Innovative techniques to evaluate the Impact of private sector development reforms: An application to Rwanda and 11 other countries

Transcription

1 Innovative techniques to evaluate the Impact of private sector development reforms: An application to Rwanda and 11 other countries Sachin Gathani Massimiliano Santini Dimitri Stoelinga 1 This version: February 06, 2013 Accepted by the MPSA Annual Conference, April 2013 Abstract The objectives of this paper are twofold: (1) to show how the synthetic control methodology can be used to measure the impact of private sector development reforms, and (2) to introduce a new technique, the proximity control methodology, that offers similar advantages but greater flexibility than the synthetic control methodology to test the validity of results. While maintaining the technical rigor of other econometric techniques used to conduct impact evaluations, these two methods are quicker costeffective alternatives that can be applied to measure ex-post the impact of policy reforms in a given country or region. They can also easily be replicated to similar reforms in other countries. We illustrate this by using both methodologies to estimate the impact of the introduction of a one-stop shop for business registration on new firm creation in Rwanda and 11 other countries. Both approaches yield similar and comparable results and show that one-stop-shops can have a very large impact: in Rwanda for example, we observe a 186% average increase in new firm creation after the reform was introduced, in Tajikistan a 132% increase, and in Belarus a 103% increase. In this paper we also propose a new way of looking at the Doing Business dataset by introducing a measure of the similarity of the business environment of a pair of countries. This metric offers a more accurate comparative representation of a country s business environment than more traditional metrics. JEL Classification: C21, C23, G18, L51, M13 Key Words: Entrepreneurship, Business Environment, Investment Climate, Impact Evaluation, Synthetic Control, Proximity Control, Doing Business, Rwanda 1 Massimiliano Santini ( msantini@ifc.org) is an Economist at the World Bank Group. Sachin Gathani ( sgathani@laterite- africa.com) and Dimitri Stoelinga ( dstoelinga@laterite- africa.com) are Partners at Laterite Ltd. ( africa.com), a research firm based in Rwanda and Malawi. This paper was prepared with assistance from Gabriela Armenta and Maria Paula Gomez. Thanks to Miriam Bruhn, Alexis Diamond, Jeffrey Grogger, Robert Lalonde, David McKenzie, and Ricardo Sabates for very helpful comments. We also thank the Rwanda Development Board for sharing updated monthly firm registration data. All mistakes in this paper are our own.

2 Table of Contents 1. Introduction Why business entry reforms matter: a brief summary of the literature Using the synthetic control methodology to measure the impact of private sector development reforms Data Rwanda: measuring the impact of a one-stop shop with the synthetic control methodology Testing the robustness of the results obtained using the synthetic control methodology Rwanda s one-stop shop and the link to new business creation Introducing the proximity controls methodology Measuring and visualizing the Doing Business similarity space The properties of the Doing Business similarity space Rwanda: measuring the impact of a one-stop shop with the proximity controls methodology Testing the robustness of the results obtained using the proximity controls methodology Replicating and comparing both methodologies to 11 other countries Conclusion Bibliography Annexes

3 1. Introduction In recent years, the development community has been facing increasing pressure from donors and client countries to show the impact of its interventions. Delivering technical assistance and lending money are not enough anymore. By isolating the intended result of the project from everything else that happened at the same time, rigorous impact evaluations tell us what works and why, making a case for successful policy interventions to be scaled up and replicated. Unfortunately, impact evaluations are usually costly and time consuming, and in the real world are implemented only in a handful of projects. As a result, all sort of assumptions need to be established to extrapolate the results obtained by one country-specific project to another country s similar intervention (for example, to which degree can we assume that conditional cash transfers that increase returns to education in Mexico will have the same effect in the Kyrgyz Republic?). While maintaining the technical rigor of other econometric techniques used to conduct impact evaluations, the synthetic control methodology is a low-cost alternative that can be applied to policy changes with aggregate country-level effects and easily replicated to similar policy changes in other countries. This methodology was first introduced by Abadie and Gardeazabal in T he objectives of this paper are (1) to show a step-by-step application of the synthetic control methodology to measure the impact of the introduction of a one-stop shop for business registration on new firms created in Rwanda, and replicate the methodology to eleven other countries; and (2) to introduce a new technique, the proximity control methodology, that offers the same advantages as the synthetic control methodology but may be easier to apply, and to show its application to the same business entry reforms in Rwanda and the eleven other countries. The principle behind both approaches is relatively straightforward: both techniques use a linear combination of control countries i.e. countries without one-stop shops to create a synthetic control region that accurately fits the reference country on certain variables of interest before the introduction of the one-stop shops. If the synthetic region closely resembles the reference country, and accurately predicts new firms creation in the reference country before the introduction of the one-stop shop, then the synthetic control region is likely to be a relatively accurate predictor of what would have happened in the reference country had the onestop shop not been introduced. 3

4 In addition, we will also show how to compare the business environments of countries by using a metric of business environment similarity developed for the proximity control methodology, which we argue presents many advantages over the Doing Business aggregate index of the ease of doing business. The paper proceeds as follows. After the introduction and a brief overview of the literature on the impact of business entry reforms, chapter three will make the case for using the synthetic control methodology to estimate the impact of private sector development reforms, and apply the methodology step-by-step to evaluate the impact of the introduction of a one-stop shop for business registration in Rwanda. Chapter four will introduce the proximity control methodology and apply it to the case of Rwanda too. Chapter five will summarize the results obtained using both methodologies estimating the impact of one-stop shop reforms in eleven other countries, before we conclude assessing the pros and cons of each methodology, and how they may be used to conduct cost-effective and rigorous impact evaluations of private sector development reforms. 2. Why business entry reforms matter: a brief summary of the literature In recent years, many countries have reformed the business entry process by reducing the time, cost, number of procedures and minimum capital necessary to start a business, with the declared intent of stimulating private sector activity. Typical reforms included the reduction of unnecessary license requirements, the streamlining of business entry processes, the reduction of overhead costs, and the improved coordination between regulatory agencies. Business entry reforms are relatively easier to implement than other private sector development reforms, and policy makers have found it straightforward to build consensus for their implementation. Approximately 80% of the 183 economies measured by the Doing Business report have made it easier to start a business since Specifically, 348 business registration reforms were introduced in 146 countries in June 2003-May 2011, about 20% of all investment climate reforms reported in the same period by the Doing Business report. 3 One of the most popular business entry reforms is the introduction of one-stop shops for business registration, which provide entrepreneurs with a single place where they can fulfill all the requirements necessary to start their businesses. A literature review on business entry reforms found that (1) the introduction of significant business entry reforms, like a one-stop shop 2 Doing Business database, Ibid. 4

5 for business registration, is directly associated with an increase in the number of firms, and that (2) a significant reduction in business registration costs affects new firm creation more in industries with low barriers to entry than in those with high barriers. 4 Klapper and Love (2010) show that business registration reforms that cut cost and/or time by more than 40% in 92 countries during the period, like the introduction of a one-stop shop, had a statistically significant impact on new business creation. Conversely, smaller reforms such as the city of Lima s (Peru) simplification of just the process to obtain a license to start a business seem to have had no significant effect on firm performance. 5 In Portugal, the introduction of a one-stop shop in 308 counties, which decreased the number of days to register a business by 91%, led to an increase of new firms created by over 17%. 6 On the other hand, a study of SMEs in Vietnam found that the decision of firms to formalize during the period led to an increase in firms gross profits and investments. 7 Preliminary evidence shows that significant business entry reforms can encourage job creation too. In Mexico, the introduction of a one-stop shop for business registration was associated with an increase of 2.2-8% in employment. 8 In Portugal, the introduction of the onestop shop led to an increase in employment by 21%. A cross-country study showed that a decrease of 61% in the number of days to register a business is associated with an increase of 0.4% in (manufacturing) employment. 9 In the next chapter, we will explain step-by-step how we can use the synthetic control methodology to estimate the impact of the introduction of the one-stop shop on new firm registration in Rwanda in The Government of Rwanda recently implemented two major waves of business environment reforms: the first in April/May 2009, when the one-stop shop for business registration was created; the second in April/May Due to limited data availability, we focus only on the impact of the first wave of reforms and in particular the impact of Rwanda s one-stop shop for business registration on new firm creation in Rwanda in We will also discuss various approaches to test the validity of the results obtained. 4 Motta et al. (2010). 5 Lorena Alcázar, Miguel Jaramillo, Panel/tracer Study on the Impact of Business Facilitation Processes on Microenterprises and Identification of Priorities for Future Business Enabling Environment Projects in Lima, Peru, Mimeo, Grade, June Branstetter, Lima, Taylor, Venancio, Do Entry Regulations Deter Entrepreneurship and Job Creation? Evidence from Recent Reforms in Portugal NBER Working Paper 16473, October John Rand, Nina Torm, The Benefits of Formalization: Evidence from Vietnamese Manufacturing SMEs, World Development Vol. 40, No. 5, pp , Bruhn, License to Sell: The Effect of Business Registration Reform on Entrepreneurial Activity in Mexico, World Bank Policy Research Working Paper No. 4538, January Kaplan, Piedra, Seira, Entry Regulation and Business Start- Ups: Evidence from Mexico, World Bank Policy Research Working Paper No. 4322, June Ciccone and Papaioannuou, Red Tape and Delayed Entry, Journal of the European Economic Association, vol.5, no.2-3, pp ,

