Female empowerment and economic growth

Faculty Economics and Business, Amsterdam School of Economics Master thesis Econometrics

Female Empowerment and Economic Growth

Abstract This study investigates the effects of female empowerment on economic growth in developing countries. This is done by performing a dynamic panel-data analysis with data from 1990-2015 and 102 countries, while making use of three different proxies for female empowerment. Female schooling, fertility and female labour participation were investigated. It is found that while their effects were not significant in the first model estimated, which includes all three simultaneously, when making use of interaction terms between the different proxies, female empowerment has a highly significant, positive effect on economic growth. Enhanced economic growth could serve as an extra incentive for governments to empower women.

Author Supervisor

Laura Schim van der Loeff dr. E. Aristodemou

Student Number Second Reader

10016155 dr. J.C.M. van Ophem

Statement of originality

This document is written by Laura Schim van der Loeff, who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been

used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Contents

1 Introduction 1

2 Theory & Previous Research 2

2.1 Economic Growth . . . 2

2.1.1 Growth models . . . 2

2.1.2 Empirical Research . . . 4

2.2 Female empowerment and Economic Growth . . . 6

2.3 Contribution of this thesis . . . 8

3 Data 9 3.1 Main Variables . . . 9

3.2 Control Variables . . . 11

3.3 Sample . . . 13

4 Research Method 14 4.1 Model on Economic growth . . . 14

4.2 The Arellano-Bond estimator . . . 15

4.3 The Blundell-Bond estimator . . . 17

4.4 Application of the Arellano-Bond and Blundell-Bond estimators . . . 18

4.5 Specification tests . . . 19

4.6 Long-term Effects . . . 19

5 Results 20 5.1 Constructing the model . . . 20

5.2 Preliminary findings using OLS, Fixed effects and Anderson-Hsiao estimators . . . . 21

5.3 Arellano-Bond and Blundell-Bond estimations . . . 24

5.3.1 Classification of the Regressors . . . 24

5.3.2 Results AB and BB2 estimations . . . 25

5.4 Extensions of the model using BB2 . . . 28

5.4.1 Estimating differenct combinations of Female Empowerment variables . . . . 28

5.4.2 Interaction terms . . . 29

5.5 Robustness checks . . . 31

5.5.1 Stronger instruments . . . 31

5.5.2 Decreasing Heterogeneity . . . 32

5.5.3 Uncollapsed BB2 estimator and yearly data . . . 36

6 Conclusions 38

7 Limitations and Future Research 39

A Bibliography 41

B Appendix 45

B.1 Appendix I List of countries . . . 45

B.2 Appendix II Descriptive statistics . . . 46

B.3 Appendix III. Collapsed versus Uncollapsed instrument set . . . 46

B.3.1 Uncollapsed Instrument set . . . 46

B.3.2 Collapsed Instrument set . . . 46

B.4 Appendix IV. Extension on the model by estimating each female empowerment vari-able separately . . . 47

1 Introduction

The question of the exact drivers of economic growth is a puzzling one which has occupied both academics as well as policy-makers for a few decades. Since the industrial revolution, the Gross Domestic Product (GDP) per capita in advanced countries has increased exponentially. This not only lead to higher living standards, but also to increased health and life expectancy, due to growth in the medical and pharmaceutical sectors. However, this has not been the case for all countries, and billions of people are still living in extreme poverty (Aghion and Howitt, 2008). Why some countries experience more and quicker growth than others remains unclear, although some pieces of the puzzle have been discovered. To understand the increase in wealth, but also in inequality, it is essential to understand what drives economic growth. Despite the fact that many studies have been conducted, researching a plethora of possible explanations, a clear-cut answer has still not yet been reached.

This thesis adds to the existing literature on economic growth by investigating the effects of women empowerment. Since women make up roughly half of the population, it seems plausible that giving them equal rights and freedom as the other half of the population, or at least an improvement of their current liberties, would lead to more growth.

In the past years, women’s development and gender equality have received a lot of attention from the media, but also from makers. As one of the millennium development goals, policy-makers are more interested than ever on how to achieve more gender equality. However, changes do not happen overnight, and women still face an incredibly challenging world every day, even more so in developing countries. Evidence suggests that women are more often the target of violence, have lower educational enrolment rates, lower labor market participation rates and earn on average lower salaries than men. Therefore, it is clear that there are humanitarian, ethical, social and legal reasons to be concerned about gender inequality. Besides these reasons, a further question that arises is whether there might also be economic reasons for a country to push for more empowerment of women. If this is the case, this might provide the extra bit of motivation to reinforce policies that promote more gender equality.

This thesis examines the effect of female empowerment on economic growth by estimating the effects of three different proxies for female empowerment on economic growth in developing countries, making use of dynamic GMM estimators on panel data.

The rest of this paper is structured as follows: The next section concerns itself with the theory on economic growth and gender development, as well as past empirical findings. Section 3 gives an overview of the data set used. The research method is then introduced section 4, followed by the results in section 5. Section 6 concludes this thesis and finally limitations and recommendations are given in section 7.

2 Theory & Previous Research

In this section theory and previous research are discussed. First, a brief overview of economic growth theories is presented, followed by influential past research. Then the focus will be shifted towards gender development, and its theoretical and empirical effects on economic growth. This section concludes by outlining the contributions of this thesis.

2.1 Economic Growth

”I do not see how one can look at figures like these without seeing them as representing possibilities. Is there some action a government of India could take that would lead the Indian economy to grow like Indonesia’s or Egypt’s? If so, what, exactly? If not, what is it about the nature of India that makes it so? The consequences for human welfare involved in questions like these are simply staggering: Once one starts to think about them, it is hard to think about anything else.” - Lucas (1988)

Before elaborating on theories and evidence about the drivers of economic growth, it is useful to define it. Economic growth is usually measured as the annual increase in a country’s GDP (Aghion and Howitt, 2008). Other indicators have been suggested, such as consumption, but increase in GDP is still the most widely used. Even though using GDP as a proxy has its shortcomings, according to Aghion and Howitt (2008), the most compelling reason for using it is that it is the main determinant of the material well-being of billions of people. As the quote above by Lucas (1988) describes, it is puzzling why some countries grow, while others stay behind. Multiple theories and models have been developed throughout the years, trying to shed light on this question. First different growth model will be discussed, followed by a review of the latest empirical evidence.

2.1.1 Growth models

The starting point for empirical studies on economic growth is the neoclassical growth model, as first constructed by Solow (1956) and Swan(1956). In this model, a Cobb-Douglas production function is assumed:

Y (t) = K(t)α(A(t)L(t))1−α, α ∈ (0, 1)

Where Y is a country’s output (and income) at time t, K stands for its capital, L represents Labor and A the technology level (all at time t). The Solow-Swan model takes the rate of Labor growth, usually modelled by the population growth, n, and technology growth, g, as exogenous. Additionally, it is assumed that a constant fraction of output, s, is invested into capital. The main predictions of this model are about the impact of savings and population growth on income per capita, y = Y_L.

The model predicts a steady state and finds the steady state income per capita to be: ln(y) = lnA(0) + gt + α

1 − αln(s) − α

1 − αln(n + g + δ)

with δ being the depreciation rate. So, the higher the savings rate, the richer the country, and the higher the rate of population growth, the poorer the country. Since technology is exogenous, it is assumed that: lnA(0) = α + , where α is a constant, and  is a country-specific shift or shock, considered to be exogenous.

