Turbulence, Firm Decentralization and Growth in Bad Times

What is the optimal form of firm organization during “bad times”? We present a model of delegation within the firm to show that the effect is ambiguous. The greater turbulence following macro shocks may benefit decentralized firms because the value of local information increases (the “localist” view). On the other hand, the need to make tough decisions may favor centralized firms (the “centralist” view). Using two large micro datasets on firm decentralization from ten OECD countries and US administrative data, we find that firms that delegated more power from the Central Headquarters to local plant managers prior to the Great Recession out-performed their centralized counterparts in sectors that were hardest hit by the subsequent crisis. Using direct measures of turbulence based on product churn and stock market volatility, we show that the localist mechanism dominates. This conclusion is robust to alternative explanations such as managerial fears of bankruptcy and changing coordination costs. Although delegation is better suited to some environments than others, countries with more decentralized firms (like the US) weathered the 2008-09 Great Recession better: these organizational differences account for about 15% of international differences in post-crisis GDP growth.


Introduction
What makes rms more resilient to large negative macro shocks? A recent literature has focused on rms' technological, nancial and governance structures as possible factors aecting their ability to cope with sudden changes in external conditions 1 , but much less is known about the role of rm organization. This paper focuses on how a specic organizational aspect of a rm: the extent to which decision-making is decentralized down from headquarters to plant managers, aects performance during an economic crisis.
This has particular relevance following the Great Recession of 2009-08, which generated a debate over how best to organize for recovery and survival. One common argument is that centralized rms were best equipped to survive the recession because of the importance of cost cutting which, due to conicting interests within the rm, is best directed from corporate headquarters (the centralist view). An alternative localist view is that recessions are periods of rapid change, and being decentralized provides rms with the necessary exibility to respond to turbulent business conditions. 2 To investigate these issues, we created two new panel datasets with explicit measures of decentralization measured prior to the Great Recession. One dataset, the World Management Survey (WMS) has rm level data across ten OECD countries (France, Germany, Greece, Italy, Japan, Poland, Portugal Sweden, the UK and US). The other dataset, the Management and Organizational Practices Survey (MOPS), is an establishment level dataset covering U.S. manufacturing plants.
3 which we constructed in partnership with the Census Bureau. We combine these datasets with rm and plant performance data before and after the 2009-08 crisis.
In order to guide the empirical analysis, we develop a model of rm decision-making building on the Aghion and Tirole (1997) approach. The model illustrates two countervailing eects of decentralization on rm performance during a crisis. On the one hand, a crisis is more likely to reduce the level of congruence between the CEO and the plant managerfor example, tough decisions on closing down projects and laying o sta will be resisted by local managers. At the same time, however, a crisis can also increase turbulence/uncertainty, thus making local information more valuable. In this case, a crisis would actually make decentralization more attractive, since it enables rms to better understand and adapt to the turbulent business environment. This result is akin to those emerging from a wide class of models where higher turbulence and uncertainty increase the value of local knowledge and the benets of decentralization. The net eect of 1 For example, see Aghion, Askenazy, Berman, Cette, and Eymard (2012) on technology; Chodorow-Reich (2014) on nancial structure and Alfaro and Chen (2012) or D'Aurizio, Oliviero and Romano (2015) on governance.
2 Support for these starkly dierent views can be found, for example, in reports by the Economist Intelligence Unit. In the depths of the Great Recession in June 2009 they wrote in favor of centralization during the crisis: Firms should be centralizing their decision-making processes. [...] In a recession investments and other decisions are scrutinized more carefully by senior management and a greater emphasis is placed on projects that provide benets across the enterprise rather than individual units. Yet three months later in August the same publication supported decentralization: Companies have to deal with dramatically more uncertainty, complexity and ambiguity in the current recession. Success does not come from centralization. True exibility arises when those who are closest to customers are empowered to respond to constant shifts in demand, preferences and attitudes.
3 We use the terms establishment and plant interchangeably throughout. 2 decentralization on rm performance is thus theoretically ambiguous.
We then turn to the empirical analysis and nd compelling evidence that, in sectors that were exogenously hit harder by the crisis, decentralized rms outperformed their centralized rivals in terms of survival chances as well as in their growth of sales, productivity and prots. We use several measures of the shock, including changes in trade patterns (exports in an establishment's industry by country cell) and a pre-recession measure of product durability to measure the shock (durable goods industries suer more in recessions as consumers can postpone purchases). Our ndings are robust to placebo tests, and a wide range of controls. Consistent with our model, we show that our empirical results are driven by the fact that the industries which had the most severe downturns during the Great Recession also had the largest increase in turbulence.

4
To show this, we employ a novel industry level measure of turbulence, the rate of new product additions and subtractions (product churn), which we built from the US Census of Manufactures ten digit product data. As shown in Bernard and Okubo (2015), product churn rises sharply during recessionsin a crisis establishments both destroy more existing products and also create more new products. 5 Using this measure on the US Census sample, we nd that decentralization signicantly protected establishments from the downturn in industries which had a bad shock, and an increase in product churn. We validate these results using an alternative measure of turbulence based on the stock market volatility, which is available for the international WMS sample as well as the US MOPS.
Alternative explanations of our results based on reduced agency problems, nancial conditions, lower coordination costs, omitted variables and other factors do not seem so consistent with the data. Finally, although organizational change is slow (we show evidence of large adjustment costs), rms subject to large negative shocks appear more likely to decentralize.
Overall, our paper suggests that the internal organization of rms may serve as an important mediating factor through which macroeconomic shocks aect rm performance and, ultimately, growth.
Our paper builds on an extensive prior literature. The benets of exploiting local knowledge harks back to a classic economic debate over economic systems between Lange (1936) and Von Hayek (1945). Lange argued that a centralized socialist economy would outperform a decentralized market economy, partly because the central planner could co-ordinate better, for example by setting prices to internalize externalities. By contrast, Hayek argued that it was impossible to aggregate all the local knowledge of agents, and it was both more ecient (and just) to allow individuals to make their decentralized choices based on the their local information. Modern organizational economics builds upon these trade-os within a rm rather than across 4 Bloom, Floetotto, Jaimovich, Saporta-Eksten and Terry (2016) shows a large variety of datasets that suggest that turbulence and uncertainty rise in downturns. 5 Contrary to Bernard and Okubo (2015), Broda and Weinstein (2010) report a pro-cyclical product churn. However, they have a a very dierent focuslooking at the net change in the product oering in retail stores (the number of new bar code products sold less current products no-longer sold)and a dierent time period (1994 and 1999-2003) spanning one mild recession.
In contrast, our measure is gross product churn (new products plus dropped products), is built on manufacturing establishment production data, and spans 15 years from 1997-2012, exploiting aggregate and industry variation.
3 the economy as a whole. On the theory side, our paper relates to the literature on decentralization within the rm (see Gibbons, Matouschek andRoberts, 2013, or Garicano andRayo, 2016 for recent surveys) and incomplete contracts (see Aghion, Bloom andVan Reenen, 2014 for surveys).
In particular, Hart and Moore (2005) analyze the optimal allocation of authority in multi-layer hierarchies. Dessein (2002) analyzes how the allocation of control can help incorporate the agent's information into decision-making in a situation where the agent has private information. 6 Our paper also relates to the existing empirical literature on decentralization and its determinants.
For example, Rajan and Wulf (2006) and Blundell et al. (2016) document a movement towards atter organizations and decentralized rms in the US and UK respectively. Caroli and Van Reenen (2001) and Bresnahan, Brynjolfsson and Hitt (2002) point at positive correlations between decentralization and both human capital and information technology. Guadalupe and Wulf (2009) argue that the Canadian-US Free Trade Agreement (FTA) in 1989 constitutes an exogenous increase in competition for US rms in the industries where taris were removed and this caused greater delayering and decentralization. Closest to our analysis is Acemoglu et al. (2007), whose model assumes rms can learn about the outcome of an investment decision from observing other rms. Hence, in sectors with more heterogeneity/turbulence or where the rm is closer to the performance frontier (so that learning is more limited) decision-making control should be more decentralized. In the contract literature, Prendergast (1982) suggested that the puzzle of performance pay in uncertain and turbulent environments (where higher risk should make the agent less willing to accept a high-powered contract) could be because of the need to exploit local information more eectively. Similarly, in the rm boundaries literature, Lafontaine and Slade (2007) also suggest that a similar puzzle over the lack of a negative impact of turbulence on franchising (vs. direct control), could again be related to the need to exploit the franchiser's superior local knowledge, which is more important in such environments. None of these papers, however, look at the interplay between rm decentralization, shocks and turbulence which is the center of our analysis.

7
The paper is organized as follows. Section 2 presents the model, Section 3 the data and methodology and Section 4 establishes our main empirical nding that in times of crisis decentralized rms outperform their centralized counterparts. Section 5 considers extensions, showing that volatility seems to matter rather than other mechanisms such as changing levels of congruence and Section 7 concludes. 6 In contrast to Aghion and Tirole (1997), there is no information acquisition eort by the agent or the principal, therefore in Dessein's model the allocation of authority is not so much a tool to motivate the agent (as in Aghion and Tirole) or give a supplier incentives to make relationship specic investments (as in Grossman and Hart, 1986). A key insight in Dessein (2002) is that in a world with asymmetric information and contractual incompleteness, the delegation of authority from a Principal to an Agent is often the best way to elicit the agent's private information.
7 Bradley et al (2011) report a positive relationship between rm independencewhich they interpret as a proxy for greater autonomy in resource allocation decisionsand rm survival during downturns using Swedish data.