6 3. Using the synthetic control methodology to measure the impact of private sector development reforms The synthetic control methodology developed by Abadie and Gardeazabal enables researchers to conduct aggregate-level impact evaluations at national, regional and sectoral levels. In the next chapter, we will show how it can be applied to estimate the impact of the introduction of a one-stop shop on new firms created in Rwanda. 10 Despite significant advantages over alternative techniques (both from a theoretical and cost-effective perspective) 11, this approach to impact evaluations has been underutilized in both policy and academic spheres. By illustrating how it can be applied in a concrete, step-by-step way, we hope to encourage its wider use by policy makers and researchers, and unlock its potential use for measuring the impact of private sector development reforms. As its name suggests, the main objective of the synthetic control methodology is to create a control region for a geographic area where a policy change (the intervention ) has taken place (the treatment region ). The control region is called synthetic because it is constructed using a linear combination of alternative regions where the intervention has not taken place (the control region ). By definition, a good control region is a region that perfectly matches the treatment region on a number of key characteristics, before the intervention takes place. 12 If two regions the treatment and the synthetic control regions are relatively similar on key characteristics and have a similar performance on the variable of interest over a certain period of time, then the synthetic control region is likely to be a good predictor of what would have happened in the treatment region had the event or intervention not taken place. The observed difference between the performance of the treatment region and the control region on the variable of interest after the intervention is our estimate of impact. For example, in their paper on the impact of terrorism on economic growth in the Basque region, Abadie and Gardeazabal (2003) construct a synthetic Basque region, using a linear combination of other Spanish regions that minimizes the difference between the synthetic Basque region and the actual Basque region on the following indicators: real GDP per capita, investment ratio, population density, sector shares as a percentage of GDP, and human capital indicators (illiteracy rate, primary and secondary education enrollment rates). The resulting synthetic Basque region closely resembles the Basque region on these economic determinants before the beginning of terrorist activity, and it also perfectly matches economic growth in the 10 For the purposes of this paper, we use the term firm as a synonym of business, company, partnership and corporation. 11 See Abadie et al. (2003). 12 Ibidem. 6

7 Basque country for a period of 20 years before the beginning of terrorist activity. The observed deviation between the Basque country and the synthetic Basque region after the start of terrorist activity is the author s estimated impact of terrorist activity on GDP growth in the Basque country. 3.1 Data In order to measure the impact of improving business entry regulation on new firms creation, we build a dataset that codes the 183 countries measured by the Doing Business 2012 report based on whether or not they have introduced a one-stop shop for business registration. The one-stop shop for business registration is defined as an organization that (i) receives documents for business registration and (ii) carries out at least one other function related to business start-up. (e.g. tax registration, social security registration, statistical agency registration, etc.). 13 In the same dataset, we indicate the year of introduction of the one-stop shop and source this information for each country. Our outcome of interest is new firms registered. We use the dataset from the 2010 World Bank Group Entrepreneurship Snapshots (WBGES), 14 which contains country-level data on new firms and entry density from 2004 to New firms are defined as private companies with limited liability, which is the same definition used by the Doing Business reports. Entry density is defined as the number of new firms per 1,000 working-age people (15-64 years old). Throughout the paper, we use the variable entry density to do all the calculations, but we often interpret the results in terms of the variable new firms because of its more common use. While the WBGES is the most comprehensive cross-country dataset currently available on firms registration, it excludes the registration of sole-proprietors. The Doing Business dataset on the ease of the business environment only includes data affecting limited liability companies, and it excludes data on sole proprietors as well. By limiting our analysis on the impact of the creation of limited liability companies, we may miss some implications on easing business regulations on micro and small firms, in particular in the context of the determinants of informality. On the other hand, we think that the results could serve as an accurate proxy for the dynamics of the formal sector as a whole. 13 Investment Climate Advisory Services, How many stops in a one- stop shop? A review of Recent Development in Business registration, Flagship report, The World Bank. 14 Publicly available from: 7

8 3.2 Rwanda: measuring the impact of a one-stop shop with the synthetic control methodology In order to measure the impact of a one-stop shop on new firm creation in Rwanda, we create a synthetic counter-factual of Rwanda following a 5-step process: Step 1. The first step in constructing a synthetic control for Rwanda is clearly defining the variable of interest. We are interested in measuring the impact of the introduction of the onestop shop in Rwanda on new firms created. Based on available data, there are two ways of measuring new firm creation: (i) by number of new firms registered; (ii) or by new business density (NBDEN), measured as the number of new firms registered per 1,000 inhabitants. In order to compare countries using the synthetic control methodology, the latter is preferable, as we can compare countries with similar characteristics keeping population constant. This is analogous to comparing two countries based on their GDP per capita rather than their total GDP. Once results have been obtained in terms of new business density, we translate them back into number of new firms registered, a more tangible metric for policy makers. Step 2. Which predictors do we select to match Rwanda to its synthetic control region, given that we are interested in new business density? Our objective is to create a region that is similar enough to Rwanda prior to the introduction of the one-stop shop, on both (i) new business density; and (ii) key characteristics that play an important role in determining the level of new firm creation. The selection of predictors should reflect our knowledge on the variables that are good predictors of new business density. Given the nature of new business creation, we choose to focus on selected macroeconomic variables that capture information on the structure and level of economic development of the economy: GDP per capita, agriculture (% GDP), industry (% GDP), services (% GDP), gross fixed capital formation (% GDP), trade balance (% GDP), and urban population (% total population). As we can see in graph 1 below, these variables are good predictors on average of 8

9 the level of new business density (R²=0.49, using average data from 105 different countries) Log New Business Registrakon (Predicted) Graph 1. Log New Business Density: Predicted vs Actual ( ) R² = Log New Business Registrakon (Actual) Step 3. The third step in constructing a synthetic control for Rwanda involves selecting the time period during which the difference between Rwanda and the synthetic Rwanda is minimized. Given that Rwanda introduced the first package of business registration reforms connected with the one-stop shop in 2009 with immediate effect we use data prior to 2009 to match Rwanda to the optimal linear combination of control countries. Data on new business density in Rwanda is only available starting in 2003, so we match Rwanda to its synthetic control using data; we call this the input period. The output of the synthetic control method is an estimate of new business creation in Rwanda s Synthetic Control both before and after the introduction of the one stop shop. We call the period the output period. We exclude the year 2010 and the impact of Rwanda s second major reform package from the analysis because business registration data is not available for some of Rwanda s comparators. 15 We predict new business registration data with a simple regression of selected explanatory variables on average values for the period. 9

10 Step 4. Then, we identify a pool of potential control countries from which the synthetic Rwanda is constructed borrowing from the statistical literature on matching, Abadie et al. (2007) call this the donor pool. We establish the donor pool using three criteria: 1. Given that the treatment is the introduction of one-stop shop, we eliminate from the donor pool all countries that already had or introduced a one-stop shop during the output period i.e. before and after Rwanda introduced its one-stop-shop. This leaves us with a pool of countries where the treatment did not take place. This means that any Synthetic Country constructed as a linear combination of any one of these control countries did not experience the introduction of a one-stop shop at any point between 2003 and We eliminate from the donor pool all countries for which we do not have the required data during the period This includes: (i) all countries for which new business density data is missing during the period; (ii) all countries for which we do not have at least one data point during the input period for each of the predictors. 3. In order to avoid biases caused by interpolating across regions with very different characteristics, 16 we eliminate from the donor pool all countries that on average during the input period had a GDP per capital level greater than USD 1,000 (constant USD), compared to USD 275 for Rwanda. The objective is to strike a balance between the size of the donor pool on one hand, and how similar the characteristics of countries within that donor pool are on the other. This leaves us with a donor-pool of seven countries for Rwanda: Cambodia, Ethiopia, Indonesia, Moldova, Malawi, Pakistan, and Uganda. Step 5. With the variable of interest, the predictors, the time-period and the donor pool now in place, we can construct a synthetic Rwanda following the methodology outlined by Abadie et al. (2003 and 2007). Synthetic Rwanda is constructed as the linear combination of countries in the donor pool that most closely resemble Rwanda in terms of the variable of interest and the predictors prior to the introduction of the one-stop shop (i.e. during the period). 16 See Abadie et al. (2007). 10

11 Table 1: Predictors (averages ) Predictors Rwanda Synthetic Rwanda GDP per capita (constant 2000 USD) Agriculture (% GDP) Industry (% of GDP) Services (% of GDP per capita) Trade balance on goods and services (% GDP) Gross Fixed Capital Formation (% of GDP per capita) Urban population (% total population) The resulting Synthetic Rwanda consists of a linear combination of Cambodia (40.5%), Malawi (32.5%) and Ethiopia (27%). As we see on table 1, this Synthetic region closely matches Rwanda on average on most of the selected predictors during the period (pre-oss): GDP per capita, agriculture (% GDP), the trade balance, and the level of urbanization are almost identical in both regions. Synthetic Rwanda is slightly more industrialized however (19.9% vs. 13.8%), less service intensive (44.2% vs. 49.5%), and has a slightly higher investment rate (20.3% vs. 17.6%). These difference are however relatively small and remain constant during the period; i.e. they do not explain the observed jump in new business registration between 2008 and In addition to fitting Rwanda on the selected explanatory variables, synthetic Rwanda predicts new business density in Rwanda during the period very well, before the introduction of the one-stop shop (see graph 2). 11

12 Graph 2. New Firm RegistraGon in Rwanda and SyntheGc Rwanda New Firm Registrakon Rwanda Synthekc Rwanda The graph shows that while new business registration in Rwanda and Synthetic Rwanda followed a relatively similar growth path between , new business registration in Rwanda greatly increased in 2009 after the introduction of the one-stop shop. Following Abadie et al. (2003), a simple difference-in-difference calculation enables us to estimate the impact of Rwanda s one-stop shop on new business density. We estimate that that the introduction of the one-stop shop and related reforms led to the registration of 2,041 new firms in 2009 alone, which is equivalent to an increase of 188% in new firms created after the one-stop shop was introduced (note that the one-stop shop was created in May 2009). In other words, after the introduction of the one-stop shop, 188% more firms registered than they would have registered had the one-stop shop not been introduced. In the next section we test the robustness of these results, before focusing in chapter 3.4 on why this increase may be attributed to the introduction of the one-stop shop and not to other factors happening in the country at the same time. 12

