Can Export Diversification Save sub-Saharan Africa from Extreme Weather?: An instrumental variable approach

DEPARTMENT OF ECONOMICS Uppsala University

Bachelor’s thesis

Authors: Joanna Arnarsdottir and Kristina Hansson Supervisor: Mohammad Sepahvand

Autumn 2020

Can Export Diversification Save sub-Saharan

Africa from Extreme Weather?

Abstract

Growth and development in the context of climate change and environmental challenges are issues of increasing importance in the economic debate. With higher levels of greenhouse gases in the atmosphere, droughts and other forms of extreme weather are expected to increase in frequency. Some of the worst affected are people living in sub-Saharan African countries. However, there are literature showing that countries who diversify their production becomes more resilient against negative shocks. This study aims to determine the relationship between precipitation anomalies and GDP per capita growth under different levels of concentration of the export portfolio, in order to understand what kind of diversification reduces economic risks connected to precipitation. Precipitation anomalies, such as abnormally heavy rainfall or droughts, is seen as a good measurement for climate change, and can thus be treated as economic shocks. We are using data on export product shares and monthly precipitation to determine whether the level of sectoral diversification in exports affects the influence precipi-tation anomalies have on GDP per capita growth. The effects are estimated using a two-stage least squares model, only targeting countries in SSA for our estimations. The results show that positive weather anomalies correlate with lower levels of GDP per capita growth. But the same negative trend cannot be seen for negative precipitation anomalies. The results also show that the level of diversification within exports does not have any significant effect on the influence that precipitation anomalies have on GDP per capita growth.

Key words: Diversification, precipitation anomalies, conditional convergence, economic de-velopment, sub-Saharan Africa (SSA), 2SLS, IV estimations, climate change

Contents

4 Data and methodology 9 4.1 Methodology . . . 9

4.1.1 The HHI and WASP indices . . . 9

4.1.2 Two-stage least squares . . . 12

4.2 Data . . . 14

4.2.1 Descriptive statistics . . . 15

4.2.2 Data Limitations . . . 18

4.3 Empirical specification . . . 19

5 Empirical results 21 5.1 Rainfall, specialisation and GDP per capita growth . . . 21

5.2 Weather anomalies under different levels of diversification . . . 24

5.3 Robustness checks . . . 28

5.3.1 Estimations using five-year means . . . 28

5.3.2 Estimations for different sub-samples . . . 29

6 Discussion 30 6.1 The economic significance of diversification and rain anomalies . . . 30

6.2 Possible explanations for the relationship between diversification and rain anomalies . . . . 32

7 Conclusion 33 References 35 Appendix 38 A Countries . . . 38 B Distribution of observations . . . 39 C Correlation of variables . . . 40

D Model with squared HHI . . . 41

E Estimations using five-year means . . . 42

Acronyms

2SLS Two-stage least squares.

3SLS Three-stage least suqares.

GDP Gross Domestic Product.

HHI Herfindahl-Hirschman Index.

ILO International Labour Office.

PDSI Palmer Drought Severity Index.

SSA sub-Saharan Africa.

UN United Nations.

UNCTAD United Nations Conference on Trade and Development.

UNIDO United Nations Industrial Development Organization.

V-Dem Varieties of Democracy Institute.

WASP Weighted Anomaly Standardized Precipitation.

1

Introduction

The issues of climate change have become increasingly apparent during recent times, causing growth in relation to climate factors to be one of the most pressing issues in development economics. Amongst researchers, the main view is that low-income countries are the most vulnerable to climate change, partly because people living in poverty are more likely to depend on climate sensitive sectors, such as agriculture (Koren and Tenreyro 2007; Davis et al. 2017). Researches also agree that the occurrence of extreme weather, such as droughts, have increased as a consequence of higher levels of greenhouse gases in the atmosphere (Oliver 2005). Previous empirical studies point out climate change and a decline of precipitation (rainfall) in recent years as part of the limitation on economic development in sub-Saharan Africa (SSA) (Brown et al. 2011; Barrios et al. 2010). Because of these findings and due to the large portion of low-income countries in SSA, it is important for policymakers to work on solutions on how to survive climate change. Studies have shown that it might be possible to reduce economic risks through sectoral diversification (Koren and Tenreyro 2007; Imbs and Wacziarg 2003). This generates the question of whether it is possible to create a resilient economy that is not affected by climate variability to the same extent as today. This important issue is at the core of our study, and therefor our research question will be How do differences in export diversification change the effect of precipitation anomalies on GDP per capita growth in sub-Saharan Africa?.

The purpose of this study is to examine how diversification in exports can create more resilient economies to one aspect of climate variability, precipitation (rainfall) anomalies. By spreading the production to different sectors one can minimise the effect of a negative economic shock. Thus, this study will attempt to examine how differences in a country’s composition of export sectors can successfully diversify the production in order to better handle negative economic shocks that follow from rain anomalies. Although precipitation anomalies are not always a climate shocks in themselves, they can be considered disturbances in the economic activities, since they are deviations from the regular weather pattern. By using export portfolio as a proxy for production in a country, we examine how GDP per capita growth is affected by precipitation anomalies, dependent on the country’s level of diversification. Our null hypothesis is that there exists no difference in how well countries absorb precipitation anomalies depending on how diversified their export portfolio is, and this study will try to give evidence of the opposite. Precipitation anomalies are measured using a Weighted Anomaly Standardized Precipitation (WASP) index to calculate deviations from average monthly precipitation. Earlier literature have examined the relationship between rainfall anomalies and GDP growth (Brown et al. 2011; Barrios et al. 2010), and diversification and GDP growth (Koren and Tenreyro 2007; Imbs and Wacziarg 2003). However, to the best of our knowledge, they have not examined how these two interrelate, something that this study aims to conduct.

Previous research analysing the relationship between diversification and economic development have so far established that diversification in production and export usually is favourable for economic development. Diversification is causing less volatile GDP growth, however at the cost of lower growth rates (Imbs and Wacziarg 2003; Cadot et al. 2011). The majority of these studies use GDP per capita growth as their outcome measurement, which is why this study uses the same (Brown et al. 2011; Barrios et al. 2010; Koren and Tenreyro 2007). It is a useful measurement of how the economic activity changes and if the welfare in a country is increasing or decreasing. Because of the trade-off between a stable, less volatile economic growth, and a higher GDP growth that is perhaps associated with higher risk (Loayza et al. 2017), it is necessary to investigate successful strategies for diversification. Without knowing how to reduce both the country-specific and sector-specific risks, the problem of successfully reducing risk becomes difficult to solve. Our results provide evidence consistent with the existing literature and theory, both concerning precipitation anomalies and diversification. However, the relationship between these two subjects and GDP per capita growth proves to be weak and it is yet to find what kind of diversification is successful to reduce risks related to climate change.

This study will use a two-stage least squares model on panel data to estimate how diversification change the effects that droughts have on GDP per capita growth in SSA. The main reason for using this method is to correct for endogeneity in our estimations. The data set is composed from different sources providing, for instance, data on weather, exports and GDP per capita. The period covered in this study is between 1992 and 2016, which are the years that data of exports sectors and precipitation levels in SSA can be retrieved. By using literature discussing conditional convergence, we will include a number of control variables in our analysis, in order to isolate the effect of diversification on GDP per capita growth. To measure the degree of diversification the Herfindahl-Hirschman Index (HHI) will be used, a widely used index for both scholars and policymakers (Koren and Tenreyro 2007; Imbs and Wacziarg 2003; Brezina et al. 2014). The index will be used to group countries in SSA and see how they react to precipitation anomalies dependent on the level of diversification.