In the Solow-Swan model, the main source of growth in the long-run is driven by technological change. To rationalize persistent cross-country differences in growth, one would need differences in the rates of technology growth. But this is considered exogenous, so the Solow-Swan model fails to give an explanation for cross-country differences. Moreover, according to Barro (1991), the model implies that a lower starting level of real income per capita leads to a higher growth rate, allowing poorer countries to catch up with the rich (a theory known as absolute convergence). However, this is not in line with the evidence, since in practice different nations have maintained different per-capita growth rates over the long-run. These critiques gave rise to the endogenous growth models. Romer(1986) and Lucas (1988) are the main contributors to endogenous-growth models, where they attempt to explain the source of growth by endogeneizing it, namely as the accumulation of knowledge. They define this knowledge very broadly, including formal education, training, research, product and process innovations. Furthermore, in endogenous-growth models there is no steady-state level of income; thus differences among countries can persist indefinitely, rejecting the idea of absolute convergence. The main difference with the Solow-Swan model is that endogenous growth models invoke constant or increasing returns to capital, in contrast to the neoclassical growth model, which assumes decreasing returns to capital.

The endogenous growth models quickly gained popularity, until Mankiw, Romer & Weil (1992) claimed that economists had rejected the Solow-Swan model too quickly in favor of the endogenous growth model. While they acknowledged that the Solow-Swan model has its shortcomings, they still found it of great value, even more after augmenting it with human capital, the main contribution of endogenous-growth models. After testing the Solow-Swan model empirically, they found that though the predicted signs were correct, the magnitudes predicted were not, which they claimed was due to omitted variable bias. The authors then augmented the Solow-Swan model with human capital, one of the main contributions of the endogenous growth models, which can be interpreted as the set of intangible resources which have improved the productivity of the labor factor (Goldin, 2016). This is often associated to knowledge and skills that are acquired through experience, education, but also health care (Teixeira and Queir´os, 2015). The revised version of the model takes the following form:

Y (t) = K(t)αH(t)β(A(t)L(t))1−α−β, α, β ∈ (0, 1)

Running a cross country regression, they found a remarkable fit. Their results indicate that the Solow-Swan model is actually consistent with the international empirical evidence, once it is augmented with human capital.

The authors argue that contrary to what Barro (1991) stated, the Solow-Swan model does not predict absolute convergence, but rather conditional convergence, where each country converges to its own steady state, which is also what Knight, Loayza and Villaneuva (1993) claim. Mankiw et al. (1992) find evidence, which suggests that there is indeed conditional convergence, a result later corroborated by most empirical research (Klasen, 2000). The extension presented by Mankiw et al. (1992), hereafter the MRW model, was quickly implemented in studies following. Even though the specification and elaborateness of the model has evolved over the years, the inclusion of human capital, along with labor and capital is a crucial part in almost all literature concerning economic growth. Throughout the literature, it has been shown that human capital is one of the main determinants of economic growth. This is because a higher educated workforce is more productive and innovative (Teixeira and Queir´os, 2015). Also, it enhances the adoption of technology from other countries. According to Hanushek (2013), human capital (proxied by schooling) has not only been found to be highly significant in empirical research, but has found its way to policy makers, as it is now a central part of the developmental strategies.

The next paragraph discusses the alterations to the MRW model in more depth, along with most recent literature on economic growth.

2.1.2 Empirical Research

As mentioned in the previous paragraph, the MRW model has been adopted by scholars ever since, evolving over the past decades. This paragraph first discusses some critiques on the MRW model, and the adaptions made as a result. This is followed by a brief summary of empirical evidence concerning the determinants of economic growth.

As Durlauf and Quah (1999) point out, though it had indeed been shown that the (augmented) Solow model has strong statistical power when explaining cross-country growth differences, suffi-ciently many problems exist with the model, and mostly with its specification and methodology. This point is also reinforced by Brock and Durlauf (2001), who state that the growth literature is notable for the large gaps that persist between theory and empirical findings. One of their main critiques is the linearity of the model, while they argue that the reality is much more complex.

Another set of issues that arises from the MRW model are related to their choice of estimation method, namely pooled-OLS method. First of all, as Islam (1995) points out, it does not correctly

account for cross-country differences. Those differences are included in the error-term  and are con-sidered exogenous. This is a very strong, albeit not very likely, assumption, but it is an econometric necessity for the MRW model, since that is the only way the OLS method is valid (Islam, 1995). Moreover, applying a pooled-OLS does not indicate a clear causal direction, which is problematic when one wants to develop a policy (Durlauf and Quah, 1999). Finally, the use of OLS quickly leads to omitted variable bias (Islam, 1995).

Later studies have taken these critiques into account, by making use of panel data, both static (Islam, 1995; Barro, 2001) as well as dynamic (Iqbal and Daly, 2014; Teixeira and Quier´os, 2015). The possible non-linearity of the equation has been estimated by Moral-Benito (2012), who used a Bayesian panel data approach while trying to identify growth determinants. Also, more interest in interaction terms between variables arose, leading to non-linear models, such as Teixeira and Queir´os (2015), who investigated the interaction effect of structural change and human capital on economic growth, making use of a dynamic panel regression.

According to Moral-Benito (2012), the main area of effort in empirical growth literature has been selecting the appropriate variables to include. The MRW model can be expanded with many determinants of growth, and more than 100 empirical studies have been devoted to finding the missing pieces of the puzzle, each one finding different, but significant determinants of growth (Moral-Benito, 2012). Besides research into many different proxies for human capital, such as schooling, health and life expectancy, many more determinants have been investigated. Quite some attention has been given to the effect of the institutional framework of countries, such as the strength of democracy and the level of corruption (Iqbal and Daly, 2014; Moral-Benito, 2012; Easterly and Levine, 1997; Mauro, 1995). Although there is not an unanimous conclusion, most studies suggest that countries with a stronger institutional framework (more democracy, higher rule of law, less corruption), have a higher growth rate. Also international openness and trade has been examined extensively as a possible growth determinant, and most literature found it to positively affect growth (Hansen and Tarp, 2001). Trade alters prices, wages and production, and therefore changes consumption and saving patterns (Aguayo-Tellez, 2011). Furthermore, Seguino (2000) points out that increased trade permits economies of scale and specialization which lead to increased productivity and output. Fischer (1993) brings forward the importance of a stable macroeconomic framework for economic growth. He claims that this is necessary, but not sufficient for sustainable economic growth. He supports this view with empirical evidence, augmenting the MRW model with inflation, and finding significant, negative results. He also points out the non-regression evidence, such as the growth crisis in Brazil coinciding with high inflation, as well as the recovery of economic growth in Chile and Mexico which followed the reduction of inflation. The author suggests that the main reason macro-economic factors matter for growth is due to uncertainty. This uncertainty affects growth in two ways; besides the reduction of investment, the price mechanism is less efficient, leading to a reduction of the level of productivity.

Other main regressors which have been found significant include religion and geography dum-mies. The determinants of growth that have been found to be highly significant in multiple studies and relevant for this thesis, are included in this thesis as control variables. They are presented and discussed in more length under ‘Data’, in section 3. Now that some more general findings on economic growth have been presented, the next paragraph concentrates on the link between gender development and economic growth.

