed balance-sheet and income statement information on non-financial corporations since 1982. - Net revenue of CADS firms account for about 70% of the total revenues of the private non-financial sector. - Firm-level capital: perpetual-inventory method (PIM) on book-value of capital, investments, divestments, and sector-level deflators and depreciation rates. - Operating value added and intermediate expenditures recorded in nominal values and converted to real using sector-level deflators from National Accounts. - Baseline labor measure: wage bill, deflated using CPI. - Expenditures on intermediate inputs deflated using a combination of sector-level deflator and regional-level CPI. - Industry classification: Nace Rev.2 two-digit (four-digit robust). - Time coverage for production-function estimation: firms sampled in CADS from 1998 to 2013. - CADS provides time-varying controls: age, cash-flow, liquidity, assets, leverage (total debt over assets). - Sample for production-function estimation: - Final sample: 76,542 firms, corresponding to 656,960 firm-year observations. - Firms included report positive revenues, capital, labor cost, and intermediate expenditures (excludes ~one-fifth of original CADS). - Credit intensity metric: total credit at end of year t−1 divided by net revenue of year t. - On average, manufacturers are granted 43 cents for each euro of revenues. - On average, non-manufacturers are granted 34 cents for each euro of revenues. ### Firm-bank matched data and other sources - Italian Credit Register (CR) (Bank of Italy): borrowers with total exposures above e30,000 (threshold wase75,000 before 2009); contains outstanding bank debt by instrument. - CR–CADS matching via unique tax identifier. - Relationship-level dataset (CR) 1997–2013: 13,895,537 observations; 852,196 unique firms and 1,008 banks per year. - For all credit relationships 1998–2013, net credit flows measured as yearly growth rate (delta-log) of total outstanding debt; focus on credit granted rather than credit used. - Additional sources: TAXIA (past interest rates, >70% of credit), supervisory reports (banks’ assets, ROA, liquidity, capital ratio), INVIND survey (IT-adoption, R&D, export; panel ~3,000 firms), PatStat (2000-2013), World Management Survey (management practices, >100 manufacturing firms). --- ### Measurement notes and assumptions - Intermediate input price assumption: arithmetic mean of national price and national price deflated by local CPI. - See Lenzu & Manaresi (2018) for PIM details (referenced). - Production functions estimated by industry; sectors with fewer than 300 firm-year observations dropped. --- ### Identification of credit supply shocks and firm-level aggregation ### Definitions and log-linear decomposition - Total credit to firm i at end of year t: C_{i,t} = ∑_b C_{i,b,t}. - Pre-existing lending relation in period t iff C_{i,b,t−1} > 0. - Bank-specific supply determinants: S_{b,t}; firm demand determinants: D_{i,t}; match covariates: X_{i,b,t}; aggregate factors: J_t. - Assumption 1 (additive growth-rate structure) and log-linearization yield: - ∆c_{i,b,t} = j_t + d_{i,t} + φ_{b,t} + ε_{i,b,t} (equation (3)); - j_t mean growth rate of credit, - φ_{b,t} bank-level supply component, - d_{i,t} firm demand component, - ε_{i,b,t} match-specific residual plus approximation error. - Assumption 2: ε_{i,b,t} ⟂ D_i, S_b and E[d_{i,t}] = E[φ_{b,t}] = 0; under this, OLS with bank×year fixed effects yields unbiased φ_{b,t} estimates. - Identification exploits firms with multiple banking relationships; robustness tests in Appendix A.1 for omitted match variables. ### Firm-level credit supply shock aggregation - Benchmark aggregation: - φ_{i,t} = ∑_b φ_{b,t} · C_{b,i,t−1} / ∑_{b′} C_{b′,i,t−1} (equation (4)). - φ_{i,t} defined for firms with credit in year t−1; histogram referenced (figure 2). - φ_{i,t} behaves as expected in response to interbank market freeze (appendix D); banks with weaker balance-sheets decreased supply more sharply. --- ### Production model with heterogeneous financial frictions and estimation ### Model overview - Production: Y_{i,t} = exp{ ω_{i,t} + f(l_{i,t}, k_{i,t}, m_{i,t}, β_s) } with industry-specific f(·; β_s). - Productivity decomposition: ω_{i,t} = ̃ω_{i,t} + ε^Y_{i,t}; ε^Y_{i,t} i.i.d. unknown when decisions made. - Intermediate inputs chosen each period to maximize variable profits; unconstrained first-order condition equates marginal product to P_{M p,t}. - Financial constraints on inputs: - m_{i,t} ≤ m^{max}_{i,t} = log 1/P^M_{p,t} K_{i,t−1} · Γ(B_{i,t−1}, φ_{i,t}, ̃ω_{i,t}), Γ strictly increasing in ̃ω_{i,t}. - Observed m_{i,t} = min{m^{max}_{i,t}, m^{unc}_{i,t}} := m(x_{i,t}, ̃ω_{i,t}, φ_{i,t}) (equation 6). - Inversion: - ̃ω_{i,t} = h(x_{i,t}, m_{i,t}, φ_{i,t}), implying y_{i,t} = Ψ(x_{i,t}, m_{i,t}, φ_{i,t}) + ε^Y_{i,t}. - Productivity law of motion allows credit supply to enter dynamics: - E_t[ω_{i,t} | I_{t−1}] = g_t(ω_{i,t−1}, φ_{i,t−1}) (equation 7). - Productivity innovation ζ_{i,t} := ̃ω_{i,t} − E[̃ω_{i,t} | I_{t−1}] generates moment conditions used in estimation. ### Two-stage estimation strategy - Stage 1: estimate Ψ_{i,t} = E[y_{i,t} | x_{i,t}, m_{i,t}, φ_{i,t}] (control function approach augmenting Ackerberg et al. (2007)). - Stage 2: use moment conditions to estimate structural parameter β_s. - Measured firm-level productivity: ω_{i,t} = y_{i,t} − f(k_{i,t}, l_{i,t}, m_{i,t}, β_s) (or va_{i,t} − f(k,l,β_s) for value-added). --- ### Empirical findings: inputs, output, credit responses (section 4) ### Baseline input/output/credit regressions (equation 9) - ∆x_{i,t} = ψ_i + ψ_{p,s,t} + γ · φ_{i,t} + η_{i,t}, where x is log total credit, log value added, log net revenue, or log input. - Fixed effects: firm and year×industry×province. ### Key regression magnitudes (Table 2; All Industries and Manufacturing) - All Industries (selected): - Credit (delta Log): φ_{i,t} = 0.949*** (0.0196). - Value Added (delta Log): φ_{i,t} = 0.123*** (0.0162). - Net Revenues (delta Log): φ_{i,t} = 0.0474*** (0.0109). - Capital Stock (delta Log): φ_{i,t} = 0.0619*** (0.0128). - Wagebill (delta Log): φ_{i,t} = 0.0154* (0.00926). - Intermediate Inputs (delta Log): φ_{i,t} = 0.0220* (0.0114). - N. of Loan Applications (delta Log): φ_{i,t} = -0.537*** (0.0796). - Any Loan Received (delta Log): φ_{i,t} = -0.0780*** (0.0173). - Manufacturing (selected): - Credit: φ_{i,t} = 0.966*** (0.0253). - Value Added: φ_{i,t} = 0.134*** (0.0201). - N. of Loan Applications: φ_{i,t} = -0.424*** (0.113). - Elasticity interpretation: - A one-percentage-point decrease in φ_{i,t} is the change necessary to lower average credit granted by one percent (elasticity ≈ 1). ### Interpretation - Firms connected to banks contracting supply receive less credit, acquire fewer inputs, and produce less output relative to peers. - Effect stronger on value added than on capital accumulation; net revenue responds almost as much as capital. - Labor and intermediate inputs less sensitive to credit supply shocks than output and capital. --- ### Credit supply effects on productivity growth (main results) ### Baseline productivity-growth regression (equation 10) - ∆ω_{i,t} = ψ_i + ψ_{p,s,t} + γ · φ_{i,t} + η_{i,t}. ### Main estimates (Table 3; selected) - All industries: - Value Added, Cobb-Douglas: φ_{i,t} = 0.0946*** (0.0155), Observations 656,960, R2 0.172. - Value Added, Trans-Log: φ_{i,t} = 0.109*** (0.0160), Observations 656,960, R2 0.185. - Net Revenue, Cobb-Douglas: φ_{i,t} = 0.0190*** (0.00477), Observations 656,960, R2 0.178. - Net Revenue, Trans-Log: φ_{i,t} = 0.0259*** (0.00491), Observations 656,960, R2 0.195. - Manufacturing: - Value Added, Cobb-Douglas: φ_{i,t} = 0.115*** (0.0178), Observations 347,990, R2 0.186. - Net Revenue, Cobb-Douglas: φ_{i,t} = 0.0303*** (0.00595), Observations 347,990, R2 0.144. ### Quantitative magnitudes and examples - A one percentage point credit supply shock induces: - ≈ 0.1 percentage point change in value-added productivity growth for the whole economy. - ≈ 0.13 percentage point change for manufacturing. - 0.02 to 0.03 percentage point change on revenue-based productivity measures. - Sample episode 2007–2009: - Drop in total growth rate of credit granted: around 12%. - Mean value-added productivity growth declined by more than 8%. - Revenue productivity growth declined by 1%. - If drop in credit fully driven by supply, it would explain between 12% and 30% of productivity drop over the same period. ### Persistence and economic relevance - Productivity effects are persistent; credit supply particularly valuable during financial turmoil. - Credit contractions both depress investments (raising MRPK) and depress TFP (lowering MRPK), creating opposing effects on MRPK distribution. --- ### Robustness, identification checks, and measurement considerations ### Robustness exercises (Table 4 and appendices) - Adding lagged controls (polynomial in assets; ratios of value