This study is organised in the following way: section 2 will present brief background knowledge of the exports and climate in SSA. Section 3 will present relevant theory that explains GDP growth and productivity as well as earlier studies on the subject. Section 4 discusses the method used in this study and presents an overview of the data and the empirical specification. Section 5 presents all the results from our regressions with a subsection presenting alternative regressions for robustness. A discussion of the results is found in section 6 and concluding remarks is found in the last section, section 7, followed by references and appendices.

2

Background

SSA is a term used by the United Nations (UN) for the African countries situated south of Sahara (UNSD 2020). It is a large group of countries spanning over a vast geographical area, why grouping them is a generalisation. There are considerable differences between the countries, but they also do share many characteristics and face many of the same problems with regards to growth, climate change and infrastructure. Most of the countries included in the group have traditionally been, and are still being, considered developing countries. (World Bank Group 2018)

A large portion of the rural population in these countries relies on agriculture as the main source of income, 8 out of 10 households are to some extent dependent on agricultural activity. There is also a significant share of the rural population that is engaged in non-farm activities, and studies have shown that the size of that group, and the share of income from non-agricultural activities, correlates positively with the level of GDP. While the share of income from non-farm activities is increasing, agriculture is still the main source of income and the sector dominating the rural economy in SSA. (Davis et al. 2017)

Although agriculture is a large source of income in SSA, the largest export products are different kinds of petroleum oils, minerals and metals. Many of the countries in Sub-Saharan Africa have natural resources in the form of minerals, making mining a large industry. Fuels is the largest export product group (41.19 percent), while stone and glass and metals only account for 13.45 and 8.84 percent respectively. Between 1992 to 2016, the total export has increased from 23 billion USD to 206 billion USD, an increase of nearly 800 percent. However the total number of different products have stayed roughly the same during this period, around 4550 product. (WITS 2020b)

According to Barrios et al. (2010) there has been a declining trend of rainfall since the 1960s in SSA. A region is affected by drought when there is a deficiency of precipitation for a longer period of time, making the water reserves insufficient to meet the human activities (Oliver 2005). In this study, like in Brown et al. (2011), a longer period of time is set to one year, which facilitates the estimations because our data set is harmonised on a one year level. However, since precipitation data is only used on country-level, it might be incorrect to assume that negative deviations of rainfall by definition indicate that a country suffers from drought. Droughts are seen as one of the greatest natural hazards, partly because the difficulties in forecasting them and their severity. However, the impacts can be reduced through risk-based management that increases the resilience. The Food and Agriculture Organization of the United Nations (FAO) have supported countries in SSA to develop proactive drought management in order to manage the challenges of drought (Food and Agriculture Organization of the United Nations 2020).

3

Theory

This section presents theories that explain determinants of GDP per capita growth and diversifi-cation in trade. Subsection 3.1 gives the fundamental framework that models the economies and their behaviour, and clearly sets the assumptions made in these areas to understand GDP growth. In subsection 3.2, recent research that is related closely to the area of this study is presented in detail.

3.1

Theoretical framework

The Solow growth model of long run economic growth, named after its developer Robert Solow, is often used to explain growth rates in the long run. It includes productivity, capital accumulation, savings rate and the population growth rate to calculate output per capita. The Solow model implies that the capital-labour ratio will develop towards a steady state, an equilibrium where there is balanced growth in both capital and labour. Depending on the initial value of the ratio, the movement towards the steady state will be of different speed and character, but eventually countries will converge to the same steady state level of capital-labour ratio. The rate of GDP per capita growth will be higher in economies with lower levels of capital per worker, and lower in economies with high levels of capital per worker. The theory assumes low capital per worker in low-income countries, causing them to "catch-up" to the richer ones which have lower levels of GDP per capita growth. If all countries have the same production function, savings level and population growth, they will in the long run have the same steady-state equilibrium. They will however move towards that level at different speed since the initial capital-labour ratio will differ between countries. (Solow 1956)

Robert J. Barro (1998) continues the work of Solow and many others in finding the key components of GDP growth. One of the main takeaways of the Solow model is that low-income countries with lower levels of capital per worker will in time converge with richer countries that already have higher levels of capital per worker, and thus lower growth rates. This implies that countries are intrinsically the same, an assumption which Barro investigates further, exploring other components affecting GDP per capita growth. His findings show that higher life expectancy, low inflation, higher education, better applications of rule of law and terms of trade are correlated with higher GDP growth. Initial level of GDP is negatively correlated with GDP growth, which aligns with the Solow model arguing that higher capital per worker increases the GDP per capita, but with diminishing returns. We use the theories of Solow (1956) and Barro (1998) to add relevant control variables to our estimations. In order to measure the true effect of rain anomalies and diversification on GDP growth, it is vital to hold previously established determinants of growth constant.

A central theory explaining the behaviour of different economies in trade is the theory of comparative and absolute advantage by David Ricardo (Krugman 1998). The theory argues that countries should specialise in the product they have the lowest opportunity cost in producing, thus producing more of the good in which the country has a comparative advantage. In turn, a trading partner specialises in other products in which they have comparative advantages in and the two parties trade their surplus goods. The total amount of goods produced increases and so does the welfare, without increasing costs. If the autarky prices are higher than the world market prices, one country can afford to buy more units of an imported product than it can afford to domestically produce a product. This implies that countries will continue to trade until the relative price of the two commodities are the same in both countries. The Ricardian comparative advantage model is simple, and consists of two economies and two goods. However, today’s globalisation and trade patterns leads to countries facing multiple partners, but by treating one country as the home-country and the rest of the world as the trade partner it is still possible to use this model. The conclusion that can be drawn from this model is that countries should specialise to achieve higher growth and incomes, which should lead to countries in the world striving to be more specialised as they engage more in trade.(ibid.)

3.2

Literature

There are multiple articles discussing the topic of diversification in relation to economic develop-ment. Koren and Tenreyro (2007) treat the relationship between output volatility and economic development using a method to assess sectoral diversification in a country. They are using industry-level panel data from United Nations Industrial Development Organization (UNIDO), depicting both developed and developing countries from 1963 to 1998. Their findings suggest that the output tends to be more volatile when a country specialises in highly volatile sectors, has a high sectoral concentration and/or specialises in sectors that are highly affected by country-specific economic shocks. Their study provides some important takeaways for our study. More developed countries tend to move towards less risky sectors, and therefore have a less volatile output than less developed countries. Sectoral concentration and development tend to have a U-shaped relationship. At early stages of development, sectoral concentration declines and sectors become more diverse. In the later stages of development, the relationship flattens out, and tends to slowly reverse for high levels of development. The study by Koren and Tenreyro (ibid.) is based on a study by Imbs and Wacziarg (2003), which come to similar conclusions of a quadratic relationship between sectoral concentra-tion and per capita income. They use different measures of concentraconcentra-tion in an economy and data from UNIDO and International Labour Office (ILO) to the determine the relationship with income levels, and discuss this in the context of earlier findings which often propose a linear relationship. This is similar to what our study does, but we add other variables in order to determine if the level of diversification has any impact on the relationship between precipitation anomalies and growth

rates.

