Abstract - Editorial Express

MEASURING THE IMPACT OF DEVELOPMENT BANKS ON LOCAL GROWTH: EVIDENCE FROM BNDES IN BRAZIL§ Maria Júlia C. Wegelin‡ and Christiano A. Coelho†

Abstract In 2006 the Brazilian national development bank, BNDES, established a credit policy designed to stimulate economic growth in municipalities with certain characteristics in Brazil. We use the fact that some municipalities borrowed through the program while others did not to estimate the effect of BNDES credit policy on local development. Using the propensity score matching (PSM) estimator we found that GDP and per capita GDP growth of treated units increased while the employment and income were not affected by BNDES credit policy. GDP and per capita GDP growth increased, on average, 0.4 percentage points per year due to the credit policy, which meant an GDP increase of only R$0.29 for each R$1.00 of BNDES disbursement. With this estimate, we found that the net social return of this policy was, on average, of -75%. This estimate is an important contribution to the recent debate relating to the limits of the development banks’ contribution to economic growth. KEYWORDS: local economic growth; development banks; policy evaluation. AREA: Applied Microeconomics JEL CODES: H11; H81; H82; O10

§

We would like to thank Marcelo Mello, Bruno Funchal and Sérgio Leão for their insightful comments. All errors are ours. ‡ CGU:[email protected]; [email protected] † IBMEC-RJ and Uerj: [email protected].

1

1. Introduction The role of state-owned banks in fostering development is extensively discussed in the economic development literature. From a theoretical perspective, it is unclear whether state-owned banks bring more or less financial and economic development. In one hand, the social view and the development view of state-owned banks, as in Atkinson and Stiglitz (1980) and Gerschenkron (1962), argues that these banks help to mitigate market failures related to positive externalities in financial markets and so help to promote financial and economic development. In the other hand, the political view, as in La Porta et al. (2002), argues that lack of accountability and transparency will make stateowned banks the perfect way to special interest groups to extract rents from the rest of society, which would make state ownership of banks detrimental to financial and economic development. Finally, the intermediate position, the agency view, as in Francisco et al. (2008), argues that state-owned banks can generate benefits through the internalization of externalities in the credit concession, but at the same time agency costs and corruption can make state-owned banks much more inefficient than their private counterparts. Given this theoretical ambiguity, it is interesting to estimate empirically the effect of state-owned banks’ credit policy on economic development. In this paper we will do that by using a specific policy designed by the Brazilian national development bank, BNDES, in 2005 and implemented from 2006 on, called PDR (Programa de Dinamização Regional). 1 As in most empirical applications in economics, we have an identification problem. We use the fact that growth of borrowing increased at some municipalities stimulated by the program while at others with similar characteristics did not to estimate the effect of this credit policy on some local development variables using Propensity Score Matching (PSM) techniques. In this way, we can estimate development banks effect on economic growth and other indicators in a much clearer way than the previous literature. Our results show that in the municipalities where there was increase in BNDES disbursements growth through the PDR program, GDP and per capita GDP increased, on 1

PDR can be loosely translated to regional development program.

2

average, 0.4% more than in the other municipalities not affected by the policy, which meant that each R$1.00 of BNDES disbursement generated a GDP increase of only R$0.29. Employment and income in turn were not affected by the policy. We proposed a metric to analyze the net social return of BNDES policy taking into account the opportunity cost for the Brazilian Treasury. Using our proposed metric we found that the net social return was of -75%. Given that, as part of a new development strategy of the Brazilian government, BNDES disbursements in real terms increased from R$113 billion in 2008 to R$ 190 billion in 2013, our results show that this kind of model already has serious limitations concerning economic growth generation. Besides, our results show how the current lack of transparency and accountability of institutions similar to BNDES can potentially cause bad policy design. This article is organized as follows. In Section 2, we review the literature pertaining to the role of BNDES in Brazilian development. In Section 3, we describe the PDR program. In Section 4, we describe the data and present some summary statistics. The empirical strategy is outlined in Section 5. Section 6 contains the main results and their analyses. Finally, Section 7 provides the conclusions.

