The Effect Of Technical Assistance On Microfinance Client Performance: A Case Study
Elvira Veijer1
University of Groningen, Groningen, The Netherlands Faculty of Economics and Business
Supervisor: Francesco Cecchi July 11, 2015
Abstract
This thesis investigates the impact of technical assistance on the economic outcomes of microfinance clients using data provided by a microfinance institution from Bolivia. The results of the OLS regressions suggest that technical assistance does not increase gross profits, income per capita, annual sales or fixed assets of clients in general. However, when including interaction variables to check for heterogeneity, several significant effects are found. Propensity Score Matching confirms the finding that rich and female entrepreneurs have higher fixed assets after they received technical assistance. This paper contributes to the current literature on business training by investigating the impact of technical assistance on Bolivian micro entrepreneurs.
Key words: Microfinance, evaluation of training programs, technical assistance JEL Codes: D1, G2, O1
2
1. Introduction
The main objective of Microfinance institutions (MFIs) is to alleviate poverty (Haynes Seawright, and Giauque, 2000). Various microcredit programs have been successful in achieving this objective by enabling micro entrepreneurs to increase profits, income and household expenditures (McKernan, 2002; Pitt and Khandker, 1998; Imai and Azam, 2012). However, concerns also emerged. Microfinance often only leads to small marginal changes in the lives of the poor. There is little evidence that microcredit programs lead to significant and sustainable increases in income levels or firm growth (Angelucci, Karlan, and Zinman, 2013). Banerjee, Duflo, Glennerster, and Kinnan (2013) find that although microcredit may help expand businesses, it does not lift entrepreneurs out of poverty. In the slums of Hyderabad, India, the provision of microfinance did not lead to an increase in monthly consumption nor business profits for the majority of the enterprises. Also basic development outcomes such as education, health or women’s empowerment did not significantly increase.
The considerable variation in the impact of microfinance may be conditional on the human capital of the borrower: various papers emphasize the importance of human capital as a key complement of financial capital (Bruhn, Karlan, and Schoar, 2010; Sayingoza, Bulte, and Lensink, 2014). Smith and Metzger (1998) find that education of micro entrepreneurs enhanced earnings after controlling for capital investment.2 Therefore, the focus of many MFIs has shifted slightly to providing both financial and human capital to clients, by pursuing strategies to teach skills through trainings next to their credit offer. Experimental evidence is scarce however, and the limited evidence that is available indicates mixed results (McKenzie and Woodruff, 2014).
In this paper, the financial impact of technical assistance training to microfinance clients in Bolivia is estimated. Currently, evidence on the effects of business training is scarce, and of technical assistance training in particular. Much tension exists in the microfinance community about whether lenders should focus on providing financial services, or should also offer
3
financial services such as training to their clients (MkNelly, Watetip, Lassen, and Dunford, 1996). To my knowledge, this is the first paper assessing the impact of technical assistance on micro entrepreneurs mainly active in the agricultural sector. Findings may help management and sponsors of microfinance institutions to make informed decisions on whether to offer technical assistance to their clients.
I will investigate the following research question:
Does technical assistance affect the financial performance of micro entrepreneurs, and is there any difference between male and female entrepreneurs?
Data is provided by a microfinance institution (MFI) from Bolivia3; the final sample consists of 354 observations over the period 2010-2014. The results of the OLS regressions suggest that technical assistance does not increase gross profits, income per capita, annual sales or fixed assets of clients in general. However, when including interaction variables to check for heterogeneity, several significant effects are found. Propensity Score Matching confirms the finding that rich and female entrepreneurs have higher fixed assets after they received technical assistance.
Section 2 surveys the literature on this topic. Section 3 explains the nature of the intervention and some background information on the MFI. Section 4 elaborates on the collected data, while section 5 presents the methodology. Results and findings are discussed in section 6, and section 7 offers concluding observations.
2. Literature Review
Training can be considered as an investment in human capital. Human capital investments require an initial cost (in the case of training: tuition and training course fees, reduced wages, and limited productivity during training period) which is hoped to be regained in the future, for example through increased earnings or productivity (Blundell, Dearden, Meghir, and Sianesi, 1999). The economic literature on human capital states that returns to investments in human capital yield positive returns, and may exceed the returns to physical
4
capital (Romer, 1990). The highest returns can be found for primary education, the general curricula, the education of women and countries with a low per capita income (Psacharopoulos, 1985).
Human capital investments are considered a key factor in the economic progress of developing countries (Tallman and Wang, 1994; Dougherty and Jorgenson, 1996; Mathur, 1999). Funkhauser (1998) argues that in developing countries, returns to human capital investments such as education are high relative to developed countries. He found that if the economy of a developing country grows, the demand for educated workers will also increase, thereby increasing the returns to education. At the same time, when the supply of educated workers increases, returns to education will decrease. At the individual level, education also positively affects entrepreneurship in terms of becoming self-employed and the success rate of new businesses (Robinson and Sexton, 1994). Johnson (1998) argued that whether the main objective of a microcredit program is social welfare or economic/employment increase, the provision of training programs should be considered as an additional service as training can be essential to reach any of these goals.
Despite the fact that experimental evidence is scarce, many MFIs are now combining investments in human and physical capital, although the content and focus of the trainings provided vary significantly (McKenzie and Woodruff, 2013). For some programs, training is required as a precondition for a loan, while other programs offer training as an additional opportunity to borrowers (Haynes, Seawright, and Giauque, 2000).
2.1. Two different perspectives
5
business development training, literacy, education and health) along with financial services. Bhatt and Tang (2001) name the former approach the Minimalist Service Delivery approach, while the second approach is referred to as the Integrated Service Delivery Approach. Proponents of the Minimalist Service Delivery approach argue that NGOs should be transformed into for-profit commercial banks, and should become self-sufficient. Although training and technical assistance might be beneficial for entrepreneurs, facilitating such services is expensive, and may therefore undermine the self-sufficiency of microfinance programs. Moreover, Grameen Bank-founder Yunus argued that the poor are creative and know how to make a living, emphasizing the importance of access to credit rather than additional services (Yunus and Yusus, 1998). Yet proponents of the Integrated Service Delivery Approach claim that additional services should also be provided to the poor. Dichter (1996) stressed the importance of providing poor entrepreneurs with market and business development services, arguing that it would be otherwise hard for entrepreneurs to use a loan productively. He states that programs that act in their own self-interest by focusing on financial self-sufficiency might lose their ability to reach the poorest of the poor. Moreover, Dichter(1996) argues that minimalist microfinance may seem to reduce poverty in the short term, but does not affect economic growth or development.
In practice, the distinction between the minimalist and integrated approach to microfinance is not clear. The Grameen Bank is often referred to as a minimalist, but the bank does provide various non-financial support services to clients. On the other hand, there are also subsidized programs that claim to have an integrated approach, while often this is mentioned to shift attention away from their ‘inefficient, minimalist operations’ (Berenbach & Guzman, 1994). As there is no clear consensus yet on whether additional services should be offered, research on this topic is highly relevant.
2.2. Empirical evidence
6
abilities (measured by a Digit Span Recall Test4) yielded higher returns. These results question the view of Yunus and Yusus (1998) that the poor are rational and that micro entrepreneurs maximize profits subject to their financial constraints.
Giné and Mansuri (2014) emphasize the importance of human capital investments for entrepreneurship. They conducted a field experiment in rural Pakistan, where microfinance clients where offered an eight day business training course focused on business planning, marketing and financial management. Their findings indicate that business training increases business knowledge, improves business practices and reduces business failure for male entrepreneurs. Moreover, household expenditures increased by $40 per year. Furthermore, the authors find little evidence of a technology-based poverty trap, which implies that large loan sizes are needed to get out of poverty: when clients where offered a much larger loan size, it had little effect on household welfare. However, the training provided was not cost-effective for the MFI.
