SlideShare a Scribd company logo
1 of 95
Download to read offline
Sciences Po
PSIA – Paris School of International Affairs
Master in International Economic Policy
Trading up or trading down?
The effect of trade openness on structural change in Sub-Saharan Africa
Eivind Breidlid
Master’s thesis supervised by Sergei Guriev, Professor of Economics
Academic Year 2015/2016
1
Acknowledgements
I would first sincerely thank my supervisor Sergei Guriev for his guidance, quick feedback, and
very helpful suggestions. His input and advice has been invaluable in this process.
Furthermore, I would like to thank Dr. Nicholas Sim who graciously shared his data on the
Baltic Dry Index. I would also like to thank Bendik Elstad for all our many discussions over a
cup of coffee outside Bibliotheque de l’École doctorale. Finally, my deepest gratitude goes to
Elyse, who not only applied her brilliant mind to proofread this paper, but also gave me
tremendous support throughout the process.
2
Abstract
The purpose of this paper is to investigate the effect of trade openness on structural
change in Sub-Saharan Africa. Structural change happens when labor reallocates from
low-productive to high-productive sectors. The effect of trade openness on structural
change is uncertain in theory. By definition, trade prompts the reallocation of resources
across sectors. On the one hand, trade liberalization allows for the expansion of the
most productive firms at the expense of the least productive firms, which raises
aggregate productivity. On the other hand, trade liberalization leads to specialization in
comparative advantage sectors, which for Sub-Saharan Africa is mainly in low-
productive agriculture. To identify the effect of trade openness empirically, I use a panel
dataset on sector employment and value added to conduct difference-in-differences
(DID) and instrumental variable (IV) analyses. I find that trade openness has a positive
effect on structural change in Sub-Saharan Africa, and that this result is quite robust to
different tests. Moreover, an overvalued exchange rate was found to have a negative
effect on structural change. The most plausible transmission channel for trade openness
is the income effect, where trade liberalization leads to increased income, which again
leads to more expenditure on services instead of agricultural products. Therefore,
structural change in Sub-Saharan Africa is driven by an expansion of services rather than
manufacturing. The findings of this paper lead to two big questions; first, whether
services can be a dynamic growth sector in Sub-Saharan Africa like manufacturing has
been for other developing countries in the past; and second, whether manufacturing
can play a bigger part in the future as opportunities for integration into global value
chains emerge.
3
Table of contents
Acknowledgements................................................................................................................................1
Abstract..................................................................................................................................................2
Table of contents....................................................................................................................................3
List of tables ...........................................................................................................................................4
List of figures..........................................................................................................................................5
List of tables and figures in the appendix...............................................................................................5
1. Introduction ...................................................................................................................................6
2. Literature review................................................................................................................................8
2.1 Structural transformation.............................................................................................................8
2.2 Stylized facts about structural transformation.............................................................................9
2.2.1 Patterns of structural transformation ...................................................................................9
2.2.2 Growth-enhancing structural change..................................................................................10
2.3 Trade and other determinants of structural transformation......................................................11
2.4 Empirical work............................................................................................................................15
3. Theoretical framework and empirical hypotheses ...........................................................................18
3.1 Growth effect of sector reallocation ..........................................................................................18
3.2 Hypotheses.................................................................................................................................19
4. Empirical strategy.............................................................................................................................22
4.1 Baseline regression.....................................................................................................................22
4.1.1 The dependent variable.......................................................................................................22
4.1.2 Difference in differences .....................................................................................................24
4.1.3 Methodological challenges..................................................................................................25
4.1.4 Sector contribution and sector shares.................................................................................26
4.2 Instrumental variables regression ..............................................................................................28
4.2.1 The gravity equation............................................................................................................28
4.2.2 Constructing the instruments..............................................................................................29
4.3 Identification assumptions .........................................................................................................32
5. Data..................................................................................................................................................34
5.1 Structural change .......................................................................................................................34
5.2 Trade liberalization episodes......................................................................................................35
5.3 Control variables ........................................................................................................................37
6. Results..........................................................................................................................................40
6.1 Descriptive statistics.............................................................................................................40
6.2 Difference in differences ......................................................................................................43
6.3 Robustness checks................................................................................................................47
4
6.3.1 Alternative dependent variables .........................................................................................47
6.3.2 Alternative trade openness variables ..................................................................................48
6.3.3 Sensitivity analysis...............................................................................................................52