2. Related Literature The main papers analyzing the effects of BNDES policy on economic performance are Lazzarini et. al (2012) and Carvalho (2013). Lazzarini et al. (2012), using public companies data between 2002 and 2009, studied the effect of BNDES loans and investments on performance and investment structure of firms. The only effect found was a reduction in financial expenditures, when companies receive subsided loans from the bank. Besides, they show that BNDES does not lend to firms with average bad performance. The recipients of these subsided loans are those with good performance and those who make donation to politicians that won elections. Therefore, their results reject the vision that development banks are used to bail out bad companies, but at the same time indicate that these loans on practice are a way to transfer subsides to big firms without any improvement on performance or increase in investment. 3

Carvalho (2013), using data of Brazilian manufacturing companies, show that Brazilian government control of banks affects significantly real decisions of firms. He shows that companies that are able to get loans from state-owned banks increase employment near elections, in regions where there is fierce competition between candidates, and decrease it in other regions. This analysis suggests that politicians in Brazil use state-owned banks loans as a mechanism to affect firms’ decisions about employment allocation.

3. The PDR program Since its creation, most part of BNDES disbursements were allocated to the south and south-east regions of the country, the richest regions.

2

Figure 1 shows BNDES

disbursements from 2000 to 2012. We can see from it that from 2006 on there was a clear increase in the disbursements growth for the poorest north and north-east regions, mainly in the north-east region. The increase in the disbursements for these regions coincides with the creation of the PDR program. The goals of the program are to reduce regional inequalities between the poorest regions (north and north-east) and the richest ones (south and south-east) and to decrease social and income inequalities as well. The regions with access to subsided credit were defined by the Ministério da Integração Nacional (Ministry of National Integration) using a two dimensional criteria: household per capita income, based on national Census of 2000, and the level of economic dynamism of the micro regions, based on the annual average GDP growth. In the Table 1 we show seven categories in which the municipalities were allocated. Most of the municipalities with low and stagnant income were in the north and north-east regions.

2

Pamplona (2011) shows that, on average, 50% of BNDES disbursements were to south-east region, while 20% were to the south region.

4

Table 1: Municipalities classification by PDR Level of household per capita income High (4º quartile) High (25% highest) Average Low (25% lowest)

∆GDP

Superior average (3º quartile)

Inferior average (2º quartile)

Low (1º quartile)

Superior average income dynamic

Inferior average income dynamic

Low income dynamic

Superior average income stagnated

Inferior average income stagnated

Low income stagnated

High income

Source: BNDES

The PDR begun on January of 2006 and the focus of the program were the low income municipalities. The program used two mechanisms to stimulate these regions: increase in the maximum volume given and decrease in the interest rate to municipalities classified as low income. We used this year as a break point to analyze the impact of this policy on the GDP growth, per capita GDP growth, employment and income. In Figure 1 below we can see yearly disbursements from 2000 to 2012 for all Brazilian regions. We can see from this graph that there was an increase in disbursements for north and northeast regions from 2006 on, but we can see that in 2008 and 2009 this trend was even more pronounced.

Figure 1: BNDES disbursements by Region in Brazil R$ million 120000 100000 80000 60000 40000 20000 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 South-east South Center-west North-east North Source: BNDES

5

In Table 2 we show that the program brought an effective change in BNDES policy. We divided municipalities in four groups: (i) those belonging to the south, southeast or center-west regions and having a high or superior average income; (ii) those belonging to the south, south-east or center-west regions and having a low or inferior average income; (iii) those belonging to the north or north-east regions and having a high or superior average income; and (iv) those belonging to the north or north-east regions and having a low or inferior average income.

Table 2: Growth of BNDES disbursements by type of municipality Incentive criteria PDR Region1 Income2 0 0 0 1 1 0 1 1 1 2

Amount of municipalities 2726 593 152 2089

Annual average growth of BNDES disbursements 2000 to 2005(A) 2006 to 2011(B) 0.57 0.41 0.65 1.06 1.07 0.88 -0.58 2.18

Variation of average growth (B-A) -0.16 0.40 -0.19 2.76

Dummy_region_pdr (1 = if unity belongs to north or north-east region; 0 = otherwise).

