Indian Labor Regulations and the Cost of Corruption

Indian Labor Regulations and the Cost of Corruption: Evidence from the Firm Size Distribution Abstract This paper investigates the effects of an important but little-researched set of Indian labor and industrial regulations. We use a novel methodology to provide a) the first objective cost estimates of any Indian labor regulations and b) evidence of their impact on misallocation of resources across firms. Our methodology takes advantage of the fact that some regulations only apply to establishments which hire 10 or more employees. Using data from India’s 2005 Economic Census, we observe that the distribution of establishments by size closely follows a power law, but with a significant drop in the distribution for establishments with 10 or more workers. Guided by a model based on Garicano, Lelarge, and Van Reenen (2013) - but augmented to allow for the possibility of misreporting - we use this drop to estimate the implied costs of the regulation. We find that there is substantial variation in our estimated costs across states, industries and ownership types, and that our costs are more robustly correlated with measures of corruption than with any other factors, suggesting that poor state implementation may be as much or more to blame for the high costs than the regulations themselves. We find further that higher costs are associated with lower rates of future employment growth in registered (but not unregistered) manufacturing, suggesting that these costs may play a role in encouraging informality. 1 Introduction India’s labor and industrial regulations have been blamed for many of the country’s ills, including low levels of aggregate productivity, slow growth of productivity, and lackluster job creation in the formal sector1 (e.g. Hsieh and Klenow (2009); Kochhar, Kumar, Rajan, and Subramanian (2006); Besley and Burgess (2004); Hasan and Jandoc (2012)). Of particular note is Hsieh and Klenow (2009)’s landmark study, in which the authors argue that aggregate total factor productivity in India could be 40%-60% higher if not for significant misallocation of resources across firms. They go on to suggest that India’s labor regulations may be to blame for the observed misallocation, although they leave the job of fully corroborating this link to others. In fact, the view that labor regulations are of primary importance is not universally held. Many argue that the laws as written are rarely enforced so that, in practice, firms are effectively unconstrained.2 Others argue that the existing evidence on the detrimental impact of labor regulations is flawed (Bhattacharjea (2006, 2009)). Still others point out that the vast majority of regulations have gone unstudied while nearly all of the attention from economists and the press has focused on a single regulation (Chapter VB of the Industrial Disputes Act) - one that is not likely to constrain any but the very largest firms (Bardhan (2014)).3 It is the goal of this paper to address the above aspects of this debate while avoiding some of the criticisms that have been leveled at previous work. In particular, we use a novel methodology and a uniquely well-suited dataset to study the behavior of firms in response to regulatory thresholds in order to determine whether and to what extent firms are in fact constrained by regulations.4 We proceed in the following steps. First, we generate the establishment-size distribution using data from the Economic Census of India (EC), which, importantly, aims to be a complete enumeration of all non-farm business units, regardless of size or status (formal or informal). From the distribution, we observe that it closely follows a power law, except for a discontinous and proportional decrease in the density of establishments with 10 or more workers. This is precisely the threshold at which a multitude of regulations become legally binding, so we take this observation as evidence that firms with 1 Here and elsewhere in the paper, the formal sector refers to business enterprises that are registered with some branch of the government. 2 For instance, in a recent paper, Chaurey (2015) provides evidence that firms seem to hire contract workers as a way of avoiding certain regulations. 3 Chapter VB of the Industrial Disputes Act (IDA) stipulates that firms in the industrial sector with 100 or more workers must obtain permission from the relevant governmental authority before laying off workers. Bardhan (2014) points out that 92 percent of firms in the garmet sector have less than 8 workers. 4 Note: for expositional purposes we occasionally refer to “firms”, although it would be more correct to refer to “factories” or “establishments”, since all of the data and most of the regulations are at the factory/establishment level rather than the firm level. Regardless, the disinction is almost moot: nearly all Indian firms are single factory/establishment firms. 2 10 or more workers do seem to be constrained in size by certain regulations - although these are not the same regulations that most others have focused on. We then develop a model of firm size choice under regulatory thresholds which is based on Garicano et al. (2013) (henceforth GLV), but augmented to explicitly allow for the possibiliy of misreporting.5 We model the regulations as causing an increase in the unit labor costs of those firms that report having exceeded the 10 worker threshold6 , and then use the observed distortion in the size distribution to estimate these costs. Under our primary estimation method at the All-India level, we find that firms behave as if operating at or above the 10-worker threshold entailed a 35% increase in their per-worker costs. Our next step is to document substantial heterogeneity in the size of our estimated costs along several dimensions including state, industry and ownership type. For example, we find that the state with the highest estimated regulatory costs is Bihar and that privatelyowned establishments have the highest costs, while government-owned establishments have the lowest. Exploring this variation further, we find that our estimated costs turn out to be correlated with some previous state-level measures of labor regulation reforms (in particular, certain measures from Dougherty (2009)), though not with others (for example, the BesleyBurgess measure from Aghion, Burgess, Redding, and Zilibotti (2008)).7 Moreover, we find strong and robust correlations between our estimated costs and two quite distinct measures of corruption8 , even after controlling for a number of factors including state GDP per capita. As further support for our state-level results, we show that industries with greater “regulatory dependence” have higher estimated costs, but only when they are located in more corrupt states.9 We take these correlations to be suggestive of the fact that the true cost of the regulations may have more to do with bureacracy and corruption, rather than the content of labor and industrial regulations themselves. Finally, we turn to a brief discussion of the possible dynamic consequences of the costs we estimate. We show that, while higher costs are associated with slower growth in employment and productivity in the registered manufacturing sector, this association is more muted 5 Misreporting was a lesser concern in GLV’s original setting, as they had access to administrative data. In contrast, the data in the Economic Census are self-reported, which makes the threat of deliberate misreporting more significant in our case. 6 This is the only way to generate a proportional decrease in the theoretical density, at least in a static model. 