The findings by Koren and Tenreyro (2007) are supported by a more recent study, Koren and Tenreyro (2013). They write on technological diversification and conclude that technological di-versification can offer an explanation to the negative relationship between output volatility and development. Their endogenous growth model builds around the idea that firms with a large vari-ety of inputs are more resilient to shocks that affect the productivity of individual inputs. Firms in more developed economies tend to use technologies with a higher variety of inputs and therefore have a higher possibility to change the ratio between the different inputs to withstand the effects of shocks. The studies by Koren and Tenreyro (2007, 2013) are used extensively in this study, because we intend to examine diversification in relationship to GDP per capita growth. They provide clear theories and results on the relationship between diversification and development, and our study tries to put this in the context of climate variability and precipitation. The main difference between our study and their studies is that they have explored diversification within production, whereas we use export data.

Brown et al. (2011) further examines the relationship between droughts and GDP growth in SSA through cross-country panel regressions with fixed effects. They compute an index for hydroclimatic variability called Weighted Anomaly Standardized Precipitation (WASP). It calculates the deviation in monthly rainfall from long-term mean precipitation and sums it in a weighted annual average of monthly deviations. By creating thresholds for values that indicate drought for different areas of a country, they calculate the area that is affected by drought relative to the total are of a country. They note that observed negative values of the WASP index are well correlated with other indices for drought (for example Palmer Drought Severity Index), meaning that WASP is a good indicator for drought. Positive values of the WASP index are not however representative for floods, since floods operate on shorter timescales than the monthly averages. Positive values of WASP only indicate levels of precipitation higher than average, but since the index provides no information on the distribution of precipitation over time, one cannot assume that this indicates flooding. They come to the conclusion that persistent negative anomalies, which indicate droughts, have a significant negative effect on GDP per capita growth. Brown et al. (ibid.) argue that their results indicate that the major concern for countries in SSA is not the rising temperature, it is rather the precipitation anomalies that have large impacts on the economy. This study is used extensively as a basis for our study, where we use the same approach to measure droughts, and combine this with the study by Koren and Tenreyro (2007).

Barrios et al. (2010) also look for trends in rainfall and GDP per capita growth, and conclude that rainfall patterns have a significant effect on GDP per capita growth in countries in SSA, compared to

other regions. In contrast to Brown et al. (2011), they use weighted monthly average precipitation on country-level in their estimations, which are estimated both using OLS and fixed-effect with interaction variables. We will follow this approach, allowing us to capture precipitation anomalies although we do not have access to more geographically detailed data. Barrios et al. (2010) conclude that the shortages of rainfall in SSA have caused GDP per capita to be 15-40 percent lower compared to other regions. They explain that this might be due to African countries being more reliant on the agricultural sector. Another reason to why droughts have a negative effect according to Barrios et al. (ibid.) is that countries in SSA relies on water for energy production. Hydropower accounts for 47 percent of total energy production in SSA , in comparison with the rest of the world where 34 percent (ibid.) of electricity comes from hydropower. Our study relates strongly both to this article and the study by Brown et al. (2011), primarily because we are examining rain anomalies as external determinants to GDP per capita growth. However, there is a gap in the literature, where the use of diversification as a means to secure growth that can withstand economic shocks that might be caused by precipitation anomalies has not been investigated. We aim to fill this gap by investigating if there is a way to affect the impact of rain anomalies on GDP per capita growth.

A study that treats the subject of diversification as an explanation, or a means, for growth is Cadot et al. (2011). They examine determinants for success for firms entering the export market. By using export data from four African countries they aim to define determinants of success when firms enter a new export market. Their findings suggest that exporters are more likely to survive beyond the first year after entering a new export market, if they sell the same product as multiple other firms from the same country. Thus, there is a networking, or synergy effect that increases the survival rate for products in a new export market. This effect could be the result of information spillovers or externalities which facilitates adaptation to changes and access to finance. With a multitude of different firms from the same country, who all export the same product to the same country, they can share information about policy changes or regulations in the destination market. Cadot et al. (ibid.) conclude that firms that enter an export market on the intensive margin are more likely to survive throughout the first year the global market than emerging exporters on the extensive margin. This gives incentives to specialise in a selected amount of sectors and products rather than diversifying (ibid.). This is linked to our study, as we are also using export data, to conclude if diversification in exports can reduce effects of precipitation anomalies. If there is evidence of export diversification reducing risks following precipitation anomalies, it would imply that diversification is beneficial despite the findings of Cadot et al. (ibid.).

The above mentioned studies show that countries experience higher productivity and efficiency at higher degrees of specialisation. This could imply that the countries in SSA with higher growth rates also have a higher degree of specialisation, which is linked to Ricardo’s theories of comparative

advantages. However, as Koren and Tenreyro (2007) point out, the volatility of GDP per capita growth decreases when a country becomes more diversified. Because we assume rain anomalies as aggregated economic shocks to GDP per capita growth, a country might be more affected if the export portfolio is less diversified, since specialised countries have higher volatility in GDP per capita growth. There is, in some way, a conflict between the aim for a resilient economy that is diversified and can withstand shocks, and an economy that experiences high growth rates and successful export markets (Loayza et al. 2017). Considering this, our study aims to add to the literature of diversification by combining previously established knowledge about GDP per capita growth, diversification and the effect of rain anomalies. To the best of our knowledge, previous studies have not yet explored the effect of diversification on GDP per capita growth through this approach.

4

Data and methodology

This chapter describes the data and methodology of this study. The methods and its advantages and disadvantages are presented in subsection 4.1. An overview of the data and its sources that the study uses is presented in the subsection 4.2. In subsection 4.3 the empirical models used for the regressions are described in detail.

4.1

Methodology

This study tries to determine the relationship between GDP growth and precipitation under different levels of concentration in exports with an instrumental variables approach. The sector composition of export are used to create Herfindahl-Hirschman Index (HHI), which is used to divide the countries in the data set in three groups depending on their level of export diversification. The three groups of HHI allows us to examine how rain anomalies affect the economic growth in Sub-Saharan Africa depending on the level of diversification in a country. If countries trade according to the Ricardian model, the exports shares makes a good approximation on the overall production in a country, since the countries are assumed to trade their surplus production.

4.1.1 The HHI and WASP indices

The Herfindahl-Hirschman Index is a central tool for measuring the diversification and is widely used in different institutions such as the World Bank and European Union, but also by researchers in the field (Koren and Tenreyro 2007). The HHI sums the squared share that every entity holds in a market (in our case each entity is for each industry in the export market). The formula for calculating HHI is HHI = n X i=1 (si)2, (1)

where sidenotes the share of the export market that industry i holds, and may hold values between

0 < si ≤ 1, the value of that share is then squared. This is calculated for all the industries and

summed up to one index. The total number of industries is denoted by n. The highest value that can be achieved is 1, indicating that the entire market belongs to one entity, the lowest value is 1/n, indicating that all entities have equal shares of the market.

