Firm Dynamics and Productivity Eric Bartelsman* July 13, 2010 Warwick University *Vrije Universiteit Amsterdam; Tinbergen Institute;.IZA. 1 Overview • • • • Identification of policy effects Cross-country firm-level data From macro to micro and back Examples: – The role of allocation and selection – Endogenous choice of innovation strategy 2 Identification of Policy Effects • What is effect of policy – Directly on firm-level decisions (ceteris paribus) – On interactions in markets • Indirectly on firm-level decisions – On aggregate economy • What methods are available for identification? 3 Data for identification • Firm-level: (natural) experiments – Micro behavior can be studied (Ceteris paribus) • Aggregate (industry) level – Identification is difficult • Cross-country panels – have variation in policy changes 4 The challenge of cross -country analysis • Macro data – e.g. SNA, PWT – Difficult to identify effects (e.g. 2 million regressions) • Sectoral data – e.g. OECD-STAN; Unido – aggregate sectors obscure causal mechanism • Meta-analysis of results from micro studies – A challenge to control for data, method, and context – Little within-country variation in policy (e.g. before and after) – No GE effects • Multi-country longitudinal micro dataset – Generally not possible (disclosure) – EUROSTAT attempting to build EU panel From firm data to macro indicators Longitudinal Micro Data Single country Surveys, Business Registers • SC LMD N.A. Multiple countries National Accounts Industry Data Macro and Sectoral Timeseries • DMD EUKLEMS/ STAN/Unido Linked Indicators 6 Mismatch between theory and empirics • Growth Theory: (e.g. Romer, 1989) Firm-level behavior, applied to macro/sectoral data (e.g. growth accounting) • IO-style theory (R&D races, competition and growth, innovation input/output/effect-CDM) has macro-policy implications but is applied to micro datasets. • New work bridging the gap (Kortum-Klette, Melitz, Restuccia-Rogerson, MortensonPissarides) 7 Why linked firm-level indicators and aggregate-level analysis may help answer questions • Heterogeneous agents at micro level Diversity in firm-level strategies Frictions, uncertainty, expectations • Market selection Sales and input growth, conditional on productivity and economic ‘environment’ • Combination of firm-level productivity impact and market share evolution gives total impact on industry productivity 8 Firm-level, cross-country comparisons • Policy environment affects all firms in country (and industry) in the same manner • Cross-country firm-level comparisons may provide means to observe/identify the impacts of policy changes • Firm-level datasets generally cannot be combined across countries 9 Distributed Micro Data Research OECD sample Demographics (entry/exit) for 10 countries Productivity decompositions for 7 countries Survival analysis 7 countries World Bank sample Same variables, 14 of Central and Eastern Europe, Latin America and South East Asia EU Sample (10 countries), updates and a few new countries Productivity decompositions Sample Stats and correlations by quartile Eurostat/ONS Sample (13 EU countries), ICT-impacts project. Linked BR-PS-EC-(CIS) Data sources • Business registers for firm demographics – Firm level, at least one employee, 2/3-digit industry • Enterprise surveys – Production Statistics – R&D Surveys – Other firm-level surveys • Countries: – About 25-30 so far. Eurostat, OECD, WB • Data are disaggregated by: – industry (2-3 digit); – size classes 1-9; 10-19; 20-49; 50-99; 100-249; 250-499; 500+ (for OECD sample the groups between 1 and 20 and the groups between 100 and 500 are combined) – Time (late 1980s – early 2000s) Indicators Collected • Moments of productivity distributions – Std Deviation, quartiles • • • • Size distribution (employment) of firms Gross job flows Firm entry and exit Decompositions of productivity levels and growth – Olley-Pakes decomposition of productivity levels – Dynamic within/between decompositions including contribution of entry and exit 12 Network Researcher Using Distributed Micro Data Analysis Policy Question Research Design Program Code Publication Metadata Cross-country Tables NSOs Network members Provision of metadata. Approval of access. Disclosure analysis 13 Steps undertaken • • • • Build network (e.g. Eurostat countries) Collect meta-data Prepare firm-level datasets Do until – Run code – Analyse cross-country datasets – Clean firm-level datasets – Fix/change code • End 14 Availability of features in firmlevel data 