Dummy_income_pdr (1 = if unity was classified as low or inferior average income by PDR; 0=otherwise). Source: Own elaboration, based on information presented by BNDES in the Portal e-SIC.

We can see from the Table 2 that after PDR, BNDES increased disbursements growth only to the municipalities who were classified as being able to receive funds through PDR program. This makes us confident to use the PDR beginning as our breakpoint date for the treatment.

4. Empirical Strategy As we have a case of a non-random experiment, we will use the policy evaluation methodology in order to measure the effect of the BNDES policy. Because in our case we do not have a well-defined treatment group, before we proceed with the discussion of our empirical methodology, we must explain how we define the treatment group. A municipality was considered belonging to the treatment group if there was an increase in the BNDES’s disbursements growth to this place after the period in which the PDR 6

program was implemented. This group is composed of 3,156 municipalities. The untreated group is composed of the municipalities where there was no increase on the BNDES’s disbursements growth after PDR. There were 2,404 municipalities in this group. We think that with this definition we can select the municipalities that were effectively benefited by the program (see Table 2 above). Having defined the treatment group, we proceed to explain our methodology choice. Figure 2 below shows the path of the GDP growth, one of the outcomes of interest, for the treatment and untreated groups, as defined above, and for the control group chosen by one of our methodologies.

Figure 2: GDP growth before treatment by group

GDP growth 0,1 0,08 0,06 0,04 0,02 0 2001

2002

2003

Treated

Control

2004

2005

-0,02 Untreated

Source: own elaboration

A visual inspection of Figure 2 shows that there are clear differences in the behavior of the outcome variable before the program between treated and untreated groups. This makes the use of difference-in-difference estimation not appropriated because of the violation of the parallel trend assumption. This is no surprise, once one takes into account that all cities, including those big cities that were not eligible for the program, were in the untreated group. So, we decided to use the propensity score matching (PSM) methodology, since with it we can select appropriated unities for the control group. Figure 2 shows clearly that the GDP growth path for the control group 7

chosen by one of the matching methods (the nearest neighbor) is much closer to the GDP path of the treated group than that of the untreated group as a whole. In order to further illustrate our point, in Table 3 below we show the t test for the difference of GDP growth before the program, by year, between treated and untreated group and between treated and the control group chosen by the PSM (nearest neighbor).

Table 3: Testing outcome average differences for treated, untreated and control Test variable: GDP growth H0: There is no mean difference between groups¥ Groups Year Treated x Untreated 2001 1.75* 2002 2.61*** 2003 6.2*** 2004 3.57***

Treated x Control§ 0.64 1.31 -1.74* -0.72

¥ - *H0 is rejected with 10% of significance;** H0 is rejected with 5% of significance;*** H0 is rejected with 1% of significance § - Using the nearest-neighbor methodology to select the control group

Source: own elaboration

The t test shows that our methodology is appropriate, since it eliminated GDP growth differences before the treatment between the treated units and the group that in the end will be used as the counterfactual. 3 To check the robustness of our results, we will use the four methods available to calculate the average treatment effect on treated (ATT) using PSM: the nearest-neighbor matching, the stratification matching, radius matching and kernel matching. The typical PSM estimator is given by the following equation:

= ∑∈ , − ∑∈ (, ) ,

(1)

In equation (1) is the number of treated units belonging to the common support area and ω(, ) is the weight given to each unit in the untreated group in order to calculate the outcome that will be used as the counterfactual. This weight will be determined by the propensity score, i.e. the probability of receiving the treatment, for

3

If we make the same visual inspection of the graphs and t-tests for the other outcomes and other matching methodologies, we will find the same results. These results are available upon request.

8

both the treated unit, ( ), and the untreated one, !, and by the matching method being used.