Another institution that uses this index as a tool is the European Commission, using it to control and analyse for changes in economic competitiveness from specialisation. They use a classification of diversification levels in order to divide different markets and apply different analysis and guidelines for their regulations of the markets within the European Union. The thresholds are calculated as

follows:

U nconcentrated if HHI < 0.1, (2a)

M oderately concentrated if HHI > 0.1 and HHI < 0.2, (2b)

Highly concentrated if HHI > 0.2. (2c)

Some markets have fewer entities operating, making it difficult to reach low values of HHI. If there are less than 5 entities operating it is impossible for a market to not be classified as Highly concentrated. In our export data there is a limited amount of different industries operating, making it difficult to create a fair classification and reach lower values of HHI. An alternative way to divide based on diversification is presented by Brezina et al. (2014) who takes the amount of entities into consideration when creating the limits for every level of diversification, the rules for the classification is therefore instead:

U nconcentraded if HHI > 1/n and HHI < 0.9/n + 0.1, (3a) M oderately concentrated if HHI > 0.9/ + 0.1 and HHI < 0.8/n + 0.2, (3b)

Highly concentrated if HHI > 0.8/n + 0.2. (3c)

The number of product classes by World Integrated Trade Solution (WITS) is 16, making it possible to use the standard HHI. However, since countries tend to be highly specialised in our data set (see section 4.2), the more allowing thresholds presented by Brezina et al. (ibid.) will be used to divide the observations in different categories. This is to try to reach a more even distribution of observations across the groups.

Similar to what Brown et al. (2011) do, this study use a Weighted Anomaly Standardized Precip-itation (WASP) index to measure the precipPrecip-itation anomalies in different countries. This index correlates well with other drought indexes, such as Palmer Drought Severity Index, however it only uses precipitation data, rather than also including temperature to measure the degree of droughts. According to Brown et al. (ibid.), precipitation is more correlated with droughts, and is seen as the dominant factor for climate variability. The advantage of using WASP, instead of other climate change indices, is that the WASP index is only dependent on precipitation data, making it easier to compute it without loosing efficiency. Brown et al. (ibid.) use thresholds of the WASP index to contextualise the values it presents. Moderate precipitation anomalies are at level -1/+1, where

a value above +1 saturates the soil to the extent that further rain might cause flooding. Values at -2 is considered extremely dry periods. Positive deviations can indicate both flooding, which operates on short periods, and precipitation anomalies just slightly higher than usual but during a longer period of time. However, Brown et al. (2011) note that the WASP index is not a good measurement for flooding, since flooding usually operates during a shorter period of time than the monthly data we retrieve. Large floods can have a large impact on infrastructure and agriculture and that might not be reflected in the monthly average precipitation if there is no rain during the rest of the month. The same positive value of WASP can be observed from just slightly more rain on a daily basis during a longer period of time, without the same devastating effect flooding has on infrastructure. Therefore, Brown et al. (ibid.) note that it is important to take precautions before drawing conclusions based on positive precipitation anomalies. Our different use of the WASP index gives us reason to be careful when drawing conclusions of both droughts and flooding. Instead we focus our analysis on the effects of more or less precipitation than the long term average.

In our study the long-term mean for precipitation is calculated between 1992 and the most recent year available, 2016. We will also include a squared variable for WASP in our estimations since the response to abnormal precipitation is unlikely to be linear, but rather exponential (ibid.). The yearly WASP values are calculated according to equation (4).

W ASP = N X t=1  Pi− ¯P σi  _¯ Pi ¯ PA , (4)

Pi is the observed average precipitation for month i, ¯Pi is the long-term average precipitation for

month i, the standard deviation of monthly average precipitation is denoted as σi. ¯PAis the annual

average precipitation for year A, which changes value depending on what year one is calculating WASP for. Since data is only available on country-level this thesis will calculate the index on country level in contrast to Brown et al. (ibid.), which calculates how much of a country’s area is affected by drought. The risk of having just one WASP index for an entire country is that it might give misleading values if the climate varies greatly within one country. If only one part of a country is severely affected by drought to the extent that it affects the welfare of the entire country, it might not be reflected in a large negative value of the WASP index. This might lead to an overestimation of the effect rain anomalies have on GDP per capita growth in the results. Other studies have on the other hand used aggregated, country-level precipitation anomalies (Barrios et al. 2010) and arrived at similar conclusions as Brown et al. (2011), therefore we do not expect large variations of using a single WASP measurement for every country.

using the WASP index rather than deviations from average precipitation allows for variability due to seasonal weather patterns, thus making it possible to separate droughts from the regular climate-cycle dry periods. The index is also favourable to use instead of a single precipitation series which just measures the level of precipitation, since the index indicates anomalies, allowing for differences in the average precipitation. The index also allows for separation of abnormally high or low levels of precipitation, which is an advantage as the responses in behaviour and, in extension, in growth are likely to differ depending on if the precipitation is anomalously high or low. (Brown et al. 2011)

4.1.2 Two-stage least squares

The estimations presented in this study is using two-stage least squares 2SLS regression models, aiming to capture the causal effect of precipitation on growth. To isolate the effect on GDP per capita growth, other variables that might affect growth need to be held constant. Many previous studies have thoroughly examined the determinants of GDP growth, and this study is mainly using the determinants found in Barro (1998) study as control variables. Other studies have used similar control variables (Barrios et al. 2010) and many of the variables Barro (1998) presents are seen as good conditions in the literature of conditional convergence (Barrios et al. 2010). Barro (1998) uses a three-stage least squares (3SLS) estimation model to allow for the endogeneity of the independent variables in the regression. He also argues that the advantage of using a 3SLS model is that it uses the cross-sectional data, in contrast to time and country fixed effects, when looking at panel data. As the correlation table in appendix C shows, the different variables that ought to explain GDP per capita growth have a high correlation between each other. This indicates that the variables are not entirely independent, making the regressions become objects to multicollinearity if those variables are not adjusted. Estimations based on 3SLS did not create smaller standard errors compared to 2SLS and thus there seems to be no gain in efficiency in the 3SLS in contrast to 2SLS1_{. This study}

will therefore use 2SLS. Using 2SLS still corrects for the endogeneity problem, however 2SLS is a more widely used and more simple method that does not require complex assumptions such as homoskedasticity. Our model will, despite Barro’s arguments, use time fixed effect. The reason for this is to correct for development that take place in the entire region due to technological advancement and global institutions, solving eventual problems created by omitted variable bias. An example of a study that use fixed effects is Brown et al. (2011) and their arguments are similar to ours. Still, country fixed effect will be excluded in the model and clustered standard errors for every country will be used instead, to use the advantages of cross-country analysis and at the same time allow for auto correlation in each country, following the same procedure as Koren and Tenreyro (2007).