15 Lessons learned • Countries with experience in firm-level require very little work to iterate • Countries which have ‘external’ micro-level users have better data quality and fewer problems • Countries where statistician/economist has computer skills have fewer problems • Investment in firm-level data preparation needed to participate in network has large payoffs to NSO – If knowledge, meta-data, and data cleaning are codified 16 Statistical Issues • How to harmonize across countries? • What about measurement error? • What are the properties of linked datasets – Assume that linking occurs perfectly through Business Register and unique firm-id • When to use sample (re-)weighting? • Who says ‘you can never be too thin…’? – Which indicators are fairly robust to sample size 17 Cross-country Comparisons • Harmonization – Sample frames; Variable definitions; Classifications; Aggregation Methods • Make comparisons that ‘control’ for errors – Exploit the different dimensions of the data (size, industry, time) – Use difference in differences techniques • Even in absence of measurement error, interpretation of cross-country indicators requires careful analysis Measurement Error • Three sources of error potentially affect comparability of indicators built from firm level data: – Classical Error of firm-level measure – Errors in observed firms (sample) – Method of Aggregation of Indicator Computation of sample re-weight 20 Results of data linking analysis • Info on first and second moments of PS and linked PS-EC are reasonably robust • Sample reweighting of linked PS-EC gives results that are similar to weighted PS • Regression coefficients are robust to smaller set of obs in linked sub-sample 21 Descriptive Statistics • Firm size • Firm demographics: • Employment and # of firms for entry, exit, continuers: by industry and size class • Firm survival : • Employment and # of survivors, by cohort, industry, year • Static and dynamic analysis of allocative efficiency: • • • Decompositions of productivity (entry/exit/continuer) Higher moments, covariances, means by quartile In presentation, focus on 2 and 4 Interpretation of Gross Turnover • Theoretical explanations – Entry explained by ‘push’ and ‘pull’ factors – Exit barriers may effect characteristics of exiting firm more than number of exits • Measurement errors – Conceptual differences in measure (e.g. labor) – Differences in underlying data sources Evidence of firm turnover Total business sector, firms with at least 1 employee Total business sector, firms with at least 20 employees • No major differences across OECD countries, especially after controlling for sector and size effects • But large differences in size at entry • Large net entry in transition economies: filling the gaps (?) Gross and net firm turnover: how the time dimension sheds light on the evolution of market forces in transition economies Entry rate by size: how the size dimension may shed light on the nature of firm dynamics • Monotonic decline in entry rate by size in US • Less clear link between size and entry rate in other EU countries; • Any role for entry costs ? Evolution of Average manuf firm-size (top quartile) Estonia Latvia 160 140 120 100 80 60 40 20 0 1994 1996 1998 2000 year 3rd quartile 2002 2004 2006 Average Firm Size Manuf Average Firm Size Manuf 160 140 120 100 80 60 40 20 0 1995 Average Firm Size, Manuf Average Firm Size, Manuf 100 80 60 40 20 0 1990 1992 year 1994 2003 2005 2nd quartile Sweden 120 1988 2001 3rd quartile 2nd quartile 140 1986 1999 year Mexico 1984 1997 1996 1998 2000 2002 90 80 70 60 50 40 30 20 10 0 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 year 27 2nd quartile 2nd quartile Overview • • • • Identification of policy effects Cross-country firm-level data From macro to micro and back Examples: – The role of allocation and selection – Endogenous choice of innovation strategy 28 Policy Example: Boosting Innovation • How does policy affect the economy at micro and macro level • How to model effects • How to take model to the data 29 Optimal Innovation Innovation, productivity and growth from micro to macro • The idea of creation of diversity (supply) and selection (demand) to model innovation is quite useful. • However, two problems: – Outcome of evolutionary process has no ‘value’. There is no mechanism to decide whether we need to intervene in process, and how. – Creation side has ‘forethought’ (Prometheus). Feedback from envisioned future outcome matters (‘General Equilibrium’, Animal