2 A simple model
To guide the empirical analysis of the relationship between rm performance and decentralization during a crisis, we develop a simple model based upon Aghion and Tirole (1997). The key idea is that there is a trade o between incentives and local information. Possible misalignment of interests between the CEO and the plant manager favors centralization. But the plant manager is likely to have better local information than the CEO, which is a force for decentralization. A negative shock may aect the returns to decentralization in two, opposite, ways. First, it may reduce the benets of decentralization, by increasing the possible misalignment of interests between the CEO and the plant manager. Second, it may increase the benets of decentralization by increasing the informational asymmetry between the CEO and the plant manager, and thus the value of local information.

Basic set up
We consider a one-period model of a rm with one principal (the CEO/central headquarters) and one agent (the plant manager). The CEO cares about the protability of the business, 8 whereas the plant manager wants to maximize private benets and is not responsive to monetary incentives. 9 Taking an uninformed action involves potentially disastrous outcomes, thus an action will be taken only if at least one of the two parties is informed. Also, the agent obtains private benets only if the rm remains in business. There are n ≥ 3 possible actions (or projects) and at any point in time only two of them are "relevant", i.e. avoid negative payos to the parties. Among these two actions, one maximizes monetary protability, one maximizes the agent's private utility. Other actions lead to very negative payos to both parties.
With ex ante probability α the agent's preferred action (conditional upon the rm remaining in business) will also be the action that maximizes prots (or monetary eciency); this variable α captures the degree of congruence between the principal's preferences and the agent's preferences. If preferences coincide, then the action that maximizes the private utility of the agent also yields monetary utility B to the principal; if preferences do not coincide, the action that maximizes the agent's private utility yields monetary payo B − k to the principal.
Informational assumptions: We assume that the principal knows about project payos with probability p, but does not know directly which action the agent actually performed. On the other hand, the agent is assumed to be perfectly informed about the project payos.
Turbulence: Suppose that the principal can obtain an early signal of forthcoming performance, e.g. a current realization of income, at some cost C, and can then possibly decide to re the agent if she believes that the signal is due to the agent's choosing a non-prot maximizing action. In the absence of turbulence, the signal reveals the bad action choice perfectly. But the higher the degree of turbulence, the more dicult it is for the principal to infer action choice from performance.
Thus, suppose that current performance is given by y = a + ε where a ∈ {a 1 , a 2 } denotes the agent's action choice (e.g. a decision whether or not to introduce a new product 10 , with a 1 < a 2 and ε is a noise term uniformly distributed on the interval [−u, u].

Solving the model
Suppose that the plant manager takes the non-prot maximizing action a 1 (e.g. a decision which delays the introduction of a new product). The CEO will infer the action choice from observing the signal realization: if and only if y ∈ [a 1 − u, a 2 − u)∪ (a 1 + u, a 2 + u] and then can correct it if she has control rights, i.e. under centralization.
By Bayes' rule the probability of the CEO guessing the action choice is: that is: The probability of guessing the correct action is clearly declining in the amount of noise parameterized by u. Hence, the probability that the prot-maximizing action will be taken eventually under centralization ( Ω ), is equal to: where P is the probability that the principal acquires the information about projects payos.
10 Equivalently, this could be whether to drop an existing product from the portfolio or to make an investment in marketing or sales that enhances the product's value to the consumer. The key thing is that the decision has to have some irreversibility.

2.3 Centralization versus decentralization
The ex ante CEO's payo under decentralization, is equal to: The ex ante CEO's payo under centralization (i.e. if the CEO delegates no authority to the plant manager), is equal to: The net gain from centralization is then given by:

Two countervailing eects of a bad shock
We think of a bad shock as reducing congruence between the principal and the agent. For example, the negative shock may require cost-cutting such as closing down plants and laying o employees which will be particularly resisted by middle managers. More formally, to the extent that the principal has invested wealth in the project whereas the agent is subject to limited liability. In other words, a bad shock is likely to increase k. For a given level of uncertainty u, this will make centralization more attractive as: There is, however, also much evidence (summarized in Bloom, Floetotto, Jaimovich, Saporta-Eksten and Terry, 2016) that negative macro shocks are usually associated with greater turbulence and uncertainty (i.e. a higher u). This makes centralization less attractive, since: If the level of turbulence does not change after the occurrence of a bad shock, the overall eect of a bad shock is to make centralization unambiguously more attractive. However, if uncertainty increases with a bad shock and k does not change, the bad shock makes centralization become less attractive. Hence, the impact of a bad shock is theoretically ambiguous. We will investigate empirically which eect dominates, and then return to the theory to help inform potential mechanisms.

3 Data Description and Measurement
We start by describing in some detail our decentralization data since this involved an extensive new survey process. We then describe the accounting and administrative data matched with the survey-based measures of decentralization and the proxies measuring the severity of the Great Recession. We describe our measures of turbulence in Section 5 when we discuss theoretical mechanisms. More details on the data are in online Appendix A.

Decentralization
Cross-country data: World Management Survey (WMS) Our international decentralization data was collected in the context of the World Management Survey (WMS), a large scale project aimed at collecting high quality data on management and organizational design across rms around the world. The survey is conducted through an interview with a plant manager in medium sized manufacturing rms.
We asked four questions on decentralization from the central headquarters to the local plant manager.
First, we asked how much capital investment a plant manager could undertake without prior authorization from the corporate headquarters. This is a continuous variable enumerated in national currency that we convert into dollars using PPPs.
11 We also inquired on where decisions were eectively made in three other dimensions: (a) the introduction of a new product, (b) sales and marketing decisions and (c) hiring a new full-time permanent shop oor employee. These more qualitative variables were scaled from a score of 1, dened as all decisions taken at the corporate headquarters, to a score of 5 dened as complete power (real authority) of the plant manager. In Appendix Table A1 we detail the individual questions in the same order as they appeared in the survey. Since the scaling may vary across all these questions, we standardized the scores from the four decentralization questions to z-scores by normalizing each question to mean zero and standard deviation one. We then average across all four z-scores and then z-score the average again to have our primary measure of overall decentralization. In the same survey we collected a large amount of additional data to use as controls, including management practice information following the methodology of Bloom and Van Reenen (2007) and human resource information (e.g. the proportion of the workforce with college degrees, average hours worked, the gender and age breakdown within the rm).
We attempt to achieve unbiased survey responses to our questions by taking a range of steps. First, the survey was conducted by telephone without telling the managers they were being scored on organizational or management practices. This enabled scoring to be based on the interviewer's evaluation of the rm's actual practices, rather than their aspirations, the manager's perceptions or the interviewer's impressions.  Table A1). Second, the interviewers did not know anything about the rm's nancial information or performance in advance of the interview.
12 Consequently, the survey tool is double blind -managers do not know they are being scored and interviewers do not know the performance of the rm. These manufacturing rms (the median size was 250 employees) are mostly privately held and too small to attract coverage from the business media.
Third, each interviewer ran 85 interviews on average, allowing us to remove interviewer xed eects from all empirical specications. This helps to address concerns over inconsistent interpretation of responses.
Fourth, we collected information on the interview process itself (duration, day-of-the-week), on the manager (seniority, job tenure and location), and on the interviewer (for removing analyst xed eects and subjective reliability score). These survey metrics are used as noise controls to help reduce residual variation.
We decided to focus on the manufacturing sector where productivity is easier to measure than in the non-manufacturing sector. We also focused on medium sized rms, selecting a sampling frame of rms with between 50 and 5,000 workers. Very small rms have little publicly available data. Very large rms are likely to be more heterogeneous across plants. We drew a sampling frame from each country to be representative of medium sized manufacturing rms and then randomly chose the order of which rms to contact.
Each interview took an average of 48 minutes and the main wave was run in the summer of 2006. We achieved a 45% response rate, which is very high for company surveys, because (i) the interview did not discuss rm's nances (we obtained these externally); (ii) we had the written endorsement of many ocial headquarters (coded as 1), or both at this establishment and at headquarters (coded as 2). There were ve choices for the question on autonomy in capital investments, starting with Under $1,000 (coded as 1) up until $1 million or more (coded as 5). Each of these six questions was then z-scored, and then averaged, and then z-scored again. The survey also included management practice questions and some background questions on the establishment and respondent. 14 Table A2 shows how our various samples are derived from the universe of establishments.

Accounting data
Cross-country WMS data We build rm level measures of sales, employment, capital, prots and materials using accounting data extracted from Bureau Van Dijk's ORBIS. These are digitized versions of company accounts covering very large samples (close to the population in most of our countries) of private and publicly listed rms. In our baseline specications we estimate in three-year (annualized) growth rates.
We are able to build rm level measure of sales growth for at least one year for 1,330 out of the 2,351 rms with decentralization data in 2006.
U.S. MOPS data In addition to our decentralization data, we also use data from other Census and non-Census data sets to create our measures of performance (growth in sales, productivity, and protability).
We use establishment level data on sales, value-added and labor inputs from the ASM to create measures of growth and labor productivity. As described in more detail in Appendix A, we also combined the plant-level capital stock data from the Census of Manufactures with investment data from the ASM and applied the perpetual inventory method to construct annual capital stocks. Finally, we measure plant protability using prots as a percent of capital stock, with plant-level prots dened as sales less total salaries and wages, material costs, and rental expenses.