McKenzie and Woodruff (2013) analyzed twenty impact studies evaluating business training programs in developing countries. They find that many studies have low statistical power, have problems with survey attrition and measurement of enterprise profit and revenues, and measure the impact over a short time period (within one year). Most evaluations show relatively low impacts of training on survivorship of existing firms, however, training programs help prospective owners launch new businesses more quickly. Almost all studies show that at least some of the practices taught in training were implemented by the firm owners. However, only a few studies find significant impact on profits and sales on the short term (Berge, Bjorvatn, and Tungodden, 2014; De Mel, McKenzie, and Woodruff, 2014; Calderon, Cunha, and De Giorgi, 2013).Berge, Bjorvatn, and Tungodden (2014) investigate the importance of a combined human capital intervention (business training) and a financial capital intervention (business grant) through a field experiment in Tanzania. The combined human capital and financial capital intervention positively affected business performance, management practices, happiness, business knowledge, and non-cognitive abilities for male entrepreneurs. The authors state that long-term finance is an important constraint for
7
microfinance entrepreneurs, but that business training is essential to transform financial capital into productive investments. Investments created by the business grant, but without the training program, did not generate any significant returns.Calderon, Cunha, and De Giorgi (2013) find that a 48-hour business skills course to female entrepreneurs in rural Mexico results in higher profits, revenues and a larger clientele. Additionally, after training entrepreneurs are more likely to use formal accounting techniques and also tend to be registered with the government more often.
Karlan and Valdivia (2011) argue that offering additional services may have beneficial effects, but also emphasize the importance of assessing whether the economies of scope outweigh the risks of having credit officers simultaneously become teachers. Credit officers may lose focus on lending and savings activities; moreover, providing detailed business advice may lead to more defaults if the borrower perceives the lender as partially accountable for any business changes that do not succeed. The authors use a randomized control trial to measure the marginal impact of adding business training to a Peruvian group-lending program for female micro entrepreneurs. They found little evidence that training helped to increase business revenues, profits or employment. However, training did help individuals to engage in certain good business practices. Besides, the improved client retention rate generates more marginal revenue than the marginal cost of providing the training, which makes it profitable for the MFI to provide training. Additionally, Karlan and Valdivia (2011) find that the clients who were initially not too interested in the training were more likely to improve retention and repayment rates, and were also more likely to implement changes to improve their businesses. Therefore, they conclude that simply charging for business training may not yield optimal results.
8
Most research investigating the impact of training on micro entrepreneurs has focused on the effects of business training, which involves business concepts such as marketing and financial management. Research on technical assistance offered by microfinance institutions is scarce. Brown, Earle, and Lup (2005) investigated the effect of technical assistance on firm growth of Romanian SMEs, where technical assistance is defined as ‘membership in a business association and training and consultancy programs from a variety of sources’. The results suggest that technical assistance is not important for firm growth, but rather the availability of loans.
To my knowledge, no specific case study has yet been published on the effects of technical assistance of micro entrepreneurs in rural areas. The objectives of this study are twofold. First, I will investigate the effect of technical assistance on the financial performance of clients of the MFI from Bolivia. This non-profit microfinance institution (MFI) is targeting clients in impoverished rural areas, mainly working in the agricultural sector. The MFI collaborates with the institute to provide clients with technical assistance, with the aim of increasing productivity. One would expect that financial performance will improve when productivity increases, in line with the findings of Berge, Bjorvatn, and Tungodden (2014) who find that training helped increase business revenues of male entrepreneurs, and the work of Calderon, Cunha, and De Giorgi (2013) who find that business skills training leads to better business performance of female entrepreneurs. Therefore I will investigate the following hypothesis:
Hypothesis 1: Technical assistance positively affects the financial performance of micro entrepreneurs.
Additionally, I will investigate whether the impact of technical assistance affects female and male entrepreneurs differently. Based on the literature, one would expect that the technical assistance will have more effect on male entrepreneurs than female entrepreneurs (Giné and Mansuri, 2014; Berge, Bjorvatn, and Tungodden, 2014):
9
3. Case study
This section provides information on the microfinance institution that is analyzed. Land-locked Bolivia is situated in the heart of South America and is one of the poorest countries of the continent: according to the Worldbank, 12.7% of the population is living on less than 2 US dollar per day. There is quite a disparity between the living standards of people living in urban areas and those living in rural areas: around 61 percent of the rural population is living under the national poverty line compared to 45% of the total population (Worldbank, 2011).
The foundation was founded in 1989 with the aim of promoting sustainable development in poor rural areas in Bolivia and operates as a non-profit Microfinance Institution (MFI). Since 2009, the foundation has expanded its credit offer with technical assistance services provided by a not-for-profit institution supporting the development of the agricultural sector. The mission of the two organizations is to ‘provide financial services with an integral approach, in particular to rural producer populations who do not have access to adequate financial and productive services and who are mainly organized in agriculture value chains’.
The foundation has 21 agencies located in seven out of nine departments of Bolivia, of which fourteen agencies are situated in rural areas and areas where agricultural activity is important.
3.1. Products/Services
The credit offer of the MFI can be divided into three main categories:
- Individual Microcredit: Designed for the agricultural productive sector and commercial/services sector for entrepreneurs that can offer personal guarantees or collateral. The loan will allow clients to purchase more raw materials, machinery, etc. - Solidary Group credit: This credit is aimed at micro entrepreneurs that cannot offer a
10
- Community Bank Credit: This credit is focused at micro entrepreneurs that wish to increase their working capital or want to start their own business. A community bank is formed when there are at least two groups of four partners that know each other and that are willing to provide guarantees for each other.
The MFI offers specialized products for each region and activity. Devices such as the ‘Production Plan’ help to organize clients’ financial statements, costs of production and structure of financing. Loans are granted both in Bolivianos and US Dollars, and payment plans can be customized for each client.
Furthermore, the MFI offers complementary financial services, providing clients access to financial services in an integral way. Examples of services offered are: pension funds, national and international money transfers, telephone and internet services, remittances, and exchange of US dollars. Additionally, the MFI has set up a collaboration with a local insurance company to develop micro-insurance products to reduce both credit risk of the MFI and client risk. Examples are death insurance (includes funeral expenses), life insurance, multi risk assurance (against material damage of the farm enterprise) and livestock insurance.
As already mentioned, the foundation also offers technical assistance in cooperation with an institute. The MFI operates as the specialized financial institution that provides access to financial services, while the institute focuses on the technical assistance area. Additionally, the institute offers market access and information. Although the institute and the MFI are separate organizations, they are ‘united by their mandate and common institutional origin’.
3.2. Technical Assistance
11
The institute provides farmers assistance in four ways: appropriate technology, participatory extension, quality inputs, and service delivery.
Appropriate technology refers to the modernization of methods in the field, which are
customized such that local producers can understand and efficiently apply new methods. These technological innovations are delivered to the producers by using trained technicians who are financed by the institute or another intermediary in the production chain. The institute forms strategic partnerships with suppliers of agricultural inputs, NGOs, universities (public and private), international agencies, and others.
Participatory extension refers to various activities that may help farmers to increase
productivity, such as technology demonstration days, formal training, field experimentation, and exhibitions.
Quality inputs: the institute works closely with suppliers of inputs that may benefit
farmers.
Service delivery is the area responsible for the structuring of good practices, training,
experience and information gathered in the field. Models are developed in such a way that they can easily replicated in different parts of Bolivia.
4. Data
To examine the impact of technical assistance this paper uses data provided by the MFI from Bolivia. Two data sets were provided: the Performance and Social Impact Measurement (MEDIS) dataset and the RTP Evaluations data set.
12
characteristics, such as annual sales, gross profits, family income and fixed assets. The RTP Evaluations data set contains 3,205 observations. As both datasets have information that is relevant to draw conclusions on the effect of technical assistance, the two dataset were merged, resulting in 2,746 observations.
