Evidence from a Regression Discontinuity Analysis in India

Formal Banking and Economic Growth: Evidence from a Regression Discontinuity Analysis in India Abstract This paper investigates the effects of formal banking expansion on economic growth. I exploit a previously unstudied reform to bank branching policy in India, which led to a large expansion in private credit to financially underserved areas. Using a regression discontinuity design at the district level, I trace the exogenous expansion of banking services through time and estimate their effect on agricultural and industrial growth. I am able to address concerns raised in other studies of policy driven banking expansion by accounting for the incentives generated by the reform in the context of India’s unique banking environment. I show first that the expansion of financial services led to significant growth in agricultural yield and output of cotton, wheat, and other important revenuegenerating crops. I then show that enterprises in states most affected by the reform experience faster growth in their total investments and capital labor ratios. Finally, I confirm the aggregate effect on local GDP growth by showing that areas with expanding banking services experienced higher rates of growth in nighttime light intensity in the years following the reform. The estimates imply that each additional private bank branch led to a 0.36% increase in local GDP. Overall, these findings offer strong causal evidence that the expansion of the financial system facilitates growth in productive activities important for driving economic development. 1 1 Introduction Access to credit expands the choice sets of households and firms, allowing them to smooth consumption and investments across time. The early literature on access to financial markets establishes the association between strong financial systems and economic growth King and Levine (1993); Jayaratne and Strahan (1996); Rajan and Zingales (1998). The intuitive force behind this connection is that productive firms, particularly small ones, often lack the ability to self-finance and rely on external resources to achieve optimal levels of investment and growth. The extent to which such firms remain credit constrained can result in hindered economic growth at the aggregate level. An obvious policy prescription to facilitate growth would be to adopt policies that broaden access to financial markets. In the context of developing economies, this can often mean a literal spatial expansion of bank networks into unbranched or under branched markets. The fundamental question is then whether policy induced branches actually affect firm and household access to credit. Though the answer may appear to be obvious and affirmative, a lower presence of the formal commercial banking system likely reflects underlying market characteristics. If banks are compelled under regulation to open branches in locations with high costs to doing business, extreme information asymmetries, or difficulties in aligning branch incentives with those of the bank, then new branches may fail to generate new bank business. Policies that do not address these underlying issues must instead focus on expanding branches in areas with a low banking presence explained by high fixed costs of entry. In either case, banks may simply be competing for market share with an informal lending sector, which can often be extensive in developing areas. If informal lenders adequately meet the needs of borrowers, formal credit expansion would be unlikely to produce additional growth. Accounting for the concerns raised above, this paper examines the effect of expanding access to the formal banking system on economic growth. To gain deeper insight as to the potential channels generating growth that may lead to poverty alleviation, I investigate responses in agriculture and manufacturing to exogenous changes in the availability of formal credit. Examining a previously unstudied branching policy reform in India from 2005, I rely on institutional knowledge of the reform and India’s banking environment to construct an empirical strategy and connect the timing of strong effects observed in the banking system to details dictating the reform implementation. The exogenous variation in access to formal banking generated by the policy reform facilitates the clear identification of the effects of banking on agriculture, manufacturing and growth. The seminal empirical paper on bank branching and development, Burgess and Pande (2005), exploits the timing of earlier reforms to show that a rapid expansion of the banking system in India during the 1977-1990 Social Banking Period corresponded with large reductions in the incidence of poverty. The analysis in the current paper identifies the potential deeper mechanisms underlying the effects of greater 2 access to formal banking. Identifying the effect of banking on specific channels of growth is generally difficult because banks tend to concentrate in profitable areas that are also likely to experience higher rates of economic growth. During times of policy intervention, banks may instead concentrate in poorer areas with slower growth. The bias from this endogeneity can vary widely, overestimating the impact of banks in the first circumstance and underestimating it in the second. Obtaining the necessary exogenous variation required to make causal inferences can be extremely challenging, particularly in development settings. Policy reforms, that may lead to natural experiments, if implemented at too fine of a geographic or demographic level may be impossible to evaluate for lack of granular enough data. Broader reforms may target areas receiving multiple interventions simultaneously making the effect of one particular mechanism inseparable from those of the others. Kochar (2011) studies the effect of banking on inequality in India during the Social Banking Period and raises just such a concern that the growth of subsidized credit distributed through the Integrated Rural Development Program (IRDP) correlates closely with branch expansion at the state level.1 The identification strategy pursued in this paper constitutes a key contribution as it allows the use of district level data across all of India, while still separating the effect of the exogenous expansion in formal banking induced by the policy reform from other interventions introduced around the same time. I exploit the selection process of the reform that designated certain districts as being under banked and encouraged entry in these districts relative to those not receiving the status. The rule, based on district population per branch relative to the national average, generated an environment exploitable via regression discontinuity analysis using the national average as a cutoff and each district’s population per branch as the running variable. Since the list of under banked districts remained essentially unchanged over the course of the policy, I am able to trace the policy effect on branching and credit, as well as agriculture and manufacturing, through time. I show the lack of a significant discontinuity in the pre-reform period, then a strong and accumulating expansion in private bank branches following the 2005 reform. I trace the effect through time by estimating the average treatment effect of the reform separately for each year from 2002 to 2012. The timing of the policy response generates additional insights into bank behavior under the regulations. Due to policy details, banks were able to delay branch openings for a period following the policy implementation. Private bank branching in under banked districts remained low during that window, after which it steadily climbed. Meanwhile, the credit extended by private banks exhibited an immediate response consistent with the revelation of information that signaled a pending reform to branch licensing. The disparity in timing be1 Kochar (2011)also discusses the timing and nature of branching policies from the Social Banking Period that are discussed in greater detail in an unpublished comment by Panagariya (2006) and a report by a working group in the Reserve Bank of India (RBI) (RBI, 2009). 3 tween branches and credit supply is consistent with the optimal strategy of profit maximizing banks that exhibit market power and anticipate future entry. Such banks may expand credit in anticipation of intensified future competition to lock in consumers who face positive switching costs. I provide a simple theoretical framework to outline the intuition for the incentives resulting from the confluence of the reform and India’s banking environment. I estimate the strongest responses to the reform from the private sector of India’s banking system which is consistent with predictions based on the incentives generated around the cutoff. I cannot rule out, however, the possibility that the public sector also increased banking services due to the reform, but did so in districts away from the cutoff. I draw on several different sources for the data in this analysis, including India’s central bank the Reserve Bank of India (RBI), the Ministry of Agriculture, India’s Annual Survey of Industries (ASI) and remote sensing data on the amount of light emitted at night and measures of rainfall. The detailed data from the RBI on bank branches and credit, from separate data sets, help provide a cross check for the two broad banking outcomes. Additionally, with credit reported annually at the district level with further disaggregation by bank group, population group and industrial sector, I am able to examine more detailed banking effects such as the expansion of credit to agricultural activities in rural and semi-urban areas following the reform. I combine separately reported data on district level crop production statistics and farm harvest prices from the Ministry of Agriculture to examine responses in agricultural productivity. I conclude that the policy reform resulted in a significant expansion of banking services by the private sector in underserved areas. The cumulative effect of the reform is estimated as an average additional 10 private bank branches per district by the start of 2012. This constitutes approximately 50% of the sample average of operating private branches per district in districts around the threshold. I find evidence that private banks exploited the timing allowed in the policy to delay branch openings in under banked areas. In contrast, credit responded to early information regarding the reform consistent with banks racing to secure market shares. In 2006, the reform had already induced an average increase of 6,725 private bank accounts for under banked districts, approximately 52% of the sample average around the threshold. Growth in the credit extended for agricultural use in rural and semi-urban areas of under banked districts demonstrates that reform effects were not solely concentrated in high population areas. This is an important finding due to the popular concern that banks, particularly those from the private sector, exclude rural areas from service. Increases to agricultural productivity that are consistent with the expansion of credit are observed. Agriculture constitutes a major employment activity in India, with over 56% of workers in 2001 engaged in agricultural endeavors. Further, policy makers placed particular importance on the availability of credit to rural and semi-urban agriculture leading up to the reform, suggesting early effects may concentrate in this sector. Positive responses in yield 4 (output per hectare) and raw output are estimated for several important crops including cotton and wheat. These results are suggestive of a positive effect of banking on agricultural productivity, though