13 New Firm Registrakon Graph 3. New Firm RegistraGon in SyntheGc Rwanda and Placebo 500 Rwanda Synthekc Rwanda Placebo Difference in NBDEN between reference country and Syntekc Control Graph 4. EsGmated New Business Density Increase in Rwanda and other controls Rwanda 3.3 Testing the robustness of the results obtained using the synthetic control methodology The final step involves inferential analysis and testing the validity of the results. Following Abadie et al. (2003, 2007) we propose two different techniques, the falsification test and the Mean Squared Prediction Error test. 17 The objective of the falsification test is to ensure that Synthetic Rwanda did not experience a shock itself in (a treatment ) as this would entail that we are either under-estimating or over-estimating the impact of Rwanda s onestop shop on new business registration. To test whether Synthetic Rwanda experienced a positive or negative shock in , we apply the synthetic control methodology described above (steps 1-4) to Synthetic Rwanda (which consists of a linear combination of Ethiopia, Malawi and Cambodia), i.e. create a synthetic region for Synthetic Rwanda. We call the newly created synthetic region, the Placebo region. As can be seen on graph 3, Synthetic Rwanda and its Placebo region do not differ significantly during the period. The match between the two regions is not perfect Synthetic Rwanda seems to experience a small bump in but the results strongly suggest that Synthetic Rwanda did not experience an impact (or shock) in that would explain the observed difference between new business registration in Rwanda and Synthetic Rwanda. 17 We define the Mean Square Prediction Error as the mean of the squared differences between new business density in one region and another over a certain period of time. 13

14 The difference between Synthetic Rwanda and the Placebo was 180 businesses in 2008, compared to 172 in 2009, which is almost identical. The observed bump in Synthetic Rwanda is due to Cambodia; in order to ensure that Cambodia (which accounts for 40.5% of Synthetic Rwanda) is not significantly skewing results, we repeat the exercise without Cambodia and find an impact of 180%, slightly lower than the estimated 188%. While this suggests we could be overestimating the impact of the one-stop shop by 8 percentage points, it does not significantly alter our overall conclusion that the one-stop shop has had an impact on new business registration in Rwanda. The second test asks the question: how unusual is the impact estimate obtained? Is it due to chance? To answer this question we conduct synthetic control tests on all the countries in Rwanda s donor pool (testing for an impact in 2009) and compare the results to the estimated impact of the one-stop shop on new business density in Rwanda. If the estimated impact for Rwanda is unusual higher than in other donor pool countries then this is additional evidence that the one-stop shop had an impact on new business density in Rwanda. If this is not the case, then the observed difference between new business density in Rwanda and Synthetic Rwanda could be due to chance rather than the one-stop shop. Abadie et al. (2007) demonstrate that this iterative effort leads to exact inference. Graph 4 clearly indicates that the estimated impact on new business density in Rwanda is unusual compared to the seven other countries in Rwanda s donor pool, thereby providing additional evidence that the one-stop shop had a statistically significant impact on new business density in Rwanda. Another way of looking at these results is to compare the ratio of the Mean-Square Prediction Error (MSPE) before and after the introduction of the one-stop shop in Rwanda to that of the other countries/controls in Rwanda s donor group. We formally calculate the MSPE for country i using the following formula: MSPE! = (Actual!,!""# Synthetic Control!,!""# )!!! (Actual!,!""#!!""# Synthetic Control!,!""#!!""# )! 6 An MSPE smaller than 1 indicates that the observed impact in 2009 is not unusual i.e. it is smaller than in other years before the intervention, while an MSPE of more than 1 indicates that the observed impact is larger than in other years. The size of the ratio enables us to compare how unusual the observed impact is across countries. Not surprisingly, we find that 14

15 this ratio is much higher in Rwanda than in other countries (see table 2) indicating that the impact we observe in Rwanda is indeed unique. We conclude that the introduction of the one-stop shop in Rwanda had a significant impact on new business density and led to an approximate increase of 188% in new firms created. Table 2: MSPE test Country MSPE Ratio Rwanda Ethiopia 1.00 Indonesia 0.71 Cambodia 0.02 Moldova 0.35 Malawi 0.60 Pakistan 1.80 Uganda Rwanda s one-stop shop and the link to new business creation Questions remain on whether the increase in new firms registered can be attributed exclusively to the introduction of a one-stop shop or instead to other investment climate reforms occurring concurrently or other changes in the economy. We argue that new business registration can be linked to the introduction of the one stop-shop but is unlikely to be linked to the introduction of concurrent reforms happening at the same time. Moreover, as we show in the Annex, we find the same substantive and significant impact on new firms created in eleven other countries after the introduction of a one-stop shop. Even before introducing major reforms in 2009 and 2010, Rwanda had been an active reformer of its business environment: in 2001, the Government introduced a new labor law; in 2002 a property titling reform; in 2004, it simplified customs procedures, improved the credit registry and undertook court reforms; in 2007, Rwanda reformed property registration and further improved customs procedures; and in 2008 certain judicial reforms came to completion, leading to the introduction of new commercial courts. 18 Despite these reforms, operating a 18 Doing Business 2009, World Bank Group. 15

16 business in Rwanda at the beginning of 2009 was not easy. Rwanda ranked 139 th in the Doing Business 2009 report (published in September 2008) and at about 0.19 firms per 1,000 people (of working age), new business registration was one of the lowest in the world. The year 2009 marked a major acceleration in the improvement of Rwanda s investment climate: on April 27 th 2009 Rwanda enacted a new Companies Act, followed by the Mortgage Law, the Secured Transactions Law and the Insolvency Law in May The Companies Act strengthened investor protection and created the one-stop shop for business registration, which opened its doors in May The creation of the one-stop shop led to an impressive reduction in the time, cost and number of procedures required to start a business: the number of procedures was reduced from 8 to 2, the time from 14 days to 3, and the cost from 108.9% of GDP per capita to 10.1%. The Secured Transactions Law improved access to credit by increasing the range of assets that can be used as collateral; the Insolvency Law eased the process of filing for bankruptcy and closing a business; while the Mortgage Law shortened the process of property registration. This mix of reforms resulted in Rwanda greatly improving its business environment. 20 This first reform package was followed by a second package in April and May 2010, when the business environment was further facilitated by the introduction of free online registration, a reduction in registration fees, new regulations regarding construction permits, a further reduction in the documents required for exports and reforms in the access to credit space. May 2009, when the first package of business reforms were implemented, also marked a turning point in new business registration (see graph 5). New business registration increased from an average of about 100 firms per month from January 2008 through to April 2009, to an average of about 300 firms between May 2009 and December In the sixteen months from January 2008 to April 2009, 1,552 firms were registered in Rwanda; an equivalent number of firms were created in just five months after these business reforms were enacted. After the second package of reforms were enacted, new business registration further increased to reach firms per month. 19 See Official Gazette n special of 14/05/ Rwanda improved its ranking in the Doing Business 2010 report from 139 to 67 in the space of a year, becoming the world top reformer. 16

17 Graph 5: New business registragon per month in Rwanda No firms per month Jan- 08 Mar- 08 May- 08 Jul- 08 Reform package 1 Sep- 08 Nov- 08 Jan- 09 Mar- 09 May- 09 Jul- 09 Reform package 2 1 Sep- 09 Nov- 09 Jan- 10 Mar- 10 May- 10 Jul- 10 Sep- 10 Nov- 10 Source: the Rwanda Development Board (2012) No macro-economic changes in April and May 2009 can justify the observed increase in new business registration. On the contrary, Rwanda, like many other developing countries, suffered a slowing down in economic growth in 2009 as a direct result of the global financial crisis. Rwanda s real growth rate in 2009 was 6.2%, compared to 11.2% in In most countries, the crisis translated into a drop in new business registration in 2009, which is what we would have expected to see in Rwanda given the 5 percentage point drop in GDP growth. 21 Instead, the timing of the increases in new business registration suggests that the implementation of investment climate reforms, and in particular the one-stop shop for business registration, are responsible for the increase in new business registration. Both accelerations in new business registration in 2009 and 2010 occurred in the months of May and June, just after the reform packages were implemented (see graph 5). While it is impossible to disentangle the respective impact of each investment climate reform given that they were all passed approximately at the same time and that we have limited data on the implementation timelines of these reforms the most likely explanation as to why new business registration accelerated immediately after the reforms were passed is the creation of the one-stop shop and the associated decrease in the procedures, cost and time of starting a new business. The one-stop shop had an immediate impact: it was widely publicized, it provided a direct window for new business to register, and offered a very efficient service to entrepreneurs. 21 Klapper, Leora and Love, Inessa, The impact of the financial crisis on new firm registration, Economics Letters, vol. 113(1), pages 1-4, October. 17

18 Other reforms certainly increased the attractiveness of starting a business, but the passthrough effect is likely to have been much more gradual: Access to credit was improved, in theory, by the passing of new laws regarding collateral, but credit to the private sector actually decreased in Rwanda between the months of May and November 2009, making it unlikely that increased access to credit was behind the rapid increase in new business registration in May and June 2009 (see graph 6). The new insolvency law may have improved insolvency procedures, but even today Rwanda remains one of the least business friendly places in the world to close a business (Rwanda ranked 165 th on closing a business in DB2012). The 2009 reform to construction permits streamlined processes by combining the applications for location clearance and the building permit in a single form and introducing a single application form for water, sewerage, and electricity connections. While this resulted in a reduction in the number of procedures and the time required for dealing with construction permits, it nevertheless still required more than 200 days and 600% of income per capita to obtain a construction permit in Kigali City s One Center for construction permits, which resulted in a further decline in the time required to get construction permits only started in operations in April Moreover, most new firms created in Rwanda since 2009 are in the retail/wholesale sector and do not necessarily involve construction. 24 Construction permit reforms in 2009 fail to explain the immediate post-reform increased in new business registration in the months of May and June IFC and World Bank, Doing Business in a more transparent world. Economy Profile: Rwanda 23 Official Gazette n 22 bis of 31/05/ Rwanda National Institute of Statistics, Establishment Census 2012, 18