In this study, I will only analyze the data of clients that have been evaluated at least two times, however the majority of the clients have only been interviewed once, and therefore many observations were lost in the process. Clients are compared through a difference-in-difference analysis, that is, difference-in-differences in outcome variables over time of clients who received technical assistance are compared with clients who indicated not having received technical assistance. The reason that only clients with two or more observations are analyzed is that it is otherwise difficult to estimate the effects of technical assistance; the results may be biased by unobservable characteristics, for example entrepreneurial ability or education. By estimating the difference in growth in outcome variables of clients who received technical assistance and those who did not, it is possible to control for unobservable characteristics that do not change over time. Difference-in-Difference analysis will be further explained in the Estimation Methods section.
The number of observations in the final sample is 354, the total number of clients is 177. All 21 agencies are represented in the sample, and the sample covers the period September 2010 to December 2014.
4.1. Assumptions on Data
13
In the rare case there are more than two observations, I have eliminated the observations between the baseline and end line value.5
- All clients that indicated that they have received technical assistance at the time of the base line and/or the end line are included in the group that received technical assistance. There is no distinction made between clients that received technical assistance only in the first period and dropped out later on, received technical assistance in both periods or received technical assistance only in the last period. Ideally, I would only analyze clients that did not receive technical assistance at the time of the first interview, but that did at the time of the second interview, such that I could perform a true difference-in-difference analysis. However, the sample size (44 observations, 22 clients) is simply too low. Additionally, it is not clear whether clients that indicated that they are receiving technical assistance have already received technical assistance, or that they will receive it in the near future. Therefore, all clients indicating they are receiving technical assistance at some point in time are included in the treatment group. For simplicity reasons, I assume that all the clients in the treatment group did not receive technical assistance at time 0, and did receive technical assistance before time 1.
- Finally, the institute provides several forms of technical assistance such as modernization of technologies, field trips, training days, etc. However, this dataset does not distinguish between the different forms of technical assistance. I assume that the effects of the different forms of technical assistance are similar in this research.
4.2. Variables
5One client had five observations, however, was interviewed three times at December 5th, 2013 and two times at December
4th 2013, with similar information for both dates. In this particular case I decided to remove four observations, as it is not
14
The following outcome variables will be analyzed: Gross Profit, Income Per Capita, Annual Sales, and Fixed Assets. All variables are estimated on a yearly basis. The main independent variable of interest is the variable indicating whether clients received technical assistance. There is no specific information on what sort of technical assistance the clients received. Rather, this variable captures a self-reported answer on the question whether clients received technical assistance in general.
To take into account the individual characteristics of the clients, the following explanatory variables will be included in the analyses: number of family members, gender, destination of loan (agricultural or non-agricultural purposes), type of loan (individual or communal), number of times a loan was taken (sequence), baseline values of amount of loan disbursed, fixed assets and annual sales, and region dummies (Chuiquisaca-Santa Cruz) indicating the regional effect compared to the base region La Paz. Additionally, I will control for the average GDP per Capita (2011) in each province, based on data provided by the Bolivian Government.6
4.3. Descriptive Statistics
Table 1 presents some characteristics of the data.The number of clients that received technical assistance is 90 (180 observations), and 87 clients (174 observations) did not receive any technical assistance. Most observations are of clients in rural areas (290 observations), and most clients have taken up an individual loan (348 observations). Additional descriptive statistics can be found in the Appendix, A.2.
Table 1. Data Description
Frequency Area Gender Type of Credit Destination of Loan
Observations Urban Rural Male Female Individual Communal Agricultural Non-agri
Non-Participants 174 34 140 120 54 172 2 102 72
Participants 180 30 150 116 64 176 4 127 53
354 64 290 236 118 348 6 229 125
This table provides an overview of how the number of observations are distributed. 'Non-Participants' refers to the clients that indicated that they did not participate in any technical assistance training. 'Participants' refers to the clients that reported that they actually participated in technical assistance training.
15
Table 2 presents some client characteristics. Families are not extremely large: the average family consists of three persons. Businesses are generally small, the average number of employees is three. At the baseline (Time=0), there is quite a difference in the average financial characteristics for entrepreneurs that did and did not participate in technical assistance. Initial Gross Profit is higher for entrepreneurs that did not participate (80,766 Bolvianos (Bs.)) than for those who did (68,772 Bs.).The average Income Per Capita prior to the training is 12,168 Bs. for those who did not participate, while the clients that received training had an Income Per Capita of 9,273Bs on average. Initial Fixed Assets are higher for the clients that received technical assistance (132,595 Bs.) than for those who did not (117,025.). Annual Sales however, are higher for those who did not receive technical assistance: 173,316Bs. compared to 141,130 Bs. for the clients that received technical assistance. Also average Family Income prior to the training is higher for those who did not participate (29,132 Bs.) compared to those who did (25,179 Bs.). The average loan disbursed is much higher for those who did participate (41,675Bs.) compared to the clients who did not (29,050 Bs.).
In summary, clients that received technical assistance had initially lower average Gross Profit, Income Per Capita, Annual Sales and Family Income. However, the loan disbursed is on average higher, as well as the value of Fixed Assets prior to the training. In section 5, I will describe the estimation methods to measure the effect of technical assistance, controlling for other factors.
5. Estimation Methods
Table 2 . Client Characteristics
Gross Profit Income Per Capita Fixed Assets Annual Sales Family Income Family Size Employees Loan Disbursed Time=0 No TA 80,766 12,168 117,025 173,316 29,132 3.3 3.3 29,050 TA 68,772 9,273 132,595 141,130 25,179 3.5 2.8 41,675 74,667 10,686 124,945 156,950 27,109 3.4 3.0 35,469 Time= 1 No TA 110,061 17,025 128,633 193,557 38,538 3.2 3.1 28,730 TA 141,719 16,963 214,326 185,852 41,828 3.3 3.1 43,936 Total 126,158 16,993 171,966 189,639 40,212 3.2 3.1 36,462
16
To assess the effect of technical assistance on the financial performance of MFI clients, one would like to compare the impact on clients that do take technical assistance and those who do not. However, such comparison is complicated, especially if there is no experimental data available. Ravallion (2007) argued that impact evaluation is mainly a problem of missing data, as it is physically impossible to measure someone in two states of nature at the same time (for example, both participating and not participating in technical assistance training).
Additionally, technical assistance programs may be targeted at a specific group of clients and not distributed randomly: the management of the MFI may have decided that if agencies have relatively poor clients, these agencies should complement their credit offer with technical assistance, while other agencies that have on average wealthier clients should not. This is referred to as program placement bias (Ravallion, 2007).
Another issue is the self-selection problem: whether a client takes up the offered technical assistance is determined by the individual himself rather than by chance. In the villages where technical assistance is offered, clients sharing similar socio-cultural backgrounds (e.g. gender, region) might have different unobservable characteristics such as entrepreneurial skills, resulting in different probabilities to participating in a technical assistance program (Imai and Azam, 2012). Therefore, it is essential to take into account the self-selection problems when assessing the impact of technical assistance. I will address difference-in-difference analysis to address these issues.
5.1. Difference-in-Difference Analysis
In this paper, I will make use of a difference-in-difference approach. This approach compares changes in the treatment group (in this case, the group of clients that received technical assistance) with changes in the control group (the group of clients that did not receive technical assistance). By comparing changes with changes, I will also control for broad economic changes, and unobservable characteristics that are constant over time.
17
For example, more capable entrepreneurs are likely to have higher earnings growth, and not just a higher base level of income (Armendáriz & Morduch, 2010). This problem can be reduced by making sure that the treatment and control group are as similar as possible. However, when comparing a group of participants in a particular village with a control group consisting of clients that did not receive technical asssistance in another village, the problem arises that the participants self-selected themselves into technical assistance training, while those in the control group had no choice to opt for technical assistance, since it was not offered to them (Armendáriz & Morduch, 2010).