2.2 Female empowerment and Economic Growth

It has proven to be less straightforward to define female empowerment than economic growth. De-spite the fact that “female empowerment” is a term used pervasively, there is no common definition of what this entails. More than 20 different definitions of empowerment have been suggested, with much common ground between them (Wyndow, Li and Mattes, 2013). In general, empowerment is conceptualized as a process, where women move from a lesser state to a higher one (Kabeer, 2005). Another commonality in these definitions is the relationship between a woman’s individual agency and the macro-social institutions and structures that restrict or enhance her ability to exercise that agency (Wyndow et al., 2013). Another aspect that is often mentioned is the equality with respect to men (Klasen, 2000). Throughout the literature, the terms ‘female empowerment’, ‘gender devel-opment’ and ‘gender inequality’ are used interchangeably, this is also the case for this thesis. This section discusses some of the literature on gender development and economic growth. There are relatively few studies that explicitly concern themselves with this relationship, and even less that do this in an empirical way (Klasen, 2000). Moreover, as far as the author of this thesis is aware, the empirical research so far has only concerned itself with one aspect or proxy for female empowerment, the most common ones being education and employment. So, first the effects of female education are discussed, followed by the effects of (in)equality in employment and wage. Also, some indirect effects are discussed.

Theoretical literature suggests that gender inequality reduces average human capital, and there-fore harms growth (Schober and Winter-Ebmer, 2011). It can reduce human capital in multi-ple ways, the most evident way being through education. Klasen (2000) mentions the selection-distortion factor. If one starts with the belief that boys and girls have a similar distribution of innate abilities, gender inequality in education thus means that the average innate ability of those who get educated is lower than it would be if boys and girls received equal education. Besides this, there are also some positive externalities related to increased female education. Research has shown that increased female education promotes the education of the next generation, reduces fertility an reduces child mortality levels (Klasen, 2000; Schober and Winter-Ebmer, 2011). Initial findings on this relationship confirm the theory (Schober and Winter-Ebmer, 2011). However, Barro and Lee (1994) find, contrary to their initial assumptions, a negative effect of female schooling (both

primary and secondary years of schooling) on economic growth. The authors suggest that a large difference between male and female school attainment might be a sign of ‘backwardness’. Following the conditional convergence theory, a country with a larger gap (so less schooling for females), will have more backwardness, and therefore a higher growth potential. Klasen (2000) points out, that their finding might also be due to multicollinearity, since male and female schooling are closely correlated in most countries. Dollar and Gatti (1999) also investigated the relationship between gender inequality in education and economic growth. In contrast to Barro and Lee (1994), they found a positive relationship between female secondary education and economic growth, a result supported by most other studies (Seguino, 2000). Furthermore, they conclude that promoting fe-male education only has a significant impact on economic growth in countries where there were already higher female education levels. In other words, countries with low level of female schooling do not benefit much in terms of growth, by increasing female schooling (Dollar and Gatti, 1999). However, studies over the last years have provided fairly solid evidence that inequality in education is detrimental to growth (Schober and Winter-Ebmer, 2011).

Another way inequality can be measured is by studying women’s access to employment. The arguments regarding gender gaps in employment are closely related to the ones mentioned above. Similar to education, Klasen (2000) identifies a selection-distortion effect. By reducing employment chances for women, the average ability of the workforce is not as high as it could be under equal employment opportunities. This is in line with his own findings, which suggest that gender bias in employment leads to lower growth rates. Additionally, Klasen and Lamanna (2009) contest that lower female labor participation is associated with higher fertility rates, leading to lower economic growth. The authors investigate the effects of employment gaps on growth, and find that low female labor participation has a major impact on economic growth, reducing it significantly. Another ar-gument presented by the authors is that increased employment and thus income, increases women’s bargaining power at home. This does not only benefit the women themselves, but can lead to a series of growth-enhancing effects, such as increased savings, more education (for themselves and their children), higher health, leading to increased human capital, now and in the future (Klasen and Lamana, 2009). In contrast to theory, Seguino (2000) found that a larger wage gap led to lower economic growth, contradicting earlier findings on this relationship. She argues that this is because lower wages for women would foster investments and exports, leading to increased output and growth. Mitra-Kahn& Mitra-Kahn (2008) emphasized these results by stating that a larger gender wage gap is even more growth promoting in developing countries (as cited in Schober and Winter-Ebmer, 2011). However, Schober and Winter-Ebmber (2011) argue that Seguino (2000) uses the aggregate wage gap, while she should have used more specific data on gender discrimina-tion. The authors replicate her study by making use of various definitions of gender discrimination, but reach the opposite conclusion, presenting evidence that more wage equality actually leads to growth. This results have also been found by Cavalcanti and Tavares (2007) (as cited in Klasen and

Lamanna, 2009).

Besides the direct effects on growth, some indirect effects can be identified. Seguino and Floro (2003) argue that men and women have a different propensity to save. More empowerment of women would lead to a shift in household savings. Since household savings consist of the most significant component of gross domestic savings in many developing countries, this would therefore indirectly influence economic growth. However, their results are not significant. Wyndow et al. (2013) investigated the effects of female empowerment on democracy, using dynamic panel data, and find that there is a positive effect. Since there is some evidence linking more democracy to higher growth rates, this could be another indirect effect. Busse and Spielmann (2006) investigate the effects of wage inequality on trade, and find a positive effect, especially when it comes to labor-intensive goods. However, when they measure gender inequality in labor force participation and educational attainment, they find a negative effect on trade.

Clearly, there is not a clear-cut answer about the effects of female empowerment on economic growth. Since interest in this topic and data availability on female empowerment is fairly new, research on the effects of female empowerment on economic growth is still in its infancy. Further research is needed in order to draw robust conclusions. The next paragraph outlines how this thesis intends to contribute to this.

2.3 Contribution of this thesis

This thesis adds to existing literature in a number of ways. First of all, there is still some ambiguity in the conclusions drawn by different studies, and more studies are needed for robustness. Second, as mentioned previously, most research focuses only on one measurement tool for female empow-erment, usually education or employment, whereas this thesis will encompass multiple measures of female empowerment, as well as interaction terms. This is done in line with Wyndow et al. (2013), who use 3 different proxies for female empowerment, in contrast to previous research. To the au-thors knowledge, this method has not yet been used to investigate the effects on growth. Finally, most research that has been done is cross sectional. This is problematic for two reasons. First of all, since the causal relationship is not directly clear, a simple cross sectional study will have limited external validity. Moreover, evidence from previous literature on economic growth (not accounting for female empowerment), suggests that there are some country fixed effects. Ignoring those will lead to omitted variable bias and thus inconsistent estimators. This thesis avoids the issues dis-cussed previously caused by OLS and FE regressions, by making use of dynamic panel data. Even though in the last few decades, there has been an augmented interest in employing panel data to estimate economic growth models, dynamic panel data has not yet been used to estimate the effects of female empowerment, as is done in this thesis.

multiple areas to quantify female empowerment. Now that a brief overview of previous literature has been given, the next section elaborates on the specific determinants included in this thesis, as well as their data source.

3 Data

In this section, the data is discussed. First, the main variables are described, starting with the dependent variable, economic growth, followed by the specification of multiple proxies for female empowerment, the explanatory variables of interest. This is followed by elaborating on other ex-planatory variables and finally the sample and dataset used are described, before moving on to the research method in the following section.

3.1 Main Variables

This thesis investigates the effects of female empowerment on economic growth. As mentioned previously, this thesis uses three distinct manners to measure female empowerment. In line with Wyndow et al. (2013), Fertility rates, Female Education and Female Labor Participation rates are used. There are many more variables which could be used, but the three mentioned seem to be the most recurring in literature, and have the highest data availability. The United Nations Develop-ment Program developed two indexes to measure female empowerDevelop-ment; the Gender DevelopDevelop-ment Index (hereafter: GDI), and the Gender Inequality Index (hereafter: GII). Both indexes encompass a range of developmental variables (including the three used in this thesis), and would theoretically have been very suitable. However they proved unavailable for many years and many countries in the dataset, unfortunately making them unsuitable. Each explanatory variable is first defined, followed by a brief explanation of why it is deemed a suitable proxy for female empowerment. The expected effect on economic growth is then discussed, based on previous literature. All data is retrieved from the World Bank Development Indicators (2017), unless stated otherwise.