19 Rwf (bn) Jan- 08 Graph 6. DomesGc credit to the private sector (by month) Source: Na0onal Bank of Rwand Mar- 08 May- 08 Jul- 08 Sep- 08 Nov- 08 Jan- 09 Mar- 09 May- 09 Jul- 09 Sep- 09 Nov- 09 Jan- 10 Mar- 10 May- 10 Jul- 10 Sep- 10 Nov- 10 The combination of these factors strongly suggests that while concurrent reforms to the one-stop shop, enacted in May 2009, contributed to improving Rwanda s general business environment, they are unlikely to have had an immediate impact on new firm creation. 4. Introducing the proximity controls methodology The proximity control methodology is largely inspired by the synthetic control methodology, but it relies on a different technique - and application - to construct the synthetic region. By using the Doing Business indicators, we show that it is possible to construct a relatively accurate control region for a reference country using linear combinations of countries with the most similar business environment to the reference region. We also highlight a different way of looking at the Doing Business indicators comparing countries not by ranking but by how similar their Doing Business indicators are. This can lead to many other interesting applications and presents a new way of representing and communicating the findings of the Doing Business reports. The logic of the approach we propose is straightforward: if we assume that a country s business environment as measured by Doing Business is one of the most important determinants of new firms creation - and that changes in a country s business environment as measured by Doing Business can significantly impact new firms creation - then it is likely that countries with very similar business environment as measured by Doing Business will have similar new business density or new business density growth. If this assumption holds, then it should be possible to estimate new business density or new business density growth in a 19

20 reference country by looking at new business density in the countries with the most similar business environments. Should these comparisons result in accurate estimates of new business density in the reference country before a policy change (e.g. the introduction of a one-stop shop), then we could use these comparator countries as control regions for the country of interest and conduct a counter-factual analysis. In the next chapters we develop and explain this approach, which is inspired by an analogous method called proximity control, where export similarity networks are used to conduct counter-factual analysis Measuring and visualizing the Doing Business similarity space Cross-country comparisons show that countries with similar overall rankings in the Doing Business report can have different business environments. To illustrate this, imagine the extreme case where country Alpha is the world s best performer on half of the Doing Business indicators, and the world s worst performer on the other half, and where country Beta is the exact opposite (i.e. the best where country Alpha is the worst, and the worst where country Alpha is the best). Then country Alpha and Beta would have the same overall ranking in the doing Business report, but in reality they would have radically different business environments. In order to avoid such scenarios and get a more accurate estimate of how similar is the business environment of a pair of countries, as measured by the Doing Business report, we introduce the new metric of Doing Business similarity (DBSim). The metric we propose is very simple, and it assumes that all Doing Business indicators (we use 32 indicators in total) are equally important in determining how similar or dissimilar are the business environments of two countries, as measured by the Doing Business report. To measure DBSim between any pair of countries a and b, we first calculate a measure of the distance (d!,! ) between their business environment as measured by Doing Business (basically a measure of how dissimilar their business environments are) by: (i) standardizing each country s score on all the Doing Business indicators; (ii) summing the squared difference between the standardized scores of both countries across all the indicators; and (iii) dividing by the number of indicators for which data is available for both countries, so that we get the average distance. Finally, to obtain a measure of similarity, rather than dissimilarity, we use the exponent of minus the distance. Formally, this can be written as: 25 See Gathani et Stoelinga (2013). 20

21 ! d!,! = 1! S N!,! S!,!!!! and, DBSim!,! = e!!!,! where S!,! is country a s standardized score on indicator i. With this measure of similarity, we can now answer questions such as: which ten countries have the most similar business environment as measured by Doing Business to country Alpha? Let us illustrate this with an example, by identifying the countries with the most similar business environment to Rwanda and Georgia based on data from the Doing Business 2010 report (see table 3). Table 3: Countries with most similar business environment to Rwanda and Georgia Similarity to Georgia in Doing Business 2010 (ranked 11 th ) Similarity rank Country d DBSim DB2010 rank Similarity to Rwanda in Doing Business 2010 (ranked 67 th ) Similarity rank Country d DBSim DB2010 rank 1 Estonia Mongolia Saudi Arabia Botswana Lithuania Kyrgyz Republic Mexico South Africa Sweden Paraguay Germany Azerbaijan Chile Burkina Faso FYROM FYROM Thailand Vanuatu UK Malawi While Georgia was ranked 11 th in the Doing Business 2010 report, its 10 closest comparators were ranked between 5 th and 51 st. As for Rwanda the spread is even larger: Rwanda was ranked 67 th in the Doing Business 2010 report, but its 10 closest comparators were ranked 32 nd to 147 th. Countries can have very different Doing Business rankings but have quite similar business environment as measured by Doing Business and vice-versa. 21

22 This way of looking at the Doing Business indicators also provides a different perspective to policy makers. Rather than aiming at increasing their Doing Business rankings, Governments should aim at becoming more similar to a certain role model or compass country. Georgia, for example, should become more similar to the UK, and Rwanda to South Africa. This approach also highlights the country s weaknesses and strengths better; while Rwanda s business environment as measured in Doing Business 2010 report was good enough to be compared to a country like FYROM (ranked 32), it was also weak enough to be compared to Burkina Faso (ranked 147). One of the side benefits of measuring the Doing Business similarity between all pairs of countries for which data exists is that we can place countries within a Doing Business similarity network. The closer countries are to each other in this network, the more similar their business environment as measured by Doing Business are and vice-versa. This provides a powerful visual tool that we can use to represent Doing Business results. We illustrate the use of this tools in graph 7-A and 7-B. Graph 7-A: Countries with similar business environments and their GDP level 22

23 Graph 7-B: with similar business environments and their depth of reforms Graph 7-A is a network representation of all countries for which data is available with links to their three closest comparators in the Doing Business similarity space (using data from the Doing Business 2010 report). Countries (the nodes in the network) are colored by their GDP per capita levels, ranging from green - for countries with the lowest GDP per capita levels - to red - for countries with the highest levels of GDP per capita. The graph reveals that countries with similar levels of GDP per capita tend to have similar business environments as measured by Doing Business all the reds are clustered together, as are the countries in orange, yellow and green. Yet there are a number of clear outliers: Georgia, for example, performs much better in the Doing Business indicators than we would expect given its level of GDP per capita (Georgia, in yellow, is surrounded by countries in red and orange); while Brazil, on the other hand, performs much worse (Brazil is in dark orange, but it is surrounded by countries in light orange, yellow and green). Graph 7-B represent the same network, only this time countries are colored not by their GDP per capita levels, but by how large their reforms to the business registration process were during (as captured by the Doing Business reports 2009 and 2010). The greater the red shading of the country (or node), the deeper the reforms. Countries in white did not 23

24 introduce any reforms during this period. This network reveals that in the countries that made the biggest reforms to business registration regulations were Rwanda and Belarus. Also, it shows that about half the countries in the world carried out some form of reform to business registration regulations during that period. 4.2 The properties of the Doing Business similarity space After having established a way to measure Doing Business similarity, we need to better understand the properties of the similarity space and how we can use them to conduct inferential analysis. In this section, we show that countries with similar business environment as measured by Doing Business are similar on a whole range of other economic indicators. We choose to focus on two properties of the Doing Business similarity space, which are relevant to our analysis. These properties highlight a clear link between the business environments of countries and their levels of GDP per capita, GDP growth, new business density, and new business density growth. These results are new to the Doing Business literature and are very strongly statistically significant. They are very much in line with the findings on export similarity space. 26 Property 1: The more similar the business environment of a pair of countries, the more similar on average are their levels of GDP per capita and their levels of GDP per capita growth As graphs 8 and 9 show, there is a very strong and statistically significant association between DBSim and the average difference in GDP per capita and GDP per capita growth between pairs of countries. What this means is that countries that have similar business environment as measured by Doing Business are likely to grow at relatively similar GDP per capita growth rates (+/- 1% on average) and are likely to have relatively similar levels of GDP per capita. While these results do not imply a causal relationship in any way, they do reveal that Doing Business similarity well captures how similar countries are on two fundamental economic variables: GDP per capita and GDP per capita growth. 26 See Gathani et Stoelinga (2013). 24

25 Average difference in log GDPpc between pairs of countries (2009) Graph 8. Average difference in GDP per capita between pairs of countries based on their DB similarity score (DB2010) R² = DBSim of pairs of countries (DB2010) Average difference in GDPpc growth between pairs of countries ( ) Graph 9. Average difference in GDP per capita growth between pairs of countries based on their DB similarity score (DB2010) 4% 3% 2% 1% 0% R² = DBSim of pairs of countries (DB2010) Average difference in log NBDEN between pairs of countries (2009) Graph 10. Average difference in NBDEN between pairs of countries based on their DB similarity score (DB2010) R² = DBSim of pairs of countries (DB2010) Average difference in NBDEN growth between pairs of countries (2009) Graph 11. Average difference NBDEN growth between pairs of countries based on their DB similarity score (DB2010) 60% 50% 40% 30% 20% 10% 0% R² = DBSim of pairs of countries (DB2010) Property 2: The more similar the business environment of a pair of countries, the more similar on average their levels of new business density and their levels of new business density growth (see graph 10-11). The results we obtain for new business density (NBDEN) and new business density growth are similar to the results we obtain for GDP per capita and GDP per capita growth. These results show a clear link between country s performance on Doing Business and new business registration. 25