Measuring the Great Recession
Our baseline measure of the intensity of impact of the Great Recession (SHOCK) at an industry by country cell level comes from the UN COMTRADE database of world trade. This is an international database of six-digit product level information on all bilateral imports and exports between any given pairs of countries.
We aggregate COMTRADE data from its original six-digit product level to three-digit US SIC-1987 level using the Pierce and Schott (2010) concordance. We deate the industry and country specic export value series by a country and year specic CPI from the OECD to measure real exports. 15 For the U.S. MOPS data we are able to construct a more detailed SHOCK variable which varies at the establishment level. Specically, we use pre-recession product level revenue data from the 2006 ASM to measure each establishment's distribution of sales across 7 digit NAICS products before the onset of the Great Recession. We then aggregate the Longitudinal Firm Trade Transactions Database (LFTTD), which contains the universe of import and export transactions for U.S. rms, to the product-year level. By matching each establishment's pre-recession distribution of sales across products to product level export growth, we are able to obtain a more precise measure of the intensity of the Great Recession which measures export growth in the products which the establishment produces. All results from the U.S. MOPS data use this establishment specic formulation of the SHOCK measure.
16 The plant-specic shock is advantageous in that it addresses an important potential bias arising from mismeasurement of the relevant economic shock for diversied plants. To the extent that diversication of product mix is correlated with decentralization, using an industry level shock introduces non-random measurement error and may bias the results. Our plant-specic shock built from plant-product data addresses this concern. Figure A1 shows the evolution of annualized export growth in the years preceding and during Great

17
Since recessions typically have a greater impact on reducing the expenditure on durable versus nondurable goods (e.g. King and Rebelo, 1999), we use as an alternative variable to capture the intensity of the Great Recession shock the average durability of the goods produced in the industry, drawn from Ramey 15 We nd similar results using other measures of the shock (such as industry sales derived from aggregating rm level data in ORBIS), but trade data is attractive as it has a large external component driven by demand in world markets and is available at a detailed level for every country and industry in our sample.
16 All of the MOPS results are robust to using the same three-digit SIC SHOCK variable which is used in the cross-country WMS analysis.
17 We also run robustness checks using discrete measure of SHOCK, in which we code an industry-country cell to be unity if exports fell over this period and zero otherwise.  22 Second, we split rms by above/below the mean level of decentralization measured before the advent of the Great Recession. Not surprisingly, all our groupings of rms experienced a drop in average sales and furthermore, 18 We also consider a discrete version using a dummy equal to 1 if the durability in the industry is greater than the median (and zero otherwise). 22 To be precise we rst divide the value of nominal exports by a country and time specic CPI. We then construct average real exports in (i) 2009 and 2008 and (ii) 2007 and 2007. We then take the log dierence between these two periods.
the drop in sales is clearly (and signicantly) larger for rms classied in industries experiencing a negative export shock (compare the two bars on the right with the two on the left). However, within the group of rms experiencing a negative shock (those on the right of the gure), the decline in sales was signicantly larger for rms that were more centralized prior to the recession. In the WMS sample, for rms in an industry-country pair hit by a greater negative shock, decentralized rms had a 8.2% fall in sales compared to about 11.8% in the centralized rms, for a dierence of 3.6 percentage points which is signicant at the 5% level (compared to an insignicant dierence of -0.1% in industries that did not experienced a shock). Panel B of Figure 1 performs the analogous exercise on the MOPS sample of US establishments. The dierence in dierences is very similar at 3.5 percentage points, also signicant at the 5% level.
The performance dierential between decentralized and centralized rms appears conned to the crisis period. Using the same four categories as in Figure 1, Figure

23
The basic nding emerging from the raw data is that decentralization was associated with relatively better performance for rms or establishments facing the toughest environment during the crisis. Moreover, the improved performance associated with decentralization is unique to the crisis period, as these rms (plants) did not outperform their peers before the crisis, and temporary, as these rms (plants) do not appear to be systematically outperforming their centralized counterparts after the crisis.

24
We now turn to more formal tests of this basic result using alternative measurement strategies and controls for many other possible confounders.
23 In the US MOPS data, although not in the cross-country WMS, centralized plants in 2012 experience a more rapid recovery in the industries most aected by the Great Recession.
24 One might ask why should centralized rms not systematically outperform their decentralized counterparts in good times?
One reason related to the model of Section 2 is that although turbulence/uncertainty spikes in deep recessions (albeit to dierent degrees in dierent industries) it does not so in other times (see Bloom et al, 2016, especially Table 2). A second reason is that, although the Great Recession is a plausibly unexpected shock to which a rm's optimal decentralization did not reect pre-recession, industry growth trends were less unusual in the pre-crisis period so rm decentralization had already been chosen endogenously to reect these trends.

Baseline regression equation
Our baseline specication is: where ∆ ln Y ijct is the sales growth rate: the three year annualized change in ln(real sales) for rm (or plant) i in industry j in country c in end-year t. 25 DEC i0 is rm (or plant) i's level of decentralization (measured in the initial year of 2006 for WMS and 2005 for MOPS); SHOCK jk is our measure of the severity of the shock of recession in the industry-country cell; x i0 is a set of controls also measured pre-recession (rm and plant size, survey noise and the proportion of college-educated employees); θ c are country dummies, φ j are industry dummies, τ t are year dummies and ε icjt and is an error term. Standard errors are clustered at the industry by country level, or just industry level depending on the variables used to proxy for the Great Recession and the specic sample used. When we use export growth as a measure of the shock the key hypothesis we examine is whether β < 0, i.e. whether decentralized rms and plants do relatively better in bad times. When we use product durability as a measure of the magnitude of the shock the equivalent hypothesis is that β > 0, as the more durable goods industries are expected to have (and do have) the largest fall in demand.
Our underlying identication assumption in equation (8) is that in the pre-Great Recession period rms were in an initial equilibrium where they had adopted their optimal degree of decentralization ( DEC i0 ) based on their current and expected environment.

Baseline results
Column (1) of Table 2 shows the results from estimating a simple specication including export growth as our recession shock indicator and a full set of country, year and three-digit industry dummies. A one percent increase in industry exports is associated with a signicant 0.07 percentage point increase in sales growth. We also nd a positive and weakly signicant association between sales growth and lagged initial decentralization (in 2006). A one standard deviation increase in our decentralization index is associated with a 0.58 percentage point increase in sales growth (e.g. growth increases from say 2.0% a year to 2.6% a year).
28 In column (2) we introduce an interaction term between decentralization and the export shock variable. The interaction term is negative and signicant (0.042 with a standard error of 0.013), which indicates that decentralized rms shrank much less than their centralized counterparts when they were hit by a negative export shock. Note that the coecient on the linear decentralization term is insignicant when the interaction term is added to the specication, which indicates that decentralized rms did not grow signicantly faster or slower in those sectors that had zero export growth.
The magnitudes of the coecients are non-trivial. Consider a macro shock causing a 1% fall in exports.
The coecients in column (2) of Table 2 suggests that the sales of an average rm (with mean decentralization score of zero) will shrink three times as much as those of a decentralized rm (with a score one standard deviation above the mean).
29 Panel A of Figure 3 shows the implied marginal eect of decentralization on sales growth as a function of export growth. These plots are obtained using the coecients reported in column (2) of Table 2. According to these estimates, decentralization has a positive association with sales growth in all industries experiencing country-industry export growth below 8%. This corresponds to twothirds of the WMS sample in the post recession period, but only 12% of rms in the pre-recession periods (this is shown in Panel B of Figure 3). In other words, the positive association between decentralization and rm growth appear to be contingent on the wider demand conditions in the aggregate environment facing the rm, which in turn may be one of the possible reasons for the heterogeneous levels of decentralization observed in 2006. It is important to emphasize that we are not claiming that decentralization is always the optimal form of rm organizationit is very much contingent on the dierent conditions that rms face.

30
The recession shock measure is industry and country specic. Therefore, in column (3) of Table 2 we include a full set of industry dummies interacted with country dummies, as well as a set of other rm controls (measured in 2006). The linear export shock is absorbed by the industry by country dummies, but 28 Note that the growth rates of both rm sales and industry exports used throughout all regressions are multiplied by 100 (i.e 1% is 1 not 0.01) 29 Assuming the eects were causal for illustrative purposes, the average rm will see a drop in sales of 0.062% (the coecient on export growth) whereas the decentralized rm will see a fall in sales of just 0.020% (0.062 minus 0.042, the coecient on the interaction).
30 In other work done using the WMS decentralization data (Bloom, Sadun and Van Reenen, 2012) we discuss other inuences on rm decentralization such as scale, human capital, complexity and culture. We exploit one source of this variation (trust) in an instrumental variable approach discussed below.

15
we can still identify the interaction of the shock with initial rm decentralization. Even in this demanding specication, the interaction between decentralization and the shock remains negative and signicant.

31
A possible concern with the estimates is that the SHOCK variable uses information dated over the same period as the dependent variable, which may give raise to an endogeneity bias. Consequently, we test for the robustness of the main results using as a proxy for the intensity of the Great Recession a measure of the durability of the products in the four-digit industry calculated prior to the recession. We include a full set of four-digit industry dummies to absorb the linear eects in column (4). Consistent with the earlier results, the interaction between decentralization and the SHOCK is positive (since more durable industries experienced greater drops in demand during the recession) and signicant.