added, cash flow, liquidity, bank debt to assets) has negligible impact on coefficients. - Excluding “important” borrowers (> 1% of any lender’s credit) does not change estimated effect. - Fixed-effect structures: results stable when varying granularity (industry×province×year vs coarser). - Oster (2016) bounding (R_max = 1, δ = 2): bounding sets never contain zero. - Alternative credit-supply measure controlling for match-level characteristics (interest rate, instrument type) yields similar effect. - Split-sample cross-validation of bank shocks yields similar baseline estimates. - Predictability and balance tests: normalized differences across quartiles of credit-supply innovation well below 0.25 (Table A.2). - Production-function robustness: - Four-digit industry classification and Olley & Pakes (1996) exit control unchanged magnitudes. - Cost-share (Foster et al., 2017) estimation in the ballpark. - Grid search: ρ varied in [0.3, 2], β_k varied in [0.01, 0.9]; γ(̃ρ,̃β_k) remains between 0.07 and 0.1 and stays positive and significant. - Bootstrap of production-function re-estimates (appendix figure A.7): bootstrapped distribution of estimated effect of credit supply on productivity all coefficients above zero. ### Measurement error, factor hoarding, and MRPK - Measurement error discussed; Table 2 mitigates concerns by showing inputs respond to credit shocks. - Factor hoarding/adjustment costs could imply short-lived productivity loss; observed persistence (at least four years) inconsistent with purely short-lived measurement artifacts. - MRPK findings (Table A.3): credit availability leads to increase in firm MRPK in panel; examples: - All industries (Value Added, log MRPK): φ_{i,t} = 0.258*** (0.0345), Observations 656,960. - Manufacturing: φ_{i,t} = 0.294*** (0.0442), Observations 347,990. --- ### Heterogeneity, persistence, and asymmetry (sections 5.2–5.4) ### Heterogeneous effects by firm and lender characteristics (Table 5) - Specification includes interaction γhet·φ_{i,t}·D_{i,t−1} with Di,t−1 indicator. - Firm size: - Top quartile of asset size distribution less affected; γhet negative; size difference significant in manufacturing only. - Lender size: - LenderSize bottom quartile → effects twice as large as other firms; interaction estimated positive and significant (column (2)). - Number of lending relationships: - Firms in bottom quartile for number of lenders show much larger TFP growth responses; fewer than 5% observations have only one lender. - Industry leverage: - Industries above median mean leverage show stronger effects on revenue productivity (interaction significant). - Firm age: - “Young” firms (bottom quartile; average young firm age = 6.3 years) estimated effect ~50% larger; difference not statistically significant. - Patent-intensive industries: - Above-median patent frequency industries appear more sensitive to credit supply shocks; focusing on manufacturing cannot reject zero effect. - Great Recession interaction: - Years 2008–2009: estimated effect about 16% larger overall and 60% larger in manufacturing. ### Persistence (section 5.3) - Dynamic specification uses unexpected credit supply shock ζφ_{i,t} with T = 3. - Empirical pattern (Figure 3): - Peak productivity effect one year after shock. - Effect remains positive and significant for at least four years. - No statistically significant pre-trend at 1% level. - Note on attenuation: within-firm estimator implies mechanical attenuation; a shock of magnitude 1·m appears as change 0.03·m to 0.06·m after 3 years due to estimator properties. ### Asymmetric effects (section 5.4) - Quintile specification shows concave relation between credit supply and revenue productivity (Figure 5). - Negative credit supply shocks drive persistence; positive shocks do not undo harm from a negative shock of same size (Figure 6). - Policy implication: stability of credit provision matters; volatility with symmetric expansions/contractions can yield net productivity loss. --- ### Interbank market collapse as natural experiment (section 6) ### Interbank context and identification - Total gross interbank funding > 13% of total assets of Italian banks at end-2006. - Total transactions among banks fell from €24bn. in 2006 to €4.8bn. at end-2009. - Euribor-Eurepo spread: practically zero until August 2007; reached over 50 basis points in the subsequent year; increased fivefold after Lehman; remained well above 20 basis points subsequently. - Focus period: 2007–2009; firm exposure INTBK_{i,2006} constructed as firms’ 2006 average of banks’ interbank exposure weighted by 2006 credit shares (time-invariant cross-sectional exposure). ### Empirical magnitudes (Table 6 and appendices) - Cross-sectional regression ∆ω_{i,t} = ψ_{p,s,t} + γ·INTBK_{i,2006} + η_{i,t}, t ∈ [2007,2009]: - A 1% increase in average bank dependence on interbank market results in: - ≈ .05% decrease in average value-added productivity growth. - ≈ .02% decrease in revenue productivity growth. - The same interbank shock that decreases credit growth by 1% decreases value-added productivity growth by 0.25% and revenue productivity growth by 0.1% for the whole sample. - These effects 2–5 times larger than baseline estimates (Table 3). - Robustness: - Appendix D placebo and pre-trend tests show no pre-crisis differences; placebo collapse in early 2000s yields no effect. - Appendix D.2: interbank dependence predicts decreases in credit granted 2007–2009 (examples: ITBK_{i,2006} coefficients −0.192*** (0.0400) and −0.241*** (0.0545)). --- ### Channels: IT, innovation, exports, and management (section 7) ### IT-intensity (Table 7) - ITi,t = (log) PCs per 1,000 euros of capital (data 1999–2001). - Regression: ITi,t = ψi + ψs,p,t + γ·φi,t + ηi,t. - Finding: positive credit supply shocks increase IT-intensity of capital (Table 7, column (1)): φ_{i,t} coefficient 0.672** (0.269), Observations 3,913, R2 0.967. ### Innovation and patents (Table 7) - Patent activity: share applying for at least one patent ≈ 2% between 2002–2007; declined to 1.5% in 2009; rose to >1.6% subsequently. - Patent regressions (lagged): PatentApp_{i,t+1} and GrowthPatentApp_{i,t+1} regressed on φ_{i,t}. - Findings: firms patent more when having easier bank credit access; Patent growth coefficient examples include 0.305* (0.159). - Interbank exposure and patents: average interbank exposure in patenting sub-sample = 13% → implied decline in patent applications ≈ 22% (back-of-the-envelope), > half of observed contraction. ### R&D and exporting - Extensive-margin LPM: Di,t = ψi + ψp,s,t + γ·φi,t + ηi,t for R&D and exporting. - Findings: firms more likely to start (and/or less likely to stop) R&D and exporting with easier external finance; cannot reject null of no effect on R&D in some specifications. ### Management practices (World Management Survey; cross-section) - MS_{i,t} = ψ + γ·φ_{i,t} + η_{i,t}; sample size 183 observations. - Finding (Table 7, column (8)): increase in credit supply stimulates adoption of superior management practices; result robust to firm controls but small sample cautioned. ### Managerial inattention hypothesis - Time and effort to secure finance distract managers from core business; evidence consistent with firms searching more for new lenders after negative credit shocks (Table 2 columns (8) and (9)). - Hypothesis may explain immediate productivity impact; more direct investigation left for future research. --- ### Conclusion (section 8): policy-relevant inferences - Empirical approach: additive growth-rate model separates credit demand from supply shocks using universe of bank-firm credit relationships 1997–2013; production model with heterogeneous credit constraints isolates idiosyncratic productivity dynamics. - Main findings: - Banks contracting credit lead connected firms to acquire fewer inputs and produce less output than competitors. - Output effects stronger than input effects → productivity affected by credit availability. - Credit supply boosts productivity growth; effects are sizable, persistent, and robust. - Effects stronger for smaller firms and industries relying heavily on bank credit. - Negative credit supply shocks produce stronger effects than positive shocks of same magnitude → credit stability matters. - Financial turmoil can have persistent effects on aggregate output by depressing firm TFP. - Activities stimulated by credit availability: IT adoption, superior management practices, export orientation, innovation. - Managerial time reallocation to secure funding may reduce effort on productivity-enhancing activities; firms seek new lending relationships more after negative shocks. *Source: wpiea2019107 - 2.1 Firm balance-sheets: The CADS dataset; sections 4–8 of the source PDF.