a measure capturing responses across crops and incorporating harvest prices is preferred. Turning to an index of crop yield across several important revenue crops in India that weights by each crop’s share of district revenue, I estimate an increase of 1,000 private bank credit accounts in a district raises average crop yield by 3%. This effect is a little less than half of the effect Jayachandran (2006) measures on crop yield from positive rainfall shocks, measured as district rainfall being above the 80th percentile of rainfall for that district. This result may reflect the ability of farmers to apply higher quality inputs purchased with credit, such as fertilizers or machinery. A redistribution of crop selection by high productivity farmers responding to loan availability may also contribute to this effect. Effects are also observed in the amount of borrowing and production activities in manufacturing using data from the ASI. To analyze the ASI data that are available at the state level, I identify a set of treatment and control states based on the share of a state’s population close to the threshold on one side or the other. I then perform a difference in differences analysis to estimate the effects of the reform. I estimate that enterprises in states with populations most affected by the reform experienced faster growth in the amount of loans they carried in the order of 23%. This result is consistent with the reform affecting the availability of credit to manufacturing. These enterprises also reported higher total investments, working capital and capital labor ratios following the reform. These responses are consistent with manufacturers being credit constrained in the pre-reform period and expanding their levels of capital given increased access to credit. Finally, I confirm the aggregate effect on local GDP growth by showing that areas with expanding banking services experienced higher rates of growth in nighttime light intensity in the years following the reform. Henderson et al. (2012) established that so called “nightlights” provide a reliable proxy for economic growth under certain caveats when regularly reported data on traditional measures are unavailable, as is the case for district level GDP in India. Taking the estimated elasticity of nighttime light to GDP from Henderson et al (2012) of 0.3, the effect estimated in the current analysis implies that each additional private bank branch led to a 0.36% increase in local GDP. Overall, these findings offer strong causal evidence that the expansion of the financial system facilitates growth across productive sectors and encourages economic development. 1.1 Related Literature These results are largely consistent with two analyses examining effects of branching in the United States. Dehejia and Lleras-Muney (2007) examine the effect of two forms of financial development in the United States from 1900-1940 on agricultural and manufacturing sector 5 development. They find that while increased bank branching encouraged growth in these sectors, deposit insurance had negative effects. Krishnan et al. (2014) show that increased branching activity in the United States, following the Interstate Banking and Branching Act of 1994, led to greater efficiency gains by previously credit constrained manufacturers. Unlike the policy reform in the current analysis that directed banks toward targeted areas and resulted in policy driven branching, the mechanism of branch expansion in these other two analyses was legislation granting banks greater ability to branch, enabling banks already wishing to expand to enter new markets. The results in the current paper confirm the positive effects of branching on both agriculture and manufacturing. The three analyses differ in their consideration of manufacturing. Dehejia and Lleras-Muney (2007) focus on expanding labor demand in manufacturing; Krishnan et al. (2014) explain increased efficiency through the adoption of more productive projects, while the current analysis focuses on capital use, showing increases in investment and capital intensity in production. To the extent that firms remain capital constrained, lower aggregate TFP due to resource misallocation as argued for the case of India in Hsieh and Klenow (2009) may be partially attributed to lower financial access. The consistency between the broad effects is not surprising given that the expectation in the current context on the private sector banks is to compete and expand access to credit conditional on entry. The differences that I observe between private and public sector banks, in their responses to the reform, likely reflects differences in the incentives and objectives generated under those respective ownership structures. La Porta et al. (2002) show that a higher incidence of government ownership in banking is correlated with slower growth looking across countries. In a series of joint and separate papers, Banerjee, Cole and Duflo examine the activity of banks from the public sector in India (Banerjee and Duflo, 2001; Banerjee et al., 2004; Banerjee and Duflo, 2014; Cole, 2009). They show evidence of under lending to productive firms, inertia in credit limits extended to firms and little difference in delivering development oriented lending resulting from government ownership. The main argument for these effects are misaligned incentives within banks, with loan officers facing few benefits from financing productive projects but punishments for loans that fail. De Quidt et al. (2013) demonstrate theoretically how market structure can greatly effect financial sector outcomes in the context of microfinance, comparing for-profit and non-profit lenders. The current analysis addresses these issues of bank ownership and incentives by analyzing private and public sector banking responses separately. The private sector, which is more likely to face profit maximizing objectives, shows a strong response to the reform around the cutoff. The public sector shows little evidence of a response near the cutoff, but this may reflect a different set of objectives that could concentrate their efforts in districts away from the cutoff, where my identification strategy does not apply. These findings highlight the importance of accounting for the institutional environment in crafting policy interventions. 6 In the next section I describe the important institutional aspects of India’s banking system and the policy reforms to the branch licensing policies utilized for analysis. In section 3 I outline a simple theoretical framework to provide intuition for potential responses to the policy reform. In section 4 I review the regression discontinuity framework and describe how I translate its principles for analyzing the manufacturing sector with difference in differences. In section 5 I describe the data used in analysis. Then in section 6 I first establish a clear response in branching behavior to the policy reforms, then identify corresponding responses in aggregate private sector credit. I then examine responses in the agricultural sector followed by manufacturing. I then present results on overall growth using nightlights. Section 7 concludes. 2 Policy Reform and Institutional Background 2.1 Policy Reform The Master Circular on Branching Authorisation Policy released September 8, 2005 implemented the policy reform on branch licensing utilized in this paper. The banking sector in India does not permit free entry of banking firms or branches. New bank licenses are granted infrequently by the Reserve Bank of India (RBI), India’s central bank, through special campaigns with recent waves in the early 1990s and again in the early 2000s. Banks must also acquire licenses prior to opening all new branches, as well as receive permission to close or shift branches in most markets. Prior to the 2005 reform, banks applied for each of these changes on a case-by-case basis through the regional office of the RBI. No broad directive with regards to the composition of markets served by the bank, such as a requirement to open branches in rural areas, existed following the end of the Social Banking period in 1990.2 The reform in 2005 changed the regulatory environment in two fundamental ways. First, the reform effectively tied new branch licenses for highly sought markets to branch entry in markets designated as under banked. Specifically, banks were issued a set of criteria by which they would be judged during the review of proposed licenses. The “nature and scope of banking facilities provided by banks to common persons, particularly in under banked areas” would be considered when granting new licenses. In addition to offering “no-frills” bank accounts, meeting priority sector lending obligations and instituting a system for receiving and addressing customer complaints, banks were encouraged to open branches in “under banked districts and rural centres.” The RBI provided a list of under banked districts with the circular. Though not stated explicitly, I will argue that a form of quota system operated requiring expansion in under banked districts for entry in rich markets. Second, the case2 The LEAD banking scheme was in operation during this time, however, by which one bank was assigned to each development block and made responsible for meeting agreed levels of branching and banking services. These banks were typically selected from the set of government owned banks. The service area approach (SAA) also operated at this time, partitioning rural areas between banks for implementing development objectives. 7 by-case application procedure for licenses was substituted with an Annual Branch Expansion Plan (ABEP) framework. Under the new system, each bank would prepare a set of proposed network changes (branch openings, closings and shifts) to be implemented over the next year. The plan would be submitted to the RBI for review, after which the bank management would meet with RBI officials to revise and finalize a set of permissions to be valid for the next year (Master Circular Branch Authorisation Policy, 2005).3 The rule governing the assignment of under banked status was based on the district average persons per branch relative to the national population per branch for India (RBI Report 2009). The spatial implications of branch licensing from the reform around the national average cutoff provide the identifying variation exploited in this analysis and is discussed in detail in section 4. Important differences exist between the above policy and those implemented under Social Banking. The degree of choice given to banks in selecting locations in which to open under the 2005 reform far exceeds that available during Social Banking. Unlike the 4:1 entitlement policy studied in Burgess and Pande (2005), that required intervention branches be opened strictly in unbanked markets, banks could choose among any markets within under banked districts to satisfy their obligation, allowing for the potential of increased direct competition between branches and banks. In stark contrast to the planned approach to district-wise branch expansion implemented in the 1980s (RBI, 2009; Kochar, 2011), banks under the current reform could choose between under banked districts for entry, as well as decide their own level of total entry, which affected their amount of entry in under banked branch districts. Finally, the banking environment differed drastically in its composition and scope of business. The private sector, largely inert under social banking, expanded and gained vitality following the deregulations beginning in 1990 and infusion of “new private” banks. Government owned banks, consisting of the State Bank of India and its Associated Banks, the set of nationalized banks, and most regional rural banks (RRBs), have traditionally dominated the banking system in India. Following reforms and deregulation after a current account crisis in 1991, a sizable private sector developed, operating alongside government owned banks. The entering new private banks were heavily vetted and selected from many candidates during a period of open applications in 1993 and again in 2001. According to RBI documents, the purpose of these new banks was to foster competition and modernize the banking system. The new private banks broadly face the same regulation as the other scheduled commercial banks, though carry the additional mandate of maintaining at least 25% of their branch network in population centers with fewer than one hundred thousand people. The other policies they face, as well as their requirements to the Priority Sector lending scheme, are identical to those on the SBI and Nationalised banks. RRBs and foreign banks face tailored regulations, including those pertaining to branching requirements. 