26 These two properties suggest that countries with similar business environment as measured by Doing Business are quite similar on the main explanatory variables that we use in the synthetic control tests. A linear combination of a country s closest neighbors in the Doing Business similarity space should therefore result in a control region with similar levels of new business density, GDP per capita, GDP per capita growth, and other relevant explanatory variables to the country of interest. In the next chapter, we show that these synthetic countries can be used as control regions to test the impact of policy changes, and we choose to test this methodology to the introduction of one-stop shops on new business registration in order to compare it to the synthetic control methodology. 4.3 Rwanda: measuring the impact of a one-stop shop with the proximity controls methodology We apply the concepts of proximity control and randomized permutations 27 to Rwanda, which introduced a one-stop shop for business registration in Our objective is to measure the impact of Rwanda s one-stop shop and related reforms on new firms creation by using a control region or proximity control - for Rwanda which fulfills the following three criteria: (i) the proximity control should accurately estimate new business density in Rwanda before the introduction of the one-stop shop; (ii) the proximity control should have similar levels of GDP per capita and other selected explanatory variables to Rwanda; and (iii) the proximity control should be composed of countries which did not have a one-stop shop during the period under consideration. The following three steps are required to construct a proximity control for the Rwanda during the period: Step 1. As in the case of the synthetic control approach, it is necessary to select a time period during which the difference between Rwanda and the proximity control Rwanda is minimized. Given that Rwanda introduced its one-stop shop in 2009, we use data prior to 2009 to construct the proximity control Rwanda. Given that new business density in Rwanda is available for the period, we match Rwanda to its synthetic control using data. 27 Ibidem. 26

27 Step 2. Next we identify a pool of potential control countries from which proximity control Rwanda can be constructed we call this the donor pool. The process we follow is different to the synthetic control approach and imposes fewer restrictions on the quality of the data: 1. We eliminate from the donor pool all countries that already had or introduced a onestop shop during the period. This leaves us with a pool of countries where the treatment did not take place. 2. From the remaining countries in Rwanda s Doing Business similarity space (using data from the Doing Business 2009 report), 28 we select Rwanda s five closest neighbors (i.e. the countries with the highest Doing Business similarity to Rwanda). According to the properties of the Doing Business similarity space, these are the five non-oss countries which in 2008 were the most likely to have similar levels of new business density, GDP per capita, and GDP per capita growth, as well as similar Doing Business indicators. Depending on the quality of the data and the quality of the resulting fit, we can choose to include more or less countries in the pool we leave this to the discretion of the researcher. The more countries are in the pool, the more likely we are to obtain a good fit for the variable of interest (in this case new business density); however, the difference between the business environments of the country of interest and the proximity control will be larger as each extra country added to the pool is farther away from the country of interest in the Doing Business similarity space. 3. Finally, to ensure the quality of the resulting fit, we eliminate from this list countries for which we have less than a minimum number of observations in time. Here we are interested in the period, so we eliminate from the sample countries for which new business density data is not available during this period. In the case of Rwanda, we remain with a donor pool of 5 countries: Malawi, Ethiopia, Niger, Uganda, and Croatia. 28 Data from the Doing Business 2009 report were collected in July 2008 June 2009, before the introduction of the one- stop shop in Rwanda. 27

28 Step 3. From this pool of five countries, we select combinations of countries to construct a proximity control for Rwanda and conduct inferential analysis. There are infinite combinations of these five countries when both weights and the different country combinations are taken into consideration. Moreover, by construction, not all combinations are good predictors of new business density in Rwanda before the introduction of the one-stop shop. To overcome these two challenges, we generate 15,000 random permutations of these five countries (in groups of 4) and select a limited number of permutations (in this case 12) that best match NBDEN in Rwanda during the period. We take the mean of these best permutations as our proximity control for Rwanda. We explain in more detail below how to select the optimal number permutations to construct the proximity control, and why in the case of Rwanda this number is 12. The resulting proximity control for Rwanda is composed of the linear combination of countries as shown in Table 4. Table 4: Composition of proximity control Composition of proximity control Weight Malawi 72.37% Niger 14.45% Ethiopia 11.11% Croatia 2.06% This linear combination of countries performs slightly better than Rwanda s synthetic control in fitting Rwanda s predictors during the , except on balance of trade which is more negative in the proximity control (see table 5). This suggests that Rwanda and its proximity control are very similar on the most relevant variables of interest. 28

29 Table 5: Predictors (averages ) Predictors Rwanda Synthetic Rwanda proximity control GDP per capita (constant 2000 USD) Agriculture (% GDP) Industry (% of GDP) Services (% of GDP per capita) Trade balance on goods and services (% GDP) Gross Fixed Capital Formation (% of GDP per capita) Urban population (% total population) Graph 12. New Firm RegistraGon in SyntheGc Rwanda and Placebo New firm reigstrakon Rwanda Proximity Control Synthekc Rwanda Also, as can be seen in the graph 12, Rwanda s proximity control is a relatively good predictor of new business density in Rwanda during the period, before the introduction of the one-stop shop. The average difference in new business density between Rwanda and its proximity control is about 125 businesses on average per year, compared to 87 for the synthetic control, making it a slightly less accurate predictor than Synthetic Rwanda. A difference of 125 new firms however remains small compared to the 2,900 new firms that were registered in Rwanda in

30 We can now estimate the impact of Rwanda s one-stop shop on new business density. While Rwanda and its proximity control had a similar performance on new business registration during the period, graph 12 shows that new business registration increased exponentially in Rwanda in 2009 following the introduction of the one-stop shop. In 2009, the difference in new business registration between Rwanda and its synthetic control was 2,260 firms compared to only 125 on average during A simple difference-in-difference calculation using 2008 as the base year enables us to estimate that the introduction of a onestop shop in Rwanda led to the creation of 1,994 firms, which is equivalent to an increase of 184% in new firm registration; not too different from our previous estimate of 188% using the synthetic control methodology. 4.4 Testing the robustness of the results obtained using the proximity controls methodology The most important questions to answer when attempting to measure the impact of an intervention is whether it is possible to produce valid statistical inference. We do it here using six different approaches. Test 1: Random permutations test The first approach consists in testing the sensitivity of the proximity control to changes in its composition. As discussed above, to construct a proximity control for Rwanda we first created 15,000 randomly generated different combinations of the 5 countries in Rwanda s donor pool, namely Malawi, Niger, Ethiopia, Croatia and Uganda, which were the 5 countries with the most similar business environment to Rwanda in Each of these 15,000 combinations is a potential proximity control with different weights assigned to each country in the pool. If Rwanda s one-stop shop, introduced in 2009, had led to no impact or little impact in that same year, then we would expect the difference-in-difference of new business density in Rwanda in 2009 compared to these 15,000 different combinations to be close to 0 on average and/or not statistically significant. Yet, as can be seen in table 6 below where we take 2008 as the base year to calculate difference-in-difference impact estimates, we estimate that the minimum impact of the OSS reform was at least 1,304 registered newly registered firms. In all 15,000 cases, the estimated impact of Rwanda s one-stop shop in 2009 was highly positive and 30

31 we can therefore reject the null-hypothesis that the difference-in-difference in 2009 was nil (tstatistic: 305). Table 6: Difference-in-difference estimates of new business density in Rwanda compared to 15,000 random combinations of control countries Difference-in-differences year on year Year Minimum number of new firms registered Maximum number of new firms registered However, as discussed, not all combinations are good predictors of new business density in Rwanda before the interventions and are hence unlikely to be good controls thereafter. In order to test whether the selected proximity control which is the average of the 12 best matching combinations is a good estimate of impact or not, we repeat this exercise with the 12 combinations that best match new business density in Rwanda between 2003 and As can be seen in graph 13, which shows the distribution of estimated impact using each of these 2 combinations as a control region for Rwanda, every single combination puts the impact in 2009 between 1,940 and 2,025 new firms, which corresponds to a minimum impact of 179% and a maximum impact of 186%. The maximum of the distribution curve comes at about 2000 firms, which is close to our estimated impact of 1,994 new firms registered. Graph 13. DistribuGon of esgmated impact using 12 closest combinagons Share of alternakve Proximity Controls 45% 40% 35% 30% 25% 20% 15% 10% 5% 0% Eskmated Impact (number of firms) 31

32 While the distribution above confirms the impact estimate, these differences could be due to marginal changes in the 12 combinations. This would be the case for example if all 12 combinations had Malawi contributing between 46% and 50% to the linear combination, Niger, between 23% and 25%, and so forth. It would imply that the 12 different combinations of controls are actually not alternative scenarios, but slightly different versions of the same scenario. It would also mean that the 7 percentage point spread we observe between the minimum and maximum impact estimates is due to very small changes in the composition of the control region therefore, the proximity control is actually quite sensitive to small changes in its composition. If this were the case, it would be necessary to increase the sample size, and focus not on the best 12 matches, but a larger sample. To check whether this is indeed an issue, we propose a measure of the composition diversity of the selected control regions, which also enables us to identify the optimal number of different combinations from which to construct the proximity control. We measure composition diversity using the following formula:!! Diversity! = (MaxShare! MinShare! )! 5 (0.25)! where MaxShare! is the maximum share of control region i in either of the N different linear combinations, MinShare! is the minimum share of control region i, and where i ( 0,5 ). If Diversity! is greater than 1 then there is more variation in the composition of the N linear combinations than if the contribution of each control region varied by 25%; if on the contrary Diversity! is smaller than 1, then there would be less variation in the composition of the N linear combinations than if the contribution of each control region varied by 25%. Diversity! is an increasing function of N: the higher the number of linear combinations selected (N), the larger the diversity of the combinations selected. With 5 control regions in the donor pool, maximum diversity is achieved when Diversity! =16, which would imply that each of the control countries is represented in the N different linear combinations with shares ranging from 0% to 100%. A diversity level of 16 however is not desirable as mathematically it is impossible for all linear combinations of control countries to be good predictors of new business density in Rwanda before the one-stop shop was introduced. In order to ensure a minimum level of variation in the selected linear combinations that make up the proximity control, we propose that the researcher select the smallest possible N such that Diversity! is greater or equal to 2. This would mean that there is at least 2 times 32