* ### 2.1 Firm balance-sheets: The CADS dataset ### 2.1 Firm balance-sheets: The CADS dataset ### CADS database: coverage and construction - The Company Accounts Data System (CADS) is a proprietary database administered by CERVED-Group Ltd. for credit risk evaluation; it has collected detailed balance-sheet and income statement information on non-financial corporations since 1982. - Net revenue of CADS firms account for about 70% of the total revenues of the private non-financial sector. - Because CADS is used by banks for credit decisions, the data are carefully controlled. - Firm-level capital series are computed applying the perpetual-inventory method (PIM) on book-value of capital, investments, divestments, and sector-level deflators and depreciation rates. - Operating value added and intermediate expenditures are recorded in nominal values and converted to real terms using sector-level deflators from National Accounts. - Baseline measure of labor is the wage bill, deflated using the consumer price index (CPI). - Expenditures on intermediate inputs are deflated using a combination of sector-level deflator and regional-level CPI. - Industry classification: Nace Rev.2 two-digit definition (robustness: four-digit definition yields very similar results). - CADS also provides firm characteristics used as time-varying controls: age, cash-flow, liquidity, assets, and leverage (total debt over assets). ### Time coverage for production-function estimation - Production functions are estimated for firms sampled in CADS from 1998 to 2013. ### Measurement notes - See Lenzu & Manaresi (2018) for details on PIM (as referenced). - Assumption on intermediate input prices: price of intermediate inputs is the arithmetic mean of national price and national price deflated by local CPI (used because some inputs might be bought on national rather than local markets). ### Firm-bank matched data: the Italian Credit Register (CR) - The Italian Credit Register (CR), owned by the Bank of Italy, collects individual data on borrowers with total exposures (both debt and collateral) above e30,000 towards any intermediary operating in the country (including banks, other financial intermediaries providing credit, and special-purpose vehicles). - The threshold wase75,000 before 2009. - CR contains data on outstanding bank debt of each borrower, categorized into loans backed by accounts receivable, term loans, and revolving credit lines. - CR data are matched to CADS using each firm’s unique tax identifier. - For all credit relationships of any Italian incorporated firm and any intermediary between 1998 and 2013, net credit flows are measured as the yearly growth rate (delta-log) of total outstanding debt. - Focus is on credit granted rather than credit used. ### Additional data sources (used for channels and robustness) - TAXIA database (Bank of Italy): past interest rates charged by banks to firms; encompasses over 70% of all credit granted to the Italian economy. Interest rates computed as ratio of interest expenditures to quantity of credit used. - Supervisory reports: banks’ assets, ROA, liquidity, capital ratio, interbank liabilities and assets (used in study of the 2007-2008 interbank market collapse). - INVIND Survey (Bank of Italy): data on IT-adoption, R&D, and export; panel of around 3,000 firms, representative of Italian firms with more than 20 employees active in manufacturing and private services. - PatStat (release prepared by UnionCamere): patent applications to the European Patent Office matched to tax identifiers for 2000-2013. - World Management Survey: management practices for more than 100 manufacturing companies. ### Sample selection and descriptive statistics - Two main datasets: - (a) Relationship-level dataset (bank-firm-year triplet) to identify credit supply shocks. - (b) Firm-level dataset (firm-year pairs) to estimate production functions; firm-level credit supply shocks are aggregated across banks. - Relationship-level dataset (from CR): all relationships between incorporated firms and financial intermediaries during 1997-2013. - Resulting dataset: 13,895,537 observations; composed of 852,196 unique firms and one 1,008 banks per year. - Firm-level dataset (from CADS) for production-function estimation: - Include firms that report positive revenues, capital, labor cost, and intermediate expenditures. - This excludes around one-fifth of original CADS dataset. - Final sample: 76,542 firms, corresponding to 656,960 firm-year observations. - Credit intensity metric: ratio of total credit granted at the end of year t−1 to net revenue of year t. - On average, manufacturers are granted 43 cents for each euro of revenues generated. - On average, non-manufacturers are granted 34 cents for each euro of revenues generated. - Additional descriptive findings (referenced): - Appendix figure A.1: credit-intense companies are larger in non-manufacturing sectors, but not in manufacturing. - Appendix figure A.2: industries with a higher capital-to-labor ratio are more credit-intensive. ### Theoretical framework overview (linking credit supply and productivity) - Empirical strategy: disentangle idiosyncratic shocks to credit supply from credit demand shocks and aggregate shocks (section 3.1); then build a production model with heterogeneous credit constraints to recover firm TFP (section 3.2). ### Credit supply shocks: definition and identification - Definition: a credit supply shock is any change in bank-specific factors affecting a bank’s ability and willingness to provide credit to firms. - Notation and aggregation: - Total credit granted to firm i at end of year t: C_{i,t} = ∑_b C_{i,b,t}. - A firm i and bank b have a pre-existing lending relation in period t iff C_{i,b,t−1} > 0. - Collect bank-specific supply determinants into vector S_{b,t}; firm demand determinants into vector D_{i,t}; match-specific covariates into X_{i,b,t}; aggregate factors into J_t. - Assumption 1: ∃ some smooth, unknown function C(·) such that C_{i,b,t}/C_{i,b,t−1} = C(J_t,D_{i,t},S_{b,t},X_{i,b,t}) / C(J_{t−1},D_{i,t−1},S_{b,t−1},X_{i,b,t−1}) (equation (1)) - This assumption limits substitution patterns amongst different lenders (rules out impact of other banks’ idiosyncratic shocks S_{b′,t} on credit by b to i). - Log-linearized decomposition (equations (2) and (3)): - ∆c_{i,b,t} = j_t + ∆d'_{i,t} c_1 + ∆s'_{b,t} c_2 + ∆x'_{i,b,t} c_3 + approx_{i,b,t} (equation (2)) - ∆c_{i,b,t} = j_t + d_{i,t} + φ_{b,t} + ε_{i,b,t} (equation (3)), where: - j_t is mean growth rate of credit in the economy, - φ_{b,t} is change in credit granted explained by bank b’s supply factors, - d_{i,t} is change in credit granted explained by firm factors, - ε_{i,b,t} is matching-specific shock (∆x'_{i,b,t} c_3) plus approximation error. - Assumption 2: ε_{i,b,t} ⟂ D_i, S_b (orthogonality of matching-specific residual with firm and bank identities), and E[d_{i,t}] = E[φ_{b,t}] = 0. - Under Assumption 2, OLS with bank×year fixed effects yields unbiased estimates of ∆s_{b,t} (bank-level supply shocks φ_{b,t}). - Identification strategy: - Focus on corporations with multiple banking relationships to exploit within-firm-and-time variability. - Relation to Amiti & Weinstein (2017): shows orthogonality can be achieved by relabeling fixed effects; paper seeks pure idiosyncratic supply factors ∆s_{b,t} and thus scrutinizes Assumption 2. - Appendix A.1: tests for omitted variables in ε_{i,b,t} (substitution/complementarity patterns and bank-firm relation characteristics) and shows main results are robust to including such controls. - Firm-level credit supply shock construction: - Lending-channel intuition: borrower-lender relationships are sticky; changes in credit supplied by an existing bank disproportionately affect the connected firm. - Benchmark aggregation: φ_{i,t} = ∑_b φ_{b,t} · C_{b,i,t−1} / ∑_{b′} C_{b′,i,t−1} (equation (4)) - φ_{i,t} defined for all firms with credit in year t−1; histogram of φ_{i,t} provided in figure 2 (referenced). - φ_{i,t} behaves as expected in response to the interbank market freeze (appendix D), indicating the additive growth rate model captures credit supply factors. - Appendix A.2: shows banks with weaker balance-sheets entering the Great Recession decreased credit supply more sharply; documents crowding out of corporate lending by sovereign debt; M&A episodes followed by contraction of credit supplied by the target. ### Production model with heterogeneous financial frictions (overview) - Objective: estimate firms’ production functions and recover idiosyncratic productivity while accounting for credit constraints and modified productivity dynamics. - Framework: - Uppercase letters denote levels; lowercase denote logs. - Firm i in industry s at time t combines capital (k_{i,t}), labor (l_{i,t}), and intermediate inputs (m_{i,t}) to generate sales Y_{i,t} according to industry-specific production function f(·; β_s). - Hicks-neutral productivity ω_{i,t} enters multiplicatively: Y_{i,t} = exp{ ω_{i,t} + f(l_{i,t}, k_{i,t}, m_{i,t}, β_s) }. - Productivity decomposition: ω_{i,t} = \tilde{ω}_{i,t} + ε^Y_{i,t}, where