3 Permissions were valid for one year with the potential for extensions. Banks accomplishing 75% of their planned expansions could submit their next ABEP regardless of the lapsed time. 8 2.2 Policy Details and Timing While the reform became official in September 2005, events leading up to its release likely provided signals as to its impending introduction. In a speech from December 2002, the director of the RBI pointed to the high share of bank investments in government securities, 39% relative to the regulatory minimum of 25%, encouraging banks to expand their commercial lending particularly in small manufacturing and agriculture (Mohan, 2002). The following November, the Vyas Committee was commissioned to investigate the flow of capital to agricultural activities. They met with several commercial banks during their investigation. In April 2004, they released an interim report followed by the final report in June, suggesting revisions to the service area approach (SAA) and encouraging greater lending by private and public sector banks. The report included a map identifying areas underserved by the formal banking sector, some of those identified as places where the “branch network of commercial bank[s] [is] below the national average (Vyas Committee Report, 2004).” The SAA program was subsequently discontinued, allowing all banks to freely apply for entry and operate in rural areas. The official list of under banked districts released in 2005 almost exactly matches selection based on district average population per branch relative to the national average, consistent with the language in the report. Thus, aspects of the Vyas Committee report could have provided solid signals to banks of the forthcoming reform. The list of under banked districts initially released in 2005 was reissued in 2006 adding a small set of districts that satisfied the under banked requirement in both years but were left off of the 2005 list. Afterward, the list was reissued each year unchanged until 2010.4 After 2010, certain states were made ineligible for under banked status, reducing the number of districts considered as “under banked districts of under banked states,” but not introducing any new districts to under banked status. Although additional reforms altered the incentives for branch expansion both within and outside under banked districts, given the lagged nature of branch openings to license issuance, I would expect to and do find lasting effects through 2012. In section 4, I discuss the algorithm used for assigning under banked status to districts in detail and how I exploit it following a regression discontinuity design strategy to identify the effect of exogenous expansion in formal banking on real economic outcomes. Although the reform became effective immediately upon its release, banks were essentially allowed a year long grace period to construct their first ABEP, with an implicit deadline for September 2006. Several banks, many of them from the private sector, waited close to the full year to submit their ABEP, during which time they were able to receive licenses in a 4 Starting in 2008, certain centers within under banked districts were made ineligible to count toward a bank’s serving of common persons. Specifically, centers within the municipal limits of state capitols, district headquarters and metropolitan centers were deemed ineligible. Further, centers within 100 km of Mumbai, New Delhi, Kolkata and Chennai, and 50 km of state capitols were ineligible. Exceptions were made for the state of Jammu and Kashmir, and the seven North Eastern states, Arunachal Pradesh, Assam, Manipur, Meghalaya, Mizoram, Nagaland and Tripura. 9 disaggregated manner. The histogram of branch license dates for a large private sector bank is shown in figure 8.5 Although annual branch expansion plans may not be observed directly, the large spikes in branch licenses set approximately a year apart are consistent with ABEPs. The figure shows the licenses from the first likely ABEP for this bank were granted in July 2006, roughly one year after the reform implementation. Similar patterns are identified for many private sector banks. Additionally, the licenses from ABEPs remain valid for a year, meaning that banks could effectively postpone the effect of the reform for nearly two years if preferable to quick entry. The optimal timing for entry from the perspective of the banks will depend on strategy and the underlying profitability of the locations. Early entry may allow banks to secure market shares, though they could delay costly entry into low profitable areas by waiting. The empirical evidence suggests most private sector banks chose to delay entry in locations of induced entry. Finally, the shifting and closing of branches, particularly in under banked districts, was heavily regulated. Branches were not allowed to shift outside otherwise unbanked centers. Given the source location was served by another commercial bank branch (other than a RRB), a branch could only shift to centers in the same or lower population group classification, and in the case of branches in under banked districts, could only shift to centers within under banked districts. Little branch closure is observed in the data, though mergers and acquisitions of banks occur during which most branches are “closed” and “reopened” under the acquiring branch, with some branches converted to satellite offices and fewer still permanently closed. 2.3 Policy Reform Discussion Incentives The 2005 branch licensing policy reform purposefully created new incentives for scheduled commercial banks to open in centers conditional on their district’s under banked status. Licenses for branches in high profit potential centers in banked districts were used to leverage bank entry into under banked districts. This mechanism works most effectively during periods of high demand for bank branches in “rich” areas, as was presumably the case experienced in India during its time of high economic growth beginning in 2003 and continuing through the decade. The branching policies and reform placed no requirements on the amount of banking required to occur at each branch. There are staffing requirements for branches, as well as minimal days and hours of operation. Banks must also offer “no-frills” accounts that carry limited fees and low minimal balances to prevent the exclusion of poor customers. Despite these requirements, though, banks could maintain staffed branches that simply minimized costs by not reviewing or approving any loan applications, not pursue new customers, and only accept deposits. 5 Known acquisitions of branches from other banks have been excluded for the histogram analysis. 10 Banks are also required to meet Priority Sector lending ratios. Banks must maintain 40% of their outstanding credit in loans to the priority sector. However, the requirement must only be met at the bank level, meaning some branches may carry heavy amounts of priority sector loans while others lend nothing to the priority sector. In 2007, new guidelines were adopted for the priority sector, reducing the set of loan categories eligible for priority status.6 The reformed guidelines concentrate lending in direct and indirect agricultural endeavors and limited the amount going to microfinance institutions and other indirect modes of lending. The priority sector reforms apply uniformly at the national level. Banks failing to meet their 40% requirement must make up the difference with loans to the NABARD RIDF fund at deterrent rates. Banks typically come very close to meeting the requirement, overshooting slightly in some years and falling short in others. 3 Theoretical Framework This section articulates a simple theoretical framework to provide intuition for the effects of the reform and the heterogeneity of responses across districts and bank groups. The theoretical framework demonstrates how the 2005 policy reform may have incentivised higher rates of entry in under banked districts and increased lending without addressing the underlying profitability conditions of those districts. Further, it rationalizes an expansion of credit in under banked districts following the policy announcement, prior to a significant increase of branches. The framework begins from a standard characterization of financial intermediation with adverse selection of borrowers common to credit markets in developing economies. Consider a single market with two periods and two types of borrowers, safe and risky. In the first period, a policy reform that will encourage entry in a (potentially unknown) set of markets beginning in the second period is announced. In the second period the reform is in effect. As in Stiglitz and Weiss (1981), each borrower has a potential project that requires a loan (normalized to size one for all borrowers) and yields the same expected return across borrowers. The borrower is assumed to have the same potential project in each period. Assume that the return from a failed project is zero, so that Ps (RsA )RsA = Pr (RrA )RrA , where RiA is the return from a successful (denoted A) project for type i ∈ {saf e (s), risky (r)} andPi (RiA ) is the probability of success for type i. Thus, safe types have projects with lower returns conditional on success but succeed with greater probability Ps (RsA ) > Pr (RrA ). If banks operate in the market, they can offer a standard debt contract with fixed repayment. Assume RiA > (1 + ri ) > 0 and that borrowers face limited liability, such that when a project is successful the borrower pays back the principle on the loan plus interest at rate ri , but that in case of failure no payment 6 The reforms to the composition of the priority sector studied in Banerjee and Duflo (2014) occurred in 1998 and 2000, prior to my analysis. 