32
An alternative exogenous shifter of the shock measure to durability is to construct a Bartik style IV where we predict the change in exports from an industry-country pair. We constructed this for every HS six digit commodity in a country by interacting the lagged (i.e., built using 2006/2007 data) export share of the commodity from country r to a partner country p with the partner country's growth in imports (of that commodity) between 2006/07 and 2008/09 from all countries except country r. Summing this across all partner countries and then aggregating to the three digit industry level gives an IV for the export shock.
The results from using this Bartik IV are very similar to those shown in Table 2.

34
Remarkably, although drawn from a distinct dataset, a single country (US) and dierent survey methodology, the results in this larger sample of plants are extremely similar to the ones reported using the cross country WMS data. The coecients on the interaction terms are of the same sign, statistically signicant and of a broadly comparable magnitude.
The results discussed so far suggest the presence of a positive relationship between rm and establishment sales growth and decentralization in the industries most aected by the Great Recession. In Table 3 we explore whether this relationship persists even when we examine Total Factor Productivity (TFP), i.e. we estimate the most general econometric model of Table 2, column (3) but also control for increases in other inputs such as employment, capital and materials on the right hand side of the equation. As discussed in the introduction, some have argued that rms need to centralize during crises, so tough cost controls and eciency-enhancing measures can be driven down throughout the company. This would imply that, although decentralized rms (or plants) may fare better on protecting sales revenue during downturns, they will do worse in terms of productivity.
Column (1) of Table 3 reports the baseline results for sales growth on the subsample of rms with data on factor inputs, while column (2) reports the productivity results.
35 Decentralization is also signicantly and positively associated with an increase in TFP during a crisis.
36 Column (3) uses the growth of protability (Earnings Before Interest and Tax divided by the capital stock) as the dependent variable and also nds a negative coecient on the interaction although it is not signicant at conventional levels. Column (4) investigates whether the positive association also extends to the extensive margin of adjustment, using an exit regression. The dependent variable is a dummy taking the value of one if the rm exited to bankruptcy between 2007 and 2011 and zero otherwise (the regression is a Linear Probability Model, and the reported coecients are multiplied by 100 for readability). This shows that more decentralized rms also had a signicantly lower probability of exit in industries that were worse hit by the crisis. Columns (5) though (7) repeat the analysis using the MOPS data, and again nds a negative and signicant coecient on the interaction term between decentralization and the shock for sales, productivity and prots growth.

Identication and robustness
A concern with the results is that our decentralization interaction is simply picking up longer term trends or proxying for some unobserved variable. To address these issues we took several steps.
Placebo test in a pre-crisis period First, we address the concern that the Decentralization * SHOCK interaction may simply be picking up some other time-invariant industry characteristic associated with the magnitude of the recession and rm decentralization. As shown in Figure 2, the raw data suggest that the dierentials in performance between decentralized and centralized rms are conned to the Great Recession.
To further probe this result, we examine the relationship between sales growth and the Decentralization * SHOCK interaction in a sample including years preceding the Great Recession in Table 4. Finding the same results in this period would raise the concern that the SHOCK dummy captures unobserved industry heterogeneity unrelated to the Great Recession such that decentralized rms always did better in certain sectors. Thus, we regard this as a placebo test. We look again at three year dierences in growth but instead pool across the 3-year dierences 2008-05, 2007-04, 2006-03 and 2005-02 to dene the pre-recession growth 35 The sample for the TFP regression is smaller due to missing data on some of the additional inputs needed for the production functions specication (in many countries revenues are a mandatory item on company accounts, but other inputs such as capital are not).
36 The sum of the unreported coecients on employment, capital and materials growth is about 0.9 suggesting decreasing returns to scale (and/or market power). Measurement error may also be responsible for attenuating the coecients on factor inputs towards zero. Note that if we calculate TFP as a residual using cost shares as weights on the factor inputs and use this as the dependent variable (dropping the factor shares from the right hand side) are results are similar to those from the estimated production function.  (2)). Column (1) shows that the coecient on Decentralization * SHOCK is actually positive, although insignicant, in the years preceding the Great Recession. Column (2) repeats the results of the specication of Table 2 Observable factors correlated with decentralization We investigated whether the Decentralization * SHOCK interaction actually reects other rm level characteristics correlated with decentralization exploiting the very rich data we have compiled. Tables 5 and 6 we augment the baseline specication of column (3) in Table 2 with interactions terms between the Great Recession indicator and a series of additional rm and plant controls. First, we include interactions with human capital of the workers as well as the overall management quality of the rm (in the WMS measured as in Bloom and Van Reenen, 2007) or the plant (in the MOPS). To control for the fact that centralization might reect rms in low prot margin market segments who always do badly in downturns (e.g. because their products are relatively homogeneous)

Specically in
we include interactions with pre-recession prot margins. We also tried including measures of scale (size of the plant and/or the rm), decentralization from the plant manager to production workers, technology adoption (data-driven decision making), union strength and plant manager characteristics (age, immigrant status and gender). Throughout these experiments the coecient on our key Decentralization * SHOCK interaction remained signicant.

39
In the MOPS data we can implement a particularly tough test. Since we measure decentralization in multiple plants within the same rm, for multi-plant rms we are able to include an interaction between 38 Ideally, we we would have an instrumental variable for decentralization, but there is no obvious candidate that credibly meets the exclusion restriction. For example, regional variation in generalized trust in the population around the rm's headquarters is strongly correlated with decentralization (see Bloom, Sadun and Van Reenen 2012). We found that rms in high trust areas outperform others in downturnsan interaction between regional trust and our export shock variable is signicantly negative in the performance regressions of Table 2. This reduced form is consistent with a mechanism whereby trust causes greater rm decentralization and therefore fosters resilience in bad times. However, there may also be other mechanisms through which higher trust helps rms outperform others during downturns, so trust cannot be reliably excluded from the second stage.
39 Although the additional variables were usually insignicant, there are exceptions. In Table 5, decentralization from plant manager to workers exhibits a similar pattern to our main decentralization measure of power between the central headquarters and plant manager. This suggests that decentralizing decision-making throughout the hierarchy is benecial during times of crisis. The management interaction is also weakly signicant, although in this case the coecient is positive. In other words, well managed rms perform relatively better in good times than in bad times.
the Great Recession indicator and average rm decentralization. 40 This means that the coecient on the Decentralization * SHOCK interaction is identied solely o dierences in decentralization across plants within the same rm. Remarkably, the results remain signicant even in the presence of the rm level of decentralization and its interaction with export growth (coecient of -0.023 and standard error of 0.010).
A further concern is that the SHOCK measure could be reecting other industry characteristics rather than the demand fall. In Appendix Table A3 we show that our key interaction is robust to including interactions of decentralization with a number of other industry characteristics such as asset tangibility, inventories, dependency on external nance and labor costs.
Validity of exports as a shock measure We have argued that trade changes are an attractive indicator of the Great Recession shock, as they are more likely to reect what is happening to demand in world markets than being a reection of country and industry specic supply factors. Furthermore, we have also shown above that our results are robust to alternative indicators of the shock such as the industry-specic durability measure, a Bartik style IV for exports or using industry output instead of exports. As a further check we estimated our models separately for exporting establishments vs. non-exporting establishments using the MOPS data (export data is not an item required in the company accounts data). As expected, the results are driven by the exporting plants who are most directly exposed to trade shocks.

41
Asymmetries We investigated whether a negative shock diered from a positive shock by allowing dierent coecients on positive than negative shocks (dened either as positive export growth or export growth above/below the median value). In all cases we found we could not reject symmetry. This is unsurprising since in the Great Recession period most rms were experiencing various degrees of a negative shock.

The role of co-ordination costs
When there are large externalities between dierent plants belonging to the same rm, decentralization is likely to be more costly (Alonso et al, 2008 41 For example, using the baseline MOPS specication in Table 2 column (5) Appendix Table A7. whether these plants are located in dierent countries or dierent states. Similarly, we looked at whether a rm was producing goods across multiple sectors (diversication dummy) or whether it was part of a foreign multinational enterprise. We also considered the degree of outsourcing (a direct question in WMS) and alternatively as measured by the ratio of intermediate goods inputs to total sales.
In all cases the main interaction between decentralization and export growth remained signicant, and in only one of the 16 cases was one of the other interactions signicant at the 5% level.
42 Although coordination costs matter in general for centralization, they do not seem to account for the better performance of decentralized rms during downturns.
5 Extensions: Understanding the Mechanism