Economic Growth: The first difference of the logarithm of real gross domestic product (GDP) per capita serves as a proxy for economic growth, as it is common in the literature (Aghion and Howitt, 2008). As a robustness check, data from the Penn World Table are used.

Fertility rates: Defined as “the average number of children that a woman gives birth to in her lifetime, assuming that the prevailing birth rate for each age category remains unchanged” (World Bank, 2017). Wyndow et. al (2013) argue that the main influence of fertility rate on female empowerment is the direct transformation of the lives of women. Lower fertility rates increase life expectancy (less mortality during childbirth) and decrease family-related workload, freeing up more time for further education and employment. So lower fertility rates

would indicate higher female empowerment. According to Galor and Weil (1996), relationships between level of fertility and the level of income per capita is one of the strongest observable correlations in cross-country data. According to the authors, it has a negative effect on economic growth, since higher fertility implies higher population growth, which in line with the MRW model and the neoclassical model, has been shown to have a negative effect on economic growth.

Female Education: Measured in average years of schooling. Education is a crucial foundation for the development which enhances a country’s human capital. Increasing women’s education leads to improved status and economic independence, due to better economic prospects, more decision-making autonomy, increased control over resources and reduced restriction on physical mobility, and therefore it serves as an excellent proxy for empowerment (Gupta and Yesudian, 2006). Educating girls and women also leads to higher human capital, which according to the MRW model, should lead to more economic growth, a result also found by Galor and Weil (1996). Though Barro and Lee (1994) found a negative effect of female schooling on economic growth, they do argue that female schooling is likely to promote growth indirectly through fertility rates. This thesis expects to find a positive coefficient, in line with theory and most empirical evidence. Whereas data on overall schooling is quite complete and widely available, data on female education proved significantly more challenging to obtain. The Barro and Lee (2014) dataset has been used for this. However, data are only available for every 5 years, and thus have been linearly extrapolated for the years in between, as done in Wyndow et al. (2013).

Female Labor Participation: This is defined by the World Bank (2017) as the proportion of the female population above 15 years that is economically active. In practice this comes down to the women in paid employment. It is included as a proxy for female empowerment because the expansion of economic rights for women has been proven an important tool for female empowerment, along with raising their status (Wyndow et. al, 2013). Labor force participation not only increases the income earned by women, but also increase their exposures to the outside world, leading to more empowerment (Gupta and Yesudian, 2006). Likewise, as stated earlier, an increase in female labor participation is expected to increase GDP. In conclusion, theory and past literature indicates that female empowerment can be expected to have a positive effect on economic growth. Therefore, the coefficients of Female Education and Female Labor Participation are expected to be positive, and the coefficient of Fertility Rates to be negative. The control variables are discussed in the next paragraph.

3.2 Control Variables

As Brock and Durlauf (2001) and Barro(2001) have pointed out, a fundamental problem with growth regressions is the open-endedness of growth theories. This entails that the validity of one causal theory does not imply the falsity of another; therefore it becomes problematic identifying the variables to include in the analysis. As a result, well over 140 variables have been proposed as potential growth determinants (Moral-Benito, 2012). However, most studies focus on about 10-20 explanatory variable. The control variables have been chosen in an attempt to be as inclusive as possible, while being mindful of not over-fitting the model. Again, all data is retrieved from the World Bank Development Indicators (2017), unless stated otherwise.

Population growth rate: Computed as the first difference of the logs of the population. Ac-cording to the MRW model a higher population growth rate should have a negative effect on economic growth, due to capital dilution (Galor and Weil, 1996). This result is indeed found by many authors, among others Bloom et al. (1998), Moral-Benito(2012) and Mankiw et al. (1992), and it is also expected in this thesis.

Investment share: Defined by the World Bank (2017) as “outlays on additions to the fixed assets of the economy plus net changes in the level of inventories”. Theoretically, an increase in savings should lead to an increase in economic growth. According to Solow (1956) this is because more savings lead to more investments which lead to more economic growth. Most empirical research uses investment share to GDP instead of savings (Barro, 2001; Teixeira& Quier´os, 2015; Moral-Benito, 2012) and do indeed find a significant, positive coefficient. This thesis also uses investment share instead of savings, and a positive coefficient is expected, in line with theory and previous empirical findings.

Human Capital: Through the vast amount of literature available on the effects of human capital on economic growth, many proxies have been suggested. To avoid multicolinearity issues, two proxies have been chosen, in line with the most recurrent ones from literature, which do not directly seem to have a lot of overlap:

• Average years of schooling is the most widely used proxy for human capital(Hanushek, 2013; Teixeira and Quier´os, 2015), and thus it is included. Hall and Jones (1999) state that an extra year of schooling will increase the productivity and efficiency of workers and thus their income. The coefficient on schooling is therefore expected to be positive. The data has been retrieved from the Human Development reports, from the United Nations Development Program (2017).

• The logarithm of life expectancy is used, in line with Bloom et al. (1998), Gallup and Sachs (2001), Miniou and Reddy (2010), Moral-Benito (2012) and many others, who argue that higher life expectancy indicates better health. This variable has been found

to have a positive and significant relation with economic growth, which is also expected in this thesis. Barro and Lee (1994) argue that this is because it not only proxies for good health, but also for other features, such as better work habits and more skills, which all reflect desirable performance in a society.

Initial GDP per capita: As it is standard in the empirical growth literature, growth during a certain period is allowed to depend on the GDP per capita at the beginning of that period, to capture conditional convergence effects (Burnside and Dollar, 2000). Barro (2001) finds a significant positive effect for this variable, but a significant negative effect for its square. Iqbal and Daly (2014) however, find a significant, negative effect, which together with a positive effect on lagged growth rates, supports the conditional convergence theory.

Government expenditure rate: This includes all government current expenditures for purchases of goods and services (including compensation of employees), as a percentage of GDP. It also includes most expenditures on national defense and security (World Bank, 2017). Among others, Barro (2001) finds a positive relationship between government expenditure rate and growth.

Inflation: As measured by the consumer price index, and included as a measure of macroeconomic policy. There are multiple indicators of macroeconomic policy, but Fischer (1993) singles out inflation as the best single indicator, and has been used repeatedly throughout literature. Fischer (1993) argues that the inflation rate serves as an indicator for the overall ability of the government, with high inflation rates indicating a government that has lost control. Fischer (1993), Burnside and Dollar(2000) and Barro (2001) all found that this variable has a marginally significant, negative effect on economic growth, as expected.

Rule of Law: Used as an indicator of the strength of institutions, in line with Barro (2001), and is defined by the World Bank (2017) as: ‘it captures perceptions of the extent to which agents have confidence in and abide by the rules of society, and in particular the quality of contract enforcement, property rights, the police, and the courts, as well as the likelihood of crime and violence’. It is measured in units of the standard normal distribution, so ranging from approximately -2.5 to 2.5. A higher score indicates a stronger Fule of Law. According to Mauro (1995), most economists would argue that efficient government institutions foster economic growth, a point also made by Barro (2001) who states that a strong legal system is central for economic growth. Data for this variable are only available from 1996 onwards. Trade: Defined by the World Bank (2017) as the sum of exports and imports of goods and services

measured as a share of GDP. It is used to measure international openness, in line with Klasen and Lannamma (2009) and Barro (2001). As discussed previously, a change in trade alters GDP both directly and indirectly. In most studies, a significant, positive effect has been found

(Barro, 2001 ; Hansen and Tarp, 2001, Klasen and Lannama, 2009), and it is also expected in this thesis.