Spirits) Optimal Innovation An aside: Innovations and taste Time Too much of a good thing? • Slippage between individual incentives and aggregate needs – Spillovers – Stepping on toes – Business Stealing • Non-transitivity – New, in cycles…. • Network externalities – Spillovers on demand side, learning/demonstration effects • Path Dependent traps – How to get rid of the internal combustion engine? Cites: Solow, Romer; Jones and Williams Innovative Activity, Knowledge Growth and Impact Innovative input Innovative output Innovative impact Innovative Activity, Knowledge Growth and Impact Crepon, Duguet and Mairesse Policy levers in Innovation Human capital formation R&D subsidies Spillovers, adoptive capacity Labor and capital markets Micro to macro Micro choices Market Selection - Entry/Exit - Innovation Strategy - Factor inputs - Product output - Competition - Policy Environment - Tech. environment Macro impact of innovation 38 Innovation and Market Interaction Policy Levers • Allocation of resources across firms • Demand and supply conditions affect firm-level input decisions • Selection • Factors affecting Entry and exit decisions • Choice of innovation strategy • Intangible investments • Technology adoption 40 Firm-level, cross-country comparisons • Policy environment affects all firms in country (and industry) in the same manner • Cross-country firm-level comparisons may provide means to observe/identify the impacts of policy changes through the three paths: • Allocation of resources across firms • Selection (entry/exit) • Choice of innovation strategy (externalities, intangibles) 41 Some micro evidence that: Source: simulated data Source: Erik Brynjolfsson Results: ICT Impacts Project If micro correlation were negative, macro effect would still be positive, if: and Cite: model calibration from Bartelsman, Haltiwanger and Scarpetta : Cross Country Differences in Productivity: The Role of Allocative Efficiency Eric Bartelsman, John Haltiwanger and Stefano Scarpetta* *Vrije Universiteit Amsterdam, Tinbergen Institute and IZA; University of Maryland, NBER, and IZA; OECD and IZA. 47 Motivation • Recent work bringing together two separate literatures: – Large and persistent differences in productivity across countries – Large and persistent differences in productivity within industries within countries • Are these connected? • Potentially – misallocation hypothesis – Distortions impact allocation of activity within industries – Our take – distortions impact size/productivity relationship within industries • Size/productivity positive covariance a core prediction of canonical models of firm heterogeneity and size distribution 48 Recent Literature • Misallocation hypothesis not new – Banerjee and Duflo (2003) discuss literature in handbook paper • New firm level datasets permit rich analysis. • Tractable model recently developed by Restuccia and Rogerson (2008) • Implemented empirically for China, India, U.S. by Hsieh and Klenow (2009) – Focus on comparisons of dispersion in productivity • Our contribution: – Multi-country, multi-moment evidence and analysis – Suggests more robust moment theoretically and empirically is size/productivity covariance – Distortions impact selection (who produces) as well as allocation among existing producers (Hsieh and Klenow focus on latter). 49 What are these distortions? • Working conjecture: – Emerging and transition economies have market structure and institutions that distort allocative efficiency • Distortions: – Restrictions on competition – Subsidies (explicit) or quotas/rationing of allocation to insiders /incumbents/favored businesses – Credit constraints to young and small businesses – Bribes and corruption (unevenly applied) – Doing Business inefficiencies: » Costly to start up a business » Costly to adjust employment » Poor or inefficient infrastructure (telephones, roads, electricity) – Again this hypothesis not new, but can now look at in richer ways with firm level datasets 50 Harmonized cross-country moments from firm level datasets • Measure STD(TFPR), STD(RLP) and Olley-Pakes (1996) decomposition within industries for 8 countries (U.S., Western European and Transition Economies). – TFPR (revenue TFP), RLP (revenue labor productivity), OP covariance – Full dataset has 20+ countries • Not all of the measures we focus on within industries are available (in particular, STD(TFPR)) 51 52 Covariance Between Size and Productivity? • Olley and Pakes (1996) static decomposition: The first term is the