Turbulence: Product churn and stock market volatility
Our empirical ndings strongly suggest that decentralization becomes more valuable in bad times. The simple model in Section 2 suggested that one reason for this was that negative shocks may be associated with greater turbulence (a higher u), which increase the benets of local information (see equation (2)). We now turn to study whether there is any direct evidence to support this idea.
Product Churn Our main measure of turbulence is changes in product churn in recession versus non recession years as a proxy. Product churn is measured using data from the US Census of Manufactures (CM). The CM, which is conducted in years ending in 2 and 7, asks manufacturing plants to list the value of annual shipments by 10-digit product code. Plants receive a list of all the product codes typically produced in their industry, along with corresponding descriptions of each code. Plants which produce products not listed on the form are instructed to write in the appropriate product code.
43 We then measure the amount of product churn at the plant level as the number of products added or dropped between the previous Census and the current Census, divided by the average number of products produced in both Censuses. That is, product churn for establishment i in year t is dened as: Our measure of industry product churn is the average plant level product churn amongst all plants within an industry (three digit US SIC-1987) which produce at least 3 products. We restrict attention to plants 42 This is the materials share in column (9) in the WMS regressions of Table 7. Two other interactions with decentralizationthe number of plants and the number of manufacturing industries in columns (4) and (8) of the MOPS regressions in Table 8are signicant at the 10% level. This could be taken as (weak) evidence that rms with more co-ordination issues with supply chains, scale or industry diversication do worse during downturns when presumably lack of co-ordination becomes more costly.
43 The ASM also has a 10-digit product trailer, but the question is formulated in a way that results in less detailed responses than the 5-yearly CMF question, so we use the CMF to measure churn. 20 with at least 3 products in order to reduce measurement error from product code misreporting. Before examining the relationship between sales growth, decentralization and turbulence (as measured by product churn), we rst examined whether decentralization really was greater in industries where turbulence was higher. Figure A2 shows that this is indeed the case: plants in the top quintile of product churn industries had a decentralization index about 0.2 of a standard deviation higher than those in the bottom quintile.
More formally, Table A4 nds a positive and signicant relationship between decentralization (the dependent variable) and product churn, particularly for decentralization of decisions regarding product introduction and sales and marketing, as the theory would suggest. Furthermore, we checked whether product churn had indeed increased more in industries that experienced a larger drop in exports during the Great Recession.
This is also the case in the data, as shown in Figure A3.
To investigate the empirical validity of the turbulence-based theoretical mechanism, we extend our basic equation (8) to include both the change in CHU RN and also its interaction with decentralization where ∆CHU RN j is the change in churn in industry j (since we estimate this regression model only in the US MOPS sample we omit the country sub-script). According to the model µ > 0 , since churn increases the value of decentralization. Moreover, to the extent that our export shock variable is proxying for rising turbulence during recessions, we would also expect β to drop in magnitude in equation (10) compared to equation (8). Table 9 shows the results of this exercise.
46 In column (1) we estimate the specication in column (4) of Table 2 for the subset of establishments for which an industry level measure of product churn could be built. This has similar results to the overall sample, i.e. the coecient on the interaction DEC i0 * SHOCK j is negative and statistically signicant. Column (2) includes the DEC io * ∆CHU RN j interaction 44 Establishments which produce the same portfolio of products in consecutive Censuses but misreport a product code in one year will be incorrectly measured as having switched products. Product code misreporting is particularly problematic for establishments with 1 or 2 products, for whom a single reporting mistake would result in very high measured product churn.
Our results are robust to using industries with plants with a lower cut-o of 2 or more products or a higher cut-o of 5 or more products.
45 Note that the measure is based on plants who survived between Census years. We also constructed an alternative measure that included plants which died and entered between Census years in the construction of equation ( instead of the DEC i0 * SHOCK j interaction. In line with the model's prediction, the coecient on the interaction with changes in product churn is positive and signicant, i.e. sales growth appears to have a positive association with decentralization in industries that experienced a greater increase in turbulence, as proxied by product churn. Column (3) includes both interactions. The coecient on the interaction between decentralization and product churn remains positive and signicant, while the coecient on the interaction between decentralization and growth in industry exports drops by a quarter in magnitude compared to column (1) and is statistically insignicant. Columns (4) to (6) repeat the same specications, this time using durability as an alternative industry level proxy for the Great Recession. The coecient on the interaction between decentralization and product churn is positive and signicant, and its inclusion again reduces the magnitude of the coecient on the interaction between decentralization and durability to insignicance.
Stock Market volatility We also use as an alternative proxy for the increase in market turbulence a measure derived from the uncertainty literature. We measure the standard deviation in monthly rm-level stock market returns in an industry by year cell over the population of publicly listed rms in each country.
The stock returns measure of uncertainty is the most standard rm-level measure and similar to those used by Leahy and Whited (1996) for example. In a stochastic volatility model based on Dixit and Pindyck (1994) the variance of stock returns will be a good predictor of the underlying level of uncertainty. These measures are then used in changes as an alternative proxy for the increase in turbulence. In the US we pool at the three digit SIC level as there are about 2,000 publicly listed rms. In the other OECD countries there are fewer publicly listed rms so we construct the measure at the SIC 2 digit level. An advantage of this measure is that it is available for the WMS as well as MOPS, but a disadvantage is that it is constructed only from rms listed on the stock market (in the same industry). Table 10 shows the results. In column (1) we reproduce the specication in column (2) of Table 2.

47
In column (2) we use the interaction between decentralization and the change in the standard deviation of stock market returns instead of our usual interaction. As expected from the theory, the coecient is positive and signicant suggesting that decentralized rms outperform their centralized counterparts in industries where stock market volatility has increased by most. In column (3)  Summary on Turbulence Taking Tables 9 and 10 together, it appears that decentralized rms did relatively better in industries where turbulence increased. At least part of the reason why decentralized 47 The only dierence is that we are using two-digit dummies instead of three-digit dummies to match the level of aggregation for the stock market volatility measures.

22
rms do better in bad times appears to be because the industries worse hit by the Great Recession were also those where turbulence also increased, consistent with our simple model.

Types of decentralization
As a related experiment to shed light on the model we looked at the dierent sub-questions which form the overall decentralization index, as shown in Table 11. Since the Great Recession was associated with a decrease in output demand, we would expect that decentralization capturing managerial discretion over outputs (sales and new products) would be more important than delegation over inputs (like labor and capital). We start in column (1) by showing the baseline result of Table 2, column (3). In columns (2) and (3) we repeat the estimation using as the decentralization index a z-scored average of the two questions capturing plant manager decentralization for hiring and investment decisions in column (2), and for sales and marketing and product introduction in column (3). In columns (4) to (6)

48
In both cases, the positive eect of decentralization in a crisis is primarily driven by the output related questions. This nding provides additional insight on the possible mechanism through which decentralization may positively aect performance during a downturn, namely the ability to better adapt to more turbulent demand conditions.

49
One concern with these ndings is the belief that in practice plant managers do not have meaningful autonomy in decisions regarding sales and marketing and product introduction, and that these decisions are typically undertaken in the marketing department of rm headquarters. It is worth recalling that while this may be the case in business-to-consumer rms which sell their goods to households directly or through retail establishments, it is less obvious in business-to-business rms which sell their manufacturing output to other rms. The latter scenario encompasses a signicant share of US and EU manufacturing activity. 49 Consistent with the previous sub-section Appendix Table A5 shows that the positive interaction between decentralization and product churn is driven primarily by the sales and marketing and product introduction questions.
50 According the Bureau of Economic Analysis, over 90 percent of US manufacturing output goes to the manufacturing sector, which will be primarily business-to-business transactions: https://www.bea.gov/industry/xls/ioannual/IOMake_Before_Redenitions_1997-2015_Sector.xlsx. This will be similar in Europe, which like the US has a higherend manufacturing sector focused more at business consumers (Chinese manufacturing output, in contrast, is more consumer focused).

23
times. There may, however, be alternative rationalizations of these results. Imagine, for example, that bad times reduce the costs of decentralization because the plant manager fears that performing the non-prot maximizing action might cause the rm to go bankrupt, and this will be more costly to the manager than CEO, as he will take a larger hit to their income (e.g. through longer unemployment). To test this idea we examine environments where the rm-specic risk of bankruptcy rose rapidly in the Great Recession.
We constructed several indicators of increased bankruptcy risk. In particular, we used the measures of exogenous increases in exposure to nancial crisis exploited by Chodorow-Reich (2014) such as exposure to mortgage-backed securities (aected by the sub-prime crisis) and a rm's pre-existing relationship with Lehman Brothers or similar at-risk banks. These are pre-Great Recession conditions relating to the supply of nance rather than product demand. We also used more conventional measures such as leverage ratios.
We found that these measures do predict negative performance in sales and other outcomes (see Appendix Table A6), as in Chodorow-Reich (2014). However, in no case did including these bankruptcy risk variables (and their interactions with SHOCK or other covariates) materially alter the coecient on the key interaction of Decentralization * SHOCK when included in equation (8). 51 This led us to conclude that the crisis was not leading to greater decentralization by fostering greater alignment between the central headquarters and plant manager.