Undoubtedly, there are many more variables which would have a significant effect on economic growth. Since this thesis makes use of fixed effects, time-invariant regressors, despite being proven significant in empirical research, such as geography and religion, are not included. As stated earlier, the list of explanatory variables has been chosen in a very comprehensive way, and the author believes it is fairly complete.

3.3 Sample

The dataset originally consisted of the years 1990-2015. Before 1990 no data is available on labor participation, and also other variables have too many missing data points, which would lead to less reliable estimators. However, since no data on Rule of Law are available before 1994, the dataset now consists of the years 1994-2015. Since an important aspect of female empowerment is the transition from a less empowered state to a more empowered state, this thesis focuses on developing countries, where this change can be more clearly measured. The countries used for this thesis are in line with Ouedrago and Marlet (2018), who start of with a sample of 94 countries. This thesis augments this to 102 countries. This is because the authors excluded some countries, due to unknown reasons, which other literature would classify as a developing country (such as Nigeria, Angola, Zambia, Zimbabwe), and have therefore been included. Countries who have more than 3 missing data points on growth are left out. The list of countries can be found under Appendix I. A summary of the statistics can be found in Appendix II. As is common in literature on economic growth, a 4 year time period is used (Rajan and Subramaian, 2008). So, 6 time points are used, from 1994-2014 with 4 yearly intervals, resulting in a total of 594 observations. However, since some data on control variables is sometimes unavailable, less observations are used in the estimations. However, the number of observations remains sufficiently large and balanced, and does not undermine the results, in the author’s opinion.

Now that the dataset and sample have been defined, the next section describes the model used, along with the estimation methods

4 Research Method

This section gives a detailed description of the research method. First, a general model for economic growth is proposed, based on the theory and literature discussed earlier. Then, the Arellano-Bond and Blundell-Bond dynamic panel data estimators are explained, which are both designed for ’small T, large N’ data sets (Roodman, 2006). This is followed by a description of the exact application of the estimators on the growth model proposed.

4.1 Model on Economic growth

The general regression model on economic growth that is used in this thesis is an extension of the MRW model. Besides the addition of explanatory variables, the specification has changed, to a dynamic model, including past realizations of economic growth as explanatory variables. As Brock and Durlauf (1999) state, a major drawback of the traditional Solow/MRW model, is their static nature. Performing a cross-country, pooled OLS regression, might give some insights into the relevance of regressors, but does not take into account fundamental differences across countries, nor does it deal with the issue of reversed causality, leading to a number of issues, that have been discussed earlier. The general model is as follows:

gi,t = α + γ1gi,t−1+ β1F Ei,t+ β2F Ei,t−q+ θ1CVi,t+ θ2CVi,t−q+ µi+ πt+ i,t (1)

Where gi,t is the economic growth of country i in period t. FE is a matrix containing the proxies

for female empowerment, along with their interaction terms: Female Education, Fertility rates, and Female Labor Force participation. The interaction terms of the 3 female empowerment variables are included, in line with Wyndow et al. (2013), to emphasize the importance of their interplay. CV represents the matrix of control variables, as discussed in section 3.2. µi represent unobserved,

time-invariant country fixed effects, such as natural resources across countries, geographical location and topography. πt are time specific fixed effects, which are measured by adding dummy variables,

1 per time period (every 4 years). These dummies equal 1 in their own time period and 0 else-where, controlling for worldwide time specific effects and preventing contemporaneous correlation (Roodman, 2006). i,t is an idiosyncratic error term. The main coefficients of interest are thus the

β’s, which reflect whether female empowerment has an effect on economic growth, and if so, its magnitude and direction. As in the MRW model, the logarithms of population growth, investment and life expectancy are used. MRW also include the logarithm of schooling, this thesis uses its actual level, for ease of interpretation (so one extra year of schooling leads to a certain percent-age increase/decrease in economic growth) . MRW focus on the percentpercent-age of the working percent-age population which is in school, whereas this thesis uses the average years of schooling.

variables, also have some lagged, varying dynamic effects. Lags of these variables (F Ei,t−q, CVi,t−q),

with q greater or equal to 1, are included in the model, to account for possible dynamic effects of these variables. In this way, they do not only work through the lagged economic growth variable, but also directly through their own lagged effect. First, two lags will be included of all explanatory variables, as well as to their contemporaneous realization. Hereafter, the amount of lags and ex-planatory variables is slowly decreased, by eliminating the (highly) insignificant ones, starting by eliminating the one with the highest p-value,to approximate the best fitting model. It is to be noted that the order of elimination may influence the final outcome of the model, though these variations are expected to be minor.

The dynamic nature of the model, due to the presence of the lagged dependent variable gi,t−1

renders the traditional ‘fixed effects’ and ‘random effects’ panel estimators inconsistent, since the regressors are correlated with the idiosyncratic error term. To illustrate this, consider the following simplified version of the dynamic DGP with idiosyncratic shocks, and fixed effects:

gi,t= αgi,t−1+ β0xi,t+ µi+ i,t (2)

Using OLS to estimate this equation leads to inconsistent estimators, due to the presence of the unobserved country-fixed effects µi. By construction, gi,t−1 and i,t are correlated, leading to

endo-geneity. By taking the first difference, subtracting the past period from the current one, the fixed effects can be eliminated, resulting in the following model:

∆gi,t= α∆gi,t−1+ β0∆xi,t+ ∆i,t (3)

However, since ∆gi,t−1 is correlated with ∆i,t through i,t−1, there is still an endogeneity issue.

Furthermore, it is not unlikely that some of the other regressors ∆xi,t might be endogenous. This

violates the strict exogeneity condition, which is required for consistent estimators of fixed effect estimations (Cameron and Trivedi, 2005). However, an Instrumental Variable variant leads to consistent estimators, which is presented in the next paragraph.

4.2 The Arellano-Bond estimator

To deal with the problem of endogeneity and fixed effects in panel data, instrumental variables can be used. In econometrics, one of the hardest issues is to find instruments that are valid, meaning exogenous and relevant (correlated with the instrumented endogenous variable). This section out-lines the first of two estimation methods, as proposed by Arellano and Bond (1991). Afterwards, the method proposed by Blundell and Bond (2001) is discussed, followed by their application on the economic growth model proposed.

as an instrument for ∆yi,t−1. Assuming the errors are serially uncorrelated, this is a valid instrument,

since yi,t−2 is not correlated with ∆i,t. Additionally, it is also a relevant instrument, since yi,t−2

is clearly correlated with ∆yi,t−1. ∆xi,t is used as an instrument for itself, thus assuming it is

exogenous. This estimator is a simple IV estimation, with one instrument per endogenous variable. According to Cameron and Trivedi (2005), the estimation becomes more efficient by using additional lags of the dependent variable as instruments.

Arellano and Bond (1991) argue that the Anderson-Hsiao estimator, fails to take all of the potential orthogonality conditions into account, even though it is consistent. They introduce a panel Generalize Method of Moments estimator (hereafter GMM), using all valid lagged values of the dependent variable as instruments. This estimator is known as the Arellano-Bond estimator, hereafter AB. Again, under the assumption of serially uncorrelated errors, all lags yi,t−q with q

larger or equal to two are correlated with ∆yi,t−1, but not with ∆i,t. The authors state that the

other independent variables, xi,t, can be partitioned into (x1,i,t, x2,i,t, x3,i,t), where x1,i,t is strictly

exogenous, x2,i,t is weakly exogenous (predetermined) and x3,i,t is endogenous. For the weakly

exogenous x2,i,t (E(i,t, xi,s) = 0 if s ≤ t), variables xi,t−q with q greater or equal to one may be

used as instruments. For the endogenous x3,i,t (E(i,t, xi,s) = 0 if s < t) variables xi,t−q with q

greater or equal to two may be used as instruments, as with the endogenous ∆yi,t−1.Besides the

inclusion of internal instruments, the AB-estimator also allows the inclusion of traditional, external instruments.