6.3.4 Two-period regression.........................................................................................................56
6.4 Sector contribution...............................................................................................................57
6.5 Sector shares as dependent variables ..................................................................................61
6.6 Instrumental variables regression ........................................................................................63
7. Discussion.........................................................................................................................................67
7.1 Does trade openness promote structural change? ....................................................................67
7.2 Why does trade liberalization lead to structural change? ..........................................................68
7.2.1 The agricultural channel ......................................................................................................68
7.2.2 The value chain channel ......................................................................................................69
7.2.3 The role of the mining sector ..............................................................................................70
7.3 Role of the exchange rate...........................................................................................................72
8. Concluding remarks..........................................................................................................................75
Bibliography .........................................................................................................................................77
A. Articles, books, journals ...........................................................................................................77
B. Datasets....................................................................................................................................80
Appendices...........................................................................................................................................81
List of tables
Table 1. Liberalization dates in SSA.....................................................................................................36
Table 2. Difference in differences.........................................................................................................43
Table 3. Difference in differences with interactions.............................................................................45
Table 4. Tariff indicator as explanatory variable.................................................................................49
Table 5. Openness as explanatory variable (1) ....................................................................................50
Table 6. Openness as explanatory variable (2) ....................................................................................51
Table 7. Placebo liberalization dates ...................................................................................................52
Table 8. Analysis excluding Nigeria.....................................................................................................54
Table 9. Two-period regression ...........................................................................................................57
Table 10. Sector contribution 1971-2010.............................................................................................57
Table 11. Sector contributions as dependent variable..........................................................................58
Table 12. Mining contribution as dependent variable..........................................................................60
Table 13. Agriculture value added share as dependent variable..........................................................61
Table 14. Two-period, Agriculture value added share .........................................................................62
Table 15. Bilateral trade equation .......................................................................................................64
Table 16. Second stage regression .......................................................................................................65
5
List of figures
Figure 1. Relationship between structural change and liberalization episodes ....................................40
Figure 2. Structural change for liberalized and non-liberalized countries.............................................41
Figure 3. Sectoral productivity over time..............................................................................................42
Figure 4. Sectoral labor share over time...............................................................................................42
Figure 5. lvr2plot SSA 1971-2010..........................................................................................................53
Figure 6. lvr2plot SSA 1981-2010..........................................................................................................53
Figure 7. Added variable plot SSA 1971-2010 without Nigeria.............................................................56
Figure 8. lvr2plot SSA 1971-2010 without Nigeria................................................................................56
List of tables and figures in the appendix
Appendix 1. Sectors in the GGDC database.........................................................................................81
Appendix 2: List of sectors and countries in the UN National Account Statistics.................................82
Appendix 3. Description of control variables.......................................................................................84
Appendix 4. Summary statistics............................................................................................................85
Appendix 5. Structural change growth with over- and undervalued currencies ...................................85
Appendix 6. Structural change and liberalization over time in individual SSA countries.....................86
Appendix 7. Alternative measures of structural change .......................................................................88
Appendix 8. Sector contribution to SC growth over time......................................................................90
Appendix 9. Industry and Services contribution to SC .........................................................................90
Appendix 10. Change in Manufacturing and Mining value added share as dependent variable..........91
Appendix 11. Service and Industry value added as dependent variable ...............................................92
Appendix 12. First stage regressions....................................................................................................93
Appendix 13. Discussion chapter regressions ......................................................................................94
6
1. Introduction
Africa needs structural transformation not structural adjustment
Carlos Lopez, Executive Director UNECA1
Economic growth is essential to reduce poverty and improve living standards in developing
countries. While high economic growth in East and South Asia has led to an unprecedented
reduction in poverty, growth in Sub-Saharan Africa (SSA) has been disappointing, with a high
incidence of negative growth rates in the 1980s and 1990s. However, since 2000 the average
per capita annual growth in SSA has been close to 3 percent. It is however, argued that the
high growth rates the past decade are due to a favorable external environment leading to high
demand for the region’s natural resources. There are therefore concerns that SSA countries
will not be able to sustain high growth rates when the benign external environment shifts. For
the growth to be sustainable, it is essential that the recent boom translated into structural
transformation and diversified economic activities (Rodrik 2014: 1-2, 5-7).
