As highlighted in the introduction, this paper investigates the impact of financing structure on MSEs performance and economic outcomes at the firm level. Moreover, by examining firm characteristics, this paper highlights the role of entrepreneurs’ gender and education, aﬀecting the relationship between financing structure and firm performance. Our empirical strategy follows a two-step analysis. First, we estimate firms’ TFP, and second, TFP is treated as a dependent variable with respect to a set of explanatory financial and economic variables.
Estimation of Total Factor Productivity (TFP)
Productivity is a key driver of long-run economic growth and indicates the performance of a firm’s use of scarce resources (Isik and Hassan, 2003), It also accounts for much of the diﬀerence in per capita income between countries (Hsieh and Klenow, 2010). We have two main methodologies for estimating TFP: non-parametric approaches (TFP index such as the Malmquist index and data envelopment analysis – DEA), and parametric approaches (production function estimation and stochastic frontier analysis – SFA).
In our study, we measure TFP estimates by fitting a Cobb Douglas production function to firm-level data. It was derived as the ratio of output produced to an index of composite inputs12. This definition is consistently applied and has been accepted in a large number of studies (Syverson, 2011). In other words, TFP is the part of firm productivity that is not explained by the quantity of inputs used. Accordingly, the relative TFP index for each firm i at time t can be generally defined as follows:
Where Yit is the output of firm i at time t, Kit is the capital input of firm i at time t, Lit is the labor input of firm i at time t. and θit indicates the central tendency of TFP. If a firm’s θ is greater than 1, it indicates high TFP relative to other firms, while a value less than 1 indicates low TFP. Rearranging
12 Output can be quantified either by revenue or by an estimator of value-added (Balk, 2009). We used the revenue in the TFP estimate because information on value-added is not available in our database.
The next feature in (1.2) is the production technology, which can be explained by diﬀerent assump-tions. Among others, translogarithmic production and Cobb-Douglas functions are the two most com-monly used methods. It is argued that both approaches have good mathematical properties. However, the elasticity of output with respect to inputs in the Cobb-Douglas function allows for easier interpre-tation than the translogarithmic output. To be more specific, the translogarithmic technique generally suﬀers from a collinearity problem between the regressors. Thus, in this study, we assumed that the pro-duction technology follows the Cobb-Douglas production function. Therefore, we can write the equation (1.2) as follows:
from the equation (1.3), the inputs were aggregated by taking the exponent of each factor from its respective production elasticity. Syverson (2011) argued that this is more generally valid as a first-order approximation of any production function. Transforming the equation (1.3) into a linear expression by taking the logarithm of both sides of the equation, we have:
According to equation (1.5), the natural logarithm of the TFP index is equal to the residual term uit in the econometric production function. In practice, this equation can be estimated using the OLS estimation technique. However, the major econometric problem in estimating production functions is the possibility that there are determinants of production that are not observed by econometricians but observed by the firm. In this case, firms may use asymmetrically observed shocks to maximize their profits or minimize their costs. More precisely, firms are expected to respond to positive (negative) productivity shocks by increasing (decreasing) their output, thereby increasing the quantity and/or quality of production inputs. Thus, estimating the equation (1.5) by OLS can lead to biased estimates, as the inputs to the production function are likely to be related to the residuals.
We now turn to the control of the endogeneity problem. To this end, we decompose the residual as uit = ωit + it, thus, we presented the production function in the equation (1.5) as follows:
as described above, in the equation (1.6), the residual has been divided into two components where ωit is productivity, and it is unpredicted shocks. In other words, this means that the eﬃciency of the firm is decomposed into a part that can be predicted by the firm, although not observable in the data, and a part due to a productivity shock that cannot be predicted either by the firm or by econometricians:
In dealing with these issues in this estimation, there are three main approaches in the literature, namely instrumental variables (IV), fixed eﬀects (FE), and control function (CF). Among the various alternatives, the standard technique for estimating the production function is the Olley and Pakes (1996) estimation (hereafter, OP). Specifically, the OP framework uses the level of investment as a proxy for unobserved productivity to control for the endogeneity problem that arises due to the correlation between observable input levels and unobservable productivity shocks. However, an important limitation of the OP approach is that investments are not decided at every point in time. Therefore, such a delay violates the monotonicity assumption (Eberhardt et al., 2010).
To address this concern, one of the most common frameworks for estimating firm-level productivity is the control function approach of Levinsohn and Petrin (2003) (hereafter, LP). LP proposes to overcome this problem by exploiting the cost of an intermediate input or electricity instead of investment as an alternative proxy for controlling knowledge of a firm’s eﬃciency. This proxy not only uses intermediate inputs that can be easily adjusted for productivity shocks and addresses concerns about the monotonicity assumption, but it is also easier to implement since it takes advantage of available data for intermediate inputs. They propose the following modified model:
On the one hand, LP performs a two-step estimation, in which the first step is to estimate the coef-ficient of labor β. However, we can see that there is a problem with LP, which is functional dependence. To be more precise, all variables are supposed to occur simultaneously using unconditional demands of intermediate inputs; this could lead to collinearity. However, in reality, the material Mit would normally be chosen after the Lit (Ackerberg et al., 2015) (hereafter, ACF). On the other side, ACF proposed the corrected function approach, which uses moment conditions very similar to those used by OP and LP, but they avoid the functional dependence problem that can occur in LP. Specifically, OP and LP as-sume that firms can adjust certain inputs instantaneously and without cost when subject to productivity shocks. However, ACF has shown that the optimal allocation of labor is also a deterministic function of TFP, and thus the elasticity of labor is not identified.