To address this issue, Intention-to-treat (ITT) analysis can be applied. ITT analysis in this case would compare two groups: one group receiving access to technical assistance and another group not having access to technical assistance. A way to group people is by comparing the clients of agencies that did not offer technical assistance at all (as control group) to those clients that were attending agencies that did offer technical assistance (as treatment group). In this way, it is possible to control for unobservable characteristics that do change over time. Instead of directly investigating the effect of receiving technical assistance, ITT basically answers the following question: What is the effect of access to technical assistance?
However, there is insufficient data to conduct an ITT analysis: only two agencies (Nataniel Aguirre and Quillacollo) did not offer technical assistance to any of their clients, and the two agencies together only served four clients (eight observations). The control group would simply be too low to draw any conclusions. Instead, I will estimate the impact of technical assistance by comparing participating and non-participating clients through using difference-in-difference (DiD) estimators in OLS regressions, and apply Propensity Score Matching (PSM) techniques as a robustness check. These two methods will control for unobservable characteristics that do not change over time. However, this research will not be able to control for unobservable characteristics that do change over time.
18
Similar to the work of Karlan and Valdivia (2011), the impact of technical assistance will be estimated by difference-in-difference (DiD) estimators through Ordinary Least Squares (OLS) regressions. The difference-in-difference estimator is obtained from comparing changes in a particular outcome variable over time between treatment and control groups.
The difference-in-difference estimator can be obtained from the following expression: 𝑌𝑖𝑗𝑡 = ∝ +𝛽1𝑇𝑖𝑚𝑒𝑡+ 𝛽2𝑇𝐴𝑖+ 𝛽3𝑇𝑖𝑚𝑒𝑡𝑇𝐴𝑖 + 𝜀𝑖𝑗𝑡 (1) where 𝑇𝑖𝑚𝑒𝑡 is a binary variable equal to one if the observation corresponds to the end-line observation, that is, after the technical assistance program has taken place. 𝑇𝐴𝑖 indicates whether a household participated in the technical assistance training (one if participated, zero if otherwise). 𝛽3 is the difference-in-difference estimator of the impact of technical assistance on outcome Y. The interaction term 𝑇𝑖𝑚𝑒𝑡𝑇𝐴𝑖 captures the effect of technical assistance over time, while the term 𝑇𝐴𝑖 only captures the effect at time 0. The term 𝑇𝑖𝑚𝑒𝑡 captures the change in the outcome variable over time. It is important to note here that I assume that technical assistance has not yet started at time 0. If technical assistance was already received before time 0, 𝛽2 + 𝛽3 would capture the effect of technical assistance. However, as currently information on when treatment has started is unclear, I will assume that at time 0 clients have not yet received technical assistance, and that the effect of technical assistance will be captured by 𝛽3.
In the tables of the results section, there will also be estimates of 𝛽3 that result from regressions that add to equation (1) various covariates such as client and enterprise characteristics, as well as regional dummies.
To measure whether treatment is heterogeneous across various client characteristics, the following model is estimated:
𝑌𝑖𝑗𝑡 = ∝ +𝛽1𝑇𝑖𝑚𝑒𝑡+ 𝛽2𝑇𝐴𝑖 + 𝛽3𝑇𝑖𝑚𝑒𝑡𝑇𝐴𝑖 + 𝛽4𝑇𝑖𝑚𝑒𝑡𝑇𝐴𝑖𝑋𝑖0 + 𝛽5 𝑇𝑖𝑚𝑒𝑡𝑋𝑖0+
𝛽6𝑇𝐴𝑖𝑋𝑖0 + 𝛽7𝑋𝑖0+ 𝜀𝑖𝑗𝑡 (2)
19
measures the impact for those individuals who do have the characteristic of interest. In this paper, the interaction will be measured between male entrepreneurs and participation in technical assistance programs. 𝛽3 would then measure the impact of financial technical assistance on financial performance for female clients, while (𝛽3+ 𝛽4) will measure the impact of technical assistance for male clients.
Additionally, I will also estimate whether there is a heterogeneity effect by splitting the sample based on how rich clients are. Rich clients are here defined as the clients that have both above median initial sales and above median initial fixed assets, while poor clients are here defined as those clients that do not have both above median initial sales and fixed assets. I will create a dummy variable Poor which will have a value of one if clients are relatively poor (that is, have below median initial sales and/or fixed assets), and zero if otherwise. Basically, the rich clients are now assessed separately from the average poor client. This dummy variable will allow us to estimate whether there are any heterogeneity effects based on profitability. 𝛽3 will then measure the impact of financial technical assistance on financial performance for rich clients, while (𝛽3+ 𝛽4) will measure the impact of technical assistance for poor clients.
Finally, I will estimate whether there are heterogeneity effects based on the amount of loan disbursed. Clients are divided in two groups: those with above median amount of loan disbursed, and those below median amount of loan disbursed. I will create a dummy variable Disbursed, which will have a value of one if clients have above median amount of loan disbursed. This dummy will help to estimate whether there are any heterogeneity effects based on the amount of loan disbursed. 𝛽3 will then measure the impact of financial technical assistance on financial performance for clients with a relatively low amount of loan disbursed, while (𝛽3+ 𝛽4) will measure the impact of technical assistance for clients that have a relatively high amount of loan disbursed.
20
As a robustness check, Propensity Score Matching (PSM) will be applied for the difference-in-difference estimators that appear to be significant in the OLS regressions. The propensity score is defined by Rosenbaum and Ruybin (1983) as the predicted probability that an individual participates. PSM matches a client that participated in the training with a non-participating client by using the propensity score, the estimated probability of non-participating in the technical assistance program. After that, the average treatment effect (ATT) of technical assistance can be obtained by comparing the averages of outcome variables for participants and non-participants (Imai and Azam, 2012).
The difference compared to OLS analyses is that participating and non-participating clients will be compared with each other based on their probability of receiving technical assistance. In other words, the financial outcomes of fairly similar clients (at least in observable characteristics) are compared. Additionally, a common support restriction is applied. The advantage is that only participating and non-participating clients are compared that are credibly similar. Clients that do not have a match are excluded, increasing the quality of the comparison. However, the sample size diminishes due to this common support requirement, thereby reducing power and the likelihood of significant results.
21
Following the work of Imai and Azam (2012), the Differences-in-Difference Propensity Score Matching (DID-PSM) method is applied, which also takes into account the longitudinal nature of the data. The DID-PSM estimator is specified as follows:
𝐸(𝑌𝑖𝑡− 𝑌𝑖𝑡′|𝑝𝑖(𝑋𝑖), 𝑇𝐴𝑖 = 1) = 𝐸(𝑌𝑖𝑡− 𝑌𝑖𝑡′|𝑝𝑖(𝑋𝑖), 𝑇𝐴𝑖 = 0)
Where t and t’ are time periods (where t=1 and t’=0, respectively after and before technical assistance training has taken place),𝑌𝑖𝑡 refers to the outcome at time t for the non-participant, 𝑝𝑖(𝑋𝑖) is the propensity score (the probability of participation) and 𝑇𝐴𝑖 indicates whether a household participated in the technical assistance training between t’ and t (1 if participated, 0 if otherwise).
When applying regular PSM analysis to cross-sectional data, the mean of the outcome of a client at one point in time is compared between participants and non-participants, conditional on the probability of participation estimated by observable characteristics (Imai and Azam, 2012). When using DID-PSM analysis, the time-series or temporal change of outcome of a household is compared after technical assistance has taken place(time t), conditional on the propensity score. Following the work of Imai and Azam (2012) and Smith and Todd (2005), DID-PSM in this paper implies that PSM is applied to the ‘first difference’ (from t’ to t) of the outcome variable of a client which received training between t and t’, and this outcome is compared with that of a client with the same characteristics (implied by the propensity score), but that did not receive any training between t and t’. PSM with Kernel weighting is applied: this implies that closer neighbors are weighted more than those neighbors with more distant PSM values. All non-participating individuals in the comparison group (within the common support limit) are matched, however, the method is very precise as various non-participating individuals are used to build a match (Chemin, 2008).