unweighted average of firm-level productivity The second term (OP cross term) reflects allocation of resources: do firms with higher productivity have greater market share. OP (1996) showed second term increased rapidly in U.S. telecommunications equipment industry after deregulation Second term is summary measure of covariance between size and productivity By construction, cross term takes out country effects in productivity levels, so abstracts from some aspects of measurement error We use log based measures of productivity in our implementation Decomposition relatively easy to compute (e.g., does not require longitudinal linkages) 53 54 55 56 57 Evidence from cross-country firm -level data • Large dispersion of productivity across firms within narrowly defined sectors. • STD(TFPR)<STD(LPR) – a puzzle in simple models • No systematic relationship in magnitudes across countries • Size and productivity distributions tend to be positively related especially in U.S. and Western Economies – Systematically lower in transition economies • Over time, dispersion moments don’t change much but covariance moments increase in transition 58 A Model of “Mis”-Allocation (Based on Restuccia and Rogerson (2008) (and similar to Hsieh and Klenow (2009)) • Households maximize utility, supply labor inelastically • Firms maximize profits – Ex ante, firms do not know productivity or distortion but know joint distribution. – Pay entry fee, learn their permanent draws of productivity and distortions – Productivity and distortion draws are firm-specific and have a permanent time-invariant component – Add two key features: • Overhead labor • Subject to transitory productivity and distortion shocks but following frictions: – Each period must make decision on how much capital and whether to produce/exit before observing transitory shock – Capital is quasi-fixed, labor absorbs transitory shocks – Very low (permanent) productivity/high distortion firms don’t produce – Size distribution due to idiosyncratic shocks, differentiated products, decreasing returns and other frictions – Dispersion in LP (correlated with TFP) due to interaction of overhead labor, quasi-fixed capital and transitory shocks 59 A Model of “Mis”-Allocation (Based on Rogerson and Restuccia (2007) (and similar to Hsieh and Klenow (2007)) Consumers supply labor inelastically and maximize utility: Firms maximize profits where: Our innovations: Overhead labor, quasi-fixed capital and transitory shocks Yields dispersion in TFPR and LPR even in absence of distortions 60 Entry/Selection Ex Ante Joint Distribution Exogenous probability of exiting in each period given by λ 61 Aggregate Relationships and Steady State Equilibrium Resources expended on entry/exit impact consumption and welfare Free entry condition and equilibrium in labor market 62 Calibration • Chosen parameters (based on literature) • Calibrated parameters: – f = overhead fraction; ce = entry cost; variance of permanent and transitory productivity shocks • Targets: – Moments from US from Table 1, and US firm-survival rates • 55 percent of U.S. Entrants survive after 5 years 63 Calibration (2) • We manage to match OP covariance and survival exactly, and get quite close on LPR and TFPR dispersion in US – Can’t match all moments exactly due to nonlinear relationship between parameters and moments • Overhead labor is crucial in matching OP covariance and getting Std(RLP)>STD (TFPR) 64 Non distorted economy responsiveness to overhead labor 65 Relationship Between Labor Productivity and Employment: No Institutional Distortions, Permanent and Transitory Shocks, Quasi-fixed capital 66 Relationship Between Productivity and Employment: Correlated Scale Distortions, Permanent and Transitory Shocks, Quasi-fixed capital 67 68 69 All of these effects are missing in Hsieh and Klenow 70 Using distortions to match OP-gap 71 Observations • Size-productivity relationship is strong in advanced economies and and has increased substantially over time in transition economies • In a model with frictions, shocks and idiosyncratic distortions we can match cross-country patterns of size-productivity relationship from firm-level data sets. • Transitory shocks, quasi-fixed capital interacting with overhead labor yields substantial labor productivity dispersion and OP(LP) gap • Idiosyncratic distortions lower consumption through selection (survival) and allocation effects • Distortions that are correlated with productivity are particularly bad for OP-gap, productivity, and welfare • While distortions should increase dispersion in TFP, the model does not match actual data – Measurement of TFP difficult – Maybe other factors (endogenous TFP dispersion, transitory dynamics) are important • For example, large differences in productivity dispersion by firm size and firm age – firm size and age distributions differ across countries for a number of reasons 72 Final Remarks • Misallocation has potential for accounting for (some) variation in productivity and consumption per capita • Are there summary measures that can serve as diagnostics? – OP decomposition shows one potentially important margin: size/productivity covariance – This is not the only relevant margin but still an important margin – Impact of distortions on selection important to take into account – Productivity dispersion is a less robust moment empirically and theoretically 73 Endogenous Productivity Dispersion • In Hsieh and Klenow, dispersion in productivity is sign of misallocation • High dispersion could be a sign of adoption of risky technology • Policy environment may distort choice of innovation and/or technology adoption 74 Source: Erik Brynjolfsson Interquartile range of firm growth distribution (pct-points) ICT and variability of outcomes Percentage of workers with broadband access Source: Eurostat ICT-Impacts Project 76 Impact of Policy on ICT and Productivity • • Firms have capabilities and desire to ‘try’ new ways of meeting market demand Since mid-1990s, this experimentation is often through ICT Firms’ desire to experiment depend on carrot and stick Leveraging of successful investment through scale increase is an enormous carrot • Policy affects resource reallocation And thus indirectly firm’s choice for innovation strategy Choice of Innovation Strategy • Experimental vs incremental/follower strategy. – Follower payoff: π – Experimenter payoff: Π • Experiment payoff with probability: p. • Relative payoff π/Π depends on uncertainty: – Π(p) with Π’<0. • When experiment fails: reconfigure, try again – Partial exit costs PEX. (Firms continue to experiment) – Total exit cost TEX (Firms give up experimentation) • Exogenous exit hazard x for all producing firms 78 Payoffs for Experimentation pΠ p (1-p)p(Π-Px) p 1-p 1-p (1-p)2(-Px-Tx) 79 Testable implications • With more experimentation average productivity is higher • With higher exit costs, experimentation is lower, especially at frontier • So: – Exit costs lower productivity more in those sectors where potential gain from experimentation is higher – Exit costs lower experimentation, and more so near frontier 80 Index of Employment Protection (source: OECD) 81 Empirical specification • Main regression: TFP effect of exit costs • Frontier indicators (by industry for US or UK): – Top quartile productivity relative to mean – Standard deviation of productivity – Adoption of Broadband 82 Exit costs and productivity Dependent var: Log(VA) Log(VA) Log(VA) Regressor: Log: Kit,Knit,Hours *** *** *** EPL .47 .34 .46 (0.02) (0.14) (0.19) --- -1.18 -1.13 (3.07) (3.08) EPL x Rank Rank variable --- Top quartile prod. relative to mean Broadband-use Num. obs. 7032 6790 7031 R-sq .97 .97 .97 Innovation strategy and employment • Mortensons-Pissarides type model with 2 sectors – 1: safe sector, known technology – 2: risky sector, draw from prod. distribution • With firing costs, option of closing down conditional on bad draw is more expensive – So, fewer jobs created in risky sector 84 Exit costs and employment Dependent var: Labor share in sector Labor share in sector Regressor: EPL .02 .02 (0.74) (0.74) -0.82 -0.84 (10.30) (10.55) Rank variable Top quartile prod/mean Broadband-use Num. obs. 5518 5518 R-sq .84 .84 EPL x Rank Exit costs and productivity • Productivity is reduced in industries that have potential gain from experimentation • Employment share in innovative industries is reduced • => High exit costs lower aggregate productivity 86 A Research agenda for innovation and productivity using micro and macro data • Discover incentives, at decision making level, of incentives of supply and use of innovative output. • Look at costs and benefits. Take into account expectations (General Eq. effects) • Look at interactions between actors • Understand process of knowledge accumulation • Depreciation, obsolescence • Spillovers, diffusion, appropriability, non-exhaustiveness • Individual vs global knowledge stocks • Estimate impact of knowledge stock on output/ welfare • Non-transitivity of ‘newness’ • Costs of churn 87