Aggregate labor productivity growth happens either because of productivity growth within
sectors, or because high-productive sectors employ a larger share of the labor force. Labor
reallocation between sectors with different productivity levels is known as structural
transformation or structural change. This paper defines growth-enhancing structural change
as the reallocation of labor from low-productive to high-productive sectors.2
Large
productivity gaps between sectors are characteristic of low-income countries. The agricultural
sector is simultaneously the least productive sector and the largest employer in the economy.
Countries that start with a large share of their labor force in low-productive agriculture
therefore have a large potential for economic growth through structural transformation.
Most Sub-Saharan African countries followed import-substitution policies in the 1960s and
1970s, aiming to nurture a domestic industry by protecting it from foreign competition.
However, following the debt crisis in the early 1980s, SSA countries accepted financial aid from
the World Bank and the IMF on the condition that they liberalize their trade regimes. Hence,
1
Lopez, Carlos (2013). β€œAfrica needs structural transformation not structural adjustment”, The Executive
Secretary’s Blog
2
If not specified otherwise, the term β€œstructural change” means growth-enhancing structural change.
7
from the mid-1980s to the early 1990s, SSA countries undertook trade liberalization reforms
under so-called Structural Adjustment Programs (Kassim 2013). There is a large body of
literature on the effect of trade liberalization on growth and incomes (e.g. Feyrer 2009).
Nevertheless, the literature on the effect of trade liberalization on the structural change
component of growth is limited. As trade openness induces countries to specialize in the
products for which they have a comparative advantage, trade liberalization naturally leads to
reallocation of production factors across sectors. However, if the comparative advantage of
SSA countries lies in low-productive sectors such as agriculture, the effect on structural change
can be negative. This study therefore asks the following questions:
What is the effect of trade openness on structural change, and through which mechanisms
does the effect work?
To answer these questions, I will use a panel dataset on employment and value added for 11
SSA countries. The main identification strategies will be difference-in-differences (DID)
regression and instrumental variable (IV) analysis.
I structure the thesis as follows. In Chapter 2, I present some important stylized facts on
structural change, before discussing the main theories on trade openness and structural
transformation. The chapter concludes with some empirical findings from the literature.
Chapter 3 then describes the theoretical framework including the hypotheses to be tested. In
Chapter 4, I explain the empirical strategy of using difference-in-differences and instrumental
variables analysis. Chapter 5 describes the datasets and variables used in the empirical
analysis, before the empirical results are presented in Chapter 6. In Chapter 7, I discuss the
results in the context of my theoretical framework and hypotheses, before Chapter 8
concludes.
8
2. Literature review
This chapter first describes the relationship between structural transformation and economic
development. I then present stylized facts on the evolution of structural change over time and
across regions. In the second part of the chapter, I will discuss theories on the relationship
between trade openness and structural change. Finally, I will show some empirical results
from the literature on the determinants of structural change.
2.1 Structural transformation
Economists have long recognized the importance of structural transformation in economic
development. Arthur Lewis recognized structural change as key to economic development. As
the capitalist (modern) sector expands, surplus labor is drawn from the mainly rural
subsistence sector. This process of economic development continues until the surplus labor
disappears, as wages between the modern and traditional sectors equalize (Lewis 1954: 8).
Simon Kuznets noted that economic growth is characterized by a high rate of structural
transformation, from agriculture to non-agricultural activities, and later from industry to
services (Kuznets 1973: 248).
More recently, Restuccia et al. (2008) find that low agricultural productivity and a high share
of labor devoted to agriculture are the main explanations behind the large income differences
between rich and poor countries (Restuccia et al. 2008: 235). Reallocating labor towards more
productive activities would hence have a major impact on economic growth. Moreover,
amidst concern that growth in SSA has failed to create sufficient productive employment
opportunities in order to absorb the growing labor force (ILO 2014, World Bank 2014),
structural change is important because it relates to the component of growth caused by
increasing employment shares in high-productive sectors. Page (2015) finds evidence that
growth originating from structural change is more effective at reducing poverty than
productivity growth within sectors. This is because growth caused by structural change
involves low-skilled workers gaining employment in sectors with higher productivity and
hence higher wages (Page 2015: 230-233).