11 is made and both borrower and bank receive zero. Borrowers face an outside option that provides utility equal to µ. Both borrowers and banks discount the future at rate δ and are risk neutral. While borrowers know their own type, banks only know the distribution of types and the parameters defining the projects. Banks prefer to lend to the safe types due to limited liability but cannot distinguish between types in the general framework. Depending on the set of parameters just described and the share of safe and risky types in the population, banks may choose to ration credit in response to adverse selection, or the market may collapse entirely (Stiglitz and Weiss, 1981). To capture the dynamic effect of the policy reform, consider the two following modifications: 1) Banks possess a screening technology that reveals a potential borrower’s type with certainty and costs amount s. 2) There exists a downward sloping demand curve among safe types. A wide set of assumptions can satisfy this condition, for example, if personal costs of marketing the successful project differs between borrowers then demand for loans will decrease in rs . To simplify the analysis, assume parameters are such that banks always choose to screen borrowers and never find it profitable to lend to the risky types. Further, the banks pass the cost of screening on to the borrower. As long as borrowers must repay the full amount of the loan conditional on a successful project, and borrowers cannot accept contracts with the potential for negative consumption, the expected default rate from safe types will remain unaffected and banks will know the demand conditional on the interest rate offered with certainty.7 This assumption greatly simplifies the game as it allows the borrower’s decision process to be considered separately for each period since current negative expected returns cannot be offset by more favorable expected lending conditions in the future. Assume banks are symmetric and profit maximizers, each facing an exogenous marginal cost of funds, including administrative costs from lending, equal to (1 + ρ), and cannot discriminate in the interest rate it offers to repeat versus first time borrowers. Since banks observe the parameters on the population defining the distribution of safe types, they know the slope of the demand curve, though do not know any particular borrower’s value of the loan. Without the threat of entry, a monopolist serving the market in the first period maximizes profits by serving the same set of borrowers in each period, increasing the interest rate in the second period to capture the additional surplus the borrowers receive from not paying the screening cost again (a sketch of the proof is given in the appendix). Knowing this, the monopolist may work backwards from the second period to determine the profit maximizing interest rates in each period. In contrast, when two banks serve a market, they compete in prices in both periods. If both enter the market in the same period, then each offers the zero profit interest rate and split the market. 7 A contract with potential negative consumption would arise when limited liability protects the borrower against a failed project, but not from a successful project for a borrower whose high marketing costs leaves them less from the project than the fixed payment owed to the bank. 12 However, if one bank acts as an incumbent, then it may choose to alter its behavior when anticipating the potential of entry. The screening cost, which may be construed as the cost to the borrower of filling out paper work or the effort of establishing a relationship with a loan officer, operates as a switching cost for the borrower.8 Borrowers will go to whichever bank results in them keeping the highest expected return from their project. For first time borrowers this is simply the bank offering the lowest interest rate. Repeat borrowers must compare their expected payoff from the incumbent’s 2nd period interest rate to that of the entrant plus the screening fee required to switch. The resulting equilibrium is intuitive: in the second period, under cutting leads the entrant to offer the zero profit interest rate and the incumbent offers an interest rate making its set of first period borrowers indifferent between switching to the entrant and staying. Since the set of first period borrowers is entirely determined by the first period interest rate, the second period interest rate is a function of the first period interest rate and the screening cost. Knowing this, the incumbent chooses the first period interest rate that maximizes profits over both periods. The threat of entry will result in the monopolist offering lower first period interest rates to secure a larger base of customers from which to earn positive profits in the second period. The set of parameters will determine how willing the incumbent is to trade off first period profits for those in the second period. The entrant will serve the remainder of the market that demands loans at the zero profit condition. Thus, credit will initially expand with the announcement of the policy reform and again upon realized entry. 3.1 Entry The effects on entry must be primarily driven through changes to the structure of fixed costs of entry as the reform did not otherwise target local market conditions. Consider multiple markets described by the framework above. Markets are differentiated by their set of parameters already discussed plus overall market size. Suppose banks each draw market specific fixed costs of entry for every market. Abstracting from the strategic considerations of entry, assume banks act myopically such that they expect to act as a monopolist if entering a market unbanked in the first period or as a duopolist when entering banked markets. Under these assumptions, expected profits for each market is known to a bank and entry will occur j for all markets j satisfying E[πB ] − Fj > 0. Markets with low profit potential or high fixed entry costs will fail to attract banks. Consider a rule that ties permission for entry in some high profit potential markets to entry in lower profit ones. Banks facing binding constraints will now open into markets where U B1 ] − F E[πB U B1 < 0 if these losses may be offset by the profit gains from the rich market, j U B1 ]−F −F E[πB ]+E[πB j U B1 > 0. This condition will be more easily satisfied in policy eligible 8 Klemperer (1987) lays the groundwork for considering the effect of switching costs. 13 districts with higher expected profits that faced unfortunately high fixed entry costs. Once entered, however, these markets may produce high levels of banking activity. In contrast, the set of markets originally served without the reform may contract if the lowest profit earning locations cannot offset the losses from policy eligible markets. Finally, the joint positive profits will be hardest to satisfy for policy eligible districts that face the lowest profit potential and highest fixed costs of entry. The reform will be unlikely to produce positive banking results for such markets. 3.2 Predictions The above framework suggests three main predictions of bank responses to the policy reform. First, the amount of credit will expand in districts where increased entry under the reform is expected to occur. This will result from banks attempting to secure market shares in these districts so as to lock borrowers in through switching costs prior to the intensive entry under the reform. Second, branch entry will increase in under banked districts with high profit potential that was offset by high fixed costs to entry which become subsidized due to the reform. Further, entry may be most profitable in locations where banks open as an entrant, with lower fixed costs making up for stronger competition for borrowers. Thus, both entry as monopolists and as competitors is possible. Finally, this framework characterizes behavior under profit maximization. Banks following other objective functions, as public sector banks may do, would be less likely to generate these responses. Comparing the behavior of private sector and public sector banks will provide a qualitative test of predictions from the framework. 4 Empirical Methodology Identifying the effect of bank branching on banking and real economic outcomes can be frustrated by classic endogeneity concerns outlined in previous work Burgess and Pande (2005), in which selection bias can overpower estimates, even changing their signs. The unique policy aspects of the 2005 branching reform create an environment facilitating the identification of banking effects on agricultural and industrial outcomes. I am able to circumvent endogeneity concerns and separately identify the banking effects from other simultaneously operating reforms by employing a regression discontinuity design that yields transparent estimates and identification founded on assumptions that are at least partially testable. First I identify and quantify the expansion of banking services in response to the policy. Next I focus on the effects of banking in agriculture, which appears to be an initial motivation for the reform and the largest employment activity in India. Then I turn to the effect of banking on manufacturing enterprises, which appeared to gain from the realized expansion of bank branches. After establishing a response in these two areas, I provide evidence of a positive effect on overall 14 growth using light emitted at night as a proxy. 4.1 Regression Discontinuity The method employed by the RBI for identifying districts as under banked in the 2005 branching policy reform, based on simple district and national averages of population per branch, yields a clear quasi-natural experiment exploitable by regression discontinuity techniques. Under banked districts were identified using two inputs. First, the national population of India, taken from the Population Census conducted in 2001, was divided by the total number of scheduled commercial bank branches operating in the country in 2005-2006 to obtain a “national average of population per branch.” Then an analogous value was calculated for each district and compared to this national average. Those districts with a calculated value higher than the national value were designated under banked. Figure 4 shows district under banked status from the 2006 list of under banked districts plotted against district population per branch. According to the rule, districts to the right of the cutoff should be assigned to under banked status, as is broadly confirmed in the graph.9 A map of the districts in India with their corresponding district averages is presented in the upper panel of figure 5.10 The above algorithm induces a cutoff at the value of the national average, treating district population per branch as the “forcing variable.” The policy generates an arbitrary difference in districts falling on the “under banked” side of the cutoff, which offer an additional value to banks opening branches within their borders: such openings count toward their requirement for “serving common persons” to gain permissions for branches in rich markets. Districts falling on the other side of the threshold do not offer this benefit, though being otherwise similar, which will be tested formally. Thus, the policy effects the probability that the districts will receive additional branches through its manipulation of bank incentives. This estimation strategy will be valid if the distribution of potential outcomes is continuous at the cutoff(Lee, 2008). A lack of perfect manipulation of the running variable so as to change a district’s treatment status, and the continuity of other factors that may affect the outcomes of interest with respect to district population per branch near the cutoff will suggest this assumption is satisfied. I verify that both of these stipulations hold below. Figure 6 presents visual results from the McCrary test for manipulation of the running variable around the threshold McCrary (2008). The distribution of districts along the running variable is shown to be smooth around the threshold. The discontinuity estimate in the log difference in height is 6.6 with a standard error of 22, thus I fail to reject the null hypothesis of continuity. The figure also highlights another ideal trait of this environment; the 9 Six districts do not follow the assignment rule, with four of them remaining in the sample used in estimation (see the section on constructing the forcing variable in the Data Appendix for details). 10 The districts with greater deficits of branches per person, denoted by darker colors, matches closely with the areas identified as being more broadly under served by the map from the Vyas Committee issued in 2004. 