ε^Y_{i,t} is i.i.d. and unknown when production decisions are made. - Intermediate inputs are flexibly chosen each period to maximize variable profits. If unconstrained, materials m^{unc} solve: ∂ exp{ f(l_{i,t}, k_{i,t}, m^{unc}, β) + \tilde{ω}_{i,t} } / ∂m = P_{M p,t} where P_{M p,t} is the price of materials faced by firm i (may depend on location p). - The model augments the control-function approach (Ackerberg et al., 2007) with credit constraints and a modified law of motion for productivity (details in appendix B). *Source: wpiea2019107 - 2.1 Firm balance-sheets: The CADS dataset* ### section 4, we provide evidence that firms acquire less inputs when they receive negative credit supply ### wpiea2019107 - section 4, we provide evidence that firms acquire less inputs when they receive negative credit supply shocks ### Financial-frictions in input demand and production-function setup - Intermediate inputs (and other inputs) may face financially generated constraints: - m_{i,t} ≤ m^{max}_{i,t} = log 1/P^M_{p,t} K_{i,t−1} · Γ(B_{i,t−1}, φ_{i,t}, ̃ω_{i,t}), where Γ is an unknown function and B_{i,t−1} is previous-period debt. - Γ is assumed strictly increasing in its third argument (productivity ̃ω_{i,t}). - Observed intermediate inputs satisfy: - m_{i,t} = min{m^{max}_{i,t}, m^{unc}_{i,t}} := m(x_{i,t}, ̃ω_{i,t}, φ_{i,t}) (equation 6). - Productivity enters the materials decision and can be inverted: - ̃ω_{i,t} = h(x_{i,t}, m_{i,t}, φ_{i,t}), implying log sales: - y_{i,t} = Ψ(x_{i,t}, m_{i,t}, φ_{i,t}) + ε^Y_{i,t}, with Ψ = h + f(l,k,m,β_s). - Productivity law of motion allows credit supply to affect dynamics: - E_t[ω_{i,t} | I_{t−1}] = g_t(ω_{i,t−1}, φ_{i,t−1}) (equation 7). - Productivity innovation: - ζ_{i,t} := ̃ω_{i,t} − E[̃ω_{i,t} | I_{t−1}], generating moment condition (8) used in estimation. - Estimation is two-stage: - Stage 1: estimate Ψ_i,t = E[y_{i,t} | x_{i,t}, m_{i,t}, φ_{i,t}]. - Stage 2: use moment conditions to estimate structural parameter β_s. - Measured firm-level productivity is residual: - ω_{i,t} = y_{i,t} − f(k_{i,t}, l_{i,t}, m_{i,t}, β_s) (or ω_{i,t} = va_{i,t} − f(k,l,β_s) in value-added case). - The paper measures firms’ ability to transform inputs into sales/value added (a revenue-based productivity measure, comparable to "regression-residual total factor revenue productivity" or tfpr_rr). ### Empirical specification for credit supply effects on inputs, output, and credit - Baseline regression (equation 9): - ∆x_{i,t} = ψ_i + ψ_{p,s,t} + γ · φ_{i,t} + η_{i,t}, where x_{i,t} is log total credit, log value added, log net revenue, or log input. - Fixed effects: firm and year×industry×province. - Key empirical findings (Table 2): - Firms connected with banks contracting supply show lower growth of credit received, inputs acquired, and output produced than peers in the same market. - The elasticity of credit granted with respect to the firm-level credit supply shock is approximately equal to 1: - A one-percentage-point decrease in φ_{i,t} is the change of credit supply necessary to lower the average credit granted by one percent. - Impact on inputs and output: - Effect is stronger on value added than on capital accumulation. - Net revenue responds almost as much as capital. - Labor and intermediate inputs are much less sensitive to credit supply shocks than output and capital (both economically and statistically). - Firms decrease loan applications to previously unconnected lenders after a positive credit supply shock (columns 8 and 9 of Table 2), supporting separation of demand and supply variation. ### Credit supply effects on firm productivity growth (main results) - Baseline productivity-growth regression (equation 10): - ∆ω_{i,t} = ψ_i + ψ_{p,s,t} + γ · φ_{i,t} + η_{i,t}, where ∆ω_{i,t} is delta log Hicks-neutral productivity and φ_{i,t} is weighted average of previous-period lenders’ credit supply shocks. - Main estimates (Table 3 and discussion): - A decrease in credit supply causes a decline in productivity growth across specifications (value-added and revenue-based, Cobb-Douglas and Trans-Log). - Quantitative magnitudes: - A credit supply shock of one percentage point induces a change in the growth rate of value-added productivity of approximately one-tenth of a percentage point for the whole economy. - The effect for manufacturing is approximately 0.13 points. - The effect on revenue-based measures of productivity is between 0.02 and 0.03 percentage points. - Interpretation and comparison: - The standard deviation of value-added productivity is more than three times that of revenue productivity in the sample, partially explaining differences in coefficients. - Example magnitude: the drop in total growth rate of credit granted between 2007 and 2009 is around 12% in the sample; over the same period mean value-added productivity growth declined by more than 8% and revenue productivity growth declined by 1%. - If the drop in credit was fully driven by supply, it would explain between 12% and 30% of the productivity drop over the same period. - Persistence and economic relevance: - Productivity effects of credit shock are persistent and credit supply is particularly valuable during financial turmoil. - Implications for MRPK and misallocation: - Credit contractions both depress investments (raising MRPK) and depress TFP (lowering MRPK), with opposing effects on MRPK distribution. ### Robustness and identification checks - Sources of identification concerns addressed: reverse causality, correlated unobservables, assortative matching. - Specific robustness exercises (Table 4 and appendices): - Adding lagged controls (polynomial in assets size; ratios of value added, cash flow, liquidity, bank debt to assets) has negligible impact on coefficients. - Excluding “important” borrowers (any firm that, between 1997 and 2013, accounts for more than 1% of any lender’s credit) does not change estimated effect (mitigates reverse causality). - Fixed-effects structure: - Comparing industry×province×year fixed effects to coarser fixed effects shows coefficient stability despite doubling R^2, suggesting correlated unobservables would have to be orthogonal to location/industry to explain results. - Appendix C.2 bounding sets (Oster (2016) method) do not contain zero. - Alternative credit-supply measure controlling for match-level characteristics (interest rate, instrument type) yields similar effect (no evidence assortative matching drives results). - Cross-validation / split-sample estimation of bank-level shocks: - Construct φ^{A}_{b,t} and φ^{B}_{b,t} on disjoint firm subsamples and compute firm shocks using bank shocks estimated on the other subsample; baseline estimates are similar (column 7). - Predictability and balance tests: - Define credit-supply innovation ζ^φ_{i,t} := φ_{i,t} − E[φ_{i,t} | φ_{t−1}] and compare lagged firm observables across quartiles of ζ^φ_{i,t}; normalized differences are all well below 0.25 (top panel Table A.2), supporting causal interpretation. - Production-function specification robustness: - Re-estimating with finer four-digit industry classification (column 8) and controlling for endogenous exit as in Olley & Pakes (1996) (column 9) yields unchanged magnitude. - Cost-share (Foster et al., 2017) estimation (column 10) produces results in the ballpark of baseline. - Grid search over alternative value-added Cobb-Douglas parameters: - ρ varied from 0.3 to 2 and β_k from 0.01 to 0.9; γ(̃ρ,̃β_k) remains between 0.07 and 0.1 across the grid and stays positive and statistically significant. - Bootstrap of production-function re-estimates (appendix figure A.7): - Bootstrapped distribution of estimated effect of credit supply on productivity: all coefficients above zero; sampling error in productivity estimation does not distort inference. - Additional concerns like measurement error, adjustment costs, and factor hoarding are discussed in appendix C.3 (not repeated here). ### Key quantitative statements and exact figures preserved from the source - One-percentage-point decrease in φ_{i,t} → credit granted falls by 1 percent (elasticity ≈ 1). - Effect on value-added productivity growth: - ≈ 0.1 percentage point (whole economy). - 0.13 percentage point (manufacturing). - Effect on revenue productivity growth: between 0.02 and 0.03 percentage points. - Credit growth drop between 2007 and 2009 in sample: around 12%. - Value-added productivity growth decline between 2007 and 2009: more than 8%. - Revenue productivity growth decline between 2007 and 2009: 1%. - Grid for robustness: ρ ∈ [0.3, 2], β_k ∈ [0.01, 0.9]; γ(̃ρ,̃β_k) ∈ [0.07, 0.1]. - Importance threshold for “important” borrower: > 1% of credit granted by any lender. - Normalized-differences rule-of-thumb threshold cited: 0.25. *Source: wpiea2019107 - section 4 (PDF chapter/section) — "section 4, we provide evidence that firms acquire less inputs when they receive negative credit supply shocks."