33 more variation in the linear combinations that contribute to the proximity control, than if the contribution of each control region in the donor pool varied by 25%. Moreover, by selecting the smallest possible N, we are selecting the best possible match between the proximity control and the reference country on the variables of interest. Table 7: Calculating Diversity 12 to test Diversity in Rwanda s proximity control Countries in donor Pool Minimum share in selected linear combinations Maximum share in selected linear combinations Squared difference between minimum and maximum Malawi 35.0% 98.7% Niger 0.0% 36.9% Ethiopia 0.0% 62.3% Croatia 1.3% 2.6% Uganda 0% 0% Sum of squared differences Diversity Based on this metric we find that the optimal number of linear combinations needed to construct Rwanda s proximity control is 12. The diversity of these 12 best linear combinations is 2.98, well above the 2-point threshold. Within these 12 linear combinations, the contribution of Malawi varies from 35% to 98.7%, the contribution of Niger from 0 to 36.9%, and the contribution of Ethiopia from 0 to 62.3% (see table 7), thereby ensuring that the 7 point spread we observe in the impact estimates is not due to small variations in the composition of the 12 linear combinations, but significant changes in their composition. Test 2: Sensitivity to changes in donor pool Another way of testing the sensitivity of the proximity control to changes in its composition is to change the countries in the donor pool. We test 21 different scenarios: (i) first replacing each of the countries in the pool with the 6 th closest country to Rwanda in the Doing Business space; (ii) second replacing each of the countries with the 7 th closest country to Rwanda; (iii) third testing the 10 possible combinations of replacing two countries out of the 5 in the donor pool simultaneously with the 6 th and 7 th closest countries; and (iv) lastly replacing all 5 countries with in the pool with the subsequent 5 countries closest to Rwanda in the Doing Business space. As the results in graph 14 indicate, the proximity control is not very sensitive to 33

34 changes in its composition. While impact estimates using the 21 alternative controls range from 1,794 to 2,001 new firms created - which is equivalent to an impact of respectively 165% and 184% - the majority of estimates are clustered around the 1,960-2,000 new firms mark. 60% Graph 14. DistribuGon of Impact EsGmates based on 21 alternagve Proximity Controls Share of observakons 50% 40% 30% 20% 10% 0% Test 3: Sensitivity to changes in time-frames One of the main risks with forcing a match between two regions, in this case Rwanda and its proximity control, is over-fitting. Over-fitting would mean that Rwanda s proximity control does not describe Rwanda, but instead fits the random noise or random errors generated by changes in Rwanda s new business registration figures between 2003 and If there were over-fitting, we would expect changes in the time frames during which we force the fit between Rwanda and its proximity control to lead to significant changes in our impact estimates. We therefore test for the possibility of over-fitting by varying the time-frames during which Rwanda is matched to its proximity control and compare the outcomes to our original impact estimate of 1,994 additional firms registered in 2009 or an impact of 184%. The results do not vary substantially with impact estimates ranging from 1,978 additional firms to 2,067 additional firms in 2009, which corresponds to +/- 3.5 percentage points from the original estimate of 1,994 firms (see table 8). Rwanda s proximity control is therefore robust to changes in the periods of fit, implying that it is unlikely that the proximity control suffers from over-fitting. 34

35 Table 8: Comparing impact estimates using 8 alternative periods Period of fit # firms Impact estimate % % % % % % % % Test 4: Falsification test To test the validity of the results obtained, it is also possible to replicate the two main tests carried out to check the robustness of the synthetic control methodology, namely the falsification test and the MSPE test. A simple way to construct a falsification test is to run the proximity control method on each of the countries that make up Rwanda s proximity control (namely Malawi, Niger, Ethiopia, Uganda, and Croatia) and compare the weighted average impact estimate on the proximity control to the estimated impact for Rwanda. Each country is weighted according to its contribution to the proximity control and the proximity control method is run using the same parameters as in the case of Rwanda to ensure comparability. If the proximity control region is a valid control region for Rwanda, then it should not have experienced any large impact during the period of interest, which is In graph 15, we compare impact estimates for Rwanda to that of its proximity control. Impact estimates for the proximity control during the period range between -36 firms to +2 firms, which is very small compared to the impact estimate for Rwanda. We conclude that the proximity control region did not itself experience any unusual changes in new business registration during the period of interest. 35

36 Graph 15. Comparing Impact EsGmates for Rwanda and its Proximity Control Number of firms Rwanda Proximity Control (Placebo) [ ] [ ] [ ] [ ] [ ] [ ] Test 5: Mean Square Prediction Error test In addition to the falsification test, we conduct a Mean Square Prediction Error (MSPE) test and compare the ratio of MSPE before and after the introduction of the one-stop shop in Rwanda to that of the other countries/controls in Rwanda s donor group. What we are testing for is an hypothetical impact in As in the case of the synthetic controls methodology an MSPE ratio smaller than 1 indicates that the observed impact in 2009 is not unusual, i.e. it is smaller than in other years before the intervention, while a ratio of more than 1 indicates that the observed impact is larger than in other years. If the ratio significantly larger than 1 then the control region might have experienced an impact during the period of interest. Not surprisingly, this ratio is much higher in Rwanda than in other countries (see table 9) indicating that the impact we observe in Rwanda is indeed unique. Moreover, for all countries except for Uganda this ratio is around 1, indicating that countries in Rwanda s donor pool did not experience a large impact in

37 Table 9: MSPE test Country MSPE Ratio Rwanda Malawi 0.1 Niger 1.5 Ethiopia 1.1 Croatia 1.8 Uganda 8.1 Test 6: Excluding Doing Business reforms in the proximity control Lastly, the results obtained could be over-estimating or under-estimating the actual impact of the introduction of the one-stop shop because of other reforms impacting time, cost, procedures and minimum capital required to start a new business. To test for this eventuality we compare average year-on-year changes in the starting a business indicators of Rwanda and its proximity control (see graph 16). We find that on average the countries that comprise Rwanda s proximity control did not implement major starting a business reforms during the period (covered by the Doing Business reports from 2004 to 2010). We define major reforms using the threshold defined by Klapper and Love (2010), who show that only business registration reforms that cut cost and/or time by more than 40% in 92 countries during the period had a statistically significant impact on new business creation. In Rwanda s proximity control the maximum average change during the period was a 10% cut in the cost and time required to start a business, achieved in Moreover, we know already from graph 15 that these reforms do not seem to have affected new business registration in any way. 37

38 Graph 16. Changes in StarGng a Business Indicators (Gme, cost, and procedures) * Change in starkng a business indicators 10% 0% - 10% - 20% - 30% - 40% - 50% - 60% - 70% - 80% - 90% DB2005 DB2006 DB2007 DB2008 DB2009 DB2010 Rwanda Proximity Control *Minimum capital requirements is excluded from the calculation of Starting a Business improvements as this has been 0 for Rwanda from 2003 through to 2009 From 2003 to 2007 (Doing Business reports from 2004 to 2008), Rwanda and its proximity control reformed at about the same pace, with gradual reductions in the cost of starting a business as a share of GDP. A large part of this relative reduction in costs was the result of high GDP growth rates (i.e. changes in the denominator), rather than actual reductions in the costs of starting a business. Business reforms accelerated in Rwanda in 2008, with a small reduction in the number of procedures and the time required to start a business (procedures were reduced from 9 to 8 and the time required from 16 days to 14 days). But the major breakthrough came in 2009 (Doing Business 2010). While the average time, cost, and procedures required to start a business in Rwanda was reduced by 81% in 2009 (Doing Business 2010), the corresponding figure for Rwanda s proximity control was just 9%, unchanged from the average of the period. Our calculations of impact estimates seem robust: they are not sensitive to (i) changes in the composition of the proximity control, (ii) changes in the donor pool, and (iii) changes in time frames used to construct the proximity control; (iv) they are not due to any large deviations nor events in the proximity control region; (v) they are unique compared to other countries in Rwanda s donor pool; and finally (v) they are not due to alternative reforms to starting a business indicators during the period of interest. 38

39 5. Replicating and comparing both methodologies to 11 other countries Table 10: Summary of impact results obtained in 12 countries 29 Country Year of Reform First year of impact estimate Estimate of # firms increase using synthetic control Estimate of # firms increase using proximity control Average estimate of # firms increase among the two methods Estimate of % firms increase using synthetic control Estimate of % firms increase using proximity control Average estimate of % firms increase among the two methods Rwanda ,041 1,994 2, % 183.5% 186% Tajikistan ,227 1,200 1, % 130.1% 132% Belarus ,800 1,749 1, % 101.8% 103% Albania % 51.8% 55% Oman % 53.2% 52% Senegal % 42.2% 45% Kyrgyz Rep ,625 1,595 1, % 43.2% 44% Georgia ,408 1,287 1, % 33.9% 36% Tunisia ,682 1,466 1, % 30.5% 33% Denmark ,681 2,988 3, % 14.2% 16% Canada ,691 14,663 14, % 9.1% 9% Netherlands , , % 2.7% 6% We replicate the synthetic control and proximity control methodologies to estimate the impact of the introduction of a one-stop shop on new business registration in 11 other countries, namely Albania (2007), Belarus (2007), Canada (2005), Denmark (2006), Georgia (2006), Kyrgyzstan (2008), Netherlands (2007), Oman (2006), Senegal (2008), Tajikistan (2008), and Tunisia (2006). 30 In some cases the synthetic and proximity control methodologies work better than in others, but overall the results are robust and comparable across countries. The magnitude of the observed impact (see table 10 or the Annexes for disaggregated results) is impressive: in Rwanda, we estimate that the introduction of the one-stop shop led to a % increase in new firms created the same year it was introduced (depending on the methodology utilized), in Tajikistan the increase was almost 132%, in Belarus it was 103%. Countries as diverse as the Kyrgyz Republic, Oman, Albania and Senegal showed an increase of 40-55% in terms of new firms created, while in Tunisia and Georgia the impact was close to 29 The results obtained for Canada and the Netherlands are not statistically significant. 30 In parenthesis, the year when the one- stop shop was introduced. 39