6. Results
22
mainly due to few observations with a large value (i.e. value above 1,500,000 bolivianos). However, the results did not change drastically when outliers were removed from the analysis. The logarithm of Annual Sales is not normally distributed, even when outliers are removed. Robust standard errors are applied where heteroskedasticity is present. Table 5 shows the results of the Propensity Score Matching Analysis also estimating the difference-in-difference effect.
6.1. Simple Analysis
Before reporting the results of the OLS regressions estimating the difference-in-difference estimator, Table 3 will first report the results of simple OLS regressions, including all clients, also those who only have one observation.
Table 3. Simple analysis of effect of Technical Assistance on outcome variables.
Dependent Variable Gross Profit (Log) Income per capita (Log) Annual Sales (Log) Fixed Assets (Log)
Model Chosen OLS with Robust S.E. OLS OLS with Robust S.E. OLS with Robust S.E. OLS with Robust S.E. OLS with Robust S.E. OLS
OLS with Robust S.E.
1 2 3 4 5 6 7 8
Explanatory
Variables Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value
Technical Assistance 0.161*** (3.44) 0.004 (0.13) 0.209*** (4.07) 0.035 (0.85) 0.184*** (3.82) 0.048 (1.15) 0.288*** (4.01) 0.000 (-0.01) Family Members -0.005 (-0.53) -0.360*** (-29.09) 0.030*** (2.81) 0.005 (0.39) Female -0.024 (-0.73) -0.022 (-0.59) -0.005 (-0.11) -0.383*** (-7.56) Agriculture 0.604*** (16.37) 0.446*** (8.68) -0.201*** (-3.09) 1.252*** (17.3) Rural -0.021 (-0.55) -0.077* (-1.66) 0.206*** (3.63) 0.027 (0.41) Comunal 0.765*** (6.36) 0.694*** (5.24) 0.410*** (3.28) 0.330 (1.64) Sequence -0.137*** (-12.38) -0.138*** (-10.14) -0.062*** (-4.73) 0.003 (0.17) Disbursed Loan 0.000*** (11.16) 0.000*** (4.09) 0.000*** (4.63) 0.000 (0.53) Annual Sales at t=0 0.000*** (11.06) 0.000*** (4.1) 0.000*** (4.46) 0.000*** (7.87) Fixed Assets at t=0 0.000*** (12.43) 0.000*** (3.93) 0.000*** (2.05) 0.000 (0.9)
Regional GDP per Capita -0.000*** (-1.44) 0.000*** (0.77) 0.000*** (2.75) 0.000 (-1.38)
() Constant 11.162*** (551.53) 10.948*** (148.03) 9.161*** (352.42) 9.959*** (111.35) 11.620*** (468.54) 11.155*** (101.72) 10.785*** (271.66) 9.896*** (75.03) Observations 2448 2448 2422 2422 2453 2453 2309 2309 R^2 0.006 0.398 0.007 0.435 0.006 0.263 0.007 0.568 Joint Significance F( 1, 2446) F( 11, 2436) F(1, 2420) F( 11, 2410) F( 1, 2451) F( 11, 2441) F( 1, 2307) F( 11, 2297) 11.83 146.64 16.57 125.07 14.60 51.45 16.08 155.38
6.2. Results OLS regressions using difference-in-difference estimators
For each outcome variable, seven regressions are performed. The first regression includes only the variables Time, Technical Assistance and the Difference-In-Difference estimator (Time*Technical Assistance). The second regression includes some client and household characteristics, while regression 3 also includes some enterprise characteristics, a variable indicating the average GDP in each region and region dummies (Chuiquisaca-Santa Cruz), comparing the regional effect against the base region La Paz. Regression 4 contains the same variables as regression 3, however, standard errors are now clustered by municipality. Regression 5-7 extend regression 4 by including interaction variables, which allows us to investigate whether any interaction effects exist between the difference-in-difference estimator and female/male clients, rich/poor clients, and a low/high amount of loan disbursed.
Clustered standard errors are applied in regressions 4-7, as errors in the same municipality could be correlated, while model errors for clients in different municipalities are assumed to be uncorrelated. If normal standard errors are applied instead of controlling for within-cluster error correlation, very small standard errors may arise, resulting in misleadingly narrow confidence intervals, large t-statistics and low p-values (Cameron and Miller, 2013). The number of clusters for the regressions on the log of Gross Profit and Income Per Capita is 38, while the number of clusters for the regressions on the log of Annual Sales and Fixed Assets is 39.
25
Table 3.A. Impact of technical assistance on Gross Profit (log)
Dependent Variable Gross Profit (Log)
Model Chosen OLS OLS OLS with Robust S.E. OLS with VCE OLS with VCE OLS with VCE OLS with VCE
1 2 3 4 5 6 7
Explanatory Variables Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value
Time 0.421 (3.36) 0.420*** (3.4) 0.420*** (3.84) 0.420*** (5.97) 0.307** (2.08) 0.066 (0.48) 0.450*** (3.87) Technical Assistance -0.118 (-0.95) -0.129 (-1.05) -0.113 (-1.05) -0.113 (-0.96) 0.117 (0.61) -0.325* (-1.92) -0.073 (-0.43) Time*Technica lAssistance 0.179 (1.02) 0.178 (1.03) 0.177 (1.14) 0.177 (1.21) 0.224 (1.01) 0.528* (1.84) 0.136 (0.80) Family Members -0.013 (-0.51) -0.022 (-0.9) -0.022 (-0.81) -0.025 (-0.94) -0.023 (-0.97) -0.021 (-0.77) Male 0.275*** (2.91) 0.193** (2.24) 0.193** (2.06) 0.308** (2.43) 0.192* (1.82) 0.222* (1.93) Agriculture 0.186** (1.97) 0.247** (2.36) 0.247* (1.74) 0.254* (1.73) 0.200 (1.56) 0.258* (1.69) Comunal 0.223 (0.65) 0.650** (2.28) 0.650 (2.24) 0.593** (2.49) 0.626** (2.23) 0.691** (2.34) Sequence -0.045 (-1.54) -0.045 (-1.16) -0.039 (-1.06) -0.036 (-1) -0.050 (-1.28) Disbursed Loan 0.000** (2.47) 0.000* (1.99) 0.000** (2.07) 0.000* (1.83) Annual Sales at t=0 0.000 (0.92) 0.000 (0.73) 0.000 (0.74) 0.000 (-0.2) 0.000* (2.01) Fixed Assets at t=0 0.000*** (6.73) 0.000*** (4.43) 0.000*** (4.42) 0.000*** (4.6) 0.000*** (4.1)
Regional GDP per Capita 0.000 (-0.67) 0.000 (-1.26) 0.000 (-0.98) 0.000 (-1.3) 0.000 (-0.55)