7 This may reflect the fact that the Besley-Burgess measures focus on the IDA, while the regulations we study are entirely different. On the other hand, if the Besley-Burgess measures are meant to capture the general effect of labor laws at the state level, one might expect the two measures to be correlated. 8 These corruption measures include a subjective, perceptions-based measured of corruption from Transparency International and a measure of the percentage of electricity that is lost in transmission and distribution as reported by the Reserve Bank of India (this latter measure has been used as a proxy for government corruption and ineffectiveness in, for example, Kochhar et al, 2006). 9 We measure “regulatory dependence” by taking the industry average of the number of inspector visits among Indian firms in the 2005 World Bank Enterprise Surveys. 3 - or even in the opposite direction - in the unregistered manufacturing sector, where the regulations are less salient. This suggests that the costs we estimate may play a role in the “informalization” of the Indian economy, by pushing workers from the formal to the informal sector. This paper aims to contribute to at least three important strands of literature. The first, which we have already mentioned, is the literature on misallocation of resources and total factor productivity (TFP), as exemplified by Hsieh and Klenow (2009). Our contribution is to provide direct evidence that at least some of the misallocation of resources across firms in India is tied to regulations or the enforcement thereof.10 In particular, we show that sizebased regulations (or at least the ways in which they are enforced) lead firms to fall short of their optimal scale, thus distorting the allocation of labor among firms in the economy and, likely, lowering TFP. Another strand of literature to which we aim to contribute relates to corruption in the enforcement of government policies. Most previous studies (eg: Besley and McLaren (1993); Mookherjee and Png (1995)) have modeled such corruption as collusion between inspectors and firms or citizens: corrupt inspectors allow firms to avoid the de jure costs of abiding by regulations in exchange for bribes. Hence, in these frameworks, corruption lowers the costs associated with regulations. However, our results suggest that the costs associated with size-based regulations are higher in more corrupt environments, and are thus more in line with an alternative framework in which corruption takes the form of extortion between inspectors and firms (i.e.: corrupt inspectors take advantage of bureacratic regulations in order to extract higher rents from firms in the form of harassment bribes).11 We present a theoretical model as well as anecdotal evidence from “ipaidabribe.com” to support this interpretation and view the support we provide for this alternative conception of corruption to be another contribution of the paper. Lastly, this paper is also clearly related to the large literature that more generally investigates the impact of Indian labor regulations on economic outcomes. The literature dates back to at least Fallon and Lucas (1993), but the more recent proliferation seems to be due to the work of Besley and Burgess (2004). In that paper, the authors first interpret state-level amendments to the Industrial Disputes Act (IDA) as either “pro-worker” or “pro-employer” and then aim to show that Indian states that amended the IDA in a “pro-worker” direction experienced slower growth in output, employment, investment and productivity in registered manufacturing. The paper, though extremely influential, has been criticized by Bhattachar10 In future work we hope to determine what portion of the TFP loss from misallocation estimated by Hsieh and Klenow (2009) can be attributed to the regulations we study. 11 This finding echoes Novosad and Asher (2012), in which it is argued that regulations can provide a means through which politicians can impose costs on businesses. 4 jea (2006) and Bhattacharjea (2009) on a number of grounds. One of Bhattacharjea’s major criticisms is that Besley and Burgess’s interpretations of amendments as “pro” or “antiworker” are subjective and debatable (ie: different people might read and code them in a different way). This criticism affects most of the subsequent academic work on this topic, since most papers use the Besley-Burgess codings, but it is a criticism we are able to sidestep with our methodology. Since our analysis is based only on firm level data and size-thresholds stated explicitly in the laws themselves, it has the advantage of objectivity. The second contribution we make to this literature is to focus on a set of regulations that have been almost entirely ignored even though they effect a much larger proportion of firms than Chapter VB of the IDA.12 The only other papers of which we are aware that study regulations that kick in at the 10-worker threshold are Dougherty (2009), Dougherty, Frisancho, and Krishna (2014) and Kanbur and Chatterjee (2013). The latter investigates the Factories Act, which applies to all manufacturing firms that use power and have 10 or more workers (or don’t use power and have 20 or more workers), but their focus is to document noncompliance under the act, which we see as complementary to our approach of estimating the costs of the regulations.13 The papers by Dougherty and co-authors employ state-level indices of labor reforms that differ from the Besley-Burgess codes in that they include consideration of non-IDA regulations such as the Factories Act, but they are constructed from surveys of industry experts and, as such, are by and large subject to similar concerns regarding subjectivity. Another way in which we distinguish ourselves from the previous literature on Indian