The main idea of these methodologies is that an intermediate input (ln Mit) such as expenditures on raw materials, energy, and electricity. It can be used as a proxy for the firm’s unobserved productivity, and unbiased production function estimates. The demand function for the intermediate input is given by :
assuming monotonicity, this explains that if the demand for the intermediate input increases monotoni-cally with ωit, it can serve as a valid indicator for the unobservable. Therefore, the demand function for the intermediate input can be inverted to give ωit as a function of capital, labor, and the intermediate input.
By introducing this function that captures the variation in the prediction of firm eﬃciency ωit in the equation (1.6), the production function can be estimated using semi-parametric methods.
the change in inputs is now unrelated to the error term it, so we have consistent parameter estimates.
We compute the TFP for each firm as the residual of an estimate of the equation (1.10).
Empirical model: Fixed eﬀect and GMM methods
After obtaining the firm-level TFP estimates, we follow the methodology employed in the Levine et al. (2000) model of financial development and growth to study the impact of credit facilitation by formal financial institutions on firm performance in the Indonesian manufacturing sector. The regression frame-work consists of a panel regression of firm performance (i) in period (t) on formal financing in the same period and a set of control variables. The econometric specification is given by the following equation:
F Pit = αi + β1 FSit + β Xit + µi + νt + it (1.11)
Where F Pi,t is a measure of firm performance. It can be measured by the TFP of the firm (i) at a time (t) estimated using Ackerberg et al. (2015) and labor productivity. Note that all variables in our regression were transformed into logarithmic form13. F Si,t is the financing structure of a firm (i) at a time (t) and measured by formal external credit or fully internal capital. Formal financing is our explanatory variable of interest; it is a dummy variable that takes the value one if the firm relies on external sources from banks, cooperatives, non-bank financial institutions (NBFIs), and zero otherwise. Meanwhile, fully internal capital is a dummy variable that takes the value of one if the MSE’s main source of capital is based on the internal source of finance (such as inheritance, savings, and asset liquidation) and zero otherwise. According to the theory discussed earlier, β1 is expected to be positive for each of these dependent variables. Furthermore, Xi,t is a vector of observable characteristics of a firm i in period t that could influence the probability of obtaining a loan. In addition, µi are firm-level fixed eﬀects controlling for unobserved characteristics of firm i that do not vary over time, and νt is a set of year fixed eﬀects. Finally, i,t can be interpreted as random shocks.
In terms of econometric methodology, we estimate the equation (1.11) using a fixed-eﬀect panel model, accounting for firm and year fixed eﬀects. Although a standard error fitted to the model can address heteroskedasticity and autocorrelation, Wintoki et al. (2012) have argued that endogeneity bias still exists. This is because fixed-eﬀects models primarily control for unobserved heterogeneity. They do not control for the endogeneity problem caused by measurement errors, time-invariant endogenous variables, and reverse causality that often occurs in financial research. As a result, the use of the FE model may still be biased, especially in the case of short panel data (Cameron and Trivedi, 2005).
To deal with this problem, some previous studies have suggested using instrument variable estimators (IV estimators) or dynamic panel GMMs. However, the problem with applying IV estimators is the diﬃculty of finding variables that can serve as valid instruments because, with weak instruments, IV estimators are likely to be biased. In other words, IV estimates with invalid instruments could oﬀer no improvement over OLS estimators. Therefore, this research applied the dynamic panel GMM explored by Arellano and Bond (1991) to address the endogeneity issue.
One of the advantages of the GMM model over the instrument estimator method is that it is much easier to have instrument variables as exogenous variables in other periods or lagged variables, which could be used as instruments for endogenous variables in the current period14. Therefore, the GMM provides an abundance of instrumentation variables, which makes it easier to achieve the conditions of valid instruments and over-identification of estimators. Besides, the Arellano and Bond estimator is suitable for short panel data that have small T and large N, i.e., few periods and many individuals. However, this research uses short panel data with large firms and only six years, so the GMM method introduced by Arellano and Bond (1991) was employed and considered the most suitable. In this case, we extend the equation (1.11) by adding the lagged dependent variable.
A problem with the original estimator Arellano and Bond (1991), called diﬀerence GMM, is that lagged variables can be weak instruments if the variables in the regressions are close to a random walk because lagged levels transfer little information about changes in the future. Additionally, the GMM diﬀerence has a weakness when there are many gaps in the unbalanced panels (Roodman, 2009). There-fore, Arellano and Bover (1995) and Blundell and Bond (2000) developed the GMM system, in which the original equation is added to the system to augment the instruments, which increases the eﬃciency of the estimators. In this estimator, the lagged diﬀerences are used as instrumentation variables for the level equations, and the lagged levels are used as instruments for the first diﬀerence equations.