40 35%. These results confirm that burdensome regulations for business registration can be a major obstacle to new firm registration. Table 10 summarizes the impact estimates obtained for each of the 12 countries (Rwanda + 11 other countries) using both the synthetic control and proximity control methodologies. While impact estimates are strongly positive, they vary from 5.5% in the case of the Netherlands to 184%-188% in the case of Rwanda. Why? One possible explanation is that in some countries the introduction of the one-stop shop led to a larger improvement in the ease of starting a business than in others. One would expect the marginal improvement in the ease of starting a business in countries such as Netherlands, Canada and Denmark, which had a good initial business environments, to be smaller than in countries such as Rwanda, Tajikistan or Belarus, where the initial regulatory burdens of starting a business were high. To test this hypothesis we compare the estimated impact of the one-stop shop with the depth of the reform induced by the one-stop shop. We proxy for the depth of the reform using the average annual percentage change in the four Starting a Business sub-indicators: the cost, time, number of procedures and minimum capital requirements to start a business. In graph 17, we plot the first year of impact estimates for all 12 countries with the corresponding improvement of the business registration process in the same year the one-stop shop was introduced. Graph 17: Depth of business entry reforms. Estimates Impact (Year 1) and Depth of Reform OMN ALB DNK CAN NLD TUN KGZ TJK BLR GEO SEN RWA Depth of reform (% change) 40

41 We find a statistically significant linear association between the depth of the reforms induced by the one-stop shop and the increase in new firm registration, suggesting that the impact of a one-stop shop is proportional to the scale of the resulting regulatory change. On average, a 1 percentage point improvement in the ease of starting a business (measured as the average percentage change on the four starting a business indicators) is associated to a 1.87 percentage point increase in new business registration (R! = 0.65). This proportional relationship between impact and reform is further evidence that what was holding back new firm registration in the 12 countries in our sample were regulatory barriers to starting a business. The more these barriers were alleviated in relative terms, the greater the impact of the one-stop shop. Increases in new firm registration can be the result of: (i) companies that were previously in the informal sector shifting towards the formal sector; (ii) new entrepreneurs deciding to start a business, as the barriers are lower; (iii) foreign investors deciding to register local businesses; or (iv) simply re-registration requirements. While additional data and research is needed to understand the exact nature of the increases in new business registration, the evidence presented here which is consistent across almost all case studies - suggests that policy makers should focus on promoting significant business entry reforms such as the introduction or improvement of a one-stop shop for business registration. The larger the reform in so far it impacts the ease of starting a business the larger the impact. Impact eskmate using Synthekc Controls (%) Graph 18. Comparing impact esgmates using Proximity and SyntheGc Controls 200.0% R² = % 100.0% 50.0% 0.0% 0.0% 50.0% 100.0% 150.0% 200.0% Impact eskmate using Proximity Controls (%) Impact estimates obtained with the synthetic control and the proximity control approaches are very similar (see graph 18), even though these methodologies can result in very 41

42 different linear combinations of control countries. The average prediction error between the two methodologies is around 3.2 percentage points for the 12 case studies, compared to an average impact of about 54%. These results suggest that both methods are interchangeable and offer equally accurate ways of measuring impact at the aggregate level. Synthetic control and proximity control share a number of features that make them attractive methodologies for certain types of impact evaluations: They enable quick evaluations at the aggregate level which are difficult to achieve and very expensive using alternative methods, in particular Randomized Control Trials which are limited by external validity issues; They do not have extensive data requirements: minimum data requirements are two observations in time one before, and one after the intervention and complete data on a number of variables for the treatment region and at least two control regions (however, the more data is available, the more likely it is that the methods work); Contingent on data availability, they can be carried out post-intervention and do not require a lot of pre-treatment planning; As we have shown in this paper, they are easily replicable to other regions/countries, thereby ensuring cross-country comparability; They are transparent: both clearly outline the weights assigned to each control region and to each variable in order to obtain the control region; 31 They are falsifiable and lead to exact inference: a number of tests enable the researcher to check the consistency and validity of the results, leading to exact inference. Yet there are some fundamental differences between the two approaches. The synthetic control methodology relies on a least squares minimization algorithm that assigns weights to control regions and selected variables, such that the resulting linear combination of control regions best fits the reference region on the selected variables before the treatment. This has a number of consequences: The number of possible variables that can be used to match the treatment and control region are limited - if more than 10 variables are included, for example, researchers will find that the algorithm often fails due to non-convexity problems; 31 See Abadie et al. (2007). 42

43 The synthetic control approach assigns weights to variables as well as control regions, which means that the weights assigned to variables are different each time the method is applied to a different case study. This limits the comparability of results when applied to several different case studies, as the observed impact will not only be the result of different control regions, but also a different combination of explanatory variables. The synthetic control methodology only produces one alternative history, or control region, which is the region that best matches the reference region on the dependent variable and the explanatory variables. As a result, the only tests that can be conducted to check the validity of the resulting impact estimates (as we show in Section 3), involves applying the synthetic control methodology to the Placebo region itself (the falsification test) as well as all the countries in the donor pool (MSPE test), or to iteratively eliminate countries from the donor pool to check the sensitivity of the results to changes in the donor pool. If the donor pool is small, then the number of possible tests is limited. It is possible for the resulting synthetic control region to be a combination of multiple countries (the limit is the number of countries in the donor pool), which can often be very different in nature, leading to biases caused by interpolating across regions with very different characteristics. 32 In order to avoid interpolating across regions with different characteristics, it is not always evident which metrics or rules to use to eliminate a control region from the donor pool. The proximity control approach solves some of these issues: It works best when Proximity is calculated using as many variables as possible (in this case 32, compared to only 7 for the synthetic control), as the resulting similarity measure is based on more information. On the other hand, this means that proximity control require more data than synthetic control, which in practice can be a problem. The weights assigned to each variable (in this case equal) can be kept constant across case studies, thereby ensuring better comparability. The underlying idea behind proximity control is that there is not only one alternative scenario but many. The paths that best match the reference country (in this case we selected the 12 alternative controls that best predicted new business registration in Rwanda) all contain information worth being used, as long as each of these are good enough predictors of the reference country before the intervention. The more diverse the 32 Ibidem. 43

44 composition of these alternative scenarios in terms of the control regions that contribute to them, the more valuable the information that is captured. Also, these alternative scenarios lead to a distribution of impact estimates, enabling the researcher to test the sensitivity of the proximity control to changes in composition. And lastly, by construction, the proximity control ensures that regions in the donor pool are the most similar to the reference region on the measure of interest. Control regions are ranked based on their similarity to the reference region and are included or excluded on that basis. In this case, we only include in the donor pool the 5 countries that have the most similar business environment to the reference country and that do not have a one-stop shop. The rule for inclusion or exclusion is very clear. These differences make the two approaches good complements for each other. 6. Conclusion In this paper we use the synthetic control and proximity control methodologies to measure the impact of introducing a one-stop shop on new business registration in twelve different countries. We show how, using easily available datasets, it is possible to conduct testable and comparable measurements of the impact of an investment climate reform in a costeffective way. Typically, research on the impact of investment climate reforms has either focused on cross-country comparisons, which face inference problems, or country-specific impact evaluations, which face external validity issues and tend to be very expensive and time consuming (see Bransmetter 2010). The synthetic control and proximity control methodologies offer a third approach, making inference possible while enabling cross-country comparisons. In chapter five, for example, we use both country-specific impact measurements and variation in cross-country outcomes to establish the link between the introduction of a one-stop shop, the depth of the related reforms, and the resulting increase in new firm creation. Moreover, the synthetic control and proximity control methodologies enable impact evaluations at the aggregate level. For example, Bransmetter (2010) measures the impact of the introduction of a one-stop shop on new firms creation in Portugal by comparing a group of counties that did introduce the one-stop shop and a group of counties that did not introduce it but the study does not answer whether or not the one-stop shop had a large impact on new firm creation at the national level. Likewise, Bruhn (2008) measures the impact of a one-stop shop 44

45 on new firm creation and employment in Mexico by using a variation in the timing of the introduction of the reform in selected municipalities. This is not an aggregate measure of the impact of the one-stop shop on new business creation in Mexico, but rather a measure of the average increase in new firm s creation on the specific municipalities. One of the main objectives of this paper is to make the case for a wider use of the synthetic control and proximity control methodologies as cost-effective alternatives to measure the impact of private sector development reforms. The World Bank Group, regional development banks and the IMF should and can make much more extensive use of these methodologies to test the impact of large-scale reforms or projects. Not only do they provide cheap, quick, comparable and testable results, but they also have a wide array of possible applications. The synthetic control methodology has been used to measure the impact of terrorism on economic growth in the Basque Country (see Abadie et al, 2003), the effectiveness of a tobacco regulatory reform in California (see Abadie et al, 2010), the impact of a World Bank health care program on aggregate health indicators in Peru (see Parodi et al, 2008), as well as the cost of the German re-unification on GDP per capita (Abadie et al, 2011). The proximity control methodology, calculated on the basis of export similarity networks, has been used to measure the cost of the political and economic crises in Kenya, Ivory Coast and Indonesia. 33 The paper also introduces a new way of looking at Doing Business data by introducing a measure of the similarity of the business environment of a pair of countries. Typically, analysis using Doing Business data is done within one country (looking at where the country ranks across variables) or across countries (looking at a specific variables of interest). However, there is additional information to be inferred by looking at measures of Doing Business similarity between country-pairs. Every year, the Doing Business team at the World Bank Group gathers data from over 180 countries and 32 different variables, totaling about 4,800 data points. By aggregating these 32 variables into one single variable of similarity between pairs of countries, the resulting number of data points in a single year increases to 32, Likewise, looking at the similarity levels of triplets of countries, the number of data points would increase to 5,725,160; and so forth. In the same way, data on pairs of countries captures much more information on the business environment than analysing the single indicators of the Doing Business database. By showing a clear link between the performance in the Doing Business indicators and new business registration, we infer that countries with similar business environments also have very similar levels of new business registration (see Graph 8), 33 See Gathani et Stoelinga (2013) x179=32,