The instruments discussed lead to the following GMM moment conditions:

E(yi,t−s∆i,t) = 0, t ≥ 3, s ≥ 2

E(x0_1,i,s∆i,t) = 0, ∀ t, ∀ s > 0

E(x0_2,i,t−s∆i,t) = 0, t ≥ 2, s ≥ 1

E(x0_3,i,t−s∆i,t) = 0, t ≥ 3, s ≥ 2

These orthogonality conditions are combined into the instrument matrix Zi, with the following

moment condition: E(Z_iT∆i,t) = 0. The AB-estimator then estimates α and β by making use

of the GMM technique, minimizing a criterion function, which includes a weighing matrix, accu-rately described by Roodman (2006). This can be done either by the one-step estimator, hereafter AB1, with a weighting matrix independent of the estimated parameters, or by a two-step estimator, hereafter AB2, in which the weighting matrix consists of a consistent estimate of the co-variance matrix of the moment conditions. It is constructed using the initial consistent estimates of the pa-rameters(Arellano and Bond, 1991). It can be shown that the two-step estimator is more efficient, and has heteroskedasticly robust standard errors, contrary to the one-step estimator (Windmeijer,

2005). However Arellano and Bond (1991) show that there exists a small downward finite-sample bias, increasing in the number of instruments, opposed to the one-step estimator, which gives unbi-ased estimates. Multiple solutions have been proposed for this, such as the Windmeijer correction term, based on the estimation of the bias, which is also used in this estimation (Windmeijer, 2005). The number of instruments and thus moment conditions increases every time period and it is the highest closest to the final time period T. Even though more instruments yield higher efficiency and thus a higher power (Kiviet et al., 2014), there are also some considerations to keep in mind, when choosing the amount of instruments. Too many instruments could lead to a number of problems, as Roodman (2009) point out. The first issue he names is that too many instruments can lead to the over-fitting of endogenous variables, which causes a higher bias in the estimated coefficients. The second issue is that the optimal weighing matrix is estimated incorrectly. Finally, he points out that too many instruments leads to weakening of the Hansen test, discussed later on. To solve these problems, less instruments can be invoked (not using all lags) and the instrument matrix can be collapsed so that the amount of moment conditions is reduced1.

4.3 The Blundell-Bond estimator

Arellano and Bover (1995) and Blundell and Bond (1998) revealed a potential weakness in the AB estimators. They noticed that the lagged levels often are weak instruments for first differences, especially if the variables are close to a random walk. Blundell and Bond (1998), basing themselves on the work by Arellano and Bover (1995), developed a method which adds extra moment conditions. They use lagged differences as instruments for the equation in levels, in addition to the lagged levels as instruments for the first differences, as under the AB estimation. They show that this improves the efficiency of the estimator. Under the AB estimation, no assumptions on the fixed effects are made, since they are simply differenced out. The Blundell and Bond estimator (hereafter: BB) relies on the extra crucial assumptions, that the country fixed effects and the variables in levels have a constant correlation over time:

E(∆gi,t|µi) = 0, E(∆xi,t|µi) = 0, (4)

1

Collapsing the instrument set is done by combining instruments through addition into smaller sets. An example of a collapsed and an uncollapsed instrument set can be found in Appendix IV.

If these assumptions hold, they lead to the following extra moment conditions:

E[∆gi,t−1(µi+ i,t)] = 0, t ≥ 2

E[∆xi,t−1(µi+ i,t)] = 0, t ≥ 2 for endogenous xit

E[∆xi,t(µi+ i,t)] = 0, t ≥ 1 for predetermined xit

(5)

If this is the case, these (lagged) differenced instruments can be used in the level equations as instruments. If the extra assumption holds, the BB estimation should be preferred, since they perform better than the AB estimations (Blundell and Bond, 1998).

Now that both estimation techniques have been outlined, the next section clarifies how they are applied in this thesis.

4.4 Application of the Arellano-Bond and Blundell-Bond estimators

After conducting a few simpler regressions, the AB2 method is applied. Model (1) is first-differenced, and ∆gi,t−1is instrumented by second and higher-order lags of gi,t. Regarding the other explanatory

variables of model (1); first all of them are treated as endogenous, and thus second and higher-order lags of these regressors are used. If a lagged realization is found to be significant and included, it is also assumed endogenous, and thus third and higher order lags of that variable are used as instruments. Later on, as a robustness check, the second lag is also included as an instrument. Initially it is not included to be sure that no endogeneity arises, however if the contemporaneous realization is endogenous, according to theory the lagged variable is weakly exogenous, with respect to the contemporaneous error-term. Similarly, the lagged value of a predetermined variable is theoretically exogenous. So, these lags could be included and provides a more powerful instrument. This is done later on, in section 5.5.1.

The furthest lag that is used is the 4th lag, and the instrument matrix may be collapsed, to limit the issues that can be caused by a high number of instruments. The reason for the 4th lag is quite arbitrary, mainly to limit the number of instruments. Since in first instance 4 year time periods are considered, the relevance of a 5th lag period can be questioned.

As assuming endogeneity for an explanatory variable that is in fact predetermined leads to weaker results, the explanatory variables are classified as endogenous or predetermined. This is done by making use of the incremental Hansen test, which tests whether a specific group of instruments is valid (Hansen, 1982). The estimation starts of with assuming endogeneity (the weakest assumption) of all regressors. Then, the instrument set (containing second and higher order lags) with the highest p-value is considered predetermined, by including the first lag of that regressor to the instrument set. The null hypothesis states that the instruments are valid. However, since a high number of instruments weakens the incremental Hansen test, a higher rejection frequency is chosen. If the test is not rejected at a 30% level, the regressors are assumed predetermined. This is subsequently

done for all regressors, the order being based on the p-value. After classifying all the regressors, these steps are repeated for the regressors that have been found to be predetermined, in order to assess if they are in fact, exogenous. The regressors that are found to be predetermined can then be instrumented by first or higher order lags. The exogenous regressors, as well as the time dummies which are assumed exogenous, are instrumented by their own contemporaneous realization.

Both one and two-step GMM estimations are performed, respectively AB1 and AB2. With the one-step estimations, the standard errors are made heteroskedasticly robust, and for the two-step estimations the Windmeijer (2005) correction is applied to correct for the possible downward bias. One could argue that the estimations using AB1 are redundant, since, as Roodman (2009) argued, the AB2 with Windmeijer correction is preferred of the AB1 with robust standard errors. However, Kiviet, Pleus and Poldermans (2017) show that this is not the case anymore after the instrument set is collapsed. Since it is probable that the error terms are heteroskedastic, as they usually are in macro-economic research, the two-step method is the standard in this thesis, and AB1 serves as a robustness check. If the p-value for the incremental Hansen test concerning the extra BB conditions is not invalid, a BB2 estimation is then used to proceed.

4.5 Specification tests

The AB-estimators are only consistent if the error-terms are serially uncorrelated (Arellano and Bond, 1991). Since the model is first differenced, there is automatically first order autocorrelation (AR(1)). However, the consistency of the AB-estimator hinges heavily on the assumption that there is no higher order autocorrelation. Therefore the Arellano Bond test for serial autocorrelation of order 2 and higher is performed, to check if this is indeed the case. If the AR(m) test, with m larger or equal to two is rejected, this might imply omitted variable bias.