4.1 Data and descriptive statistics We used the annual series of BNDES disbursement by municipality between 2000 and 2011 collected at Portal e-SIC (Electronic System of the Citizens Information Service). However 20% of total disbursements were removed of our data basis, because there is no information about which municipality those disbursements refer at the data basis shown by BNDES. The outcome variables are: GDP growth, GDP per capita growth and employment and income index. GDP and population were obtained from the databases available at the Brazilian Institute of Geography and Statistics (IBGE), associated with information from the 2000 and 2010 Census. Employment and income index were obtained from the annual survey of Rio de Janeiro industries federation (Firjan). The study is based on official government statistics, provided by the Labor, Education and Health Ministries. The index ranges between 0 and 1, and the closer to 1, the greater the municipality development. The following variables were used to calculate propensity score: infrastructure level of the municipalities; non-earmarked credit related to GDP in 2005; income level, as defined by PDR; income dynamism, as defined by PDR; annual GDP growth between 2001 and 2005; annual growth in GDP per capita between 2001 and 2005; employment and income index from 2000 to 2005; and GDP in 2005. To measure the level of infrastructure of municipalities, we used as a proxy the percentage of municipality households who had sewage access in 2000, according to IBGE and data of 2000 Census. To measure non-earmarked credit, we used as a proxy the fluctuation of loans and financing stocks, published monthly by the Brazilian Central Bank.

9

Table 4: Descriptive statistics of the variables Variable bndes gdp gdp_growth pop gdp_per_capita gdp_per_capita_gro wth employment_income infraestructure_2000 non_earmarked_cred it_2005

Description Annual BNDES disbursement in municipality (R$ 2011) GDP of municipality (R$ 2011) GDP growth Total population of municipality (inhabitants) Per capita GDP of municipality (R$/inhabitant 2011)

Per capita GDP growth Employment and income index (value between 0 and 1) Percentage of municipality households who had sewage access in 2000 (%) Fluctuation of loans and financing stocks (R$)

Income level (1=low income; 0= others) inferior_average_inc Income level ome (1=inferior average income; 0= others) average_ Income level superior_income (1=superior average income; 0= others) Income level high_income (1=high income; 0= others) Income dynamism dynamic_income (1= dynamic income; 0= stagnated income) Source: own elaboration low_income

Average (Std. Error) 1.05e+07 (1.80e+08) 4.38e+08 (5.02e+09) 0.058 (0.139) 32,890.78 (196,215.4) 8,406.73 (10,535.71) 0.051 (0.143) 0.40 (0.16)

Source BNDES IBGE IBGE IBGE IBGE IBGE Firjan

0.81 (0.22)

IBGE

5.38e+07 (1.54e+09)

Brazilian Central Bank

0.24 (0.43) 0.24 (0.43) 0.26 (0.44) 0.26 (0.44) 0.46 (0.5)

BNDES BNDES BNDES BNDES BNDES

5. Results 5.1 First stage: p-score estimation The first step in our estimation procedure consisted in defining the p-score of each municipality in our sample (5,560, being 3,156 belonging to the treatment group and 2,404 belonging to the untreated group). We use all the available characteristics considered relevant to determine the probability of being treated and the outcomes being analyzed. We used the following variables: percentage of municipality households who had sewage access in 2000 (infrastructure_2000), amount of non-earmarked credit granted to the municipality in 2005 (non_earmarked_credit_2005), level and type of 10

municipality income in 2005, as defined by PDR (low_income, inferior_average_income, superior_average_income, high_income and dynamic_income), annual GDP growth for years from 2001 to 2005 (gdp_growth_2001, gdp_growth_2002, gdp_growth_2003, gdp_growth_2004 and gdp_growth_2005), annual per capita GDP growth for years from 2001

to

2005

(gdp_per_capita_growth_2001,

gdp_per_capita_growth_2003,

gdp_per_capita_growth_2002,

gdp_per_capita_growth_2004

and

gdp_per_capita_growth_2005) and GDP in 2005 (gdp_2005). Results of the logit model estimated in this first stage can be seen in Table 5.