15 cutoff is located near the peak of the density, meaning most districts fall close to the cutoff, suggesting the generalization of the effect for most districts may be reasonable. The lack of manipulation around the cutoff, beyond passing the McCrary test, is extremely defensible on intuitive grounds. Even if banks and districts were able to perfectly anticipate the criteria for assigning under banked status, their ability to manipulate assignment would be limited. The population level in the current equation was taken in 2001, four years prior to the policy. Thus, agents attempting to influence district status could only do so through altering the number of operating branches within district boundaries, which results from the collective branching decisions of all banks and conditional on RBI permissions, making manipulability extremely unlikely. Figure 7 presents a series of plots of district baseline characteristics, with dots reporting local averages for districts falling within 200 persons per branch non-overlapping bins. A local linear regression of the data is shown with flexible slope on either side of the cutoff. While the figures constitute a visual RD testing for continuity at the cutoff centered at zero, they also summarize broader trends in branching at the time of the policy reform. Districts left of the cutoff enjoyed more branches per person by definition. These districts also tended to be places with higher populations living in large cities, exhibited higher literacy rates, had lower populations of scheduled caste and tribe persons and had a lower percentage of main workers engaged in agriculture. Each of these characteristics appears to be smooth at the cutoff, suggesting proper randomization of districts around the cutoff. The continuity is tested formally by performing RD analysis with the baseline characteristics as the dependent variable. The tests fail to reject the null hypothesis of continuity at the threshold, with reduced form results presented in table 1. 4.1.1 Technical Details of RD The identification of local average treatment effects through regression discontinuity analysis is now well established in the literature (Black, 1999; Angrist and Lavy, 1999; Van der Klaauw, 2002; Lee et al., 2004), with the theoretical work on identification in Hahn et al. (2001) and the origins of the method in Thistlethwaite and Campbell (1960). To reduce bias from including observations far away from the cutoff where the identification does not hold, I use local linear regressions, dropping observations outside a set bandwidth of the cutoff (Hahn et al., 2001; Lee and Lemieux, 2010). I restrict all analysis to local linear and local 2nd degree polynomial regressions as recommended in (Gelman and Imbens, 2014). I set the bandwidth at 3.5 thousand persons per branch for all regressions, which falls within the range of optimal bandwidths selected for individual years by the Imbens and Kalyanaraman (2011) method.11 I fix the bandwidth to provide transparency for tracing the evolution of the policy effect across 11 Results are robust to different bandwidth selections, and 2nd degree polynomials typically perform better with wider bandwidths than linear specifications as in the example from Lee and Lemieux (2010). 16 years, as this fixes the set of included districts across regressions. In Figure 5 the lower left map indicates districts included in the local linear regressions. The districts are spread out geographically across most of the country, with under banked districts typically not far from banked counterparts. The map in the lower right panel identifies districts close to the cutoff on either side. Again, districts are distributed geographically and tend not to bunch by under banked status. For each year, I first estimate the local linear regression of the reduced form equation, yi = α + Di τ + f (P opP erBranch − Cutof f ) + δXi + i (1) using a uniform kernel. yi denotes a banking or economic outcome of interest in district i, such as the number of operating bank branches or crop yield. Di = 1[P opP erBranchi − Cutof f ≥ 0] is an indicator for satisfying the rule for assignment to under banked status, P opP erBranchi is the population per branch for district i, f (·) is a flexible functional form, Xi is a set of controls, τ is the coefficient of interest measuring the discontinuity at the threshold, and i is an idiosyncratic error. In all regressions, I include the pre-random assignment value of the dependent variable from 2001 to improve precision and reduce sampling variability (Imbens and Lemieux, 2008; Lee and Lemieux, 2010). In addition, I include the 2001 district population and its square to further improve precision. The described method constitutes the reduced form estimate, with the probability of under banked status instrumenting for actual assignment. These values are reported graphically. I report the “fuzzy RD” results implementing the regression discontinuity using Calonico, Cattaneo and Titiunik’s “rdrobust” package with a triangular kernel. I use the fuzzy RD with banking outcomes because the rule I use to assign under banked status does not perfectly match the list.12 When considering agricultural and other economic outcomes, their most interesting relationship is to realized banking behavior, rather than district assignment to under banked status. For these outcomes, in addition to the reduced form that presents the general effects of receiving under banked status, I estimate the effects with the fuzzy RD instrumenting for banking outcomes. That estimate will inform the effect of the specific banking outcome on the outcome of interest. However, that effect should be treated with caution as reform status will effect several dimensions of banking outcomes. To implement the fuzzy RD analysis I first “residualize” the data, regressing yi on the set of controls Xi from equation 4.1.1, then estimating equation 4.1.1 replacing the left hand side variable with the residuals obtained from the first regression and dropping the controls from the specification (Lee and Lemieux, 2010). Conventional estimates of the RD are reported, as are bias-corrected 12 I fail to match 6 out of 572 districts to their realized under banked status from the 2006 list. See the data appendix for details. 17 estimates and the robust standard errors from Calonico et al. (2014). I will focus on the conventional estimates in discussing results. 4.1.2 Dynamic Strategy The identification of the policy effect on banking outcomes is bolstered by the ability to regularly estimate the effect of the reform through time, both before and following its implementation. In the pre-reform period, no discontinuity should exist at the cutoff. In the post-reform period, the effect of the policy should be expected to grow according to the timing set in place by the rules of the reform and its revelation. To demonstrate the timing of the reform effects, I estimate equation (4.1.1) separately by year for banking outcomes, agricultural outcomes and outcomes measured through remote sensing including rainfall and nighttime light emitted into space. Given that the set of under banked districts remained essentially unchanged in the reform period considered, this captures the short and medium term effects as they emerge. 4.2 Effects observed in Manufacturing To examine the effect of increased financial access on the manufacturing sector, I use ASI data available at the state level. The level of aggregation prevents conducting the regression discontinuity just described. Instead, I follow a difference in differences approach, utilizing the institutional knowledge of the reforms to construct sets of treatment and control states. I select the set of “under banked treatment states” in the following way. Using population census data at the district level, I construct the shares of state population in under banked districts. For the population of each state in under banked districts, I calculate the share of that population belonging to districts falling within a close bandwidth of the national average of population per branch, generally within 4 thousand persons per branch. Those states with large shares of their population in under banked districts close to the threshold are selected as the treatment group. I then construct a control group using a comparable procedure from districts with banked status. “Banked States” include Haryana, Uttarakhand, Punjab, Mizoram, Daman and Dimiu, Karnataka, Puducherry, and “Under Banked States” include Rajasthan, Tripura, Jharkhand, Orissa, Dadra and Nagar Haveli. For each treatment and control group pairing, I estimate the following, yit = α + ξpost06t ∗ treati + ϕpost06t + ψtreati +β1 yeart ∗ statei + β2 yeart + β3 statei + β4 Xit + ωit (2) where post06t indicates financial years 2006 and later, treati indicates that the state belongs to the treatment group, and the remaining terms indicate controls for state fixed effects and state specific time trends, as well as a matrix of additional controls in Xit with an 18 idiosyncratic error ωit . The coefficient of interest will be on the interaction term post06t ∗treati , which will give the difference of within-state differences between the states receiving under banked status and those not. In addition to controlling for post 2006 and treated state individual effects, the regressions include the logged number of manufacturing units in the firm and the logged number of employees in the enterprise to control for enterprise size. Plant age and its square are also included as controls as these may influence the firms’ access to credit and capital markets. Although this identification strategy is not as ideal as the RD, the careful selection of the treatment and control states should help in eliminating potential threats and I will take the estimate as suggestive of the effect from the policy reform on manufacturing. 5 Data The primary data on banking are from data sets maintained by the RBI. The Master Office File (MOF) provides a detailed record of bank branch locations and characteristics, from which detailed branch network information by bank may be constructed. I have also matched most branches to approximate geocoded locations based on postal codes (PIN) and center names. In the present analysis, the MOF is used to construct the running variable and national average based on population per branch at the district and national levels respectively. The Basic Statistical Returns 1, 2 and 7 provide time series data on credit and deposits at various levels of aggregation. The empirical methods and analysis pursued in this work is greatly determined by the level of data availability. Although branch location data are available in detail through time by bank, much of the credit and deposits data are only available annually as aggregates to bank group level by district. Thus, although it is possible to observe which banks are operating within a district, assigning a certain number of accounts or amount of credit from any particular branch is impossible. Fortunately, the policy reform applied at the district level, allowing analysis directly at the level of the reform. Utilizing the time dimension further helps to disentangle some effects of the reform from changes to bank group classifications. The availability of credit limits, amounts and accounts by the intended geographic utilization of loans constitutes a strength of the data used in this analysis. The use of Call Reports from banks do not typically allow for this level of geographic precision of where loans are actually directed. This feature strengthens the arguments that loans reported in certain areas are not financing projects in neighboring districts. To conduct the analysis on agriculture, I develop a new data set by processing and combining separate annually available data from the Ministry of Agriculture, Directorate of Economics and Statistics on crop production statistics and crop farm harvest prices. By matching district production levels to farm harvest prices by crop, I am able to construct an index of 19 crop yields similar to that in Jayachandran (2006) for crop years 2002 - 2008. The use of an index circumvents certain concerns arising from differences in crop suitability across districts. Data on manufacturing enterprises are from the Annual Survey of Industries, reported annually for registered firms. Measures from enterprises with fewer than 100 employees are taken from a 20% sample of firms representative at the state level. The ASI data used in this analysis does not report the district of the enterprise. As described in the empirical strategy section above, I adjust for the level of the data being broader than the level of the reform so as to best capture the spirit of the RD design. District level data on several measures of interest, local GDP for example, are unavailable or available only sporadically. To overcome the lack of traditional measures, I consider data recorded from remote sensing on rainfall and the amount of light emitted at night from the TRMM satellite and DMSP-OLS Nighttime Lights Time Series, respectively. The nighttime light data are used to proxy for changes in local GDP, as prescribed in Henderson et al. (2012). See the Data Appendix for greater detail on all data used in the analysis. 