Chiquisaca -0.287 (-1.44) -0.287*** (-4.6) -0.215*** (-2.79) -0.189** (-2.3) -0.204** (-2.25) Cochabamba 0.276** (2.48) 0.276*** (3.01) 0.276*** (2.98) 0.284*** (3.42) 0.280*** (3.19) Oruro -0.041 (-0.19) -0.041 (-0.39) -0.065 (-0.62) -0.001 (-0.01) -0.082 (-0.71) Potosí 0.092 (0.66) 0.092 (0.7) 0.091 (0.9) 0.136 (1.11) 0.101 (0.74) Santa Cruz 0.247 (1.5) 0.247* (1.72) 0.221 (1.51) 0.309** (2.06) 0.372*** (3.04) Time*Technical Assistance*Male -0.062 (-0.24) Time*Male 0.164 (0.82) Technical Assistance*Male -0.355* (-1.79) Time*Technical Assistance*Poor -0.496* (-1.81) Time*Poor 0.499** (2.92) Technical Assistance*Poor 0.302 (1.59) Poor -0.665*** (-5.2) Time*Technical Assistance*Disbursed 0.096 (0.36) Time*Disbursed -0.070 (-0.29) Technical Assistance*Disbursed -0.099 (-0.54) Disbursed (dummy) 0.206 (1.03) Constant 10.930*** (123.34) 10.675*** (77.92) 10.578*** (42.7) 10.578*** (34.59) 10.458 (35.22) 11.089*** (37.06) 10.471*** (30.2) Observations 352 352 352 352 352 352 R^2 0.092 0.128 0.326 0.326 0.338 0.368 0.320 Joint Significance F( 3, 348) F( 7, 343) F( 17, 333) F(15, 37) F( 18, 37) F(19, 37) F(18, 37) 11.77*** 7.19*** 10.23***
26
Table 3.B. Impact of technical assistance on Income Per Capita (log)
Dependent Variable Income per Capita (Log)
Model Chosen OLS OLS OLS OLS with VCE OLS with VCE OLS with VCE OLS with VCE
1 2 3 4 5 6 7
Explanatory Variables Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Time 0.534*** (3.12) 0.500*** (3.44) 0.506*** (3.77) 0.506*** (5.48) 0.648*** (3.69) 0.117 (1.1) 0.676*** (4.16) Technical Assistance -0.190 (-1.11) -0.135 (-0.92) -0.136 (-0.93) -0.136 (-1.18) 0.384 (1.59) -0.444* (-1.86) 0.042 (0.26) Time*Technica lAssistance 0.165 (0.69) 0.158 (0.77) 0.140 (0.75) 0.140 (0.72) -0.066 (-0.19) 0.607 (1.62) 0.024 (0.09) Family Members -0.352*** (-11.15) -0.348*** (-11.63) -0.348*** (-11.42) -0.346*** (-11.58) -0.350*** (-12.29) -0.343*** (-10.42) Male 0.344*** (3.06) 0.251** (2.32) 0.251*** (2.93) 0.664*** (3.76) 0.242*** (2.69) 0.262*** (2.74) Rural 0.002 (0.02) 0.144 (1.2) 0.144 (1.18) 0.157 (1.16) 0.101 (0.87) 0.136 (1.01) Comunal 0.335 (0.84) 0.650 (1.63) 0.650** (2.1) 0.569** (2.45) 0.621* (1.96) 0.652** (2.09) Sequence -0.044 (-1.24) -0.044 (-0.86) -0.040 (-0.86) -0.032 (-0.68) -0.042 (-0.76) Disbursed Loan 0.000 (1.28) 0.000 (0.69) 0.000 (0.87) 0.000 (0.61) 0.000 (1.44) Annual Sales at t=0 0.000 (0.89) 0.000 (0.53) 0.000 (0.47) 0.000 (0.24) Fixed Assets at t=0 0.000*** (5.37) 0.000*** (5.01) 0.000*** (4.9) 0.000*** (5.61) 0.000*** (4.31)
Regional GDP per Capita 0.000 (-0.96) 0.000** (-2.64) -0.000** (-2.06) -0.000*** (-2.82) -0.000* (-1.88)
Chiquisaca -0.970*** (-2.8) -0.970*** (-9.81) -0.851*** (-8.94) -0.886*** (-10.95) -0.874*** (-5.7) Cochabamba -0.184 (-1.36) -0.184* (-1.81) -0.193* (-1.97) -0.176* (-2) -0.183* (-1.83) Oruro -0.109 (-0.44) -0.109 (-0.73) -0.138 (-0.89) -0.070 (-0.5) -0.143 (-0.94) Potosí -0.269 (-1.43) -0.269*** (-2.73) -0.269*** (-2.94) -0.238*** (-2.75) -0.283** (-2.29) Santa Cruz -0.014 (-0.08) -0.014 (-0.11) -0.071 (-0.56) 0.032 (0.23) 0.061 (0.39) Time*Technical Assistance*Male 0.292 (0.85) Time*Male -0.196 (-1.01) Technical Assistance*Male -0.763*** (-2.7) Time*Technical Assistance*Poor -0.657* (-1.77) Time*Poor 0.547*** (2.78) Technical Assistance*Poor 0.435 (1.58) Poor -0.643*** (-3.86) Time*Technical Assistance*Disbursed 0.315 (1.04) Time*Disbursed -0.406* (-1.83) Technical Assistance*Disbursed -0.422* (-1.78) Disbursed (dummy) 0.453* (1.76) Constant 8.698*** (71.13) 9.608*** (57.98) 9.704*** (32.14) 9.704*** (35.21) 9.346*** (30.41) 10.185*** (39.99) 9.485*** (30.29) Observations 341 341 341 341 341 341 R^2 0.076 0.339 0.458 0.458 0.474 0.477 0.464 Joint Significance F( 3, 337) F( 7, 333) F( 17, 322) F(15, 37) F( 18, 37) F(19, 37) F( 18, 37) 9.26*** 24.44*** 16.05*** - - -
27
Table 3.C. Impact of technical assistance on Annual Sales (log)
Dependent Variable Annual Sales (Log)
Model Chosen OLS OLS with Robust S.E. OLS with Robust S.E. OLS with VCE OLS with VCE OLS with VCE OLS with VCE
1 2 3 4 5 6 7
Explanatory Variables Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value
Time 0.101 (0.73) 0.105 (0.71) 0.104 (0.92) 0.104 (1.29) 0.162 (1.17) 0.055 (0.52) 0.119 (1.51) Technical Assistance 0.039 (0.28) 0.069 (0.57) 0.001 (0.01) 0.001 (0.02) -0.006 (-0.06) -0.097 (-1.09) 0.079 (0.8) Time*Technica lAssistance 0.050 (0.26) 0.052 (0.27) 0.054 (0.37) 0.054 (0.53) 0.042 (0.37) 0.063 (0.32) 0.093 (0.63) Family Members 0.043 (1.61) 0.033 (1.64) 0.033 (1.41) 0.034 (1.46) 0.034 (1.55) 0.039 (1.51) Male -0.129 (-1.27) -0.188** (-2.22) -0.188 (-1.63) -0.156 (-0.95) -0.191* (-1.84) -0.169 (-1.64) Agriculture -0.368*** (-3.18) -0.071 (-0.66) -0.071 (-0.47) -0.072 (-0.48) -0.133 (-1.24) -0.069 (-0.44) Comunal -0.111 (-0.51) 0.076 (0.33) 0.076 (0.25) 0.079 (0.26) 0.038 (0.13) 0.107 (0.35) Sequence 0.057** (2.05) 0.057** (2.18) 0.056** (2.18) 0.070** (2.23) 0.054* (1.96) Disbursed Loan 0.000* (1.67) 0.000 (1.27) 0.000 (1.25) 0.000 (1.32) () Annual Sales at t=0 0.000 (0.85) 0.000 (0.65) 0.000 (0.64) 0.000 (-1.06) 0.000 (1.57) Fixed Assets at t=0 0.000*** (5.07) 0.000*** (3.4) 0.000*** (3.39) 0.000*** (3.74) 0.000*** (3.35)
Regional GDP per Capita 0.000 (-0.2) 0.000 (-0.44) 0.000 (-0.49) 0.000 (-0.5) 0.000 (0.21)
Chiquisaca -0.575*** (-2.7) -0.575*** (-6.57) -0.579*** (-6.92) -0.436*** (-4.58) -0.491*** (-5.06) Cochabamba 0.097 (0.9) 0.097 (0.76) 0.097 (0.75) 0.102 (0.92) 0.098 (0.81) Oruro 0.014 (0.06) 0.014 (0.1) 0.016 (0.1) 0.073 (0.51) -0.031 (-0.23) Potosí 0.083 (0.65) 0.083 (0.77) 0.084 (0.76) 0.142* (1.85) 0.080 (0.66) Santa Cruz 0.230 (1.61) 0.230 (1.3) 0.231 (1.28) 0.314* (1.86) 0.319* (1.84) Time*Technical Assistance*Male 0.013 (0.07) Time*Male -0.083 (-0.5) Technical Assistance*Male 0.013 (0.09) Time*Technical Assistance*Poor -0.017 (-0.07) Time*Poor 0.070 (0.43) Technical Assistance*Poor 0.141 (1.35) Poor -0.630*** (-6.62) Time*Technical Assistance*Disbursed -0.061 (-0.26) Time*Disbursed -0.032 (-0.2) Technical Assistance*Disbursed -0.188 (-1.26) Disbursed (dummy) 0.254** (1.84) Constant 11.545*** (117.1) 11.711*** (71.68) 10.911*** (47.86) 10.911*** (45.07) 10.890*** (45.91) 11.397*** (51.17) 10.776*** (43.15) Observations 354 354 354 354 354 354 354 R^2 0.006 0.049 0.465 0.465 0.466 0.520 0.468 Joint Significance F( 3, 350) F(7, 346) F( 17, 336) F( 15, 38) F(18, 38) F( 19, 38) F(18, 38) 0.73 2.86*** 11.39*** - - -