regulations is that we explore the effect of regulations in all non-farm segments of the Indian economy - not just in registered manufacturing, on which nearly all previous academic studies have focused. A final contribution of the paper is to provide suggestive evidence that improper government enforcement of regulations may play a role in shifting employment from the registered to the unregistered sector. In the next section (Section 2), we provide an overview of the relevant institutional details regarding Indian labor and industrial regulations. Section 3 introduces the data and covers some basics about the size distribution of enterprises in India. In Section 4 we go over the theoretical model and our corresponding empirical strategy. Section 5 provides the main results. In Section 6, we interpret the findings, explore the multiple dimensions of variation 12 Chapter VB of the IDA only applies to manufacturing firms with 100 or more workers. In contrast, the regulations we study affect all firms with 10 or more workers and are thus relevant for a much larger share of firms. We have also tried analyzing Chapter VB of the IDA using the same methodology we employ for the regulations with the 10 worker threshold, but find no effects. I.e.: there does not seem to be a proportional decrease in the density of establishments with more than 100 workers. We also fail to observe “bunching” of firms at sizes just below 100, although the presence of rounding may make such bunching impossible to discern even if it exists. 13 Our estimated costs are robust to the possibility of noncompliance. 5 in our results, and investigate the connection between our estimated costed and corruption. Section 7 concludes. 2 Institutional Background: Size-Based Regulations in India In this paper we attempt to investigate the effects of certain size-based industrial and labor regulations in India. These are regulations that only apply to establishments that exceed a certain size, measured either in terms of a firm’s revenue, the amount of fixed capital invested, or the number of workers employed. One of the most significant such thresholds occurs when establishments employ 10 or more workers, after which they must register with the government and meet various workplace safety requirements (under the Factories Act14 , for example), pay social security taxes (under the Employees’ State Insurance Act), distribute gratuities (under the Payment of Gratuity Act) and bear a greater administrative burden (under, e.g., the Labor Laws Act). Not only are the laws numerous, it has been argued that certain components of the laws are antiquated and/or arbitrary. For example we read in the “India Labour Report” that “Rules under the Factories Act, framed in 194815 , provide for white washing of factories. Distemper won’t do. Earthen pots filled with water are required. Water coolers won’t suffice. Red-painted buckets filled with sand are required. Fire extinguishers won’t do... And so on” TeamLease Services (2006). Firm owners who choose not to comply with such regulations may face costs if discovered and convicted.16 In addition to - or in lieu of - the explicit costs of complying with the regulations, establishments with 10 or more workers may be subject to implicit costs associated with increased interaction with labor inspectors, et al, who may have the power to extract bribes and tighten (or ease) the administrative burden firms face. Indeed, inspectors in India have a large amount of discretion regarding the enforcement of administrative law. For example, in some cases, the definition of what constitutes a “day” is at the discretion of the inspector, and it is a commonly held view that “[w]hile grave violations are ignored, minor errors become a scope for harassment” (TeamLease Services (2006)). This kind of behaviour has been referred to as “harassment bribery” (Basu (2011)). Anecdotal evidence of inspectors using the complexity and sheer amount of paperwork as 14 Technically the Factories Act applies for 10-plus worker establishments only if they use power. For establishments that do not use power, the Factories Act does not apply until they employ 20 workers. 15 The Factories Act itself dates to 1948, but the origins of the law go back another 100 years at least, to Britain’s first Factory Acts. 16 These costs may include fines and/or prison sentences. 6 a way to extract bribes is easy to come by. For example, we have included a selection of citizen reports from “ipaidabribe.com” in Appendix 2, which demonstrate just this kind of behaviour.17 Interestingly, some of the reports suggest that the size of the bribe paid is a direct linear function of the number of employees - which will be relevant to our estimation procedure later. As we alluded to earlier, the 10 worker threshold is not the only one relevant; there are other cutoffs at which different regulations become binding. For example, the threshold that seems to have received the most attention, both from academics and the press, is that of 100 workers, at which enterprises in most states become subject to Chapter VB of the Industrial Disputes Act, under which they must be granted government permission to lay off workers. There are other cutoffs still,18 but in this paper we will focus on estimating some of the costs and effects associated with the regulations that come into force at the 10 worker cutoff. One important limitation of our analysis is that we will not be able to address issues regarding the efficacy of any regulations in promoting worker welfare. 3 3.1 Data and the Size Distribution in India Data The data we rely on to investigate the 10-worker threshold comes from the Economic Census (EC) of India. The EC is meant to be a complete enumeration of all (formal and informal) non-farm business establishments19 in India at a given time, regardless of their size. It is this last clause that makes the EC different from every other data source available and precisely suited to our needs. Although the 2005 dataset contains a large number of observations (almost 42 million), there is not very detailed information collected on each observation. For each establishment in the data, there is only information on a handful of variables including the total number of workers usually working, the number of non-hired workers (such as family members working alongside the owner), the registration status, the 4-digit NIC industry code, the type of ownership (private, government, etc) and the source of funds 17 We thank Andrew Foster for this suggestion. For example, firms with 20 or more workers must abide by the Provident Funds Act. Firms with 50 or more workers must comply with Chapter VA of the Industrial Disputes Act, which requires them to provide compensation and notice to employees prior to lay-offs. 19 The EC refers to these as “entreprenuerial units” and defines them as any unit “engaged in the production or distribution of goods or services other than for the sole purpose of own consumption.” As is common in the literature, we occasionally refer to them as “firms” even though the unit of observation in the data is actually a factory or an establishment, rather than a firm (i.e.: multiple establishments may belong to the same firm). We do this for expositional purposes and justify our use of this convention with the observation that the proportion of establishments that belong to multi-establishment firms is minute. 