1_{The estimations of 3SLS were computed with the same instrumental variables and the same endogenous variables}

According to Abadie et al. (2017) it is important to look at the sample process before deciding if the standard errors should be clustered or not. If the samples taken from a certain population are expected to correlate within certain groups, the standard errors should be adjusted according to those groups in order not to be misleading. Therefore it is suitable to cluster standard errors on country level since the natural distribution of the observations are clustered and expected to correlate within each country. The downside of using IV estimations with robust clustered standard errors is that the model becomes less precise and the standard errors might grow large in our estimations, and for this reason it might become more difficult to receive significant results. It is more important to adjust for heteroskedasticity even if the models becomes less precise since calculations based on homoskedasticity have a risk of creating bias in the model.(Baum et al. 2003)

In this thesis there are multiple variables (GDP per capita, fertility rate, life expectancy and rule of law) that are treated as endogenous variables. This is mainly based on the reasoning of Barro (1998), but also by looking at the correlation between the different variables (see appendix C). By using their own lagged value as instruments, the variation of the variables used in the second-stage equation is only dependent on earlier levels of the same variable and does not depend on other variables, hence the endogeneity problem has been corrected. The earlier values of for example fertility rate is not affecting today’s value of GDP per capita growth if not through more recent values of fertility rate. Moreover, this approach also solves for reverse causality since today’s value of GDP per capita growth cannot affect earlier values of the control variables. Thus, we assume that GDP per capita growth is only subject to more recent changes in fertility rate. To reduce the risk of not fulfilling the exclusion restriction we use five-year lags of the variables we are instrumenting. Other papers have shown that a five-year lag is enough to decrease the dependency ratio from the instruments (Swamy and Fikkert 2002; Barro 1998). It is straightforward to show that the five-year lags fulfil the relevance condition through a correlation test. To further examine if the variables should be treated as endogenous, Wooldridge’s robust test score is computed. It is a similar test as the common Sargan-Hansen endogeneity test except that the test is adjusted for models with robust standard errors (Baum et al. 2003). The main drawback of using the adjusted test is that the individual endogenous variables cannot be tested.Thus, the test becomes less precise in targeting true endogenous variables. A way to solve for this problem is through adding more endogenous variables with each model, to see how the χ2 _{-value and its p-value is affected. Nonetheless, if}

the 2SLS models produce significant p-values on the endogeneity test the estimations needs to be instrumented since a regular OLS-estimation runs the risk of being inconsistent (ibid.).

4.2

Data

This thesis uses panel data retrieved from numerous data bases. The data set consists of observations from 38 countries in sub-Saharan Africa (SSA) from a period of 25 years, 1992-2016. The period is based on the availability of export data from the World Integrated Trade Solution (WITS) database, which is a collaboration between the World Bank and the United Nations Conference on Trade and Development (UNCTAD), and the availability of precipitation data from the Climate Change Knowledge Portal. The countries included in the data set are described in appendix A. The countries are all SSA countries and have been chosen based on the availability of export data. Most of the countries are mainland countries, although there are a few islands included. No countries have been excluded based on any characteristics, but solely based on availability of data. The reason that we are using data from 1992 and onward is because climate change has become a larger subject in society in general, and it is argued that the impact from climate change has been growing during recent years (Oliver 2005). Thus, it is suitable to have the most recent data available to analyse our research question.

A majority of the data comes from the World Bank, where we have retrieved data on different measures of GDP2_{and the variables life expectancy}3_{, fertility rate}4_{, public expenses}5_{, inflation rate}6

and terms of trade7_{. The rule of law index has been retrieved from the Varieties of Democracy}

Institute at University of Gothenburg8. Data on exports, including export product shares for different sectors have been retrieved from the WITS database9. Monthly precipitation data has been retrieved from Climate Change Knowledge Portal, which is a part of the World Bank Group10. Since this study is using data from a number of different data bases, the data have been harmonised in terms of the earlier mentioned variables, years and countries, creating a new, uniform data set. This has mainly consisted of choosing variables, years and countries to be included and transferring them from existing data sets. All data used have been collected on a one-year basis, except for the data on precipitation which have been collected on a monthly basis. This facilitates the use of the WASP index, which is calculated based on monthly observations. The data processing, creation of new variables and regressions have been performed in Stata.

Table 1: Summary statistics

VARIABLES Obs. Mean Sd Min Max

GDP/capita growth (%) 940 1.606 4.946 -47.5 37.54

GDP/capita capita, PPP (2017USD) 941 4,213 4,697 436.7 26,422

Life expectancy 950 55.88 8.237 26.17 77.64 Fertility Rate 946 5.080 1.328 1.360 7.752 Public expense (%) 386 20.86 9.269 7.244 51.46 Inflation Rate(%) 817 8.878 13.58 -60.50 183.3 Rule of Law 950 0.487 0.250 0.0570 0.924 WASP 950 -0.698 5.658 -41.48 12.54

Terms of Trade (change) 904 0.786 15.082 -99.194 74.768

HHI export 599 0.353 0.193 0.0890 0.993

Note: This table shows descriptive statistics of variables used in our estimations. Data has been retrieved from The World Bank Database, WITS and the Varieties of Democ-racy Institute between 2020-11-16 and 2020-11-20.

4.2.1 Descriptive statistics

Table 1 shows summary statistics of the variables that are used in our analysis. The dependent variable is GDP per capita growth, with 940 observations. Notable about this variable is the fluc-tuations it exhibits, with a mean of 1.6 but a standard deviation three times the size of that, 4.8. Only one variable has a standard deviation that exceeds that, WASP, with a standard deviation eight times higher than the mean value. Both GDP per capita and inflation rate exhibits stan-dard deviations larger than their mean values, but not as significantly different as the previously mentioned variables. These variables all show both negative and positive values, but even so, they show great volatility. The fluctuations displayed in GDP per capita growth, ranging from -47.50 to 37.54 show that the countries in our data set have high peaks and deep valleys in their growth. Figure 1 is depicting the distribution of GDP per capita growth across the period of observation. Although the variable seems somewhat normally distributed and the vast majority observations are in the range of -10 to +10, there is a significant number of observations with values below or above those values. The most extreme values are very few, but indeed extreme, while growth rates of 5-10 (positive and negative) percent are to be considered normal in this data set. This further shows that low-income countries tend to exhibit a high level of volatility in their growth rates.

The influencing variables are WASP and the HHI. The distribution of WASP observations is

de-2_{GDP per capita growth is retrieved from World Bank Data (2020c). GDP per capita is retrieved from World}

Bank Data (2020d)

3_{World Bank Data (2020f)} 4_{World Bank Data (2020b)} 5_{World Bank Data (2020a)} 6_{World Bank Data (2020e)} 7_{World Bank Data (2020g)}

8_{Varieties of Democracy Institute (2020)} 9_{WITS (2020a)}

Figure 1: Distribution of GDP per capita growth

picted in figure B.1 in appendix B. WASP seems to be showing a somewhat skewed distribution of precipitation in SSA. This can be explained by the fact that WASP, calculated using monthly precipitation data, captures droughts better than flooding since flooding can occur during a short period of time. For there to be a drought, there needs be lower than average precipitation levels for a longer period of time, while heavy rain for only one or two days can cause flooding. Heavy rain during only a few days, causing flooding, might therefore have a small effect on the WASP index if the remaining days of the month have little or no rain, making the monthly average normal. The distribution of the WASP values could also simply be explained by dry periods occurring more frequently than flooding in SSA. While the mean is close to 0, which is to be expected from an evenly distributed population, the minimum value, -41.48, is far more extreme than the maximum value, 12.54. Although the minimum value is an outlier, the observations seem to be to be more spread out below 0, which is the value of WASP if precipitation has been normal during a year. There seems to be an approximately equal amount of observations ranging from 0 to +10, and 0 to -10, but there are multiple observations with values below -10, while only a few above +10. There is no pattern to distinguish for this variable, the observations seem to be somewhat random. This is to be expected, since the WASP index depicts deviation from the monthly average, and not average

precipitation which would be expected to show more obvious patterns.