Changes in decentralization over time
Recall that our identication assumption is that pre-recession decentralization is weakly exogenous and that there are some adjustment costs which mean that after the Great Recession shock rms do not immediately adopt the new optimal (more decentralized) organizational form. A corollary of our theory, however, is that rms will start moving to a more decentralized form (to the extent that they believe the shock is likely to be long-lasting). Hence, we should expect to see some increase in decentralization for rms and establishments more exposed to the shock. increasing the benet of centralization. On the other hand, it makes the plant manager's local information more valuable, and so implies that decentralized rms will perform relatively better in unexpected downturns.
To empirically investigate these issues we collected new data on a panel of rms in 10 OECD countries (WMS), and plants in the US (MOPS) and exploited the negative shock of the Great Recession which reduced demand across industries and countries in heterogeneous ways. Using our pre-recession data on decentralization we nd that negative shocks hurt growth in centralized rms and establishments signicantly more than in their decentralized counterparts. This is true whether we use export shocks which vary at the industry by country (WMS) or establishment (MOPS) level, or exogenous predictors of these negative shocks like product durability. Further, as the localist model suggests, this eect is driven by the industries which experienced a greater increase in the turbulence (as measured by product churn and stock market volatility) that accompanied the crisis.
As discussed above, the eects are not trivial in size at the micro level, but we can also perform some very rough calculations at the macro level (see Appendix B and Table A9 for details). We estimate how much of the post crisis dierences in GDP growth performance across countries are related to the dierent levels of pre-crisis decentralization. Under this view, the fact that the US had relatively more decentralized (and therefore exible) rms meant that it could weather the global economic storm better than many more centralized countries. Assuming the nancial crisis was a common cross country negative shock, we can trace out the implied post shock growth performance depending on each nation's average decentralization. We calculate that greater decentralization in the US could account for about 15% of the US 's superior GDP growth post 2012-2007 compared to the other OECD countries in our sample.
We see our paper as a rst attempt to unravel the relationship between growth and the internal organization of rms using micro data with observable measures of decentralization. There are many directions to take the research. First, we need to look at the ways in which, in the longer-run, rms change their organizational forms. For example, as the eects of the Great Recession recede, how will the growth eects and degree of decentralization change? Second, we would like to go deeper into the relation between the debt structure of companies (and so their bankruptcy risk) and the incentives for rms to change. Finally, it would be valuable to examine the macro-economic implications of our modeling framework in more detail. Do the eects we identify matter in terms of thinking about business cycles and how economies and companies can be resilient to these adverse events?   Table 2 column (2) as a function of the shock (export growth in cell). Panel B shows the distribution of firms in industry-country cells with different levels in cell). Panel B shows the distribution of firms in industry-country cells with different levels of export growth before and after the Recession.

Table 2 -Decentralization and Sales Growth -Main Results
(1) (3)  (4)), country and year dummies and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, WMS also includes analyst dummies Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry by country level in columns (1) (4)), country and year dummies and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, WMS also includes analyst dummies and MOPS whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(% of employees with a college degree).  Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry by country level in columns (1)-(3) and just industry in columns (5)-(7). Sales growth is the annualized three-year change of firm ln(sales). TFP growth is the same as sales growth except we include the growth of employment, capital and materials on the right hand side of the regression. Profit growth is EBIT/capital for WMS and gross profits/capital for MOPS (profits measured as plant sales -wage bill -materials -  (5)-(7). All columns include three digit industry by country and year dummies and controls for firm and plant size, skills and "noise" controls.

U.S. Census Data (MOPS)
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry by country level in columns (1)-(3) and just industry in columns (5)-(7). Sales growth is the annualized three-year change of rm ln(sales). TFP growth is the same as sales growth except we include the growth of employment, capital and materials on the right hand side of the regression. Prot growth is EBIT/capital for WMS and gross prots/capital for MOPS (prots measured as plant sales -wage bill -materials -rental expenses). For all these dependent variable we  (7). All columns include three digit industry by country and year dummies and controls for rm and plant size, skills and "noise" controls.

TFP Growth
Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry by country level. Sales growth is the annualized three-year change of firm ln(sales). TFP growth is the same as sales growth except we include the growth of employment, capital and materials on the right hand side of the regression. For columns (2) and (5) (1) and (4)  columns include dummies for year and for three digit industry by country pairs, and controls for firm and plant size, skills and "noise" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration and analyst dummies).
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry by country level. Sales growth is the annualized three-year change of rm ln(sales). TFP growth is the same as sales growth except we include the growth of employment, capital and materials on the right hand side of the regression. For columns (2) and (5) (1) and (4) we use long dierences 2008-05, 2007-04, 2006-03 and 2005-02. Columns (3) and (6)   Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry by country level in all columns. Specifications are the same as Table 2 column (3) except augmented with additional variables from the WMS (linear and interacted with export growth). Management is the zscored average of 18 z-scored management questions (see Bloom and Van Reenen 2007 for details). "Log(% employees with a college degree)" is the natural logarithm of the percent of employees with a bachelors degree. Worker decentralization is the z-scored average of 2 questions on worker autonomy. Foreign/Male plant manager=1 if plant manager is from a foreign country or male, respectively.
Notes: WMS Data. *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry by country level in all columns. Specications are the same as Table 2   Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry level in all columns. The specification is the same as Table 2 column (5) except augmented with additional variables from the MOPS (linear and interacted with export growth). Management is the z-scored average of 18 z-scored management questions (see Bloom et al. 2013 for details). "Data-Driven Decision-Making" is the z-scored average of 2 questions on the use and availability of data in decision-making. "Log(% employees with a college degree)" is the natural logarithm of the percent of employees with a bachelors degree. "Union" is the percent of employees that are members of a labor union.
Notes: MOPS Data. *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry level in all columns. Specications are the same as Table 2 column (5) except augmented with additional variables from the MOPS (linear and interacted with export growth). Management is the z-scored average of 18 z-scored management questions (see Bloom et al. 2013 for details). Prot margin is the pre-recession level of prot over sales. Data-Driven Decision-Making is the z-scored average of 2 questions on the use and availability of data in decision-making. Log(percentage of employees with a college degree) is the natural logarithm of the percent of employees with a bachelors degree. Union is the percentage of employees that are members of a labor union.       Notes: **signicant at 10%; ** 5%; *** 1% level. Specication are the same as Table 2 (exports) at the product level (HS7) for the products the plant produced just prior to the Great Recession in 2006. All columns include three digit industry dummies, firm and plant size, skills and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, and whether the survey was answered online or by mail). "PRODUCT CHURN" is the three digit industry of value of the average change in the (number of products added between t and t-5 plus the number products dropped between t and t-5)/(average number of products between t and t-5).
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry level. The dependent variable is the annualized three-year change of rm ln ( digit industry dummies, rm and plant size, skills and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, and whether the survey was answered online or by mail). "PRODUCT CHURN" is the three digit industry-level value of the average change in the (number of products added between t and t-5 plus the number products dropped between t and t-5)/(average number of products between t and t-5).
(3) )" is defined analgously using the log of the standard deviation. All columns include three digit industry dummies and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, and whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(% of employees with a college degree).
Notes: *signicant at 10%; ** 5%; *** 1% level. The dependent variable is the annualized three-year change of rm ln ( Table 2. 40 (3) in columns (4)-(6). All columns include three digit industry, country and year dummies and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, WMS also includes analyst dummies and MOPS whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(% of employees with a college degree). "Decentralization -Hiring & Investment" is the z-scored average of the z-scored measures of plant manager autonomy in hiring and capital investments (and also pay increases in the MOPS data). "Decentralization -Sales & New Products" is average for product introduction and marketing. . All columns include three digit industry, country and year dummies and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, WMS also includes analyst dummies and MOPS whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(% of employees with a college degree). "Decentralization -Hiring & Investment" is the z-scored average of the z-scored measures of plant manager autonomy in hiring and capital investments (and also pay increases in the MOPS data). "Decentralization -Sales & New Products" is average for product introduction and marketing.

41
Web Appendices -Not Intended for Publication A Data Appendix

A.1 Industry-level variables Exports
We measure changes in exports in an industry by country cell using the UN COMTRADE database of world trade. This is an international database of six-digit product level information on all bilateral imports and exports between any given pairs of countries. We rst aggregate the COMTRADE value of export data (in US dollars)from its original six-digit product level to three-digit US SIC-1987 level using the Pierce and Schott (2010) concordance. We deate the industry and country specic export value series by a country and year specic CPI from the World Bank (2010 base year) to measure real exports. The Export growth variable is dened as the logarithmic change in exports in 2008-09 (the average in a cell across these two Great Recession years) relative to 2006-07 (the average across the two years immediately prior to the Great Recession). The real export growth variable is winsorized at the 5th and the 95th percentile.

Durability
Data on the average durability of the goods produced in the industry are drawn from Ramey and Nekarda (2013). This is a continuous cross-sectional measure at the 4-digit industry level.

Bartik Instrument
The Bartik IV for export growth in a country-industry cell is constructed as the change in world import demand (WID) for commodity m in country r between time and t (2008 and 2009) and t − 1 (2006 and 2007), is dened as: zmr,t = p s mpr,t−1 * W ID mpr ,t where s mpr,t−1 denotes the share of exports of commodity m from country r to partner country p at time t − 1; W ID mpr ,t is the log change in total imports of commodity c in partner country p between t and t − 1 from all countries excluding country r (hence the r' sub-script). Consider, for example, the Bartik IV for changes in US exports of anti-ulcer drugs. For a given partner, like the UK, we calculate the share of all American anti-ulcer drugs exported that were exported to the UK in t − 1 , s drugs,U K,U S,07−06 , and then multiply this by the change in the imports of anti-ulcer drugs into the UK from every country (except the US), W ID drugs,U K,U S ,09−08 . This is a prediction of what the demand for US exports to the UK will be coming from exogenous world demand (rather than US specic factors). We repeat this for every country in the world (not just the UK) and then sum over all the US partner countries in the world.
Commodity m is measured at the 6-digit level of the Harmonized Commodity Description and Coding System (HS).
Commodity level measures are then mapped into Industry j three-digit Standard Industry Classication (SIC) codes using the Pierce and Schott (2010) concordance.