18 7 for the establishment. There is no information on capital, output or profits, and the data is cross-sectional. The EC has rarely been used in academic papers - possibly because it is cross-sectional, contains a significant amount of measurement error, and only contains information on the handful of variables just enumerated, so that better data sources exist for most purposes. The EC is ideal for our purpose, however, since it includes information on employment size and covers the entire universe of establishments. Other more commonly used datasets, such as the CMIE’s Prowess Database, the Annual Survey of Industries (ASI) or the National Sample Survey’s (NSS) Unorganized Manufacturing Surveys cover only certain parts of the distribution and thus cannot be used for our purpose. The ASI, for example, only covers factories in the manufacturing sector that have registered with the government under the Factories Act. However, registration under this Act is only required for establishments with 10 or more workers if the unit uses power (20 or more workers if the factory uses no power). Therefore, the selection of the ASI varies discontinuously at precisely the point of interest. Similar limitations on coverage make the other datasets - other than the EC - unsuitable. Aside from the Economic Census, we also supplement our analysis with data from a variety of other sources. From the ASI we get employment and productivity in the registered manufacturing sector. We generate those same variables for the unregistered sector with data from the Ministry of Statistics and Programme Implementation (MOSPI) and the Reserve Bank of India (RBI). We get data on state and industry level corruption from a) Transparency International’s “India Corruption Study 2005”, b) the RBI, and c) the World Bank Enterprise Survey for India (2005). Data on State-level regulatory enforcement come from the Indian Labour Year Book.20 Other measures of state-level regulations come from Aghion et al. (2008) and Dougherty (2009), while industry-level measures of exposure to trade liberalization come from Ahsan and Mitra (2014). 3.2 The Size Distribution of Establishments in India Figure 1 below shows the distribution of establishments by the number of total workers (hired and non-hired workers) for establishments with up to 200 total workers in 2005. Perhaps the most striking feature of figure 1 is the extraordinary degree to which the distribution is right-skewed. Indeed, about half of all establishments are single person enterprises, while the densities for establishments with 10 or more workers are almost imperceptible.21 Figure 2 shows the drop in density for establishments with 10 or more workers in detail and figure 20 We would like to thank Anushree Sinha and Avantika Prabhakar for their considerable and generous help in obtaining these data. 21 The densities for establishments with more than 200 workers are also imperceptible. We have omitted them only for clarity in the figure. 8 Figure 1: Distribution of establishment size for establishments with 1-200 total workers, 2005 3 shows the full distribution of establishment size frequencies according to a log scale. Each point represents one bar in the earlier histograms. Two things are most striking about figure 3. First, the natural log of the density is a linear function of the natural log of the number of total workers. This implies that the unlogged distribution follows an inverse power law in the number of total workers. This pattern will be important for the analysis that follows but it is not very surprising in and of itself: power law distributions in firm sizes have been documented in many countries (e.g. Axtell (2001) and Hernández-Pérez, Angulo-Brown, and Tun (2006)). The second and more unique feature of the distribution is that there appears to be a level shift downward in the log frequency for establishment sizes greater than or equal to 10. Figure 4 shows this effect for establishments with fewer than 100 workers by running an OLS regression of the log density against log firm size and allowing the intercept to vary for firms with 10 or more workers. To the best of our knowledge, ours is the first paper to document this phenomenon in India. Also of note from the figures above is that there appears to be a significant amount of nonclassical measurement error, seemingly due to rounding of establishment sizes to multiples of 5 and 10. The existence of rounding is not surprising given that the data are self-reported and that respondents are asked to give the “number of persons usually working [over the last year]”. Partially to alleviate concerns that the non-classical measurement error due to rounding might bias our results (and partially for other reasons to be made explicit shortly), we will employ an estimation procedure which first smooths the data non-parametrically. 9 Figure 2: Distribution of establishment size for establishments with 5-25 total workers, 2005 Figure 3: Distribution of establishment size, 2005, log scale Figure 4: Downward shift at the 10-worker threshold in the distribution of establishment size, 2005, log scale (omitting establishments with more than 100 workers) 10 4 4.1 Model and Empirical Strategy Basic Model To interpret the downward shift from Figure 4 in economic terms, we turn to the model in GLV. In their framework, size-based regulations are assumed to increase the unit labor costs of firms that exceed the size threshold, which results in a downshift in part of the theoretical firm size distribution. From the magnitude of the downshift they observe in the empirical distribution they attempt to estimate the additional labor costs imposed by the regulations. GLV begin with a distribution of managerial ability (α ∼ φ(α)) as the primitive object, following Lucas (1978). As is common in the literature (e.g. Eaton, 2011), they assume that the distribution of managerial ability follows a power law (e.g. φ(α) = cα α−βα ). It is this that will generate a power law in