6 Results 6.1 6.1.1 Banking Bank Branches The analysis focuses attention on the response from banks in the private sector. The notion that these banks introduce a new banking technology and the rapid expansion of their branch networks during this period makes them particularly likely to drive innovation and a transformation of the banking environment in affected districts. As profit maximizers, the theoretical framework suggests they are also the most likely to respond strongly to the reform around the cutoff. Responses from the public sector will be noted to provide contrast. To motivate the primary set of empirical results, I first consider a visual example for two years. Figure 9 presents the standard visual RD for operating private bank branches for the pre-reform year 2000 and the post-reform year 2012. The y-axis denotes the number of operating private bank branches per district, with dots reporting the local averages of districts falling within 200 persons per branch non-overlapping bins. The horizontal axis is the forcing variable of district population per branch centered on the national average and scaled to thousands of persons per branch. Considering the figure from year 2000, districts do not appear to vary systematically in their number of bank branches prior to the reform. In the post reform year, under banked districts show higher numbers of operating branches relative to banked branches just on the other side of the cutoff. The discontinuity of the number of branches estimated at the cutoff from either side yields the local average treatment effect of the reform on private branches. Next, I make the analysis more precise by presenting the annual results 20 from estimating equation 4.1.1 with operating private branches as the dependent variable. The ability to observe the number of branches across time, and the fact that the list of under banked districts did not change yearly, allows the effect of the reform to be identified not only by spatial variation between districts, but through time as the reform became implemented and branches were able to accumulate. The right panel of figure 10 plots the intercept points at the cutoff from annual local linear regressions from the banked and under banked sides of operating private sector branches in a district. Districts maintain the same value of the forcing variable across years so the set of districts remains unchanged.13 The red dashed line provides the estimated intercept from approaching the threshold along the under banked side as in the classic RD graphical representation. The solid blue line reports the corresponding intercept approaching from the banked side. The vertical distance between the two, reported for each year, corresponds to the discontinuity at the cutoff estimated as τ in equation 4.1.1. A vertical red line between the two points indicates a positive discontinuity, with under banked districts exhibiting a higher value at the threshold than banked districts, with significance at least at the 10% level.14 These figures not only present the average treatment effect, but place the level of the intercepts vertically so that overall growth and decline may be easily recognized. The figure identifies important policy aspects. In the years leading up to the reform, there is little difference in the estimated number of branches from the banked and under banked districts at the cutoff. This in itself acts as a partial validation test of the randomization of districts around the cutoff. A strong response to the policy does not occur until after 2006, which was a likely possibility given the timing of the reform. Still, the small increase in the positive discontinuity in 2005 and 2006 is not inconsistent with some banks working to establish market share in under banked districts. The strongest effects in branches occur beginning in 2008 and are estimated precisely at the 5% and 1% confidence levels, which is consistent with banks waiting until mid 2006 to submit their first ABEP and opening their branches just before their licenses expire in mid 2007. Estimates are presented in table 4. The steadily growing discontinuity is consistent with a response from private banks to the branching policy. In the figure on the left, I report the estimated effect on operating and granted licenses. The most important feature from this graph is the first statistically significant positive effect on licenses measured one year earlier than branches on January 1st, 2007. Turning to the estimation results from licenses and branches in table 4, the effect from licenses precedes a similar response in branches beginning in 2007 through 2010. The policy starting in 2010 13 New districts since 2001 that claimed territory from more than one source district are dropped along with the source districts. In addition, Thane and Pune districts in Maharashtra are dropped, as is Varanasi district in Uttar Pradesh after 2002. See the Data Appendix for details. 14 Thanks to Johannes Schmieder for help in clearly displaying the dynamic nature of the effect graphically. Note that these figures rely on estimation using a uniform kernel. 21 was amended such that banks could open in lower population centers without a prior license, resulting in licenses for such openings being issued on the day of branch entry, despite its presumed reporting in ABEPs. The combined timing of the licenses and operating branches, as well as the pre-reform and post-reform pattern demonstrates the exogenous differential change in branch reach in districts belonging to under banked districts near the cutoff relative to the otherwise similar districts on the banked side. The cumulative average effect of the policy in 2012 is estimated at approximately 10 more private sector branches in under banked districts at the cutoff relative to the banked districts. The effect is a little less than 50% of the sample mean reported in the table for 2012 at 20 private sector branches in districts around the cutoff. The size of the private sector presence increased for the sample overall in this time from an average of 10 branches per district in 2006 to 20 in 2012. 6.1.2 Credit The 2005 policy reform on branching permissions directly cites opening branches in under banked districts as a condition affecting total permissions to a bank. However, the other terms mentioned, offering no-frills accounts and meeting priority sector requirements, apply at the bank level rather than by district. Thus, there is little direct pressure from the reform on bank credit and deposit behavior, particularly around the threshold. Recall, however, that the theoretical framework predicted that banks would expand credit in under banked areas in anticipation of future competition. Figure 11 shows the annual discontinuities in total district credit from private banks, analogous to the figure presented for operating branches discussed above. Similar to the early pre-reform years in private bank branches, the number of credit accounts in thousands shown in the left panel of the figure displays little difference between the banked and underbanked districts at the cutoff. However, consistent with the timing of the Vyas Committee commissioning and report, the number of credit accounts began increasing in under banked districts in 2004 and 2005. Underlying this change is also a change in the composition of banks in these districts, with fast growing branches opening in these districts as more inert banks were acquired by nationalised banks. This behavior is consistent with aggressively growing banks acting preemptively on the expectation of reforms by expanding in areas likely to be more heavily contested in the future. Turning to the estimates in table 5, the response from this preemption is estimated at 6,725 additional credit accounts in the under banked districts at the cutoff, which is 52% of the sample mean for districts around the cutoff. The discontinuity in accounts continues to grow over the next few years and is estimated precisely for all years except 2009 and 2011. The slight retraction of the discontinuity in 2008 may be explained 22 by the exit of a private bank through acquisition by the public sector in 2007.15 The decline in the discontinuity in 2009 may reflect the tightening of restrictions regarding specific cities eligible as under banked within districts based on their proximity to major metropolitan areas or being metropolitan themselves. Unlike the branches data, credit cannot be broken out by bank within a district to form a clearer picture as to the exact channels driving the aggregate responses. The last two years again show increased expansion in credit accounts consistent with the growth in branches in these years. The results from outstanding credit amounts in millions of rupees show qualitatively consistent results. The amounts data are measured with less precision, which may result from many large investment projects being lumpy in nature, leading annual district levels to fluctuate more than the number of accounts. 16 As noted by RBI Deputy Governor Mohan in a 2006 speech regarding financial inclusion, the expansion of retail credit after 2003 accounted for a major source of increased lending (Mohan, 2006). Breaking credit out by personal loans, figure 12, confirms that the response in the growth of personal loans in under banked districts near the cutoff relative to banked districts was significant. The initial jump in personal loan accounts in 2004 corresponds with the changing composition of banks as aggressive private banks slowly expanded their branch presence in under banked districts. Also at this time, the interest rates on consumption loans were liberalized, allowing interest rates to dip below the bank’s self reported cost of funds plus profit margin. Personal loan amounts largely mirror the expansion of accounts, though the tightening after 2008 may correspond to a change in priority sector lending requirements making the requirement more stringent. An implication of the theoretical framework behind the hypothesis of a preemptive response in credit by profit maximizers is that public sector banks, which follow less clear objective functions, are unlikely to show the same pre-reform response as private sector banks. Figure 13 and the corresponding table 7 confirm a lack of response prior to the policy implementation around the cutoff, as well as a muted response during the reform years as well. These results are consistent with the incentives generated by the reform operating most strongly on private sector banks aiming to grow in reform years. 