28
Table 3.D. Impact of technical assistance on Fixed Assets (log)
Dependent Variable Annual Sales (Log)
Model Chosen OLS OLS with Robust S.E. OLS with Robust S.E. OLS with VCE OLS with VCE OLS with VCE OLS with VCE
1 2 3 4 5 6 7
Explanatory Variables Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value Coefficient t value
Time 0.175 (0.84) 0.182 (1.01) 0.192 (1.23) 0.192** (2.43) -0.078 (-0.42) -0.089 (-0.52) 0.255** (2.23) Technical Assistance -0.093 (-0.45) -0.166 (-0.94) -0.022 (-0.13) -0.022 (-0.12) -0.161 (-0.48) 0.155 (0.69) -0.010 (-0.04) Time*Technica lAssistance 0.187 (0.64) 0.186 (0.71) 0.173 (0.79) 0.173 (1.33) 0.441* (1.85) 0.341* (1.71) 0.317* (1.8) Family Members 0.011 (0.27) 0.012 (0.34) 0.012 (0.23) 0.010 (0.2) 0.009 (0.19) 0.022 (0.44) Male 0.627*** (4.34) 0.353*** (2.75) 0.353** (2.59) 0.162 (0.72) 0.329** (2.57) 0.337** (2.31) Agriculture 1.0639*** (6.68) 0.645*** (4.62) 0.645*** (2.75) 0.647*** (2.74) 0.541** (2.15) 0.633*** (2.9) Comunal -0.250 (-0.32) -0.126 (-0.27) -0.126 (-0.16) -0.124 (-0.15) -0.189 (-0.24) -0.071 (-0.09) Sequence -0.037 (-0.91) -0.037 (-0.65) -0.037 (-0.66) -0.029 (-0.53) -0.038 (-0.66) Disbursed Loan 0.000 (-0.26) 0.000 (-0.2) 0.000 (-0.19) 0.000 (-0.66) Annual Sales at t=0 0.000*** (8.31) 0.000*** (4.27) 0.000*** (4.25) 0.000*** (5.07) 0.000*** (4.34) Fixed Assets at t=0 0.000 (-0.36) 0.000 (-0.4) 0.000 (-0.39) 0.000 (-0.73) 0.000 (-0.72)
Regional GDP per Capita -0.000*** (-3.03) -0.000*** (-3.59) -0.000*** (-3.42) -0.000*** (-3.77) -0.000*** (-3.69)
Chiquisaca -0.405 (-1.07) -0.405*** (-2.7) -0.405** (-2.16) -0.276* (-1.75) -0.320* (-1.8) Cochabamba -0.257 (-1.63) -0.257* (-1.78) -0.257* (-1.74) -0.243 (-1.65) -0.232 (-1.58) Oruro -0.512* (-1.69) -0.512 (-1.62) -0.516 (-1.57) -0.366 (-1.16) -0.558* (-1.74) Potosí -0.427* (-1.96) -0.427** (-2.14) -0.427** (-2.12) -0.342** (-1.99) -0.409* (-1.87) Santa Cruz -0.538*** (-2.73) -0.538* (-1.9) -0.536* (-1.86) -0.452 (-1.66) -0.546** (-2.07) Time*Technical Assistance*Male -0.372 (-1.19) Time*Male 0.375 (1.53) Technical Assistance*Male 0.194 (0.68) Time*Technical Assistance*Poor -0.256 (-1.00) Time*Poor 0.402** (2.07) Technical Assistance*Poor -0.196 (-0.84) Poor -0.695*** (-3.9) Time*Technical Assistance*Disbursed -0.232 (-0.8) Time*Disbursed -0.134 (-0.59) Technical Assistance*Disbursed -0.087 (-0.4) Disbursed (dummy) 0.302** (2.17) Constant 11.093*** (74.71) 9.947*** (42.65) 11.009*** (30.66) 11.009*** (22.23) 11.151*** (22.73) 11.665 (27.88) 10.855*** (22.08) Observations 339 339 339 339 339 339 339 R^2 0.011 0.220 0.469 0.469 0.471 0.512 0.474 Joint Significance F( 3, 335) F( 7, 331) F( 17, 321) F(15, 38) F(18, 38) F(19, 38) F(18, 38) 1.28 11.28*** 16.66*** - - - -
29
The difference-in-difference estimator for female entrepreneurs (Time*Technical Assistance, regression 5) is negative, and is positive for male entrepreneurs (Time*Technical Assistance*Male). However, the signs are not significant. When splitting the sample into rich and poor entrepreneurs (regression 6), the difference-in-difference estimator is significant for rich entrepreneurs (0.528) and poor entrepreneurs (0.528 + -0.496 = 0.032) at the ten per cent significance level. Rich entrepreneurs benefit significantly more from receiving technical assistance than poor entrepreneurs. Rich entrepreneurs that take up technical assistance have initially significantly lower profit than rich entrepreneurs that do not receive technical assistance (-0.325 at the ten per cent significance level), however, the difference-in-difference estimator (0.528) offsets the initial lower gross profit. The poorer entrepreneurs have significantly lower profits (0.655 at the one per cent significance level), but over time gross profit of the poor without technical assistance increases significantly by 49.9%. Additionally, poor entrepreneurs that take up technical assistance experience an additional 3.2% growth in gross profit. The amount of loan disbursed (regression 7) does not significantly affect the effect of technical assistance, although both signs are positive.
Table 4.B presents the results of the OLS regressions on Income Per Capita (Log). The difference-in-difference estimator is positive but insignificant for regressions 1-4. Time is significantly positive, suggesting that Income Per Capita increases over time, regardless of whether clients receive technical assistance. When including an interaction variable for male entrepreneurs (regression 5), there is a negative (but insignificant) sign for female entrepreneurs and the difference-in-difference estimator, and a positive but insignificant sign for the difference-in-difference estimator and male entrepreneurs. In regression 6, the heterogeneity effects for rich and poor clients are estimated. The difference-in-difference estimator for rich entrepreneurs is positive but insignificant, but the sign for poor entrepreneurs is significantly negative (0.607 + -0.657= -0.05), suggesting that poor entrepreneurs do not benefit from receiving technical assistance, but that their income per capita decreases by 5% if they receive technical assistance. The difference-in-difference estimators in regression 7 are both positive but insignificant, suggesting that the amount of loan disbursed does not affect Income Per Capita.
30
significant effect of technical assistance. Time has a positive sign, but is also insignificant. In regression 5, the difference-in-difference estimator is positive for both female (Time*Technical Assistance) and male (Time*Technical Assistance*male) entrepreneurs, but is insignificant. The difference-in-difference estimators for rich and poor entrepreneur are positive, but also insignificant. Regression 7 reports the results with the interaction for the amount of loan disbursed. Both difference-in-difference estimators are insignificant.