the theoretical firm size distribution. A firm with productivity or managerial ability α faces the following profit-maximization problem: π(α) = max αf (n) − wτ̄ n n where n is the number of workers a firm employs, f (n) is a production function (with f 0 (n) > 0 and f 00 (n) < 0), w is a constant wage paid to all workers, and τ̄ is a proportional tax on labor that takes the value 1 if n ≤ N and τ if n > N , where τ > 1. , one can see that From the first order condition on this maximization problem, α = f wτ 0 (n) higher productivity establishments/managers will employ more workers, and that firms which cross the threshold (N ) and must therefore pay higher labor costs will hire fewer workers than they would otherwise. This latter feature is built to match the observed “downshift” in the actual firm size distribution to the right of the regulatory threshold. One can informally characterize the solution as follows: one set of managers with particularly low productivity (below some threshold α1 ) will be effectively unconstrained. These managers would have chosen to hire fewer than 10 workers whether or not the regulation was present. Another set of managers with slightly higher productivity (between some thresholds α1 and α2 ) would, in the absence of the regulation, have chosen to hire 10 or more workers - but who, in the presence of the regulation, obtain higher profits by hiring only 9 workers to avoid the discontinuous increase in costs implied by crossing the threshold. These mangers should be “bunched up” at 9. The last set of managers are those with high enough productivity (α > α2 ) that it is not worth it to avoid the regulation and so they choose to exceed the threshold and pay the tax. However, these managers face higher marginal costs than they would in the absence of the regulation and therefore employ fewer workers by a constant proportion (resulting in a “downshift” in the logged firm size distribution). An exact expression for the distribution of firm size, χ(n), can be recovered as a trans11 formation of the distribution of managerial ability, φ(α), since the first-order conditions on the firms’ maximization problems imply a monotonic relationship between α and n. The key result is that a function of the tax enters multiplicatively in the expression for the density of firms size n (for all n > 9). Therefore, the function of the tax enters additively in the log density for all firms large enough to be subject to the tax. Formally, the density of firms with n total workers, χ(n) is given by: χ(n) =           
1−θ 1−β (β − 1)n−β θ
β−1 1−θ 1−β (N 1−β − τ − 1−θ n1−β u ) θ 0
1−θ 1−β θ β−1 (β − 1)τ − 1−θ n−β if n ∈ [nmin , N ) if if n=N n ∈ (N, nu ) if n ≥ nu where θ measures the degree of diminishing returns to scale, capturing both features of the production function and market power, β represents the negative slope of the power law and τ is the implicit per worker tax. Taking logs and combining the first and last cases22 leads to: # " 1−β β−1 1−θ (β − 1) − β log(n) + log(τ − 1−θ )1{n > 9} log(χ(n)) = log θ This leads to an estimating equation: log(χ(n)) = α − β log(n) + δ1{n > 9} (1) We can identify τ according to: 1−θ τ = exp(δ)− β−1 τ is thus a function of θ, β and δ. We get estimates for α, β and δ from equation 1. Knowing α and β pins down θ, which allows us to identify τ . 4.2 Concerns Regarding Misreporting Before proceeding further, we must consider how our results might be affected by the possibility of misreporting. This is important because one of the underlying assumptions of the analysis above is that the size distribution of firms as observed in the Economic Census is accurate. However, since the data are self-reported, it is possible that plant managers may misreport information to Economic Census enumerators. Specifically, if the managers are aware of the increased regulatory burden that is associated with employing 10 or more 22 In other words, we ignore the bunching at N and the valley directly after, since these are features that are not easily observable in the data. Instead we focus on the ranges n ∈ [nmin , N ) and n > nu . 12 workers, and if they believe that the EC enumerators will relay information to government regulatory bodies, they may wish to hide the fact that their actual employment exceeds the threshold. To see how this type of behavior might affect our results, we model it explicitly in the following subsection. A further reason to be concerned about the possibility of misreporting is due to the fact that Economic Census enumerators were required to fill out an extra form containing the address of any establishment that reported 10 or more workers. It is conceivable that enumerators might have found it preferable to under-report the number of workers for establishments with 10 or more workers in order to avoid the extra burden of filling in the “Address Slip”. Although we do not model this type of problem explicitly in what follows, the implications are nearly identical to those of the model we do explicitly analyze.23 4.3 A Theoretical Model of Misreporting Our model of misreporting starts with the theoretical model from Section 4.1, and amends it to allow firms to choose not only their true employment (n), but also their reported employment (l). Then, a firm with productivity α faces the following profit-maximization problem: π(α) = max αf (n) − wn − τ l ∗ 1(l > 9) − F (n, l) ∗ p(n, l) n,l where α, f (n), w and τ are all defined as they were previously. The problem is identical except that now firms pay the extra marginal cost, τ , only on their reported employment, and not on their true employment. Furthermore, they only pay this cost if their reported employment exceeds the threshold.24 There is now an incentive for firms to misreport their employment in a downward direction (i.e.: to set l < n). Counteracting this incentive is that misreporting firms may be caught by the authorities with probability p(n, l), and made subject to a fine, F (n, l). As written above, both the probability of being caught and the magnitude of the fine may in general depend on n and l in an arbitrary way. However, if one is willing to make the assumption that the expected cost of misreporting (F ∗ p) is an increasing and convex function of the degree of misreporting, n − l, it will be possible to use an estimation technique that will be only minimally biased by the presence of misreporting. Fortunately, based on our understanding of the context in which firms make these decisions,25 23 The only difference is that higher fixed costs would replace higher marginal costs. It is, moreover, easy to show that if our estimation strategy is robust to the model of misreporting we do analyze, it is also robust to this second type of misreporting as well. 24 In point of fact it is most likely that firms’ answers to Economic Census enumerators have no impact on their regulatory burden, but it is possible that firms believe otherwise, and that is what is relevant. 