9
2.2 Stylized facts about structural transformation
2.2.1 Patterns of structural transformation
Herrendorf et al. (2013) investigate the evolving relationship between economic development
and sectoral shares of employment and value added, using historical time series for currently
rich countries, and panel data for a broader set of countries. An increase in GDP per capita is
associated with a continuous decrease in agriculture’s employment and value added shares,
and a continuous increase in the service sector’s employment and value added shares.
Manufacturing, on the other hand, follows an inverse U-shaped development. At low levels of
development, income growth is accompanied by a rapid expansion of manufacturing shares,
before the relationship flattens. At high levels of income, manufacturing shares decrease with
income growth. The value added and employment shares of services accelerate around the
same income level manufacturing shares peak, which is about US$8000 in 1990 international
dollars (Herrendorf et al. 2013: 9-16).
Rodrik (2015) shows that countries’ manufacturing shares are increasingly peaking at earlier
stages of development. Controlling for GDP per capita and population, the average country’s
manufacturing employment share is 11.7 percentage points lower post-2000 than in the
1950s, and 8.8 percentage points lower than in the 1960s. The only region without a negative
time trend in manufacturing employment controlling for income and population is Asia.
Moreover, countries’ manufacturing shares peak at a lower level of income. Since 1990,
countries have reached peak manufacturing at incomes that are around 40 % of the levels
experienced before 1990 (Rodrik 2015: 8-15). Felipe et al. (2015) find that global productivity
growth in manufacturing has been similar to non-manufacturing between 1970 and 2010.
Although manufacturing productivity grew faster within countries, this was offset by a
reallocation of manufacturing jobs to less productive countries. Furthermore, because
manufacturing shifted to countries with more labor-intensive production, the manufacturing
employment share has not declined globally. Nevertheless, manufacturing productivity
growth has accelerated the last decades. Since 1990, it has been 50 % higher than aggregate
productivity growth, coinciding with the increased contribution from China and other East-
Asian countries. As global manufacturing employment is no longer expanding,
industrialization is more difficult for low-income countries today because of growing
10
international competition from mainly labor-abundant Asian countries (Felipe et al. 2015: 11-
23).
De Vries et al. (2013) analyze the evolution of structural change in Sub-Saharan Africa. They
find that rate of structural change was strong between 1960 and 1975, with a large decrease
in agricultural employment share, and a corresponding strong increase in manufacturing’s
share of employment and value added. Despite the strong increase in employment share,
manufacturing managed to increase their relative productivity from 1960 to 1975, showing
few signs of decreasing marginal productivity. The period between 1975 and 1990, however,
showed negative growth and very slow reallocation of labor from agriculture. The 20 years
following 1990 have been a period of rapid reallocation of labor across sectors. Agricultural
employment share fell from 61.6 % in 1990 to 49.8 % in 2010. However, this reallocation of
labor from agriculture did not benefit the manufacturing sector, as its share of employment
decreased slightly. Market services on the other hand nearly doubled their share of
employment from 24.1 % to 36.8 %. Unfortunately, the relative productivity of market services
fell considerably, indicating low marginal productivity (De Vries et al. 2013: 10-12).
2.2.2 Growth-enhancing structural change
The authors then analyze whether labor reallocated from low-productive to high-productive
sectors (static structural change), and whether labor reallocated from low productivity growth
to high productivity growth sectors (dynamic structural change). They find that Sub-Saharan
Africa experienced a strong reallocation of labor from low-productive to high-productive
sectors between 1960 and 1975, while the reallocation gains were small between 1975 and
1990. Between 1990 and 2010, labor moved from lower to higher productive sectors,
producing a static growth effect. At the same time, labor reallocated to sectors that
experienced lower productivity growth, resulting in a large negative dynamic reallocation
effect. Thus, the aggregate contribution from structural change to growth is thus negligible in
this period. These results point to low marginal productivity in the expanding sectors. For
example, reallocation of labor to market services has a strong static effect on labor
productivity (1 %), but this positive impact is offset by the negative dynamic effects (-0.96 %)
(De vries et al. 2013: 15-22).