* ### 5.2 Heterogeneity ### 5.2 Heterogeneity ### Heterogeneous effects by firm characteristics - Baseline heterogeneous specification: - ∆ωi,t = ψi + ψs,t,p + ψd · Di,t−1 + γ·φi,t + γhet·φi,t·Di,t−1 + ηi,t - Di,t−1 = 1 if firm i in year t−1 belongs to a specified part of the distribution; 0 otherwise. - Firm size: - Firms in the top quartile of the year-specific asset size distribution are less affected by credit supply shocks. - γhet is estimated negative; the size difference is economically and statistically significant in manufacturing only. - Lender size: - LenderSizei,t−1 = (∑b Assetsb,t−1 · Cb,i,t−1) / (∑b′ Cb′,i,t−1). - Firms connected to smaller banks (bottom quartile of LenderSize distribution) experience credit supply effects twice as large as other firms (columns (2) of Table 5). - The impact remains statistically larger than zero for the rest of the sample. - Number of lending relationships: - Firms in the bottom quartile for number of lenders in the previous period show much larger TFP growth responses to credit supply shocks (column (3) of Table 5). - Note: less than 5% of observations have only one lender, so distinguishing single-lender firms is not reliably estimated. - Industry leverage: - Industries classified above the median mean leverage (debt over assets) exhibit stronger effects of credit supply shocks on revenue productivity (column (4) of Table 5). - Firm age: - "Young" firms = bottom quartile of year-specific age distribution; average young firm age = 6.3 years. - Estimated effect for young firms is approximately 50% larger, but the difference between age groups is not statistically different from zero. - Patent-intensive industries: - Industries are divided by 2-digit patent frequency (share of firm-year observations with at least one patent application). - Firms in industries with above-the-median patent frequency (e.g., R&D services and pharmaceutical manufacturing) appear more sensitive to credit supply shocks (column (6) of Table 5). - When focusing on manufacturing, null of zero effect cannot be rejected. - Great Recession (2008–2009) interaction: - Allowing coefficients to differ for years 2008 and 2009 yields an estimated effect about 16% larger during the Great Recession (60% larger in manufacturing) (column (7) of Table 5). - Productivity impact of credit supply is present outside the financial crisis as well. --- ### 5.3 Persistence ### Persistence of credit supply effects on productivity - Dynamic specification: - ωi,t = ψi + ψp,s,t + ∑τ=−T^T γτ · ζφi,t + ηi,t, with ζφi,t = unexpected credit supply shock (eq. (11)); set T = 3. - Empirical findings (graphically in Figure 3): - Peak in productivity occurs one year after the shock. - Effect remains positive and significant for at least four years. - Temporary credit contraction can have persistent effects on productivity. - No statistically significant pre-trend at 1% confidence level. - Methodological note: - Within-firm estimator induces mechanical negative correlation between observation means at different lags; expected attenuation: a shock of magnitude 1·m shows as change in productivity of only 0.03·m to 0.06·m after 3 years. --- ### 5.4 The asymmetric effect of credit supply shocks ### Nonlinear and asymmetric relationship - Quintile specification: - ∆ωi,t = ψi + ψp,s,t + ∑q=1,q≠3^5 γq · 1(φi,t ∈ q) + ηi,t; median quintile q = 3 omitted (γ3 = 0). - Key patterns (Figure 5): - Relation between credit supply and revenue productivity is concave. - Firms linked to banks with relatively low credit supply experience lower revenue productivity growth than competitors. - Firms linked to banks with particularly strong increases in credit do not grow at a particularly high rate. - Positive vs. negative shocks (Figure 6): - Re-estimating persistence by differentiating positive and negative shocks shows effects are driven by negative credit supply shocks (coefficients for negative shocks shown with negative values). - Increase in credit supply cannot undo harm from a negative shock of the same size; volatility in banking sector supply is detrimental. - Policy implication: - Stability of credit provision matters: a credit crunch followed or preceded by an expansion of the same magnitude leads to a net loss in average firm productivity. --- ### 6 The Interbank Market Collapse as a Natural Experiment ### Interbank market context and measurement - Interbank funding importance and collapse: - Total gross interbank funding accounted for over 13% of total assets of Italian banks at end-2006. - In Italy, total transactions among banks fell from €24bn. in 2006 to €4.8bn. at end-2009. - Euribor-Eurepo spread: practically zero until August 2007; reached over 50 basis points for all maturities in the subsequent year; increased fivefold after Lehman; remained well above 20 basis points subsequently. - Identification strategy: - Focus period: 2007–2009 (credit dried up most); ECB interventions partially offset shock after 2009. - Firm exposure measure: banks' interbank exposure at end-2006 averaged at firm level using firms’ 2006 credit shares; INTBKi,2006 is time-invariant; use cross-sectional variation. - Cross-sectional regression (three-year window, t ∈ [2007,2009]): - ∆ωi,t = ψp,s,t + γ·INTBKi,2006 + ηi,t - ψ includes province×industry×year fixed effects. - Empirical findings (Table 6): - Firms whose lenders relied more on the interbank market in 2006 had significantly lower revenue and value-added productivity growth during the credit crunch. - A 1% increase in average bank dependence on the interbank market results in: - approximately .05% decrease in average value-added productivity growth; - approximately .02% decrease in revenue productivity growth. - The same interbank shock that decreases credit growth by 1% decreases value-added productivity growth by 0.25% and revenue productivity growth by 0.1% for the whole sample. - These effects are between two and five times larger than the baseline estimate from Table 3. - Robustness and validity: - Appendix D robustness exercises suggest bank exposure was not correlated with borrowers’ pre-crisis characteristics. - Firms harder hit by interbank collapse were not more sensitive to the business cycle before the 2007–08 crisis and had similar productivity growth rates. - Placebo test using a fake interbank collapse in the early 2000s recession finds no effect. --- ### 7 Beyond Measurement: Channels ### Overview - Aim: assess how credit supply shocks affect productivity-enhancing activities using INVIND survey information on R&D investment, export, IT-adoption, and self-reported “obstacles to innovation”. - Timing and horizon: - Some mechanisms (IT-adoption, management) can affect productivity immediately; others (innovation) may take years. - Section 5.3 shows immediate productivity effects lasting at least two years and persistent effects for at least four years. ### 7.1 IT-intensity of capital - Measure: - ITi,t = (log) number of PCs per 1,000 euros of capital (PC purchases counted as investment; data available 1999–2001). - Specification: - ITi,t = ψi + ψs,p,t + γ·φi,t + ηi,t - Finding (Table 7, column (1)): - Positive credit supply shocks increase IT-intensity of capital stock. - Suggests financial frictions reduce the quality of capital inputs captured in the productivity residual. ### 7.2 Innovation and exporting - Patents: descriptive facts - Share of firms applying for at least one patent: approximately 2% between 2002 and 2007; declined to 1.5% in 2009; rose to a bit more than 1.6% in the following two years. - Approximately 5 patent applications per 100 firms per year before the Great Recession; around 3.4 in 2009. - Patent regressions (lagged specification): - PatentAppi,t+1 = ψi + ψp,s,t + γ·φi,t + ηi,t - GrowthPatentAppi,t+1 = ψi + ψp,s,t + γ·φi,t + ηi,t - GrowthPatentAppi,t+1 uses Davis et al. (1996) formula: GrowthPatentAppi,t+1 = 2·(PatentAppi,t+1 − PatentAppi,t−1) / (PatentAppi,t+1 + PatentAppi,t−1). - Findings (Table 7, columns (2) and (3)): - Italian firms patent more when they have easier access to bank credit. - Interbank exposure and patents: - PatentGrowthi = ψp + ψs + γ·INTBKi,2006 + ηi, where PatentGrowthi = 2·(Patentpost,i − Patentpre,i)/(Patentpost,i + Patentpre,i). - Pre/post windows: pre ends 2006; post starts 2007; post ends 2010; pre starts 2001 (results robust to alternative boundaries). - Finding (Table 7, column (4)): negative impact of interbank exposure on patents. - Average interbank exposure in patenting sub-sample = 13% → implied decline in patent applications ≈ 22% (back-of-the-envelope), more than half of observed contraction. - R&D and exporting (extensive margin linear probability models): - Di,t = ψi + ψp,s,t + γ·φi,t + ηi,t, where Di,t = 1 if firm i engages in R&D or exporting in year t. - Findings (Table 7, columns (5) and (6)): - Firms are more likely to start (and/or less likely to stop) conducting R&D and exporting when they have easier access to external finance. - Cannot reject null of no effect of credit shocks on R&D (statistical uncertainty). - Self-reported financial constraint on innovation: - FinConi,2010 = ψs,p + γ·φi,2010 + ηi, where FinConi,2010 = 1 if difficulty in getting external funds reported as “somehow important” or “very important” obstacle to innovation (2011 survey wave; question refers to 2010). - Finding (Table 7, column (7)): firms receiving positive credit supply shocks are less likely to consider external funds a substantial obstacle to innovation. - Caveat: single-year question prevents panel variation, but provides indirect evidence that financial frictions dampen innovative efforts. *Source: wpiea2019107 - 5.2 Heterogeneity (IMF Working Paper content unit)* ### 7.3 Management practices ### 7.3 Management practices ### Evidence and data - Management practices matter for firm performance, as shown by Bloom et al.