Herfindahl-Hirschman Index, is an index for concentration, ranging from 1/n (in our case 0.0625) to 1, where 1/n indicates a diverse economy and 1 a perfectly concentrated economy. The HHI export variable holds only 599 observations, due to lack of data from certain countries and sectors, especially in the early years of the data set. The observed values range from 0.09 to 0.99, indicating great variations, but the mean indicates that most countries and years exhibit high levels of specialisation. Figure B.2 in appendix B depicts the distribution of HHI in the data set, and shows that the distribution over the different diversification levels is very uneven. Out of 599 observations in total, only 59 are considered to have unconcentrated markets, 179 are moderately concentrated and 361 are highly concentrated. Half of the observations show highly concentrated export markets, where there are only 10 percent of the observations indicating that the markets are diverse. It is important to note that the thresholds for classification, as depicted in equations 3a, 3b and 3c, are not evenly distributed between 0 and 1. The thresholds are marked in figure B.2 and are calculated to be set at 0.156 and 0.25, based on the discussions of Brezina et al. (2014).

The remaining variables are control variables derived from the World Bank database. When choos-ing what variables to include, the works of Barro (1998) have been used extensively. Barro (ibid.) works with panel data to define the main determinants of GDP per capita growth. The variables used in our models are all based on the findings of that study, but they are not completely identical due to problems with data availability. For example, we do not use any education related variable, as Barro do in his estimations, due to lack of data. Because we do not have access to the exact same data as Barro has, alternative variables were explored but none held enough observations to be useful in our estimations. For the same reasons, the models of this study do not include a democracy variable. Share of public expenditure, that is included in the estimations of Barro (ibid.), is presented as ”public expense” in table 1 but is not used in the estimations, because of the low number of observations. Excluding these variables might create problems with omitted variable bias. However, this bias seems to be small as Barro (ibid.) only presents small effects from these excluded variables, such as the democracy index. It is very difficult to estimate what effect these variables would have on the data if we could observe them, so we will abstain from doing so. The remaining variables included in our study show large ranges, which goes to show that the countries observed in the data exhibit large differences, both within and between each other.

GDP per capita has a mean of 4,213 and a standard deviation slightly larger, 4,697, also showing a large range and variation. However, this variable exhibits a much smaller within-country vari-ation, but a striking between-country variation. The variables depicting human capital (e.g. life expectancy, fertility rate) also show great variations, mainly between countries. Rule of law is an

index showing the degree of transparency, independence, predictability and equal enforcement of laws in a country, and to what extent the actions of government officials comply with the law. The index range is 0-1, and the variable range is almost as broad, 0.05-0.92. Terms of trade change is depicting the net barter terms of trade, which is calculated as the percentage ratio of the export unit value indexes to the import unit value indexes, relative to the reference year 2000. Thus, the ratio indicates the relative prices of exports to imports, and changes that increase the ratio indicate that export prices have risen which can be considered to be good for the economy. The terms of trade change variable has been created through differentiating the terms of trade, thus the variable is showing the absolute change in the export to import price ratio. The variable is exhibiting large variances with a range of -99.195 to 74.769, and some countries are exhibiting large within-country variation. There seems to be no general trend for the changes between different years.

4.2.2 Data Limitations

The data consists of 25 years and 38 countries, giving a maximum of 950 observations for each variable. Working with data from low-income countries, one always faces the problem of miss-ing data. Especially in the earlier years of the data set, there are multiple variables with many observations missing. The worst case contains only half of the possible observations (e.g. public expenses). Because of the nature of the missing values, which are not systematic and in many cases not individual, but for multiple consecutive years, we have not imputed any values. A limitation with this study is that we do not observe a full panel for all countries. These missing values result in a smaller data set which could be less representative of the true values for the variables observed. This might create distorted results that describes some countries better than others since there is only sufficient data from certain countries. For example, there is data on almost every year and relevant variable for Madagascar, yet almost no data for Angola until 2007.

The share of observations that is missing may affect the robustness of the results, if being large, which needs to be taken into consideration when interpreting the different results and their signifi-cance. An example is that fewer observations make the analysis more sensitive to extreme values. By examining the GDP per capita growth in table 1, there are indications that there are extreme values since the minimum value is approximately 10 times bigger than standard deviation. It is possible to reduce the effect of such values by taking the logarithm of those variables. Therefore, the logarithm of certain variables will be used in the regressions, as done in many earlier studies and essays, such as Koren and Tenreyro (2007) and Barro (1998). Another advantage of using the logged values is that the interpretation of the coefficients in models is done in terms of percent, rather than units. For example, it might be more relevant to see how one percent increase of life expectancy affects GDP per capita growth rather than looking at an increase of one year in life expectancy.

4.3

Empirical specification

In this section, we are describing the equations that constitute the models used in our estimations. All of the estimations presented in this study are performed using a two-stage least squares regression model to correct for endogeneity. The WASP index is lagged by one year for two reasons. Firstly, the effect of rainfall anomalies is expected to be delayed, because water reserves and vegetation usually survive a short period of less precipitation before the water levels become critical. The second reason is that the rainy season in SSA countries is between November and March, making the precipitation levels rise towards the end of the year, after the harvest. It is reasonable to expect rain falling in November to affect the harvest the following year, rather than the most recent harvest. One of our models also includes a squared coefficient for WASP, because both positive and negative anomalies could have similar effect on GDP per capita growth.

Equation 5 is depicting the first model presented, including only the main variables (HHI in exports and the WASP index) and logarithmic GDP per capita. This model is used to determine whether there is an initial effect of HHI and WASP, and to maximise the number of observations included. Equation 6 shows an extended version of equation 5, with GDP per capita growth as the dependent variable and a number of influencing control variables. The outcome variable Yit, represent the

GDP per capita growth in country i in year t. The models are estimated with an intercept α, however the point estimates of the intercept is excluded from our tables. A time fixed effect is included and it is written as λt, in this model. This means that the coefficient for the years is

different for every year, t. Koren and Tenreyro (2007) and Imbs and Wacziarg (2003) find that the relationship between diversification and growth likely is quadratic, where concentration tends to decline with earlier stages of development and then reverse slightly. Therefore, we have computed a model with a quadratic term for HHI in appendix D. However, these models seem insignificant and do not change the coefficients and explanatory power of the models. Thus, this study will only look at the linear effect of HHI in further analyses. Equation 7 is the same equation as equation 6, with the exception that it is also including a squared WASP term.

Yit= α + β1HHIit+ β2W ASPi(t−1)+ β3log(GDP/capita)it+ it (5)

Yit= α + +λt+ β1HHIit+ β2W ASPi(t−1)+ β3log(GDP/capita)it

+β4log(Lif e expectancy)it+ β5log(F ertility rate)it+ β6Rule of Lawit

+β7Inf lation rateit+ β8T oT changeit+ it

Yit= α + λt+ β1HHIit+ β2W ASPi(t−1)+ β2W ASPi(t−1)2

+β3log(GDP/capita)it+ β4log(Lif e expectancy)it+ β5log(F ertility rate)it

+β6Rule of Lawit+ β7Inf lation rateit+ β8T oT changeit+ it

(7)

Because all of the models are estimated using 2SLS, there are instrumented variables for several of the variables found in the equations above. Exactly what instruments are used for which variable is not clearly stated in Barro (1998). Therefore this study is only assuming endogeneity for those variables that have high correlation with GDP per capita and have strong instruments when lagged for five years. Partially to keep the model as simple as possible, partially since GDP measures all the economic activity for a country so it might be endogenous with several variables. There are in total four endogenous variables that have IV regressions: logged GDP per capita, logged fertility rate, logged life expectancy and rule of law. As explained in subsection 4.1.2, all the variables that are endogenous are lagged 5 years in order to fulfil the assumption of exclusion restriction.