11
McMillan et al. (2014) find that between 1990 and 2005, structural change made a negative
contribution to productivity growth in both Sub-Saharan Africa and Latin America. Meanwhile,
structural change made a positive contribution to growth in Asia, explaining most of the
productivity growth gap between Asia and the two other regions. This is disappointing
considering the large structural change potential in Sub-Saharan Africa, with its high share of
labor in low-productive agriculture. The effects are not, however, uniform across the region.
Ghana, Ethiopia and Malawi all experienced growth-enhancing structural change as the
employment share in agriculture fell, while manufacturing employment increased. In Nigeria
and Zambia, on the other hand, manufacturing and tradable services contracted, while the
employment share in agriculture increased (McMillan et al. 2014: 18-23).
While structural change is negative between 1990 and 1999, structural change made a positive
contribution to labor productivity after 2000. Nigeria and Zambia then had an expansion in
manufacturing and a contraction in agriculture and services. Overall, half of the African
countries in the sample experienced an increase in the employment share in manufacturing
post-2000, although the size of the changes is not enough to transform these economies
rapidly (McMillan et al. 2014: 23-25).
2.3 Trade and other determinants of structural transformation
Fundamentally, international trade allows sectoral expenditure shares to deviate from
sectoral production. This allows each country to run a net export surplus in its sector of
comparative advantage. Thus, patterns of specialization induced by trade openness directly
affect the labor share. Moreover, as trade affects relative prices, sectoral expenditure shares
change, shifting sectoral production and labor shares further (Uy et al. 2013: 668).
In classic models of trade, the movement of labor and capital across sectors are what allows
countries to reap the benefits of openness. Countries gain from trade by moving resources to
their comparative advantage sectors; which is defined by relative technological differences in
the Ricardian model or by relative factor endowments according to the Heckscher-Ohlin
model. Models with increasing returns to scale predict geographical agglomeration of
production, which can lead to observable reallocation of labor across sectors. However, new
trade models that focus on intra-industry trade do not have clear predictions on labor
12
reallocation across broad economic sectors, as specialization due to intra-industry trade might
be highly disaggregated within economic sectors (Wacziarg & Wallack 2004: 411-412). In the
Melitz model of heterogeneous firms, trade openness leads to both export opportunities and
import competition. The most productive firms take advantage of the export opportunities to
expand their production. As profits increase, the most productive firms demand more labor,
thereby driving up wages. The least productive firms shrink or exit due to increased product
market and labor market competition that reduces their profits (Melitz 2002: 23-26, Janiak
2006: 2). Openness therefore leads to reallocation of labor from the least to the most
productive firms.
Theoretical studies of structural transformation focus on two economic mechanisms that
drive the reallocation of labor across sectors. One type of models focuses on demand factors,
with income effects driving the process of structural transformation. The other main type of
models focuses on supply factors, as differential productivity growth drives changes in relative
prices and sectoral labor allocations (Dabla-Norris et al. 2013: 5).
Herrendorf et al. (2013) present a three-sector model consisting of agriculture, manufacturing
and services. In the model preferences are non-homothetic, meaning that the relative share
of consumption does not stay constant as income changes. Specifically the income elasticity
is below one for agriculture, constant for manufacturing and above one for services. Keeping
relative prices constant, an increase in income will lead to more expenditure devoted to
services at the expense of agriculture. This leads to corresponding changes in sectoral
employment and value added shares. Alternatively, structural change is driven by changes in
relative prices, due to differential technological progress across sectors. The sector that has
the highest productivity growth will experience a relative price decrease. Assuming price
elasticity of demand is below one (inelastic), the quantity increase in demand will not fully
compensate the fall in price, reducing the nominal value added of that sector. As the sector
can produce the same amount of goods with fewer workers, labor shifts away from this sector.
Correspondingly, the sector with the lowest productivity growth will experience higher
relative prices, and increase their share of employment and nominal value added share
(Herrendorf 2013: 41-46).