6.2 Agriculture Agriculture constitutes the primary economic activity for the majority of Indians. The 2001 Population Census reports that over 56% of India’s workers were engaged in agricultural or related activities at the time of the census which, due to the exclusion of marginal workers, likely provides a lower bound. Given that concern over the availability of credit in rural areas 15 Bharat Overseas Bank was acquired by Indian Overseas Bank that already held a 30% interest in the bank. The large dip in credit to banked districts in 2008 appears to be driven by outliers, as changes in districts affected by the above mentioned merger in the previous year do not show strong responses in credit amounts. 16 23 led to the commissioning of the Vyas Committee that catalyzed a reform to rural branching and presumably the broader policy reform in 2005, early policy effects concentrated in agriculture could be expected. The analysis below indeed shows an early response in increased agricultural lending that wanes as the details of the reform become known, and increases again with greater emphasis placed on rural and semi-urban markets by later refinements to the reform. Attention is then turned to the effect of expanded banking services on agricultural performance. 6.2.1 Credit to Agriculture Figure 14 shows the reduced form RD in the district percentage change in credit amount to direct and indirect agricultural activity in rural and semi-urban areas from their 2001 levels.17 A positive and statistically significant response in under banked districts is first detected for credit to direct agricultural activities in 2005, the year following the Vyas Committee Report. The effect continues into 2006 after which it diminishes for a few years until emerging again in 2009. The greater relative expansion of credit in under banked districts at the cutoff is significant in magnitude as well, exceeding the sample means in 2005 and 2006, and remaining above 50% of the sample mean in the later years. The analogous effect estimated for lending to indirect agricultural activities is imprecisely estimated until 2009 after which the discontinuity is positive, statistically significant, and large in magnitude exceeding the sample mean. Results are reported in table 8. The early expansion of credit beginning in 2005 is consistent with the timing of the Vyas Committee and emphasis placed on agricultural lending by policy makers, as well as with the incentives of private banks to increase lending to secure their market share of profitable loans in anticipation of increased competition. The decrease after 2006 may be attributable to banks learning that the branching policy reform was less directly tied to agricultural lending than initially anticipated. Alternatively, a subsidized credit program to farmers exclusively through public sector banks initiated around that time may have drawn away demand for private loans, washing out the private bank effect in direct agricultural credit.18 The growth in agricultural lending after 2008 in under banked districts has several potential explanations. New branches opening as a result of the policy are growing in strength during these years. A refinement to the branching policy in 2008 made metropolitan areas ineligible for under banked status in 2008, creating greater incentive to move into lower populated areas. The adoption of a new branching policy in 2010 that placed greater emphasis on rural and semi-urban branch entry in all districts with a bonus for under banked areas also explains the faster growth 17 The percentage change is approximated using the difference in logs of credit amounts from the 2001 reported levels. 18 The Credit Subvention Scheme operated through NABARD and exclusively distributed through government sector banks was initiated in 2006-2007. 24 in the sample mean relative to the discontinuity for direct agricultural lending in 2010 and 2011. Alternatively, a reform to priority sector lending in 2007 also placed greater emphasis on agricultural lending. Required investment in the Rural Development Infrastructure Fund for failing to meet priority sector quotas, first coming due in the 2009 financial year, was accounted as indirect agricultural lending by banks. Finally, the categorization of loans by the RBI was revised in 2008, making direct comparisons by sector pre- and post- 2008 less accurate. Without finer data on loans, disentangling the exact causes is likely not possible. Rainfall Annual rainfall is undoubtedly an important input for agricultural performance in India. Figure 15 presents estimated discontinuities in the district averaged percentage deviations of rainfall measures from their mean levels across the points of measure within a district. Since rainfall is random and unaffected by the policy reform at the cutoff or anywhere else, this analysis also serves as a falsification test of the RD design. As anticipated, rainfall does not show significant discontinuities at the cutoff. This suggests the response from credit and agricultural performance is not discontinuously effected by exogenous productivity shocks around the cutoff in the years considered. Crops Figure 16 shows the reduced form regression discontinuity analysis for yield and output for two major crops in India, cotton and wheat. I present discontinuity analysis for crop yield (tonnes per hectare of cultivated land) and output (tonnes). Each specification controls for the district averaged percentage deviations of rainfall, district population and its square, and the 2001 pre-randomization value of the dependent variable. The analysis for the output is the most striking for cotton, while the effect on yield is greater for wheat. The response from these individual crop statistics suggests the branching policy reform positively impacted output and/or productivity. However, considering crops individually, and absent price data for the crop output, makes interpreting the results difficult. Not every district produces each crop, or is well suited for every type of agriculture. Farmers may be moving in or out of crops based on their prices. Yields may decrease if farmers enter high paying crop markets with plots of land poorly conditioned for those crops. Alternatively, yields may rise if farmers invest more in productive and profitable crops. Thus, a measure better incorporating the incentives faced by the farmer is needed. To address these concerns, I compute an index of crop yields similar to that used in Jayachandran (2006). The index is constructed as a weighted average of crop yields for rice, wheat, jowar and groundnut, using crop revenue shares for the district as weights (see Data Appendix for details). I am able to construct the measure for the July-June years 20012002 to 2007-2008 from data on crop prices and production statistics collected at the district level. The price data for crops is available for a slightly smaller set of districts and generally restricted to crops for which the particular district produces in greater volumes. The index 25 carries the added benefit, however, that a wider set of districts in India produce at least one of the crops in volume, meaning the set of districts through time will change less than considering output from a single crop. The results from the reduced form RD analysis are shown in the top panel of table 11. The estimates show positive discontinuities after 2005, though are estimated imprecisely except in 2008. To estimate the effect of banking activity on average crop yield, I estimate a fuzzy RD of the crop yield index on total private sector credit accounts, instrumenting for credit accounts with the discontinuity. In the bottom panel, I present the fuzzy RD results for the pre-reform and post-reform periods. No effect is estimated in the pre-reform period. In the post reform period, I estimate an average effect of 0.03 with 10% significance, which may be interpreted as every thousand private bank accounts increases the crop yield by an average of 3%. This is a little less than the average effect of a positive rainfall shock, for rainfall above the 80th percentile for that district, on crop yield estimated in Jayachandran (2006). 6.3 Industrial Activities Though the initial drive of the policy reform may have been to increase financial inclusion in low population areas and increase the credit flow to agriculture, many of the populated centers of under banked districts benefited from increased branch entry. This section investigates to what extent manufacturing enterprises benefited from the expanded bank presence by receiving loans and being able to invest in productive assets. Credit to Manufacturing and Processing Figure 17 presents the reduced form RD effect for the percentage change in credit amount to manufacturing and processing. The effect after 2007 resembles the expansion of bank branches, with a steadily growing positive effect in under banked districts. Turning to the fuzzy RD results, the estimates are all quite imprecisely estimated. This may be due to the lumpiness with which capital is acquired in manufacturing, with large projects arriving infrequently per district but constituting a large share of credit. Multi-plant manufacturing enterprises may also have been able to secure funds for plants in unbranched areas through their headquarters. As such, large manufacturing investments may be less reliant on local branch access. While less response in credit appears to be earmarked for manufacturing, it is possible that credit lent as personal loans, that shows a strong response to the reform, may end up funding capital investments or freeing up other resources that then get invested in enterprises. To investigate this possibility, the next section examines input decisions from registered manufacturing firms. 26 6.3.1 Evidence from the ASI In table 13 I present the results from difference in differences analysis using data from the ASI. The analysis uses years 1999-2010. In column (1) I estimate the effect on logged assets excluding land and inventory. The average treatment effect is positive but imprecisely estimated at a value of 17%. The effect on logged working capital, in column (2), is estimated at 0.264 with significance at the 10% level. The effect on the amount of outstanding loans held by the firm is estimated to increase 24% with statistical significance at the 10% level. Total investment increased by 19.7%, with statistical significance at the 10% level. The capital labor ratio is estimated to increase by 3.4 in response to the policy and is also estimated with precision at the 10% significance level. The sample mean for the under banked states sample was 10.88 post reform, making this a sizable effect. The estimates are quite robust to considering other ranges of years around the reform. In each regression I control for the rural status of the enterprise, the age of the plant as measured by years since opening, the number of total enterprises in the firm to which the enterprise belongs, the logged number of employees at the enterprise to control for size, and state fixed effects with state specific time trends. I exclude industry fixed effects as new NIC codes were adopted in 2008, potentially making some industry codings inconsistent through the time series. In practice, the inclusion of 3 digit NIC codes has little effect on the estimates. The significant increase in loans carried by enterprises from under banked districts in the post reform years would indicate that the increased banking activity is finding its way to the industrial sector. The increases in working capital as well as total investments suggests firms are expanding the use of productive inputs with the expansion of credit. Further, the increase in the capital labor ratio is consistent with previously credit constrained firms making investments in capital as those constraints are relaxed with the inflow of new formal credit. These adjustments to the productive technologies of the firm are likely to result in changes in efficiency. If credit rationing resulted in the misallocation of credit, the expansion of credit may produce large impacts if it helps correct inefficient dispersions of marginal products of capital across firms. 