(2013) for India and by Giorcelli (2016) for Italy. - Firm-level data on management practices come from the World Management Survey (WMS - http://worldmanagementsurvey.org/). - WMS scores management from one (worst practice) to five (best practice) across eighteen key management practices used by industrial firms. - Merged WMS data on Italian companies by name yields a sample of 183 observations. ### Empirical specification - Cross-sectional model estimated: - MS_{i,t} = ψ + γ·φ_{i,t} + η_{i,t} - MS_{i,t} is the overall management score for firm i surveyed in year t. ### Key finding - Results (column (8) of Table 7) indicate that an increase in credit supply stimulates the adoption of superior management practices. - The relation between credit supply shock and management is largely unaffected by the inclusion of a large set of firm-level controls. - Caution: the small sample size might cast doubt on the robustness of this result. --- ### 7.4 Managerial inattention ### Context and motivation - Dealing with investors and creditors takes a substantial share of executive time. - Bandiera et al. (2011) study the use of time by 94 CEOs of top-600 Italian companies and document that finance is the topic on which the CEO spends the most time talking with others in the firm. - Of the outsiders with whom CEOs spend the most time, investors and bankers are, respectively, third and fifth. ### Mechanism - For smaller private companies (the bulk of the sample), time and effort required to establish and maintain relations with lenders may be even more demanding. - If managerial delegation is difficult (Akcigit et al., 2016), the more difficult or time-consuming it is to find external funds, the less managers can work on improving their core business. - Entrepreneurs connected to lenders who contract their credit supply might need to spend more time and energy to establish new lending relations and therefore exert less effort on improving firm productivity. ### Supporting evidence - Anecdote: an author’s aunt managing the family business during the credit crunch said, “I barely have time to go to the factory, I spend most of my mornings at banks trying to get some money.” - Results on firms’ effort to search for new lenders (columns (8) and (9) of Table 2) are consistent with this mechanism: firms receiving positive credit supply shocks are less likely to try to establish new lending relationships. - The “managerial inattention” hypothesis may explain why the impact of credit supply on productivity growth is partly immediate (within a year). - A more direct and complete investigation of this hypothesis is left to future research. --- ### 8 Conclusion ### Summary of empirical approach - Study period and data: universe of bank-firm credit relationships over the period 1997-2013. - Methodology: estimate an additive growth rate model to separate credit demand from credit supply shocks; use estimated bank-level supply shocks and stickiness of lending relationships to build a measure of firm-specific shocks to credit supply. - Production model: build a model of production with heterogeneous credit constraints to estimate an industry-specific production function and isolate firm idiosyncratic productivity dynamics. ### Main findings - Firms connected to banks contracting their supply of credit acquire less inputs and produce less output than their competitors. - The effect on output is stronger than the effect on inputs, suggesting that productivity is affected by credit availability. - Credit supply boosts productivity growth; these effects are sizable, persistent, and robust. - Effects are stronger for smaller firms and for industries relying heavily on bank credit. - A negative credit supply shock produces much stronger effects than a positive one of the same magnitude; stability of credit supply matters in addition to quantity. - Financial turmoil can have a persistent effect on aggregate output because it depresses firms’ TFP in the short and long run. - Financial frictions harm beyond allocative efficiency. ### Activities stimulated by credit availability - Adoption of IT. - Sound management practices. - Export orientation. - Innovation. ### Additional conjecture and evidence - Reduction of credit supply might force borrowers (notably, managers and entrepreneurs) to consume time and energy to establish connections with additional lenders and consequently exert less effort in improving business performance. - Firms’ attempts to create new lending relationships are more frequent when they experience negative credit shocks. *Source: wpiea2019107 - 7.3 Management practices (PDF chapter/section) — canonical source URL: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019107.pdf* ### References ### References ### Key empirical findings (from figures and tables) - Credit supply shocks (φi,t) significantly affect firm outcomes: - A 1% increase in φi,t is the supply shock needed to increase the credit granted to firm i by 1% (definition repeated across tables). - Credit, inputs, outputs, and loan application responses to credit supply shocks (Table 2, All Industries): - Credit (delta Log): φi,t = 0.949*** (std. error 0.0196). - Value Added (delta Log): φi,t = 0.123*** (0.0162). - Net Revenues (delta Log): φi,t = 0.0474*** (0.0109). - Capital Stock (delta Log): φi,t = 0.0619*** (0.0128). - Wagebill (delta Log): φi,t = 0.0154* (0.00926). - Number of Employees (delta Log): φi,t = 0.00608 (0.00889). - Intermediate Inputs (delta Log): φi,t = 0.0220* (0.0114). - N. of Loan Applications (delta Log): φi,t = -0.537*** (0.0796). - Any Loan Received (delta Log): φi,t = -0.0780*** (0.0173). - Similar magnitudes for Manufacturing subsample (Table 2, Manufacturing): - Credit: φi,t = 0.966*** (0.0253). - Value Added: φi,t = 0.134*** (0.0201). - Net Revenues: φi,t = 0.0436*** (0.0143). - Capital Stock: φi,t = 0.0610*** (0.0169). - N. of Loan Applications: φi,t = -0.424*** (0.113). - Any Loan Received: φi,t = -0.0583** (0.0242). - Credit supply shocks and productivity growth (Table 3): - All industries, Value Added, Cobb-Douglas: φi,t = 0.0946*** (0.0155), Observations 656,960, R2 0.172. - All industries, Value Added, Trans-Log: φi,t = 0.109*** (0.0160), Observations 656,960, R2 0.185. - All industries, Net Revenue, Cobb-Douglas: φi,t = 0.0190*** (0.00477), Observations 656,960, R2 0.178. - All industries, Net Revenue, Trans-Log: φi,t = 0.0259*** (0.00491), Observations 656,960, R2 0.195. - Manufacturing subsample shows larger coefficients: - Value Added, Cobb-Douglas: φi,t = 0.115*** (0.0178), Observations 347,990, R-squared 0.186. - Net Revenue, Cobb-Douglas: φi,t = 0.0303*** (0.00595), Observations 347,990, R-squared 0.144. - Robustness (Table 4, Cobb-Douglas revenue productivity, All Industries): - Baseline: φi,t = 0.0190*** (0.00477), Observations 656,960, R-squared 0.178. - Firm Controls added: φi,t = 0.0248*** (0.00534), Observations 483,665, R-squared 0.184. - Alternative match controls, split-sample, 4-digits, endogenous exit, and cost-share methods produce consistent positive estimates; examples: - Split sample (column 7): φi,t = 0.0197*** (0.00503), Observations 656,960, R-squared 0.178. - Cost controls (column 10): φi,t = 0.0256*** (0.00736), Observations 545,162, R-squared 0.185. - Heterogeneity (Table 5, Cobb-Douglas revenue productivity): - Baseline All Industries: φi,t = 0.0208*** (0.00520), Observations 656,960, R-squared 0.179. - Interaction effects show variation: - Small lenders dimension interaction: Interaction = 0.0169* (0.00926). - Few lenders interaction (manufacturing): Interaction = 0.0380*** (0.0124). - High sectoral leverage interaction: Interaction = 0.0253*** (0.00951). - Manufacturing baseline: φi,t = 0.0391*** (0.00719), Observations 347,990, R-squared 0.144. - Exposure to interbank market and productivity growth (Table 6): - All Industries, Value Added, Cobb-Douglas: ITBKi,2006 = -0.0412* (0.0220), Observations 110,070, R2 