25 This understanding is informed by informal interviews with small businesses in Chennai and our reading 13 we believe that this is the most reasonable assumption on the functional form of the expected cost that one could make. One plausible way to obtain convex misreporting costs is to suppose that firms are caught with a probability that is linearly increasing in the degree of their misreporting (i.e.: n − l) and subject to a fine if caught which is also a linear function of their misreporting. Another possibility is that the probability of being caught is itself an increasing and convex function of the degree of misreporting and the fine if caught is fixed. In what follows we will assume the latter for clarity of exposition, but the analysis is identical for any assumption that yields convex costs of misreporting. 2 , Specifically, suppose that misreporting firms are caught with probability p(n, l) = (n−l) 100 and pay a fixed fine, F , if caught. Then their profit maximization problem is: π(α) = max αf (n) − wn − τ l ∗ 1(l > 9) − F ∗ n,l (n − l)2 100 The solution to this problem can be informally characterized as follows. The lowest productivity firms (those with α below some threshold, α1 ) will be unconstrained, choosing n ≤ 9 and reporting truthfully (l = n). Higher productivity firms, with α ∈ [α1 , α2 ], will choose n > 9, exceeding the regulatory threshold, but will find it profitable to misreport their employment, setting l = 9. These firms will only appear to be “bunched” up at 9, but will in fact have higher employment. The last category of firms are those with α > α2 , which are productive enough to warrant hiring work forces so large that they cannot avoid detection with reasonable probability and must report l > 9. Even these firms, however, with both n > 9 and l > 9 do not find it profit-maximizing to report truthfully. They can save on their unit labor costs by shading their reported employment, and will choose τ . Note that the degree of misreporting is by a constant amount, rather than a l = n − 50 F constant proportion.26 More formally, the log of the density of firms with true employment n, logχ(n), is given by:   logA − βlog(n) if n ∈ [nmin , 9)     log[ξ(n)] if n ∈ [9, nm (α2 )] logχ(n) =  0 if n ∈ (nm (α2 ), nt (α2 ))     logA0 (τ ) − βlog(n) if n ≥ nt (α2 ) while the log of the density of firms with reported employment l, logψ(l) is given by: of the secondary literature. 26 This outcome is a result of the convex cost assumption. 14 Figure 5: Theoretical Model of Misreporting, log scale (thick line = true distribution; thin line = reported distribution; dashed line = counterfactual distribution) 0.200 0.100 0.050 0.020 0.010 0.005 0.002 0.001 2 logψ(l) =            5 logA − βlog(l) log(δl ) 0 0 logA (τ ) − βlog(l + 10 20 50 100 if l ∈ [lmin , 9) if l=9 if n ∈ (9, lt (α2 )) 50 τ ) if l ≥ lt (α2 ) F where terms have been simplified and collected.27 Both of these densities are graphically represented in Figure 5 under specific values of the parameters (the true distribution, χ(n), is represented by a thick line, and the reported distribution, ψ(l), is in blue). The key things to note are the following. First, for the range l ≤ 9, the true distribution coincides with the reported/observed distribution. Second, there appears to be bunching at 9 in the reported distribution, but these firms in fact have greater than 9 workers. Third, compared to the distribution for n < 9, the true distribution and the reported distribution for n  10 are downshifted (A0 (τ ) < A), just as was the case in the model without misreporting.28 Fourth and most significantly - the reported distribution converges to the true distribution for large τ ) − βlog(l) → 0. This is due to the fact that the misreporting is by a l: lim βlog(l + 50 F l→∞ constant amount (as noted earlier), rather than by a constant proportion. To conclude this subsection on the theoretical implications of allowing for misreporting, we make two observations. First, misreporting may lead us to observe “bunching” in the firm size distribution when in reality there may be none. This is irrelevant for us since our estimation strategy does not rely on the bunching in any way. Second, misreporting may lead the reported/observed distribution to understate the true distribution close to the cutoff. However, because the reported distribution converges to the true distribution for large l/n, misreporting is not able to induce a downshift in the reported distribution that 27 28 Derivation to be added in an appendix. As before, the downshift is a function of τ . 15 differs from the downshift in the true distribution at large values of l. Therefore, if we use an estimation strategy that focuses mostly on values far from the cutoff, our estimate of the downshift using the observed distribution is likely to reflect the real downshift and thus we are likely to avoid this source of bias. We develop such an estimation strategy in the following subsection. Before proceeding, however, we should note again that the above analysis assumes that the expected costs of misreporting are strictly convex. There exist non-convex functional forms of the cost function which may lead one to observe a downshift in the reported distribution that is greater than the one in the true distribution, thus biasing any estimates of τ upwards. 4.4 An Empirical Strategy Robust to the Possibility of Misreporting Since convex misreporting costs imply that misreporting will only distort the distribution of reported establishment size versus the true distribution of establishment size close to the cutoff, we estimate the model on the full distribution of establishment size. Since estimating equation 1 treats each establishment size as one observation, using the full distribution of establishment size will mean that the model is primarily estimated using data far from the 10-worker cutoff.29 However, estimating equation 1 on the full distribution of establishment size introduces two complications. First, we cannot perform the estimation on the empirical PMF for large firm sizes, since the empirical probability mass is truncated at the reciprocal of the number of observations (see figure 6 below), while the underlying density continues to diminish in establishment size. Second, respondents appear to round their reported number of workers to the nearest multiple of 5 (see figure 2), a phenomenon that is more pronounced for larger establishments and that could bias our