6.4 Economic Growth and Light emitted at Night The final analysis following the RD design examines discontinuities in changes of the emission of light into space at night. Henderson et al. (2012) established that so called “nightlights” provide a reliable proxy for economic growth under certain caveats. Important among these is the prescription to compare changes in light through time for one area to those in another, rather than comparing levels of light only across places or levels of light only across time. There are several reasons for this: the time series is composed of readings taken by different satellites in different blocks of years. The instruments between satellites vary, and their 27 precision changes with age. The raw data are also processed in ways undisclosed to researches and vary across years. Part of this processing includes decisions on interpreting very low levels of light. Thus, any errors or idiosyncrasies generated by these processes get accentuated by differences in the degree of urbanization across locations. This analysis accounts for these concerns by estimating the discontinuity in the difference of logged average district light since 2004. Thus, the dependent variable can be interpreted as the approximate percentage change in average light emitted in a location from its level in 2004. The RD compares these changes in estimating the discontinuity at the threshold. Figure 18 graphically reports the discontinuities estimated with a 2nd degree polynomial, which better captures the underlying data. Since the level of light is reported from measurements taken during the calendar year, 2005 is the first year with months under the enacted reform. Estimates are presented in Table (14). A slight negative discontinuity is estimated in the first year and is a small fraction of the average percentage change in light for districts in the sample. The discontinuity is small and positive again in 2006 though the average change in districts was negative overall. A positive jump in the discontinuity to 9.4% appears in 2007 and is estimated significantly at the 1% level, with the average change in light for districts in the sample increasing as well to 11.4%. A similar response is found in 2008 with 2009 showing low levels of light emitted in general for the sample around the threshold and a smaller discontinuity. The last three years show similar discontinuities in light to 2007 and 2008, with 2011 estimated with precision at the 10% confidence level. To estimate the effect of expanding branch presence on the change in the amount of light emitted in districts at night, I perform a fuzzy RD of the change in light on private bank branches for the pre-reform period, which in this case is only 2005, and the post reform period constituting years 2006-2012 pooled together controlling for year fixed effects in addition to district population and its square. I run the estimation using local linear regressions because these better fit the bank branching data and offer a strong first stage. The pre-reform effect reported in the lower panel is negative and small, consistent with the reduced form estimate for 2005. The conventional estimate reported in the post reform column is estimated positive, but small and insignificant. However, the bias-corrected measure which accounts for a local quadratic estimate with a wider bandwidth, better capturing the quadratic relationship in the night light data, yields a positive and significant coefficient. This estimate is significant at the 1% level and has a value of 0.012. The coefficient may be interpreted as the average effect of each bank branch from the reform period is to increase nighttime light by 1.2%. Taking the estimated elasticity of nighttime light to GDP from Henderson et al. (2012) of 0.3, this implies that each bank branch raises local GDP by approximately 0.36%. The average increase in bank branches in the post reform period is estimated at approximately 5, implying the total effect was an average increase of local GDP in the districts by 1.79%. 28 6.5 Robust to NREGA A competing explanation for the change in the spatial allocation of bank branches, increased banking activity, and subsequent responses in economic outcomes is the introduction of the Mahatma Gandhi National Rural Employment Guarantee Act (NREGA) that closely coincided in time with the branching policy reform. The act constitutes a public works program aimed at relieving poverty in rural areas by providing 100 days of guaranteed work to individuals from rural areas. The implementation of NREGA occurred in three stages, with 200 districts selected to begin the program in the fiscal year April 2006 through March 2007, with 130 new districts introduced in 2007-8 and the remaining 263 districts introduced in 2008-9. Zimmermann (2012) and Klonner and Oldiges (2014) analyze the effect of NREGA and provide background on the program. Of particular importance to the current analysis, NREGA benefits were distributed through bank accounts. One may conclude that this would increase the demand for formal banking, potentially increasing both the geographic reach and level of banking services. While likely true, to confound the current results there must also be a discontinuous break in the implementation of the program and disbursement of benefits at the “under banked” cutoff used for the regression discontinuity. Districts were assigned to the various roll-out phases based on a composite index on district “backwardness” from the National Planning Commission (2003). In table 15 I test whether a discontinuity in phase assignment can be detected at the cutoff. A significant discontinuity would suggest a correlation with the NREGA program. The test fails to reject the null hypothesis of continuity at the cutoff for all three phases. Thus, NREGA phase assignment and therefore likely its benefits as well, would be unexpected to differ at the cutoff. In analysis not shown, I perform a visual RD of the district composite index at the under banked cutoff. No discontinuity is observable at the cutoff. Further, the general notion that persons per branch is generally increasing with worsening district conditions is confirmed by the trend of the index on “backwardness.”19 7 Conclusions Greater access to formal financial markets through policy driven branch expansion can relax credit constraints, allowing productive firms to invest at their optimal rates and households to smooth consumption across time. However, due to information asymmetries, possibly misaligned incentives within bank hierarchies and high costs to serving certain areas, expanding branch networks need not necessarily result in greater credit access or economic growth. 19 Out of concern that the omitted districts are disproportionately from one side of the cutoff or the other, I repeat the McCrary test only including districts missing the composite index value. I fail to reject the null hypothesis of continuity in the density of districts at the cutoff with the discontinuity estimate in the log difference in height at -31 and a standard error of 38. 29 Identifying the effect of formal bank access also proves challenging due to classic endogeneity concerns of branch location choice. This paper utilizes exogenous variation in formal banking access generated by a recent and previously unstudied branching policy reform in India from 2005. Using a regression discontinuity design based on the assignment rule of districts to under banked status, the analysis demonstrates an expansion of banking services consistent with the timing and incentives of the reform. Agricultural productivity and the capital intensity of manufacturing are shown to increase in areas receiving higher credit due to the reform. I estimate that an increase of 1,000 private bank credit accounts in a district raises average crop yield by 3%. This effect is a little less than half of the effect Jayachandran (2006) measures on crop yield from positive rainfall shocks. Manufacturing enterprises in areas with increased access to banking exhibited higher growth in total investments, working capital and capital labor ratios. Finally, I confirm the aggregate effect on local GDP growth by showing that areas with expanding banking services experienced higher rates of growth in nighttime light intensity in the years following the reform. The estimates imply that each additional private bank branch led to a 0.36% increase in local GDP. Overall, these findings offer strong causal evidence that the expansion of the financial systems facilitate growth in productive activities important for driving economic development. They further confirm the potential effectiveness of policy reforms in producing this expansion. The evidence suggests that the competition anticipated to be generated by the reform led profit maximizing banks to increase the quantity of credit ahead of additional branch entry to secure market shares in profitable areas. This effect from competition enhanced the response to the reform, beyond what could be expected from branch entry alone. Importantly, the expansion of credit does not seem to have been restricted to urban areas which is a common concern in developing areas. Rural and semi-urban markets in underserved areas also exhibited increases of credit from private banks. Though caution must be used in examining the credit data at disaggregated levels, the timing and effects in rural areas are consistent with banks responding to priority sector requirements by lending to target groups through their newly opened branches. This would suggest the importance of harmonizing incentives across policies to effectively meet policy objectives. Future work should examine the effects of interactions between the policy interventions. The results of this analysis suggests that private sector banks respond strongly to the incentives generated by regulations. Areas targeted by the policy reform demonstrated a variety of benefits. While this speaks to the effects of financial development, little can be said in terms of overall policy evaluation. Resources were redistributed across districts due to the reform. Without measures for the opportunity costs of these resources, a full welfare analysis is beyond the scope of this paper. The introduction of bank or branch level data on credit, as well as farmer level data of characteristics, borrowing, inputs and crop selection would allow for a closer examination of these potential costs, give greater insights as to the 30 specific mechanisms driving productivity gains and provide a basis for a more complete welfare analysis. Future work should also concentrate on a deeper understanding of the interactions between private sector, public sector and informal lenders in facilitating financial development. 31