0.112. - Manufacturing, Value Added, Cobb-Douglas: ITBKi,2006 = -0.0675** (0.0311), Observations 57,986, R2 0.134. - Net Revenue measures also show negative and significant ITBK coefficients in some specifications. - Credit supply and productivity-enhancing activities (Table 7): - φi,t positively associated with PCs per unit of capital: coefficient 0.672** (0.269), Observations 3,913, R2 0.967. - φi,t associated with No. of Patents: 0.163 (0.284), and Patent Applications growth: 0.305* (0.159) in some panels. - φi,t associated with Export Growth: coefficient 2.166* (1.116) in cross section (column 5) with Observations 2,868, R2 0.872. - φi,t−1 effect: coefficient 0.0418** (0.0195) in patent growth specification. ### Tables and descriptive statistics (Table 1 and figures) - Table 1 (Descriptive statistics, main firm-level variables, All Industries and Manufacturing): - Value Added (All Industries): Mean 5,312; Std. Dev. 33,819; Median 1,641; N 656,960. - Net Revenues (All Industries): Mean 27,073; Std. Dev. 156,638; Median 8,813; N 656,960. - Wagebill (All Industries): Mean 3,377; Std. Dev. 19,693; Median 1,062; N 656,960. - Capital Stock (All Industries): Mean 8,636; Std. Dev. 153,346; Median 1,545; N 656,960. - Intermediate Inputs (All Industries): Mean 21,888; Std. Dev. 137,390; Median 6,873; N 656,960. - Credit Granted (All Industries): Mean 7,924; Std. Dev. 3,644; Median 52,737; N 650,664. - Employees (All Industries): Mean 804; Std. Dev. 722; Median 865; N 6,960. - Notes: One observation is one firm for one year, between 1998 and 2013 (unbalanced panel). All variables (except number of employees) expressed as thousands of 2010 euros using sector-level deflators from national accounts. Source: CADS and Credit Register. - Figures summarize distributions and dynamics: - Figure 2: Histogram of credit supply shock and productivity growth; productivity estimated as residual from (log) revenues production function assuming Cobb-Douglas or Trans-Log; normal distribution superimposed. - Figure 3: Productivity before and after an unexpected credit supply shock with 99% confidence intervals; top panel all industries, bottom panel manufacturers; productivity estimated from (log) revenue production function; Cobb-Douglas or Trans-Log functional forms. - Figure 4: 3D view of Credit Supply Shock and VA Productivity for different Cobb-Douglas parameters ρ and βk; regression ∆ωi,t(ρ,βk) = ψi + ψs,t,p + γ·φi,t + ηi,t shown; one observation is one firm-year between 1998 and 2013. - Figure 5: Productivity growth in manufacturing by quintile of credit supply shock; third quintile normalized to zero; Cobb-Douglas and Trans-Log revenue productivity shown. - Figure 6: Revenue productivity before and after a credit supply shock - negative vs positive shocks. ### Mechanisms and channels (evidence from regressions and appendix) - Credit supply shocks affect real activity: - Positive φi,t increases credit granted and inputs (capital, intermediates) and is associated with positive productivity growth. - Negative association of φi,t with newly observed loan applications and any loan received suggests credit-tightening reduces access to new lenders. - Heterogeneous effects: - Manufacturing firms exhibit larger productivity responses to credit supply shocks than the full sample. - Firms with few lenders, high sectoral leverage, or operating in certain dimensions show heterogeneous interactions (Table 5). - Interbank channel: - Firms whose lenders had greater exposure to the interbank market in 2006 (ITBK i,2006) experienced negative productivity growth during 2007-2009 (Table 6). - Innovation and management channels: - φi,t correlates with higher PCs per capital, patent applications growth, R&D investment and export activity in some specifications (Table 7). - Management score (World Management Survey) included as outcome in cross-section specifications. ### Methodology and robustness (appendix A and B) - Construction of credit supply shock φb,t and φi,t: - φi,t = sum over lenders b in Bi,t−1 of φb,t · wcb,i,t−1 where weights wc are proportional to previous-period credit share. - A 1% increase in φi,t corresponds to the supply shock needed to increase credit granted to firm i by 1%. - Tests for substitution and match-specific shocks (Appendix A.1.1–A.1.2): - Augmenting first-stage with main substitute bank shock yields correlation ≈0.99 between φ estimates, suggesting limited bias from ignoring substitution. - Including match-specific controls (loan size relative to borrower or lender totals, interest rate, length of lending relationship, instrument type, past non-performing loans, collateral share) produces supply shocks with correlations above 94% with baseline, mitigating concerns about omitted match shocks. - Interest rates imputed for roughly a third of observations. - Drivers of bank-level supply shocks (Appendix A.2): - Banks more reliant on wholesale funding in 2007 (Loans/Deposits) decreased credit supply more during 2007-2009. - Better-capitalized banks (capital adequacy ratio) decreased credit supply less. - Banks increasing exposure to sovereign assets during 2010-2013 decreased credit supply to corporates. - Lenders’ M&A episodes reduce credit supply to pre-existing borrowers. - Production function estimation (Appendix B and B.1): - Revenue production function: Yi,t = exp{ωi,t} F(Li,t, Ki,t, Mi,t, β); value added analog provided. - Productivity decomposed into structural component ̃ωi,t and iid shock εY,i,t. - Control function approach (Ackerberg et al. 2015) used: estimate Ψ as Et[yi,t | li,t, ki,t, mi,t, ki,t−1, φi,t, cpip,t, Jt], approximate by third-order polynomial in xi,t plus year fixed effects. - Two functional forms for f(·): Cobb-Douglas (first order) and Trans-Log (second order); results are similar across both. - Production functions are industry-specific; sectors with fewer than 300 firm-year observations dropped. - Time discretization and input-specific credit constraints ΓM, ΓK, ΓL incorporated to allow materials, capital, or labor to be constrained differently by credit availability. - Estimation details and inference: - One observation is one firm for one year between 1998 and 2013 (unbalanced panel) unless otherwise noted. - Firm fixed effects and province×industry×year fixed effects included in main specifications; singletons dropped. - Standard errors are (two-way) clustered at firm and main-lender×year level in most regressions. - Significance notation: *** p<0.01, ** p<0.05, * p<0.1. - OLS computations use algorithms from Correia (2016). *Italic: Source: References and tables/appendices from the PDF content unit "wpiea2019107 - References".* ### section 4. Therefore, if one excludes credit supply shocks from the model, past inputs are correlated ### section 4. Therefore, if one excludes credit supply shocks from the model, past inputs are correlated with the productivity innovation, and there are no valid instruments to identify the parameter of interests. ### Results: production function elasticities and descriptive findings - Mean elasticity of value added to capital (to labor) for the whole economy: ≈.17 (≈.64). - Mean elasticity of value added to capital (to labor) for manufacturing: ≈.19 (≈.62). - Mean elasticity of net revenue to capital (to labor) for the whole economy: ≈.07 (≈.14). - Mean elasticity of net revenue to capital (to labor) for manufacturing: ≈.04 (≈.13). - Mean elasticity of net revenue to intermediate inputs: ≈.81 for both manufacturing and all industries. - Quantity-elasticities for manufacturing (computed using sector-level σ): ≈.05 for capital, ≈.17 for labor, and ≈1.06 for intermediate inputs. - Notes on estimation: averages are weighted for number of observations in the sample of the main specification. Revenue production function estimates can be translated into quantity production function parameters depending on the competitive structure of the product market. ### Alternative model and identification concerns - Critique addressed: inclusion of credit supply changes as factors of the level of the credit constraint. - Concern: researcher needs to observe exact value of credit supply shocks; the credit shifter estimated in section 3.1 may be a proxy. - Authors’ response: an alternative model (first-order log-linearization) was provided in a previous version; it offers a firm-specific estimate of productivity growth and produces qualitatively and quantitatively similar impacts of credit supply shocks to the baseline. ### Appendix C.1 — Revenue vs quantity productivity (TFP, TFPR and TFPQ) - Production setup: - Quantity produced: Q_{i,t} = exp{ω^q_{i,t} + f(l_{i,t},k_{i,t},m_{i,t},β^q)} = exp{ω^q_{i,t} + β^q_l·l_{i,t} + β^q_k·k_{i,t} + β^q_m·m_{i,t}}. - Quantity sold (CES demand): Q_{i,t} = (P_{i,t}/P_t)^{-σ} · exp{θ_{i,t}}. - Deflated revenues: Y_{i,t} = P_{i,t}·Q_{i,t}/P_t = Q_{i,t}^{σ−1/σ} · exp{θ_{i,t}}. - Log-linear revenue equation: - y_{i,t} = 