A.2 World Management Survey (WMS) International Data Firm-level Accounting Databases
Our sampling frame was based on the Bureau van Dijk (BVD) ORBIS which is composed of the BVD Amadeus dataset for Europe (France, Germany, Greece, Italy, Poland, Portugal, and the United Kingdom); BVD Icarus for the United States, BVD Oriana for Japan. These databases all provide sucient information on companies to conduct a stratied telephone survey (company name, address, and a size indicator). These databases also typically have accounting information on employment, sales and assets. Apart from size, we did not insist on having accounting information to form the sampling population, however.
Amadeus are constructed from a range of sources, primarily the National registries of companies (

The Organizational Survey
In every country the sampling frame for the organization survey was all rms with a manufacturing primary industry code with between 50 and 5,000 employees on average over the most recent three years of data. Interviewers were each given a randomly selected list of rms from the sampling frame. More details are available in Bloom, Sadun and Van Reenen (2012) where we compare the sampling frame with Census demographic data from each country and show that the sample is broadly representative of medium sized manufacturing rms. We also analyzed sample selection -the response rate was 45% and respondents appear random with respect to company performance, although larger rms where slightly more likely to respond.
We collected a detailed set of information on the interview process itself (number and type of prior contacts before obtaining the interviews, duration, local time-of-day, date and day-of-the-week), on the manager (gender, seniority, nationality, company and job tenure, internal and external employment experience, and location), and on the interviewer (we can include individual analyst xed eects, time-of-day, and subjective reliability score). We used a subset of these noise controls (see text) to help reduce residual variation.
In analyzing organizational surveys across countries we also have to be extremely careful to ensure comparability of responses.
One step was the team all operated from two large survey rooms in the London School of Economics. Every interviewer also had the same initial three days of interview training, which provided three calibration exercises, where the group would all score a role-played interview and then discuss scoring together of each question. This continued throughout the survey, with one calibration exercise every Friday afternoon as part of the weekly group training sessions. Finally, the analysts interviewed rms in multiple countries since they all spoke their native language plus English, so interviewers were able to interview rms from their own country plus the UK and US, enabling us to remove interviewer xed eects.
The construction of the degree of decentralization measures (from Central Headquarters to Plant Manager) is discussed in some detail in the text. The questions are addressed to the plant manager. We only keep observations where at least two of the four decentralization questions were answered (and we include a control for the number of non-missing questions in the set of noise controls). We drop observations where the plant manager is also the CEO (5% of rms). In cases were the Central Headquarters is on the same site as the plant we interviewed we add a dummy variable to indicate this (one of the noise controls) to reect potentially greater monitoring. We use the data from the 2006 wave in all cases except when we analyze changes in decentralization as an outcome where we exploit the fact that we ran another wave in 2009 and 2010 for a sub-sample of rms.
As a check of potential survey bias and measurement error we performed repeat interviews on 72 rms in 2006, contacting dierent managers in dierent plants at the same rm, using dierent interviewers. To the extent that our organizational measure is truly picking up company-wide practices these two scores should be correlated, while to the extent the measure is driven by noise the measures should be independent. The correlation of the rst interview against the second interviews was 0.513 (p-value of 0.000), with no obvious (or statistically signicant) relationship between the degree of measurement error and the decentralization score. That is to say, rms that reported very low or high decentralization scores in one plant appeared to be genuinely very centralized or decentralized in their other plants, rather than extreme draws of sampling measurement error.

Firm-level variables
Our rm accounting data on sales, employment, capital (xed assets), prots and intermediate inputs came from BVD ORBIS.
Whether the variable is reported depends on the accounting standards in dierent countries. Sales are deated by a three digit industry producer price index. BVD has extensive information on ownership structure, so we can use this to identify whether the rm was part of a multinational enterprise. We also asked specic questions on the multinational status of the rm (whether it owned plants aboard and the country where the parent company is headquartered) to be able to distinguish domestic multinationals from foreign multinationals.
We collected many other variables through our survey including information on plant size, skills, organization, etc. as described in the main text. We also collected management practices data in the survey. These were scored following the methodology of Bloom and Van Reenen (2007), with practices grouped into four areas: operations (three practices), monitoring (ve practices), targets (ve practices), and incentives (ve practices). The shop-oor operations section focuses on the introduction of lean manufacturing techniques, the documentation of processes improvements, and the rationale behind introductions of improvements. The monitoring section focuses on the tracking of performance of individuals, reviewing performance, and consequence management. The targets section examines the type of targets, the realism of the targets, the transparency of targets, and the range and interconnection of targets. Finally, the incentives section includes promotion criteria, pay and bonuses, and xing or ring bad performers, where best practice is deemed the approach that gives strong rewards for those with both ability and eort. Our management measure uses the unweighted average of the z-scores of all 18 dimensions.
Our basic industry code is the U.S. SIC (1997) three digit levelwhich is our common industry denition in all countries.
We allocate each rm to its main three digit sector (based on sales). Both at this establishment and at headquarters which is assigned a value of one-half; Only at headquarters which is assigned a value of zero. We then standardize each question to have a mean equal to zero and standard deviation equal to one, take the mean over all six standardized questions, and then standardize this mean so that it has a mean equal to zero and standard deviation equal to one.

Exports
Our proxy for the Great Recession is a plant-specic export shock constructed by matching the product les of the 2006 ASM which disaggregate establishment revenues by product class to the Longitudinal Firm Trade Transactions (LFTTD) data which contain the universe of export shipments at the rm-product level. To construct our measure, we rst match the product categories from LFTTD (ten-digit Harmonized System categories, or HS10) to the 7-digit NAICS product classes contained in the ASM using the Pierce and Schott (2009) concordance. Next, we aggregate exports to the 7-digit NAICS level and calculate the change in exports in each product over the Recession, dened as the logarithmic change in exports in 2008-09 (the average in a cell across these two Great Recession years) relative to 2006-07 (the average across the two years immediately prior to the Great Recession). Finally, we construct our plant-specic export shock as the weighted average of product export growth in the crisis, where fore each plant, the weights assigned to each product category is that plant's share of sales revenue in the product as measured before the crisis in the 2006 ASM.

Product Churn
Product churn is constructed using data come from the US Census Bureau's Census of Manufactures (CM). The CM asks establishments to list the dollar value of annual shipments by 10-digit product code. Establishments receive a list of all the product codes typically produced by establishments in their industry, along with corresponding descriptions of each code.
We start by calculating the total number of 10-digit products by each establishment in a given year, as well as the number of added products and the number of dropped products for each establishment compared to the previous CM 5 years earlier.
This of course restricts the sample to manufacturing establishments which were alive ve years earlier. We further restrict the sample by dropping establishments producing fewer than 3 products in both Censuses. Product churn at the establishment level is measured as the number of products added or dropped between the previous Census and the current Census, divided by the average number of products produced in both Censuses. That is, product churn for establishment i in year t is dened as: Product Churn i,t = Products Added i,t + Products Dropped i,t 0.5 (# Products i,t + # Products i,t−5 ) Industry product churn in year t is the average establishment-level product churn amongst establishments within an industry

ASM variables
Directly from the ASM we obtain material inputs, shipments (deated by a three digit price deator) as our sales measure and the headcount of employees for labor. Real capital stocks are constructed using the perpetual inventory method, following the methodology in Bloom, Floetotto, Jaimovich, Saporta-Eksten and Terry (2016). In particular, we combine detailed data on the book value of assets every 5 years from the CM with annual investment data from the ASM. We rst convert CM capital stocks from book to market value using BEA xed asset tables. We then deate capital stocks and investment using industry-year price indices from the NBER-CES Manufacturing Industry Database. Finally, we apply the perpetual inventory method, using the formula Kt = (1 − δt)K t−1 + It . This procedure is done separately for structures and for equipment. However, since the ASM contains investment broken down into investment in equipment and investment in structures, but the CM does not break down capital stocks into these two components, we must apportion plant capital stocks into each component. We do this by assigning the share of capital stock to equipment and structures which matches the share of investment in equipment and structures.