Table 5: P-score logit regression of the first stage Variable

Coefficient

infrastructure_2000

0.002

non_earmarked_credit_2005

3.073

Standard P>|z| Error 0.196 0.993 0.909 0.001***

low_income

0.89

0.126

0***

inferior_average_income

0.375

0.111

0.001***

superior_average_income

-0.086

0.095

0.366

dynamic_income

-0.126

0.074

0.089*

gdp_growth_2001

0.182

0.946

0.848

gdp_growth_2002

-1.191

1.669

0.475

gdp_growth_2003

5.619

3.071

0.067*

gdp_growth_2004

-0.413

1.36

0.761

gdp_growth_2005

-5.001

2.63

0.057*


-1.186

0.968

0.22


0.15

1.661

0.928


-7.036

3.067

0.022**


-1.178

1.352

0.383


3.41

2.622

0.193

4.37E-12

8.48E-12

0.607

gdp_2005 2

R : 0.039 Number of observations: 5,531 * significant at 10%; ** significant at 5%; *** significant at 1%. Source: own elaboration.

Despite some variables not being statistically significant, we decide to maintain them on the estimation, given their economic significance. With these controls we were able to satisfy the balance property needed to guarantee the randomness of treatment given the observable characteristics. 11

5.2 Second stage: the matching estimators for the Average Treatment Effect on Treated (ATT) In the second stage we calculated the ATT in three outcomes of interest: GDP, per capita GDP and an index of employment and income calculated by Rio de Janeiro industries federation (Firjan). Tables 6 to 8 show the results for the methodologies used to do the matching: nearest neighbor, radius, kernel and stratification.

Table 6: ATT on GDP growth Outcome

GDP growth

Method

Average outcome for treated

Average outcome for control

ATT

Standard Error

t

Nearest neighbor

0.0721

0.0692

0.003

0.002

1.472

Radius

0.0721

0.0673

0.005

0.002

2.628***

Kernel

-

-

0.004

0.001

2.954***

Stratification

-

-

0.004

0.001

2.772***

* significant at 10%; ** significant at 5%; *** significant at 1%. Source: own elaboration.

Table 7: ATT on per capita GDP growth Outcome

Per capita GDP growth

Method



ATT

Standard Error

t

Nearest neighbor

0.0656

0.0632

0.002

0.002

1.088

Radius

0.0657

0.0611

0.005

0.002

2.666***

Kernel

-

-

0.004

0.002

2.635***

Stratification

-

-

0.004

0.002

2.452***


12

Table 8: ATT on Employment and income index Outcome

Employment and income index

Method



ATT

Standard Error

t

Nearest neighbor

2.55E-11

-7.90E11

0

0.002

0

Radius

2.37E-11

0

0.001

0

Kernel

-

-1.09E11 -

0

0.001

0

Stratification

-

-

0

0.001

0


Results show that in three out of four methodologies the ATT was positive and statistically significant for GDP and per capita GDP. For the GDP, the effect of BNDES program was between 0.3 and 0.5 percentage points and for per capita GDP was between 0.2 and 0.5 percentage points. For the employment and income index the effect of treatment was not statistically significant.

5.3 Cost-benefit analysis: calculating the social return of BNDES program To better understand the economic effect of the BNDES policy we have to compare the benefits generated on GDP growth with the opportunity costs of these funds for the government and society. In the last few years the importance of this kind of analysis increased, since most part of the increase in BNDES disbursements in this period was financed by funds from the National Treasury. Treasury finance of BNDES increased from R$13.7 billion in 2006 to R$319.9 billion in 2011, and to R$413 billion in 2013. 4 As the Treasury must issue bonds to give these funds to BNDES, there is a direct financial cost for society, since BNDES pays for the Treasury an interest rate lower than that the Treasury pays when issuing the bond. So this direct financial cost is equal to the 4

All values are measured in R$ of 2013.

13

difference between the interest rate that the Treasury has to pay on these bonds and the interest rate that it charges from BNDES. Besides the direct financial costs, there is the opportunity cost for the Treasury (and society as a whole) of the use of these funds. We can express the Total Social Return (TSR) for the country of this Treasury policy of financing BNDES as: "# =

($%&'() *$ + )*,-./0-12134%&'() ,-./0-12134%&'()

= (5$,6 − 57 ) − 1 (2)

In Equation 2 5$,6 is the total economic benefit generated by the BNDES, 57 is the total benefit that would be generated if these funds would not have been lent to BNDES. We will use the following assumption: 57 = 9:; ?@;@A;B CD