Yit= α + λt+ β1HHI_unconcit+ β2W ASPi(t−1)+ β3W ASPi(t−1)∗ HHI_unconcit

+β6log(GDP/capita)it+ β7log(Lif e expectancy)it+ β8log(F ertility rate)it

+β9Rule of Lawit+ β10Inf lation rateit+ β11T oT changeit+ it

(8)

Yit= α + λt+ β1HHI_moderately concit+ β2W ASPi(t−1)

+β4W ASPi(t−1)∗ HHI_moderately concit+ β6log(GDP/capita)it+

β7log(Lif e expectancy)it+ β8log(F ertility rate)it+ β9Rule of Lawit

+β10Inf lation rateit+ β11T oT changeit+ it

(9)

Yit= α + +λt+ β1HHI_concit+ β2W ASPi(t−1)+ β5W ASPi(t−1)∗ HHI_concit+

β6log(GDP/capita)it+ β7log(Lif e expectancy)it+ β8log(F ertility rate)it

+β9Rule of Lawit+ β10Inf lation rateit+ β11T oT changeit+ it

(10)

Equation 8, 9 and 10 above are similar to equation 6, with the only exception that they include interaction variables between the WASP index and dummy variables for the different levels of diversification. This way we can examine how the effect of WASP differs at different levels of diversification. This approach also allows us to use all the observations in one regression which produces more reliable results.

5

Empirical results

In this section, the results of the different models described in subsection 4.3 are presented. Subsec-tion 5.1 describes the relaSubsec-tionship between rain anomalies and GDP per capita growth with three different models. Subsection 5.2 presents the linear model, with interaction variables presented in equation 8, 9 and 10. It also presents these models but adding a quadratic term for WASP. Subsection 5.3 examines the robustness of the results in two different ways. First by using five-year means of all the variables, which also examines if long-term droughts have a different impact on GDP per capita growth, and secondly by dividing the observations into sub-samples based on their level of export diversification.

5.1

Rainfall, specialisation and GDP per capita growth

Table 2 presents the results from the estimations done using equations 5, 6 and 7, estimating the relationship between GDP per capita growth, HHI in exports and WASP. Model (1) is a very simple model used to determine if there are signs of a relationship, while model (2) is the main linear model, using multiple control variables. Model (3) is a non-linear model, including all the control variables used in model (2) while also adding a quadratic term for the WASP variable. The effect of the WASP variable is negative and statistically significant on the five percent level in all three models, indicating that there is a negative relationship between precipitation anomalies and GDP per capita growth. The relationship, being very similar in the models, is however quite small compared to many of the other variables. The effect of an increase of one unit in the WASP index will lead to an decrease in the growth rate with 0.05 percentage points, while an increase with only one percent in for example fertility rate will decrease growth rates with 3-4 percentage points, which in comparison is a much larger effect. Considering thresholds of Brown et al. (2011) for the WASP index, with the effect depicted in table 2, it would require extremely heavy rainfall (or extremely low levels of precipitation in model (3)) in order to have an economically significant impact on GDP per capita growth. The quadratic term indicates that there is a small increasing negative effect of WASP on GDP per capita growth, but this effect is not statistically significant.

The WASP variable, having negative values in all models, indicates that GDP per capita growth rates decrease with positive deviations from average rainfall. While flooding is expected to lower growth rates due to the devastating effect it can have on infrastructure, buildings and agriculture, the WASP index is, as previously explained in subsection 4.1, not a good measure of flooding. Positive rainfall anomalies would be expected to have none or a slightly positive effect on growth rates, due to the benefit for the agricultural sector, which the results contradicts. However, the effect WASP have on GDP per capita growth can be considered quite small in contrast to the other variables.

The data shows that there are great anomalies among the observations, which might make an important difference between countries that experience weather anomalies and countries that do not, even if the effect of every unit of WASP is small. Our results, as depicted in table 2, are consistent with Brown et al. (2011) regarding the size of the effect. The major difference is that Brown et al. (ibid.) separate the anomalies in two coefficients, one for droughts and one for heavy rainfall. In their study both the coefficients are negative, showing that both positive and negative precipitation anomalies have negative impacts on growth rates. Brown et al. (ibid.) do include a linear model in their estimations as well, but find no effect on GDP per capita growth. Our model (2) with only one linear coefficient, has a negative coefficient which implies a higher GDP per capita growth as the negative precipitation anomalies grow larger. Model (3) with squared WASP shows results that are more consistent with the models of Brown et al. (ibid.) which include two coefficients, because the squared value of WASP is negative. Although it is not divided in two different coefficients, the squared value indicates a negative effect on GDP per capita growth from both positive and negative weather anomalies. WASP has therefore increasingly negative effects on GDP per capita growth as the rain anomalies grows. The results are similar despite that we use the WASP index differently in this study compared to Brown et al. (ibid.), which indicates that the different use of the WASP index does not lead to biased results.

Further, table 2 also depicts the HHI variable, showing the effect that increased export specialisation has on GDP per capita growth. HHI is exhibiting an effect pointing in a different direction in model (1) compared to model (2) and (3). The effect is negative in model (1), while it is positive, and stronger in model (2) and (3), and it is only showing statistical significance in model (3). The effect might be considered rather large since an increase in specialisation with 0.1 units increases the growth with 0.14 percentage points. A low but stable growth rate can be considered more favourable than a fluctuating, unstable growth rate. Having a lower total growth rate, an increase of 0.14 percentage points can possibly account for a large portion of the total growth.11 Initial GDP per capita is exhibiting effects moving in the same direction in the three models, however it is not showing statistical significance in model (1), while the effect is significant on the one percent level in model (2) and (3).

In terms of other variables depicted in model (2) and (3) in table 2, there are a number of significant coefficients; initial GDP per capita, fertility rate and rule of law. Rule of law is however statistically significant on a higher significance level in the quadratic model, in which HHI in exports is also significant, although on the ten percent level. This does indicate that the third model is the most advantageous, but to facilitate a deeper analysis, both the quadratic and linear model are presented

11_{The observations in our data exhibit a large variation of GDP per capita growth with the majority of observations}

between -10 percent and +10 percent growth rates. When comparing the ”regular” fluctuations of GDP per capita growth in our observations, an increase of 0.14 percentage points due to specialisation might seem small.

further. Model (1) is useful to determine any signs of an initial relationship, but it proves inadequate in explaining the majority of the variations and to base further analysis on.