B Magnitudes
In Table A9 we consider some simple calculations of cross-country magnitudes. Our thought experiment is to consider the Great Recession as a global shock as reected by a fall in trade. We use the US value of the shock from COMTRADE of a fall in exports of 7.7 percent. This is the empirical dierence between 2009-08 vs. 2007-06 that we use as our industry-country specic shock measure elsewhere in the paper.
We take the 2006 average levels of cross-country decentralization by country (column (1) of Table A9) and the empirical estimates in column (2) of Table 2 to estimate the average annual implied eect of GDP of the shock (column (2) of Table A9).
We express this relative to the US in column (3). For all countries except Sweden there is a negative relative implied eect because decentralization in the US is greater than every other country except Sweden. Column (4) displays the actual annual change in GDP growth since the start of the global nancial crisis (from World Bank data) for each of our countries and then again expresses these relative to the US base in column (5). Every country except Poland (which is still in a strong catch-up phase of development) experienced a slower growth performance than the US over this period, averaging just over a third of a percentage point (base of column). Column (6) divides the column (3) into column (5) which is the fraction of relative economic performance accounted for by decentralization (note that since we are assuming a common shock, none of this dierence is due to the magnitude of the crisis being worse in some countries than others).
Overall, column (6) of Table A9 suggests that an average of 15% of the post-crisis growth experience between countries is accounted for by decentralization. This is non-trivial as mentioned in the text, but it is worth noting that there is a large degree of heterogeneity between countries underneath this average. Almost all of the dierential growth experience of France and Japan compared to the US can be accounted for by decentralization (96% and 95% respectively), whereas decentralization accounts for virtually none of the UK's performance. In particular, as noted above, because Sweden is more decentralized than the US we should expect it to have outperformed the US, whereas it grew about half a percentage point more slowly. If we drop Sweden, the importance of decentralization doubles to accounting for almost a third of the dierence (32%). Note that the contribution is also negative for Poland, because although Poland is more centralized than the US, it grew more quickly Notes: MOPS data. Industry product churn is the average of plant product churn. Plant product churn = (# products added from '02 to '07 + # products dropped from '02 and '07)/(0.5*# products produced in '02 + 0.5*# products produced in '07). Avg. decentralization z-score

Quintile of product churn
Notes: MOPS data. Industry product churn is the average of plant product churn. Plant product churn = (# products added from '02 to '07 + # products dropped from '02 and '07)/(0.5*# products produced in '02 + 0.5*# products produced in '07).   Notes: The electronic survey, training materials and survey video footage are available on www.worldmanagementsurvey.com

Question D4: "How much of sales and marketing is carried out at the plant level (rather than at the CHQ)"?
Probe until you can accurately score the question. Also take an average score for sales and marketing if they are taken at different levels.
Scoring grid: None-sales and marketing is all run by CHQ Sales and marketing decisions are split between the plant and CHQ The plant runs all sales and marketing Probe until you can accurately score the question-for example if they say "It is complex, we both play a role," ask "Could you talk me through the process for a recent product innovation?" Scoring grid: All new product introduction decisions are taken at the CHQ New product introductions are jointly determined by the plant and CHQ All new product introduction decisions taken at the plant level

Question D3: "Where are decisions taken on new product introductions-at the plant, at the CHQ or both"?
For Questions D1, D3, and D4 any score can be given, but the scoring guide is only provided for scores of 1, 3, and 5.

Question D1: "To hire a FULL-TIME PERMANENT SHOPFLOOR worker what agreement would your plant need from CHQ (Central Head Quarters)?"
Probe until you can accurately score the question-for example if they say "It is my decision, but I need sign-off from corporate HQ." ask "How often would sign-off be given?"

Sales Growth
Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry by country level in all columns. Specifications are the same as Table 2 column (3) except augmented with additional variables. "Asset Tangibility" is the ratio of tangible assets, i.e. net property, plant and equipment, to total assets for the corresponding industry in the US over the period 1980-1989, computed at the ISIC 3 rev 1 level (inverse measure of credit constraints). "Inventory/Sales" is measured as the inventories to total sales for the corresponding industry in the US over the period 1980-1989 (measure of liquidity dependence). "External finance dependency" is measured as capital expenditures minus cash flow divided by cash flow for the corresponding industry in the US over the period 1980-1989 (measure of credit constraint). "Labor costs" is measured as the total labour costs to total sales for the corresponding industry in the US over the period 1980-1989 (another measure of liquidity dependence).
Notes: WMS Data. *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry by country level in all columns. Specications are the same as Table 2 column (3) except augmented with additional variables. "Asset Tangibility" is the ratio of tangible assets, i.e. net property, plant and equipment, to total assets for the corresponding industry in the US over the period 1980-1989, computed at the ISIC 3 rev 1 level (inverse measure of credit constraints). "Inventory/Sales" is measured as the inventories to total sales for the corresponding industry in the US over the period 1980-1989 (measure of liquidity dependence). "External nance dependency" is measured as capital expenditures minus cash ow divided by cash ow for the corresponding industry in the US over the period 1980-1989 (measure of credit constraint). "Labor costs" is measured as the total labor costs to total sales for the corresponding industry in the US over the period 1980-1989 (another measure of liquidity dependence).   (1) and (2) is overall decentralization z-score. The dependent variable in columns (3) and (4) is the z-scored average of the z-scored measures of plant manager autonomy in hiring, capital investments, and pay raises. The dependent variable in columns (5) and (6) is the z-scored average for product introduction and marketing questions. "Product Churn" is the three digit industry level value of the average change in the (number of products added between t and t-5 plus the number products dropped between t and t-5)/(average number of products between t and t-5).

All
Capital Expenditure, Hiring, and Raises

Product Introductions and Sales and Marketing
Notes: MOPS Data. *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry level. The dependent variable in columns (1) and (2) is overall decentralization z-score. The dependent variable in columns (3) and (4) is the z-scored average of the z-scored measures of plant manager autonomy in hiring, capital investments, and pay raises. The dependent variable in columns (5) and (6) is the z-scored average for product introduction and marketing questions. "Product Churn" is the three digit industry level value of the average change in the (number of products added between t and t-5 plus the number products dropped between t and t-5)/(average number of products between t and t-5).   by three digit industry cell. All columns include three digit industry dummies and controls for firm and plant size, skills and "noise" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, whether the survey was answered online or by mail).
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry level. The dependent variable is the annualized ve-year change of rm ln ( and controls for rm and plant size, skills and "noise" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, whether the survey was answered online or by mail).   . "Lender exposure to housing bubble" is the . "ABX exposure" is the correlation of the firm's lender's daily stock returns with the return on the ABX AAA 2006-H1 index, which follows the price mortgage-backed securities issued with a AAA rating. "Lender health" is an aggregation of lender balance sheet variables including trading account losses, real estate charge-offs, and the deposits to liabilities ratio. We combine these variables into one lender health measure by normalizing each to have mean 0 and standard deviation 1, taking an average, and then normalizing this average to have mean 0 and standard deviation 1. All columns include "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(% of employees with a college degree).
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered by the rm's primary lender. The dependent variable is the annualized three-year change of rm ln(sales) from 2009-06. Decentralization is measured in 2005. "EXPORT Growth" is average change (2008/2009 average compared to 2006/2007 average) in ln(exports) at the product level (HS7) for the products the plant produced just prior to the Great Recession in 2006. "Lender exposure to housing bubble" is the . "ABX exposure" is the correlation of the rm's lender's daily stock returns with the return on the ABX AAA 2006-H1 index, which follows the price mortgage-backed securities issued with a AAA rating. "Lender health" is an aggregation of lender balance sheet variables including trading account losses, real estate charge-os, and the deposits to liabilities ratio. We combine these variables into one lender health measure by normalizing each to have mean 0 and standard deviation 1, taking an average, and then normalizing this average to have mean 0 and standard deviation 1. All columns include "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(percentage of employees with a college degree).

U.S. Census Data (MOPS)
Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry. Sales growth is the annualized three-year change of firm ln(sales). "EXPORT Growth" is change in ln(exports) in country by three digit industry cell between the 2008 and 2009 average (the main Great Recession years) compared to the 2006 and 2007 average (the latest pre-Recession years). All columns include three digit industry dummies and controls for firm and plant size, skills and "noise" controls.
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry. Sales growth is the annualized three-year change of rm ln(sales). "EXPORT Growth" is the average change (2008/2009 average compared to 2006/2007) in ln(exports) at the product level (HS7) for the products the plant produced just prior to the Great Recession in 2006. All columns include three digit industry dummies and controls for rm and plant size, skills and "noise" controls.

Yes
Year

Yes
Yes

Yes
Yes

Yes
Yes Cluster SIC3*Cty SIC3 Notes: *significant at 10%; ** 5%; *** 1%. Estimated by OLS with standard errors clustered at three-digit industry by country level in column (1) and just industry in column (2) (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, WMS also includes analyst dummies and MOPS whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(% of employees with a college degree).
Notes: *signicant at 10%; ** 5%; *** 1% level. Estimated by OLS with standard errors clustered at three-digit industry by country level in column (1) (2). All columns include two digit industry, country and year dummies and "noise controls" (plant manager's tenure and hierarchical seniority and the interview's reliability score, day of the week and duration, WMS also includes analyst dummies and MOPS whether the survey was answered online or by mail). Firm and plant size are ln(employment) are skills is the ln(percentage of employees with a college degree).   (1) and an assumed shock of 7.7% (the empirical fall in aggregate US exports in the Great Recession as in our model). Actual GDP growth in column (4) is taken from the World Bank market sector GDP series. Relative values in column (3) and (5) are the simple differences from the US base. Sweden has a negative value in column (6) because it is the only country more decentralized than US, but had a weaker GDP performance. Poland has a negative value because it had faster growth than the US despite being more centralized (it is still in a catch up phase of growth).

Actual annual average GDP growth (2012-2008)
Difference in actual GDP growth relative to

US % of growth difference accounted for by Decentralization
Notes: All GDP growth numbers in percentage points. Implied GDP growth in column (2) uses the coecients on the model of column (2) Table 2 combined with the value of decentralization from (1) and an assumed shock of 7.7 percent (the empirical fall in aggregate US exports in the Great Recession as in our model). Actual GDP growth in column (4) is taken from the World Bank market sector GDP series. Relative values in column (3) and (5) are the simple dierences from the US base. Sweden has a negative value in column (6) because it is the only country more decentralized than US, but had a weaker GDP performance. Poland has a negative value because it had faster growth than the US despite being more centralized (it is still in a catch up phase of growth).