46 This has practical implications for policy makers. While Rwanda was ranked 67 th in the Doing Business 2010 report on the overall ease of doing business, it was also more similar to Burkina Faso (ranked 147 th ) than it was to Macedonia (FYROM, ranked 32 nd ). These differences and similarities between countries can guide policy makers towards better targeting policies and understanding how reform packages as opposed to individual reforms aimed at moving up the rankings will impact their position in the Doing Business similarity network. Furthermore, comparisons also create an additional motivation for governments to change; while Rwanda can pride itself for being one of world s top reformers several years in a row, we showed that in 2010 its business environment was still too similar to Malawi, Burkina Faso and Paraguay, which all ranked well below the 120 th place in the Doing Business report. Finally, we showed that introducing a one-stop shop for business registration is an effective policy to increase the size of the private sector and that the deeper the extent of the reform, the greater the impact. For countries where the one-stop shop leads to large improvements in the time, cost, procedures, and minimum capital required to start a business, this type of reform can mark a true turning point. That is how in Rwanda, Tajikistan and Belarus, the introduction of the one-stop shop led to more than doubling the creation of new businesses after only one year in operation. 46

47 Bibliography Abadie, Alberto and Gardeazabal, Javier The Economic Costs of Conflict: A Case Study of the Basque Country, American Economic Review, vol. 93(1), pages , March. Abadie, Alberto, Diamond, Alexis, and Hainmueller, Jens, Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California s Tobacco Control Program, Journal of the American Statistical Association, vol. 105(490), pages Abadie, Alberto, Diamond, Alexis, and Hainmueller, Jens, Comparative Politics and the Synthetic Control Method, MIT Political Science Department Research Paper Alcázar, Lorena, and Jaramillo, Miguel Panel/tracer Study on the Impact of Business Facilitation Processes on Microenterprises and Identification of Priorities for Future Business Enabling Environment Projects in Lima, Peru. Mimeo. Grade. Peru, Lima. June. Bennett, A. and C. Elman Qualitative Research: Recent Developments in Case Study Methods. Annual Review of Political Science. 9: Branstetter, Lima, and Taylor, Venancio Do Entry Regulations Deter Entrepreneurship and Job Creation? Evidence from Recent Reforms in Portugal NBER Working Paper 16473, October. Bruhn, Miriam License to Sell: The Effect of Business Registration Reform on Entrepreneurial Activity in Mexico, World Bank Policy Research Working Paper 4538, January. Ciccone and Papaioannuou Red Tape and Delayed Entry, Journal of the European Economic Association, vol.5, no.2-3, pp Doing Business database, The World Bank Group. Fearon, J.D Counterfactuals and Hypothesis Testing in Political Science. World Politics. 43(2):

48 Gathani, Sachin and Stoelinga, Dimitri Export Similarity Networks and Proximity Control Methods for Comparative Case Studies, Forthcoming, Journal of Globalization and Development. Investment Climate Advisory Services, How many stops in a one-stop shop? A review of Recent Development in Business registration, Flagship report, The World Bank. Kaplan, Piedra, Seira Entry Regulation and Business Start-Ups: Evidence from Mexico, World Bank Policy Research Working Paper 4322, June. Klapper, Leora, and Love, Inessa, The impact of business environment reforms on new firm registration, Policy Research Working Paper Series 5493, The World Bank. Klapper, Leora and Love, Inessa, The impact of the financial crisis on new firm registration, Economics Letters, vol. 113(1), pages 1-4, October. Motta, Marialisa, Ana Maria Oviedo, and Massimiliano Santini An Open Door for Firms. The Impact of Business Entry Reforms. Viewpoint Note No. 323, The World Bank Group. Washington, DC. June. Newey, Whitney K. and McFadden, Daniel. Large Sample Estimation and Hypothesis Testing, in Robert F. Engle and Daniel L. McFadden, eds. Handbook of econometrics, Vol. 4. New York: Elsevier Science, 1994, pp Rand, John and Torm, Nina The Benefits of Formalization: Evidence from Vietnamese Manufacturing SMEs, World Development Vol. 40, No. 5, pp Rubin, D.B Formal Modes of Statistical Inference for Causal Effects. Journal of Statistical Planning and Inference. 25: Sims, Christopher A. and Zha, Tao. Error Bands for Impulse Responses. Econometrica, September 1999, 67(5), pp World Bank Group Entrepreneurship Snapshots (WBGES). World Bank,

49 Annexes 35 Albania Year of Reform: 2006 Year of Impact: 2007 Notes Synthetic Controls Proximity Controls Donor pool: 6 countries; GDP per capita >USD1000 & <USD2500 Donor pool: 5 countries; Peru and Bolivia eliminated due to impact Impact estimates Year 1 Year 2 Year 3 Estimated Impact (%) 55% 39% -27% Estimated Impact (#firms) New Business Density in Albania and its Proximity and Synthetic Controls ( ) 2500 Number of new firms Albania Synthetic Albania Placebo Proximity Control Indicator Albania Synthetic Albania Proximity Control (6) GDP per capita (constant USD2000) Agricture (%GDP) Industry (%GDP) Services (%GDP) Trade balance (%GDP) Gross fixed capital formation (%GDP) Urbanization (% Population) MSPE Synthetic Albania Proximity Control (6) Albania Minimum MSPE Controls Maximum MSPE Controls Proximity Control: Sri Lanka (58.7%), Argentina (21.1%), Philippines (15.87%), Uruguay (4.3%) Synthetic Control: Guatemala (55.3%), Sri Lanka (44.7%) 35 The Estimated Impact is the average estimated impact among both methodologies. 49

50 Belarus Reform 2007 Year of Impact: 2007 Notes Synthetic Controls Proximity Controls Donor pool: 25 countries; minimum GDP per capita 300 Donor pool: 5 countries; Bolivia eliminated from pool due to impact Impact estimates Year 1 Year 2 Year 3 Estimated Impact (%) 103% 44% 17% Estimated Impact (#firms) New Business Density in Belarus and its Proximity and Synthetic Controls ( ) Number of new firms Belarus Synthetic Belarus Placebo Adjusted Proximity Control Indicator (mean ) Belarus Synthetic Belarus Proximity Control (606) GDP per capita (constant USD2000) Agricture (%GDP) Industry (%GDP) Services (%GDP) Trade balance (%GDP) Gross fixed capital formation (%GDP) Urbanization (% Population) MSPE Synthetic Belarus Proximity Control (606) Belarus Minimum MSPE Controls Maximum MSPE Controls Proximity Control: Philippines (75.73%), Argentina (23.3%), Moldova (0.83%), Uruguay (0.12%) Synthetic Control: Philippines (68.3%), Pakistan (20%), Argentina (11.7%) 50

51 Canada Year of Reform: 2005 Year of Impact: 2006 Notes Synthetic Controls Proximity Controls Donor pool: 3 countries; minimum GDP per capita Donor pool: 5 countries; Peru and Ireland eliminated due to impact Impact estimates Year 1 Year 2 Year 3 Estimated Impact (%) 9% -1% Estimated Impact (#firms) New Business Density in Canada and its Proximity and Synthetic Controls ( ) Number of new firms Canada Synthetic Canada Placebo Proximity Control Indicator (mean ) Canada Synthetic Canada Proximity Control (166) GDP per capita (constant USD2000) Agricture (%GDP) Industry (%GDP) Services (%GDP) Trade balance (%GDP) Gross fixed capital formation (%GDP) Urbanization (% Population) MSPE Synthetic Canada Proximity Control (166) Canada Minimum MSPE Controls Maximum MSPE Controls Proximity Control: Hong Kong (47.6%), Malaysia (32.3%), Italy (14.2%), Austria (5.8%) Synthetic Control: Italy (50.7%), Hong Kong (49.3%) 51

52 Denmark Year of Reform: 2006 Year of Impact: 2006 Notes Synthetic Controls Proximity Controls Donor pool: 5 countries; minimum GDP per capita 5000 Donor pool: 5 countries; Peru and Ireland eliminated due to impact Impact estimates Year 1 Year 2 Year 3 Estimated Impact (%) 16% Estimated Impact (#firms) New Business Density in Denmark and its Proximity and Synthetic Controls ( ) Number of new firms Denmark Synthetic Denmark Placebo Proximity Control Indicator (mean ) Denmark Synthetic Denmark Proximity Control (122) GDP per capita (constant USD2000) Agricture (%GDP) Industry (%GDP) Services (%GDP) Trade balance (%GDP) Gross fixed capital formation (%GDP) Urbanization (% Population) MSPE Synthetic Denmark Proximity Control (122) Denmark Minimum MSPE Controls Maximum MSPE Controls Proximity Control: Austria (42.83%), Hong Kong (37.8%), Iialy (12.2%), Malaysia (7.1%) Synthetic Control: Croatia (39.6%), Uruguay (33.4%), Hong Kong (25%), Argentina (1.9%) 52