Table 2: Multiple models for GDP per capita growth

(1) (2) (3)

VARIABLES GDP/capita GDP/capita GDP/capita

growth (%) growth (%) growth (%)

HHI export -0.277 1.732 1.381*

(1.247) (1.348) (0.837)

WASP (1yr lag) -0.0509** -0.0534** -0.0662**

(0.0252) (0.0247) (0.0326)

WASP2_{(1yr lag) -0.00210}

(0.00313) GDP/capitaPPP(log) -0.175 -1.731*** -1.321*** (0.384) (0.483) (0.298) Life expectancy(log) 0.888 1.610 (4.022) (1.522) Fertility rate(log) -4.364*** -3.127*** (1.250) (0.816) Rule of Law 2.024* 2.571*** (1.120) (0.695) Inflation rate (%) -0.0283 -0.0205 (0.0339) (0.0224) Terms of trade(diff.) -0.000761 -0.00110 (0.0102) (0.0116)

Time FE YES YES YES

Observations 574 541 541

R-squared 0.068 0.139 0.150

χ2_{-value 196.93 425.08 379.93}

P-value <0.01 <0.01 <0.01

Robust regression scores 212.397 66.2689 66.2675

Woolridge’s P-value >0.01 >0.01 >0.01

Note: The models shows the different models for the relationship between GDP per capita growth, HHI in exports and WASP. All the models have included a time fixed effects and are presented with clustered robust standard errors. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Table 2 also shows the results of a Wooldridge’s robust score test. All of the values presented are statistically significant which indicates that it is correct to adjust the model for endogeneity for the chosen variables. Since the scores drop dramatically when we add control variables, and hence more endogenous variables, there could be reason to doubt that some of the assumed endogenous variables are in fact exogenous. However, in models using clustered errors and Wooldridges’s robust score, it is not possible to see which of the assumed endogenous variables are showing evidence of being exogenous. Hence, all the models are assumed to be correct since the overall p-values are very low. In addition to this the partial R2_{-values for each endogenous variable included is above}

75 percent, and have F-statistic well above the critical value of ten. These results makes it more difficult to sort out the weaker instruments.

5.2

Weather anomalies under different levels of diversification

Table 3 presents the linear models from equations 8, 9 and 10 in subsection 4.3. It presents the relationship between GDP per capita growth, HHI in exports and WASP, with an added interaction variable for each model. Model (1) in table 3 includes an interaction variable which combines WASP and a binary variable taking the value 1 if a country has an unconcentrated export portfolio and 0 otherwise. Model (2) and (3) includes the same interaction variable but for moderately concentrated and concentrated portfolios respectively. The WASP variable exhibits a weak negative effect on GDP per capita growth in all models, model (1) and (3) shows results with statistical significance. The effect is robust across all models, being somewhat larger in model (3). The effect of having an unconcentrated export portfolio is negative according to model (1) and the effect becomes positive if the portfolios becomes more specialised. The largest effect is seen in model (3), where highly specialised countries have 0.348 percent points more growth relative to the other observations. This is consistent with the results presented in subsection 5.1.

The signs of the effects of both WASP and the interaction between binary HHI and WASP are pointing in the same direction in model (1) and (2), but the effect of the interaction variable is larger in model (2). However, in model (3), the effects move in opposite directions and the effect of the interaction variable is positive. This indicates that the effect of rain anomalies while having a specialised export, is smaller compared to having an unconcentrated or moderately concentrated export. This is because of the initial negative effect that rain anomalies have on GDP per capita growth balances out the positive effect the effect WASP has in countries with concentrated export12. The total effect from rain anomalies have the largest impact in countries with moderately concen-trated exports. The control variables show very similar results in all three models and are mainly consistent with the results presented in table 2, with the same levels of statistical significance. All of the models are exhibiting high χ2_{-values and low p-values, indicating that the models in their}

entirety have high statistical significance.

Table 4 includes the interaction variables from equation 8, 9 and 10 with the exception that the models in table 4 also includes the squared WASP variable, like in equation 7. The initial effect of the WASP variable is negative in all of the models, showing very similar effects as the linear models presented in table 3. The effect of WASP is largest in model (3), -0.0993, compared to -0.0588 in model (1) and -0.0448 in model (2). These are similar values as in the models of table 3 with the same levels of statistical significance, which is to be expected. The quadratic effect, which points towards an decreasing negative effect for higher values of WASP in all models presented in table 4, shows very similar sizes of the effect in all models.

12_{In model (1), the total effect of WASP on GDP per capita growth in countries with unconcentrated export is}

calculated to be −0.0493 − 0.0115 = −0.0608. In model (2), the total effect is −0.0343 − 0.0575 = −0.0918. In model (3), the total effect is −0.0900 + 0.0638 = −0.0262.

The binary HHI variables and interaction variables in table 4 are showing very similar effects as in table 3. The interaction variables between HHI and WASP are increasing the negative effects of the WASP variable in all models except for model (3). This balances out the negative effect from both the initial linear WASP variable and squared WASP variable, which indicates that the effect of WASP on GDP per capita growth is smaller for countries with concentrated exports13. However, when looking at HHI and the interaction variables as a whole, the effects are robust over the models presented in table 4. This shows that the total effect of rain anomalies is similar under different concentrations of export. Since the effects of HHI and the interaction variables do not prove any statistical significance (except for the effect of HHI in model (1)), there is no evidence that countries are affected by rain anomalies to different degrees. Because of this, it is difficult to draw conclusions based on the results in table 3 and table 4. The fact that the results are very similar for linear and non-linear models shows that the results are somewhat robust, and that there might not be a large need for a quadratic model. This is further emphasised by the small values of the coefficients for the quadratic effects.

13_{The effect of the linear and the interaction variables for countries with unconcentrated export is −0.0588 −}

0.00298 = −0.06178. The total linear effect for countries with moderately concentrated export is −0.0448 − 0.0614 = −0.1062. The total linear effect for countries with concentrated export is −0.0993 + 0.0633 = −0.036.

Table 3: The effect of WASP with interaction variables

(1) (2) (3)

VARIABLES GDP/capita GDP/capita GDP/capita

growth (%) growth (%) growth (%)

WASP (1yr lag) -0.0493* -0.0343 -0.0900**

(0.0281) (0.0311) (0.0389) HHI(unconc.) -0.999** (0.446) HHI(unconc.)*WASP -0.0115 (0.0781) HHI(moderately conc.) 0.0265 (0.513) HHI(moderately conc.)*WASP -0.0575 (0.0662) HHI(conc.) 0.348 (0.460) HHI(conc.)*WASP 0.0638 (0.0543) GDP/capitaPPP(log) -1.507*** -1.515*** -1.523*** (0.471) (0.468) (0.475) Life expectancy(log) 0.503 0.527 0.671 (4.097) (4.230) (4.103) Fertility rate(log) -4.074*** -3.982*** -3.976*** (1.288) (1.275) (1.269) Rule of Law 1.816 1.688 1.751 (1.183) (1.234) (1.195) Inflation rate (%) -0.0230 -0.0195 -0.0221 (0.0358) (0.0350) (0.0346)

Terms of trade(diff.) -0.000356 3.06e-05 -0.000340

(0.0102) (0.0101) (0.0101)

Time FE YES YES YES

Observations 541 541 541

R-squared 0.143 0.137 0.140

χ2 _{464.77 540.62 692.54}

P-value <0.01 <0.01 <0.01

Note: The table shows the same baseline model for the relationship between GDP per capita growth, HHI in exports and WASP, with a different interaction variable for each model. All the models include a time fixed effect and the standard errors are robust. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1