Table 4.D reports the results on the OLS regressions on Fixed Assets (Log). Similar to the previous analyses, the difference-in-difference estimator has a positive sign, but is not significant. Time has a positive sign, and is significant in regression 4. In regression 5, the difference-in-difference estimator for female entrepreneurs is significant (at the ten per cent significance level) and positive (0.441). The difference-in-difference estimator for male clients is negative, but insignificant. Regression 6 estimates the heterogeneity effects based on how rich clients are. The difference-in-difference estimator for rich entrepreneurs is significantly positive (0.341 at the ten per cent significance level), indicating that rich entrepreneurs invest more in fixed assets after they have received technical assistance. The sign for poor entrepreneurs is also positive but insignificant. In regression 7, the regression includes an interaction term referring to the clients that had a relatively high amount of loan disbursed. The difference-in-difference estimator for the clients that had a relatively low amount of loan disbursed is significantly positive (0.317 at the ten per cent significance level), while the sign for the difference-in-difference estimator for clients with a relatively high amount of loan disbursed is negative, although insignificant. This finding suggests that clients with a relatively low amount of loan disbursed benefit from receiving technical assistance, while those who have a relatively high amount of loan disbursed do not.
31
the findings for gross profit, fixed assets increase for rich entrepreneurs that receive technical assistance, by 34.1%. Additionally, clients with a relatively low amount of loan disbursed benefit from technical assistance, their fixed assets increase by 31.7%.
6.2. Results Propensity Score Matching
Table 5 presents the final results of DID-PSM, which are performed as a robustness check on the significant results of the regression analyses discussed in the previous paragraph. After controlling for the propensity score, DID-PSM compares the first difference of a certain economic outcome variable (Gross Profit, Income Per Capita, Annual Sales and Fixed Assets) for households that did receive technical assistance and those households who did not. As the outcome variables are expressed in log terms, the treatment effect (the effect of receiving technical assistance) denotes the growth of the economic outcome variable achieved by receiving technical assistance.7 The propensity score is based upon the following variables: Family Members, Agriculture, Communal loan, Sequence. The variables Annual Sales and Fixed Assets at the time of the first interview/observation are not included for calculating the propensity score, but are used as additional controls next to the propensity score, to make the estimation more robust while not including too many variables for the calculation of the propensity score. 8 Standard errors are clustered at the municipality level.
8 To calculate the difference-in-difference estimation based on Propensity Score Matching, I used the following Stata Code
32 Table 5. Propensity Score Matching Analysis on log of Economic outcome variables
Dependent Variable
Gross Profit (Log) - Only Poor
Gross Profit (Log) - Only Rich
Income Per Capita (Log) - Only Poor
Fixed Assets (Log) - Only Female
Fixed Assets (Log) - Only Rich
Fixed Assets (Log) - Only Low Amount
Disbursed
1 2 3 4 5 6
Baseline Mean t value Mean t value Mean t value Mean t value Mean t value Mean t value
Control 10.443 (96.56) 10.749 (64.66) 8.172 (60.15) 10.126 (20.26) 11.309 (91.55) 9.828 (41.57) Treated 10.513 (143.74) 10.591 (44.66) 8.293 (53.12) 10.054 (28.28) 11.257 (72.16) 9.838 (62.94) Difference (baseline) 0.071 (0.71) -0.158 (-1.29) 0.122 (0.64) -0.071 (-0.21) -0.052 (-0.59) 0.01 (0.06) Follow up () Control 11.122 (85.23) 10.958 (78.63) 8.954 (57.73) 10.014 (16.82) 11.28 (75.12) 10.211 (45.05) Treated 11.066 (82.15) 11.213 (29.41) 8.907 (56.1) 10.391 (24.57) 11.617 (50.78) 10.257 (41.14)
Difference (follow up) -0.056 (-0.37) 0.256 (0.82) -0.048 (-0.29) 0.378 (0.88) 0.337* (2.07) 0.046 (0.25)
Difference in Difference -0.127 (-0.76) 0.414 (1.58) -0.17 (-0.65) 0.449* (1.85) 0.389** (2.95) 0.035 (0.24)
No. Of Treated 126 38 117 59 34 69
No. Of Controls 122 38 116 46 30 86
R^2 0.206 0.367 0.167 0.391 0.750 0.499
33
Table 5 presents the results of the difference-in-difference estimation using Propensity Score Matching. Estimation 1 investigates the impact of technical assistance on Gross Profit (log) for the poor. The sign for the difference-in-difference estimator is negative, but insignificant. Estimation 2 is similar, but investigates the impact for rich entrepreneurs. The difference-in-difference estimator has a positive sign, but is insignificant. Estimation 3 investigates the effect of technical assistance on Income Per Capita (log) for the poor: the effect is negative, but insignificant. Estimation 4 estimates the effect on Fixed Assets for female entrepreneurs: the sign is significanty positive at the ten per cent significance level, suggesting that female entrepreneurs increase their assets by 44.9% after they have received technical assistance. Rich entrepreneurs (estimation 5) also significantly have higher assets (38.9%) after they have received technical assistance. For clients that have a relatively low amount of loan disbursed (estimation 6), there is also no significant difference-in-difference effect.
In conclusion, Propensity Score Matching analysis confirms the OLS findings that female entrepreneurs have significantly higher fixed assets, as well as rich entrepreneurs. The other findings of the OLS regression analyses cannot be confirmed.
6.3. Discussion
The results of the OLS regression analyses suggest that participation in technical assistance training does not directly affect Gross Profit, Income Per Capita, Annual Sales and Fixed Assets, although it is worthwhile to mention that in the OLS regressions, the difference-in-difference estimator does have the expected positive sign for all outcome variables.
34
Nonetheless, Propensity Score Matching do confirm the findings that female and richer entrepreneurs have significantly higher fixed assets after they have received technical assistance. Since there are no significant results found for the other outcome variables that are also confirmed by Propensity Score Matching, the increase in fixed assets cannot be explained from an increase in sales, income or profits. This may be related to the way the data are collected. Measuring income or profits usually yields more noisy results as opposed to measuring fixed assets. Measuring fixed assets is a rather objective proxy of wealth, and may therefore be the only variable where the effect of technical assistance can be detected. Another explanation is that technical assistance training may encourage female clients aned richer clients to invest more, however, this investment has not (yet) led to an increase in profitability or sales. It is quite remarkable that female entrepreneurs seem to benefit from technical assistance training through increases in their fixed assets, while men do not. This finding suggest that Bolivian women do not have such a low decision making power as suggested by Giné and Mansuri (2014). A possible reason that only rich entrepreneurs experience an increase fixed assets is that they may be able to make large investments, while poor entrepreneurs may find it harder to gain access to sufficient funds.
6.4 . Statistical Power
35
(positive) impact at all, however due to the low power and small sample size, it is difficult to find any significant results.
Female entrepreneurs will experience an increase of 44.1% according to OLS estimations, and of 44.9% based on Propensity Score Matching. If we take the most conservative estimate, this suggest that their fixed assets will increase by 44.1% if they receive technical assistance. Since the average value of fixed assets of female entrepreneurs is 95,470 Bs. (much lower than for male entrepreneurs who have an average initial value of fixed assets of 150,104 Bs.), an increase of 44.1% would mean an increase of 42,102 Bs. (approximately 6,111 US Dollars). Rich entrepreneurs will on average have 31.7% higher fixed assets after participating in technical assistance training, according to OLS estimates. Propensity Score Matching indicates that fixed assets will increase by 38.8%. Since the initial fixed assets of rich entrepreneurs has an average value of 171,392 (the average initial value of relative poor entrepreneurs is 31,565), an increase of 31.7% (the most conservative estimate of the two) refers to an average increase of 54,331 Bs. (approximately 7,886 US Dollars).
7. Conclusion and Limitations