The Sargan-Hansen test is performed to test the validity of the instruments (Sargan, 1958; Hansen, 1982). Since the null-hypothesis states that the instruments are valid, a high p-value is desired. As Roodman (2009) points out, the test is weakened by the inclusion of too many instruments, and one should be careful when interpreting the coefficients. The Hansen test is also used in an incremental way, as described above, to test the subsets of instruments, in order to classify them. It is also used to assess whether the extra assumptions needed for the BB estimation are valid.

4.6 Long-term Effects

The model described seeks to explain the possible effect of (twice) lagged and contemporaneous realizations of female empowerment on economic growth. However, there can also be a long-term effect, where an increase in one of the female empowerment variables leads to a change in economic

growth many years or decades from now. Considering the simplified version of model (1):

gi,t = α + γgi,t−1+ βF Ei,t+ i,t (6)

The Female Empowerment measures directly affect economic growth, but they also have an effect through lagged growth, of size γβ. The effects on growth z periods from now (gi,t+z) is γzβ.

Sum-ming up these effects, with z going to infinity, leads to the cumulative effect of female empowerment : _1−γβ . This cumulative effect can be interpreted as the total economic growth caused by a singular change in F Ei,t. By subtracting gi,t−1 from both sides of equation (5), it can be re-written as

follows:

∆gi,t = α − (1 − γ)gi,t−1+ βF Ei,t+ i,t (7)

In case of a unit-root; γ = 1, there is no convergence in growth. On the other hand, if γ = 0, there is no dynamic adjustment process. If 0 < γ < 1, then limγ→∞γzβ = 0, in other words, a one-time

shock will slowly deteriorate over time.

These effects can be estimated, along with 95% confidence interval, using nlcom in Stata, which makes use of the delta method.

This can also be applied when more lagged variables are included in the model, for instance: gi,t = α + γ1gi,t−1+ γ2gi,t−2+ β1F Ei,t+ β2F Ei,t−1+ i,t (8)

The cumulative effect then becomes : β1+β2

1−γ1−γ2. Depending on the number of lags found to be significant, the long term effects are calculated, along with their 95 % confidence interval.

5 Results

This section discusses the results of the Arellano-Bond and Blundell-Bond estimators. After con-structing the model with the AB2 estimator a few simpler regressions are conducted as a first step. Then the explanatory variables are classified as endogenous, predetermined or exogenous and the estimation results of the AB2 estimator are presented. As robustness check, the AB1 estimator is also included in that section. After having tested the additional conditions necessary for the BB2 regression, and finding them valid, the estimation results of BB2 are also presented. Sub-sequently, two extensions of the model concerning female empowerment are presented. Hereafter, some additional robustness checks are discussed, followed by the discussion of the long-term effects. 5.1 Constructing the model

The model is constructed with the AB2 estimator, starting by estimating model (1), after taking the first difference. So, at first all variables are included as well as their first and second lag. The

high number of variables and instruments led to highly insignificant results. The model was then refined by omitting variables with the highest p-value. Since this is can be quite arbitrary, especially if all p-values are quite high and close to each other, this has been repeated many times with a different order of elimination. Moreover, as a robustness check, the model has also been estimated the other way around, starting with only economic growth, its lag and the contemporaneous female empowerment variables, adding control variables and their lags one by one. This too has been done in many different ways, since the order of adding variables affects the model. None of the twice lagged variables proved to have a significant effect, so they all have been left out of the model. Also, despite the none significant first lags being omitted, they still have an indirect effect through the lagged dependent variable. By eliminating highly insignificant variables through trial and error, and with the robustness checks discussed, the following model is constructed:

∆gi,t= γ1∆gi,t−1+ β1∆F Ei,t+ θ1∆CVi,t+ θ2∆CVi,t−1+ ∆i,t (9)

None of the Female Empowerment variables had a significant lagged effect. CVi,t contains all

variables except government expenditures. CVi,t−1 contains population growth rate, schooling and

investment. Government expenditures have been left out altogether from the model, since in no different specification did it seem to have a significant effect on economic growth. This might be due to the sample selected, as Devarajan, Swarop and Zou (1996) find that developing-country governments have been misallocating public expenditures, which therefore do not contribute to the economic growth of that country. Rule of law also proved insignificant in all estimations except OLS, but omitting it from the model led to insignificant results of inflation, schooling, investment and trade, so it is still included. Year dummies were initially included, however they were not significant (neither individually nor jointly (p-value: 0.14)), and are therefore not included in the model. Before proceeding with the classification of the regressors and the results of the AB and BB estimations, some preliminary estimations are discussed.

5.2 Preliminary findings using OLS, Fixed effects and Anderson-Hsiao estimators Before diving into more sophisticated estimation methods, an OLS-regression is performed, ignoring the dynamic panel nature of the model, but using panel-clustered standard errors. Even though in this regression potential correlation between the error terms of countries over time have been accounted for by the clustered standard errors, the country-level unobserved heterogeneity has not been accounted for, leading to endogeneity. In order to do this, a within transformation is applied to take this into account, and the model is re-estimated with fixed effects (FE). Furthermore, the FE estimator does not take into account the endogeneity that has been introduced in the model, as explained in section 4.1. Before proceeding to the more efficient AB and BB estimators in the next section, the Anderson-Hsiao estimator (AH) (see section 4.2) is applied, instrumenting the lagged

dependent variable with its twice lagged level, while the other explanatory variables are considered exogenous and used as instruments for themselves. The results of those three estimations can be found in table (1). Before briefly discussing the results of these estimations, it is to be reiterated that all these estimation techniques have some clear shortcomings. The OLS-estimator is biased since it is ignoring dynamic effects, as well as country fixed effects, additionally to erroneously assuming exogeneity. The FE estimation suffers from Nickell bias (Nickell, 1981), which might be severe since there are relatively few time periods, and do not take into account endogeneity. The AH estimator should lead to more consistent estimators, however it is very imprecisely estimated. In general the estimations under AH have a higher significance level than in the other two estimations. It is to be noted that the AH estimator has less observations than the other two, losing degrees of freedom. This is because the first observation for all countries is lost when applying the first difference transformation. Even more observations would be lost, if not only the second lag , but also the third lag of the dependent variable would be used as an instrument.

Except in the FE estimation, lagged economic growth has a positive coefficient, and initial GDP has a negative coefficient. According to Iqbal and Daly (2014), this supports the conditional convergence theory.

The female empowerment variables all have the expected sign, though only labour participation in the AH regression is significant. This result might be due to multicollinearity of the variables, as mentioned in section 2.2. Also inflation is highly significant in all estimations and has the predicted negative sign. Trade and rule of law are positive in the FE and AH estimations, as expected. However, some results are in opposition to theory and previous research. Population growth rate has a positive coefficient, and Life expectancy has a negative one in all three estimations, in contrast to theory and previous research.

Moreover, schooling and investment seem to have a lagged effect with the opposite sign of their contemporaneous effect in all three estimations.

Since all three estimations techniques have their shortcomings and the model has not yet been improved by accounting for possible endogeneity of the other explanatory variables, as is done in the next section, only a brief overview of the results has been given in this section. A more in-depth analysis and possible explanations are provided in the next section, once the results of the AB and BB estimations have been presented. Also, some comparisons are made between those estimations and the ones performed in this subsection.