results. To address these two problems, we first estimate the density associated with each number of workers χ(n) non-parametrically using the method of Markovitch and Krieger (2000), which addresses the econometric issues arising in nonparametric density estimation of heavytailed data. We then use the nonparametric density estimates as a basis for fitting the model in equation 1, augmented by dummy variables for having 1, 2, 8, 9 and 10 - 20 workers.30 29 The largest establishment in the 2005 EC has 22,901 workers. The rationale for flexibly modeling the density at 1 and 2 workers is that own account enterprises and 2worker enterprises are likely to be household enterprises and may therefore differ fundamentally in character from their larger counterparts. The rationale for flexibly modeling the density at 8 and 9 workers is that the theory above predicts that the reported density just below the cutoff will be biased upwards by any misreporting effects. Similarly, the theory also predicts that values above - but close to - the threshold may also be biased (downwards). Therefore we also flexibly control for such values (10 to 20) as well, although doing so has only a very small effect on the estimates: as explained above, the estimates are driven mostly 30 16 Figure 6: Downward shift at the 10-worker threshold in the distribution of establishment size estimated on nonparametric density estimates, 2005, log scale (including all establishments). Black points = actual data; Grey = smoothed data. Figure 6 depicts the strategy. The black dots show the raw data. The grey dots represent the result of the first step: nonparametric density estimates associated with each establishment size. The line shows the fit of the model in equation 1, augmented by the dummy variables, to the nonparametric density estimates. Figure 6 above provides some evidence for the model described in section 4.3. The observed establishment size distribution appears to converge back to a power law with the same slope as for establishments with fewer than 10 workers, but deviates slightly from that slope at sizes just above the 10-worker cutoff. In the next section we report the results of the estimation. 5 Preliminary Results In this section we apply the estimation procedure described above to the 2005 Economic Census of India and report the results. Standard errors obtained from a wild cluster bootstrap procedure with 200 replications are given in parentheses.31 In the tables below, we first report estimates for τ − 1 at the All-India level and for a selection of States, Industries and Ownership Types. Estimates for all States, Industries and Ownership Types are reported in the Appendix. The All-India estimate on τ − 1 is .35 and is statistically significant. This by observations relatively far from the threshold. 31 We cluster at the firm size level to allow for the possibility that reporting errors may be correlated by firm size. 17 means that, on average, establishments in India that hire more than 9 workers act as though they must pay additional labor costs of 35% of the wage per additional worker. In the tables and figures it can be seen that there is substantial variation in the magnitude of our estimates of the per-worker tax by State, Industry and Ownership Type. For example, the point estimate on τ − 1 for the State of Kerala is .14, while the estimate for Bihar, on the other hand, is .70, implying that establishments in Bihar act as though they must pay a tax of 70% of the wage for each additional worker they hire past 9 workers. We also observe substantial differences in the size of τ by industry: it appears that the effective tax is highest for establishments in construction and retail. As one might expect, the tax is nonexistent for establishments in the public administration sector (in fact it is negative, but this seems to result from the fact that the assumed power law does not fit the distribution of establishments in this sector well). Similarly, when looking at the differences by ownership type, we find that the estimates for τ are highest for private firms (especially unincorporated proprietorships), and nonexistent (or negative) for government-owned firms, where presumably the regulatory burden is less than in the private sector. Estimates of τ by State Using the Full Distribution of Establishment Size τ −1 Level All-India .347 (0.059) By State Bihar .693 (0.069) 0.165 (0.047) 0.138 (0.033) 0.502 (0.069) Gujarat Kerala Uttar Pradesh By Industry Construction 18 .478 (0.047) Manufacturing Wholesale, retail Public admin., social security By Ownership Type Government and PSU Unincorporated Proprietary 6 6.1 .268 (0.039) .637 (0.115) -.311 (0.031) -.092 (0.028) .490 (0.005) Discussion and Investigation of Mechanisms Interpretation of Results Thus far we have argued that the observed downshift in the distribution of establishments with 10 or more workers is related to the existence of certain labor and industrial regulations that become binding at that point. But if the regulations are responsible for the observed effect, then differences in the substance or application of the regulations should explain (at least part of) the great variation we observe across States and Industries.32 In this section we explore these dimensions of variation with the goal of reaching a deeper understanding regarding the causes and consequences of the costs we have tried to estimate. The regressions we do are cross-sectional and the variables used are endogenous, so the results cannot be given a causal interpretation, but we find them instructive nevertheless. To preview our results, we do observe a correlation between our estimated costs (τ ) and certain measures of the substance of the regulations. Moreover, we also find a robust and independent correlation between our estimated costs and several different measures of 32 The variation across ownership types is straightforward to explain: the regulations are clearly not applied in the same way to privately owned enterprises and government enterprises. An additional explanation is that government establishments are not profit maximizing and thus would require a different motivational theory altogether to produce the observed power law distribution. 19 corruption/poor state governance, suggesting that it is not only the regulations themselves but also their enforcement and application that is responsible for the high costs we estimate. We also sketch a theoretical framework of bribery and extortion which casts light on the proper interpretation of our empirical results. Finally, we present some suggestive evidence that our costs may have significant negative dynamic implications, as they are associated with lower growth in employment and productivity in registered manufacturing - and higher growth in employment in unregistered manufacturing. 