To ensure the validity of the set of instruments, Arellano and Bond (1991) proposed two key tests to check the validity of the GMM model. The first test is the Sargan test or Hansen test for overiden-tification. The GMM requires that the overidentification restrictions be valid. The second test is the Arellano-Bond test for autocorrelation errors. The residuals in the first diﬀerence AR (1) are expected to be correlated, but there should be no serial correlation in the second diﬀerence AR (2). The condition for second-order serial correlation is that any historical value of the dependent variables beyond certain lags, which control the dynamic aspects of an empirical relationship, is a valid instrument because it will be exogenous to the current shocks to the dependent variables (Wintoki et al., 2012).
We further verify the robustness of our results. This research performed regressions in which industry and year dummy variables were included to capture industry- or year-specific FE. Besides, alternative measures of the dependent or independent variables were applied to retest the results.
Results and discussion
In this section, we analyze the data according to our empirical strategy. We first generate the baseline result by assessing the relationship between formal financing and firm productivity. Then, fixed-eﬀect regression and GMM are used to investigate our main hypothesis that external sources of financing from formal entities have a positive eﬀect on MSEs performance. Next, we assess how women entrepreneurs may have a distinct impact on the mentioned hypothesis. Finally, we check with other measures of firm performance to examine the varying impact of financing structures on firm performance.
Table of contents :
1 Financing Structure, Micro and Small Enterprises’ Performance, and Woman Entrepreneurship in Indonesia
1.2 Related Literature
1.3 MSEs & financing sources in Indonesia: stylized facts
1.3.1 Development of women entrepreneurs in Indonesia
1.3.2 Main difficulties experienced by manufacturing MSEs
1.3.3 Employment structure and average wage by the gender
1.3.4 Market structure of the Indonesian manufacturing sector
1.3.5 Heterogeneity of geographical attributes
1.4 Data and variables of interest
1.4.1 Data description
1.4.2 Descriptive statistics
1.5 Empirical strategy
1.5.1 Estimation of Total Factor Productivity (TFP)
1.5.2 Empirical model: Fixed effect and GMM methods
1.6 Results and discussion
1.6.1 Formal financing and firm productivity
1.6.2 Women entrepreneurs, formal financing and productivity
1.6.3 Alternatives firm performance measures
1.6.4 Robustness checks
1.A Sample and data definition
1.B Additional figures
2 Gender and Access to Finance on Business Development: Evidence from the World Bank Enterprise Surveys
2.2 Related Literature
2.3 Stylized facts
2.3.1 Female ownership and economic development
2.3.2 Perception on access to financing by gender
2.4 Data and descriptive statistics
2.4.1 Data description
2.4.2 Dependent variables: firm growth
2.4.3 Independent variables
2.4.4 Control variables
2.5 Empirical strategy
2.6 Results and discussion
2.6.1 Gender of the business and employment growth
2.6.2 Female ownership, access to finance and employment growth
2.6.3 Further analysis by firm characteristics and across countries
2.7 Robustness checks
2.7.1 Alternative measure of firm growth: sales growth
2.7.2 Use an alternative measure of external financing
2.7.3 Instrumental variable estimation
2.7.4 Firm growth and crisis
2.7.5 Size of the informal sector
2.9 Conclusion and policy implications
2.A Sample and variable definition
2.B Additional figures
3 Productivity, Resource Misallocation and Manufacturing TFP: Evidence from Indonesian Firm-Level Data
3.2 Related Literature
3.3 Theoretical framework: Misallocation and Productivity
3.3.1 Marginal revenue products and sources of misallocation
3.3.2 Physical, revenue productivity, and aggregate TFP
3.3.3 Decomposition analysis of industry TFP
3.4 Data and stylized facts
3.4.1 Data description
3.4.2 Calibration of parameters
3.4.3 Stylized facts
3.5 Results and discussion
3.5.1 To what extent are resources misallocated in Indonesia
3.5.2 Implications for the size distribution of firms
3.5.3 Potential gains of resource misallocation elimination
3.5.4 Decomposition of potential gains
3.5.5 Selection and productivity
3.5.6 Policy implications and changes in allocative efficiency
3.6 Robustness checks
3.6.1 The role of distortions in firm exit and exporting status
3.6.2 Alternative specifications for TFP gain
3.6.3 Different parameter values for the capital share
3.6.4 Productivity evolution
3.6.5 Alternative measure of capital stocks
3.A Model details
3.A.1 Optimization problems
3.A.2 Profit maximization problem
3.A.3 TFPR, TFPQ and industry TFP
3.A.4 Decomposition of aggregate TFP
3.B List of variables included in the analysis
3.B.1 Construction of misallocation measures for dependent variables
3.B.2 Independent and control variables
3.C Additional tables
3.C.1 Descriptive statistics of the firms and productivity in the sample
3.C.2 Industries classifications based on three-digit international ISIC
3.D Additional figures