1/σ · θ_{i,t} + (σ−1)/σ · ω^q_{i,t} + β_l·l_{i,t} + β_k·k_{i,t} + β_m·m_{i,t} with β_x = (σ−1)/σ · β^q_x. - Productivity growth decomposition: - ∆ω_{i,t} = 1/σ · ∆θ_{i,t} + (σ−1)/σ · ∆ω^q_{i,t}. - Implication: revenue-based productivity growth can reflect changes in technical efficiency (ω^q) or market appeal (θ). Product innovations mainly affect θ; process innovations mainly affect ω^q. - Main empirical specification restated: - ∆ω_{i,t} = 1/σ · ∆θ_{i,t} + (σ−1)/σ · ∆ω^q_{i,t} = ψ_i + ψ_{p,s,t} + γ·φ_{i,t} + η_{i,t} (equation (27)). - Identification concern highlighted: potential correlation between output demand (∆θ) and credit supply factors; evidence in sections 5.1 and 6 is described as reassuring. - Pricing channel note: with general inverse demand P_{i,t} = D(Q_{i,t},θ_{i,t},P_t), credit supply shocks could alter pricing incentives and measured productivity via ∆p_{i,t}. However, section 4 shows positive credit shocks increase input acquisition and, empirically, a positive effect of credit on productivity growth is observed. ### Appendix C.2 — Robustness to unobservables (Oster, 2016) - Implementation details: - Define R_un and γ_un (unrestricted regression with full FEs) and R_con and γ_con (restricted regression with only province and sector and year FEs). - Use Oster formula for approximated bias-adjusted treatment effect γ(δ,R_max) = γ_un − δ·(γ_con − γ_un)·(R_max − R_un)/(R_un − R_con). - Choices: R_max = 1; δ = 2 (conservative). - Bounding strategy: bounding set for γ constructed using γ_uc and γ(δ=2,R_max=1) as extreme points. - Result: bounding sets (Table A.4) never contain 0; authors conclude results on effect of credit shocks on productivity growth are robust to unobservable shocks. ### Appendix C.3 — Measurement error concerns - Acknowledged that inputs are measured with error in practice. - Rewriting of equation (10) emphasizes potential residual effect of credit supply shocks if inputs are mismeasured: - ∆y_{i,t} = ψ_i + ψ_{p,s,t} + ∆f(k_{i,t},l_{i,t},m_{i,t},β) + γ·φ_{i,t} + η_{i,t}. - Authors’ mitigating points: - Table 2 (inputs on the LHS) mitigates concerns that mismeasurement in inputs drives findings. - Classical measurement error on dependent variable worsens precision but not consistency. - Factor hoarding and adjustment costs could generate non-classical measurement errors (e.g., time-to-fire workers, capital utilization), but these stories imply short-term productivity loss; section 5.3 shows credit supply shock effects last for at least a few years, inconsistent with short-lived measurement artifacts. ### Appendix C.4 — Marginal Revenue Product of Capital (MRPK) - MRPK formula under Cobb-Douglas sales-generating production: - MRPK := ∂Y_{i,t}/∂K_{i,t} = β_k · exp{ω_{i,t}} · L^{β_l}_{i,t} · M^{β_m}_{i,t} · K^{1−β_k}_{i,t} (equation (28)). - Allocation intuition: in frictionless world marginal products equalized within industry; credit constraints can prevent efficient reallocation, generating within-industry MRPK dispersion. - Ambiguity: when credit constraints loosen, firms can acquire more capital (MRPK should decrease) but credit expansions also raise TFP and other inputs, so net sign ambiguous. - Empirical finding (Table A.3): credit availability leads to an increase of firm MRPK in the panel dimension; detected positive or null effect on MRPK growth depending on specification. This implies firms appear more hungry for capital when constraints get slacker and that exogenous credit availability affects marginal productivity by altering TFP. ### Appendix D.1 — Interbank shock: placebo and robustness tests - Pre-crisis test: estimating equation (15) for t∈[2004,2006] shows firms more exposed to the interbank freeze did not have statistically different productivity growth before the credit crunch (Table A.7, columns (1)—(4)). - Placebo collapse test: hypothetical interbank freeze in 2003 (t∈[2003,2005]) using ∆ω_{i,t} = ψ_{p,s,t} + γ·INTBK_{i,2002} + η_{i,t} yields non-significant predictors of subsequent productivity growth (Table A.7, columns (5)—(8)). - Business cycle sensitivity check: firm-specific sensitivity α_i to grGDP_t estimated from ∆y_{i,t} = ψ_i + α_i·grGDP_t + ε_{i,t} using pre-2006 years. No statistically significant correlation found between α_i and lenders’ reliance on the interbank market in 2006 (Table A.7, column (9)). - Balancing tests: normalized differences across quartiles of interbank dependence reported in Table A.2; authors do not reject homogeneous distribution of observable characteristics. ### Appendix D.2 — Interbank exposure and credit granted - Estimated relations for firms active in industries and provinces over t∈[2007,2009]: - ∆credit_{i,t} = ψ_{p,s,t} + γ·INTBK_{i,2006} + η_{i,t} - φ_{i,t} = ψ_{p,s,t} + γ·INTBK_{i,2006} + η_{i,t} - Empirical magnitudes (Table A.6): - An increase of dependence from the interbank market of 1% leads to a decrease of the growth rate of credit granted between a quarter and a fifth of a percentage point (see columns (2) and (4)). - Columns (1) and (3) show the credit supply shock measure φ_{i,t} responds negatively to interbank shocks. - Conclusion: firms more exposed to the collapse of the interbank market decrease credit received relative to peers in the same industry and location. ### Selected quantitative estimates from supplemental tables and robustness - MRPK regressions (Table A.3), all industries (Value Added, log MRPK): - φ_{i,t} coefficient: 0.258*** (standard error 0.0345). - Observations: 656,960. R-squared: 0.9150. - MRPK regressions, manufacturing (Value Added, log MRPK): - φ_{i,t} coefficient: 0.294*** (standard error 0.0442). - Observations: 347,990. R-squared: 0.9270. - Oster bounding sets for φ_{i,t} effect on productivity growth (Table A.4), All industries: - Cobb-Douglas Value Added: [0.043 ; 0.095]. - Trans-Log Value Added: [0.057 ; 0.11]. - Cobb-Douglas Net Revenue: [0.019 ; 0.066]. - Trans-Log Net Revenue: [0.026 ; 0.071]. - Exposure to interbank market and credit supply (Table A.6): - φ_{i,t} = ψ_{s,t,p} + γ·ITBK_{i,2006} + η_{i,t}: ITBK_{i,2006} coefficient examples: -0.144*** (std. err. 0.0278) and -0.164*** (std. err. 0.0257) across specifications. - ∆cred_{i,t} regressions: ITBK_{i,2006} coefficients: -0.192*** (std. err. 0.0400) and -0.241*** (std. err. 0.0545). - Observations range: 110,070; 108,267; 57,986; 57,349. R-squared examples: 0.191; 0.093; 0.197; 0.089. - Placebo tests (Table A.7) show non-significant ITBK_{i,2002} coefficients in placebo windows and no pre-trend differences in productivity growth for units later hit by the interbank freeze. - Baseline productivity effect and robustness (Table A.8), Cobb-Douglas Value Added productivity: - Baseline φ_{i,t} coefficient: 0.0946*** (std. err. 0.0155), Observations: 656,960, R-squared: 0.172. - Manufacturing baseline φ_{i,t}: 0.115*** (std. err. 0.0178), Observations: 347,990, R-squared: 0.186. - Robustness specifications (lagged controls, sample restrictions, pooled estimator, alternative measures, split-sample IV, 4-digit sector definition, endogenous exit) yield qualitatively similar positive estimates. - Trans-Log revenue productivity robustness (Table A.9), All industries baseline: - φ_{i,t} coefficient: 0.0259*** (std. err. 0.00491), Observations: 656,960, R-squared: 0.195. - Manufacturing baseline φ_{i,t}: 0.0323*** (std. err. 0.00649), Observations: 347,990, R-squared: 0.180. - Additional bank-level and M&A associations (Table A.5): - Example coefficient: (loans/deposits)_{b,2007} on bank-level Credit Supply: -0.0189*** (std. err. 0.00533). Observations: 1,635. R-squared: 0.086. ### Summary of core empirical inferences - Exogenous firm-level credit supply shocks (φ_{i,t}) are positively associated with subsequent firm productivity growth across specifications, functional forms (Cobb-Douglas and Trans-Log), and output measures (Value Added and Net Revenue). - Results are robust to Oster (2016) bounding analysis (R_max = 1, δ = 2), placebo tests around the interbank freeze, additional controls, alternative estimators, split-sample procedures, and checks for pre-trends. - Credit supply shocks also affect input acquisition and MRPK: credit expansions raise investment and are associated with higher MRPK, indicating that credit availability affects both input allocation and marginal productivity through TFP changes. - Exposure to the 2006 interbank market dependence predicts subsequent declines in credit granted during the 2007–2009 period and is associated with productivity declines during the credit crunch episode; placebo and balancing tests alleviate concerns of pre-existing differences. *Source: wpiea2019107 - section 4. Therefore, if one excludes credit supply shocks from the model, past inputs are correlated with the productivity innovation, and there are no valid instruments to identify the parameter of interests.* --- _Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019107.pdf_