6.2 τ and Corruption: Evidence from the Interstate Variation We start by regressing our state-level estimates of τ against other established measures of the regulatory environment (see Table 2).33 These measures include the “Besley Burgess” (BB) measure of labor regulations from Aghion et al. (2008) and several measures from Dougherty (2009). The first is a measure of the number of amendments that a state government has made to the Industrial Disputes Act in either a “pro-worker” or “pro-employer” direction, as interpreted by Aghion et al. (2008), who update the measure to include amendments up to 1997.34 Positive values indicate more “pro-worker” amendments, which are assumed to imply a more restrictive environment for firms operating in those states. Dougherty (2009) also provides state level measures that reflect “the extent to which procedural or administrative changes have reduced transaction costs in relation to labor issues” Dougherty et al. (2014). Higher values therefore indicate an improved environment for firms. Dougherty’s measures are unique in that they cover a wide range of labor-related issues - not just the IDA. In the analysis below, we will focus on measures from Dougherty (2009) that cover reforms regarding 1) the Factories Act and 2) an overall measure of reforms. All relevant variables in our analysis have been rescaled to have mean zero and standard deviation one, with the goal of allowing comparability between regression coefficients in different specifications. In Table 2, correlations are reported between τ and the three measures both by themselves and while controlling for other factors (notably state GDP per capita and the state’s share of employment in manufacturing). The Besley Burgess (BB) measure does not seem to be correlated with τ (although our power is limited by the very small number of observations) while the two measures from Dougherty (2009) are significantly correlated after applying controls (though not all the correlations are strongly significant) and have the “correct” sign: states that saw more “transaction cost reducing” reforms have lower τ s. 35 On the one 33 Note that the estimates of τ we use in all the analysis below were generated using the procedure in Section 4.4 that we have argued is robust to possible misreporting and non-classical measurement error. 34 Since there have been few state-level amendments to the IDA between 1997 and 2005, this measure should be largely the same in 2005. 35 In this and most of the analysis ahead, we focus on the 18 largest Indian States, for which data are most 20 hand the lack of correlation between τ and the BB measure is not surprising, as the latter capture variation only due to state amendments to the Industrial Disputes Act, which does not vary over the ten person threshold. On the other hand, if the Besley Burgess measure is meant to capture the general regulatory environment (which is how it is used in countless studies), we might well expect it to correlate with our measure of regulatory costs. That the correlation does not hold may therefore be of interest. While the prediction regarding the correlation between τ and BB 97 may be ambiguous, that is not the case for Dougherty’s measures of transaction-cost reducing reforms related to the Factories Act. We should expect our measure of τ to correlate negatively with the latter, since the Factories Act does vary across the 10 worker threshold, and indeed we see that it does. τ is also correlated with Dougherty’s more comprehensive measure of reforms, one which aggregates reforms across all areas, although it does not appear to correlate with any other subcomponents (which are not depicted here). Table 2: Tau vs Other Measures of Regulations Besley-Burgess measure (regs) (1) tau -0.0734 (0.204) (2) tau -0.00841 (0.192) (3) tau 0.0481 (0.172) Dougherty measure (all reforms) (4) tau (5) tau (6) tau -0.289 (0.201) -0.407∗∗ (0.174) -0.408∗ (0.198) Dougherty measure (FA reforms) (7) tau (8) tau (9) tau -0.211 (0.187) -0.318∗ (0.166) -0.304 (0.181) log of net state domestic product pc -0.432∗ (0.226) -0.592∗∗ (0.214) -0.604∗∗ (0.206) -0.575∗∗ (0.241) -0.590∗∗ (0.217) -0.605∗∗ (0.257) share of employment in manufacturing -3.490 (4.886) 2.657 (5.090) 3.899 (3.281) 2.962 (4.934) 3.486 (3.435) 4.215 (5.163) share of privately owned establishments -11.66 (6.768) 2.003 (5.895) 0.529 (6.221) share of registered establishments 3.042∗ (1.530) -0.0440 (1.476) 0.654 (1.469) Constant Observations 0.641∗∗∗ (0.204) 15 5.315∗∗ (2.165) 15 15.79∗∗ (6.707) 15 0.653∗∗∗ (0.187) 18 6.207∗∗∗ (1.962) 18 4.259 (6.113) 18 0.631∗∗∗ (0.189) 18 6.091∗∗ (2.068) 18 5.481 (6.517) 18 * p¡0.10, ** p¡0.05, *** p¡0.01, Standard errors in parentheses Only including Major Indian States In addition to the above measures regarding state-level changes to the statutory, proceconsistently available and which offer the most precise estimates of τ (the power law relationship breaks down in smaller states when there are not enough observations). 21 dural and administrative aspects of the regulations, we also regress τ against certain other measures of the labor environment. Table 3 reports the results of τ regressed against per capita measures of strikes, man-days lost to strikes, lockouts and man-days lost to lockouts. One might imagine that strikes and lockouts capture relevant features of the regulatory and labor environment,36 but we do not find them to be robustly correlated with τ . One might also expect τ to be correlated with aspects of the regulatory enforcement. To test this hypothesis we regress τ against state level variables related to enforcement such as the number of inspections, convictions, and fines levied under various regulations.37 The results of the regressions for a subset of the enforcement related variables are shown in Table 4. In short, the only enforcement variable that is even close to being significantly correlated with τ is the percentage of factories registered under the Factories Act that have been inspected. However, as can be seen from the table, the enforcement data are only available for a small subset of the major states, leaving very little power in the regressions. Furthermore, the regressions shown exclude Uttar Pradesh, which is a substantial outlier in the enforcement data. 36 For example, some industrial regulations explicitly undermine or support the rights of parties to engage in strikes or lockouts. 37 These data were obtained from the 2005 Indian Labour Yearbook, which we were able to attain with the generous help of Anushree Sinha and Avantika Prabhakar of NCAER. 22