Table 1: Dependent Variable: Economic Growth

Variable Lag OLS FE AH

Economic growth 1 0.07 -0.11** 0.06 (0.05) (0.05) (0.07) Fertility 0 -0.8 -0.11 -0.7 (0.34) (0.84) (1.4) Labor participation 1 0.07 0.06 0.16** (0.01) (0.06) (0.07) Female Education 0 0.07 0.08 0.2 (0.35) (1.14) (1.4) Initial GDP 0 -0.0002** -0.0006 -0.002*** (0.0001) (0.0004) (0.0007) Population growth rate 0 1.15** 0.5 1.64

(0.47) (0.94) (1.41) Population growth rate 1 -0.3 0.54 0.44

(0.43) (0.35) (0.37) Log(Life Expectancy) 0 -2.9 -6.06 -9.14 (2.44) (7.5) (11.41) Inflation 0 -0.008*** -0.01*** -0.006*** (0.002) (0.002) (0.002) Schooling 0 -1.3 0.31* -0.29 (0.77) (0.71) (0.86) Schooling 1 1.3* 2.4*** 2.35*** (0.75) (0.75) (0.80) Investment 0 0.24*** 0.17*** 0.18*** (0.04) (0.05) (0.06) Investment 1 -0.2*** -0.24*** -0.20*** (0.05) (0.05) (0.06) Trade 0 -0.004 0.004 0.02* (0.004) (0.013) (0.013) Rule of Law 0 -0.14 0.2 0.55 (0.38) (1.17) (1.98) Constant 17.1 12.8 -(11.1) (30) -Time-dummies used No No No Number of observations 312 312 230 ***_{p < 0.01,}**_{p < 0.05,}*_{p < 0.1}

5.3 Arellano-Bond and Blundell-Bond estimations 5.3.1 Classification of the Regressors

The explanatory variables are classified with the incremental Hansen test, which tests whether each subset of instruments is valid, as discussed previously. Table 2 reports the p-values and nature of each independent variable, classified as discussed in section 4.4. Since too many instruments can greatly weaken the Hansen statistic, all instrument sets are collapsed, and up until the third lag is used, since using the 4th lagged led to p-values of (close to) 1. All the while, the auto-correlation test and Sargan test for over-identifying restrictions were satisfied. It is to be noted that there is no overall consensus about the manner of classification. The rejection region of 30% can be questioned, as well as the order of testing. As a robustness check, the instruments have also been classified while using the full instrument sets and once collapsing them one by one, after classifying them. Not collapsing the instrument set led to p-values close to one. This was even more so after the first lag had been added in a few instrument sets, after classifying them as predetermined (and thus increasing the total number of instruments), and have been deemed unsuitable to draw conclusions from. Collapsing the instrument sets one at a time, after classifying them, led to the same conclusions about the nature of the instruments, though the order of classification, and the p-values differed.

Table 2: Results incremental Hansen test to determine nature of variables

Variable p-value stage 1 p-value stage 2 Nature of variable

Inflation 0.922 0.495 exogenous Labor Participation 0.524 0.592 exogenous Log Life Expectancy 0.802 0.740 exogenous Female Education 0.837 0.764 exogenous

Trade 0.655 0.584 exogenous

Investment 0.057 - endogenous

Rule of Law 0.529 0.763 exogenous Population growth 0.452 0.170 predetermined Fertility 0.761 0.091 predetermined Log average GDP 0.081 - endogenous Government Expenditures 0.031 - endogenous

Schooling 0.049 - endogenous

The variables are presented in order of the first classification

Table 2 shows that in the first classification round, which classified regressors as endogenous or predetermined, investment, average GDP, government expenditures and schooling were found to be endogenous. The second round of classification concerned itself only with the regressors which had been found to be predetermined, to check if they might be exogenous. As can be seen in table 2, the

majority of the regressors can be classified as exogenous. This classification is used to proceed with the estimations. The endogenous variables are instrumented with their second and third lags, the predetermined variables are instrument by their first, second and third lag. However, the variables whose first lag is included in the model are also instrumented by lag 2 and 3. As discussed in section 4.4, in section 5.5 a robustness check is conducted, instrumenting those predetermined variables by their first lag as well. The exogenous variables are treated as ordinary instruments. Likewise, geography dummies (African, Asian and Latin American countries) are also included as ordinary instruments, although they are not included in equation (4), since they would simply be differenced out. They are included as instruments to somewhat control for regional differences. Both the year and geography dummies are assumed exogenous, and this is not contradicted by the incremental Hansen test (p-value 0.612). They are instrumented by their own contemporaneous realization. 5.3.2 Results AB and BB2 estimations

After fitting the best model through trial and error with AB2, it has also been estimated by AB1, for reasons explained earlier. The incremental Hansen test for the extra conditions for the BB estimator was not rejected (p-value: 0.720), therefore the model found under AB2 has also been estimated with the BB2 estimator. To limit the instrument count, especially under the BB2 estimations, the instrument sets have been collapsed. As explained previously, this is a trade-off, since collapsing the instrument set does embody the same expectation, it does contain less information and is less efficient, because less moment conditions are generated (Roodman, 2009). However, this is done to avoid over-fitting of the variables, and to decrease the bias, as explained by Bun and Sarafidis (2013). As a robustness check, it is also estimated without collapsing the instrument set, which is discussed later on. The results of the three estimations can be found in table 3.

Even though year dummies have been left out of the estimation, they are still included as instruments. As can be seen from the table, the specification tests all lead to satisfactory results. It can be seen that the AB2 and AB1 estimation results are very similar, in sign and order of magnitude. Since some instrument sets are collapsed, it is indeed not evident that AB2 would outperform AB1, as Roodman (2009) pointed out. It can be seen that the standard errors of the AB2 estimation are (relatively) larger than the ones of the AB1 estimation, in line with what Windmeijer (2005) states, that the standard error of AB2 is always at least as larger as the one under AB1.

Table 3: Dependent Variable: Economic Growth

Variable Lag AB1 AB2 BB2

Economic growth 1 -0.07 -0.12* 0.06 (0.06) (0.07) (0.06) Fertility 0 0.71 1.61 -0.48 (1.72) (1.8) (0.47) Labor participation 0 0.14 0.05 -0.007 (0.09) (0.09) (0.02) Female Education 0 1.23 1.44 0.45 (1.55) (1.56) (0.81) Initial GDP 0 -0.0006 -0.001* -0.0002 (0.0009) (0.0008) (0.0002) Population growth rate 0 -0.59 -1.49 0.93

(1.09) (1.38) (0.99) Population growth rate 1 0.59 0.79*** -0.35

(0.38) (0.28) (0.56) Log(Life Expectancy) 0 2.71 -3.49 1.94* (12.6) (14.2) (1.011) Inflation 0 -0.01*** -0.01**** -0.009*** (0.002) (0.003) (0.002) Schooling 0 -0.34 -1.39 -4.8* (1.88) (2.09) (2.6) Schooling 1 2.34 4.72*** 4.8* (1.73) (1.7) (2.5) Investment 0 -0.03 0.02 0.21** (0.106) (0.1) (0.08) Investment 1 -0.38*** -0.35*** -0.31*** (0.09) (0.07) (0.08) Trade 0 0.03** 0.03** 0.02* (0.017) (0.01) (0.009) Rule of Law 0 0.28 -0.66 -0.3 (2.21) (2.47) (0.6) Time-dummies used No No No Number of observations 230 230 312 Number of Instruments 42 42 44 p-value AR(1) 0.027 0.032 0.002 p-value AR(2) 0.606 0.415 0.848 Sargan/Hansen p-value 0.463 0.463 0.332 ***_{p < 0.01,}**_{p < 0.05,}*_{p < 0.1}