13
In the absence of trade, income growth will lead to a higher share of employment in services,
and the sector with the highest productivity growth will experience a declining share of
employment. However, Matsuyama (2009) introduces international trade where countries’
comparative advantage follows relative sector-productivities (Ricardian specialization). If the
Home country has larger productivity growth in manufacturing than in services, labor should
shift to services. However, if productivity growth in manufacturing is higher in Home country
than abroad, comparative advantage will shift and increase manufacturing employment at
Home. Total manufacturing employment will fall, and manufacturing employment abroad will
unambiguously decline. The overall effect on manufacturing employment at Home is
ambiguous. While the trade effect boosts manufacturing employment, the relative price effect
is negative. Allowing for openness can help explain the finding that countries with low
productivity growth in manufacturing are experiencing a stronger fall in manufacturing
employment than countries with high productivity growth (Matsuyama 2009: 482-484).
Uy et al. (2013) include evolving trade costs in the model. In the model, the manufacturing
and agriculture sectors are tradable, while the service sector is non-tradable. Opening the
economy up to trade has three main impacts on structural change. First, declining trade costs
will lead to net exports in the comparative advantage sector. Hence, labor will move from its
comparative disadvantage sector to its comparative advantage sector. As trade costs decline
over time, each country’s comparative advantage is increasingly revealed, leading to increased
specialization and increased labor reallocation to the comparative advantage sector. Second,
a country with relative higher productivity growth within a sector will supply an increasing
share of the world demand; this counteracts the negative effect of relative prices on the labor
share caused by high productivity growth. However, as productivity increases and export
demand slows in the long run, the relative price effect dominates the trade effect, leading to
a decline in domestic manufacturing employment. This could explain the non-linear
relationship between income and manufacturing employment share previously discussed.
Third, lower trade costs leads to higher income growth, which in turn reinforces the effect of
non-homothetic preferences on labor reallocations, leading to a stronger reallocation of labor
from agriculture to services (Uy et al. 2013: 671-673).
Using a similar framework, Rodrik (2015) notes that most developing countries are pure price
takers on global markets. If a small open economy has high productivity growth in the
14
manufacturing sector, the relative price of manufacturing goods will not fall, and the effect on
manufacturing employment and value added shares is unambiguously positive. However, as
price takers, developing countries also import relative price declines caused by relative high
productivity growth in the manufacturing sectors of other countries. For manufacturing
employment to expand in developing countries, productivity growth differentials between
manufacturing and non-manufacturing, must not only be positive, but also need to exceed the
decline in relative prices on the world market. Higher productivity growth in manufacturing
than non-manufacturing does therefore not directly lead to increased share of manufacturing
in the economy. Only the countries with the highest productivity growth globally will expand
their labor force in the manufacturing sector (Rodrik 2015: 15-22).
One clear prediction from these models is that the labor shares of sectors that produce
tradable goods should differ across countries that have different sectoral productivities.
Countries reallocate labor to the sector where they have a comparative advantage, and
continue this process as long as productivity growth is sufficiently high and global demand is
not saturated. However, comparative advantage is not static and labor could reallocate to new
sectors with relative high productivity growth. A decline in the relative price of manufacturing
experienced by small developing countries could be the result of technological progress
elsewhere, large increase in supply (e.g., the rise of China), or domestic trade liberalization.
Countries that lag behind global productivity growth in manufacturing could experience a
reallocation of labor away from manufacturing when opening up to trade. This could explain
why countries in Sub-Saharan Africa have failed to expand their manufacturing sector the last
20 years, as the region’s manufacturing productivity has fallen behind the technological
frontier (De Vries et al. 2013: 13-15)3
.
However, Baldwin (2011) argues that globalization has undergone radical change from the
mid-1980s, giving developing countries new opportunities to industrialize. Until the 1980s,
globalization mostly concerned falling trade costs, while since the 1980s globalization is driven
by lower transmission costs. Developing countries can now expand manufacturing by joining
global and regional supply chains. Countries are no longer required to build the whole supply
chain at home, they can simply produce the components for which their technological
3
Relative productivity actually increased in the 1970s, but experienced a rapid fall in the 1980s and has
continued downwards since.
15
disadvantage is least marked, and export it into an international supply chain. Industrialization
is both easier and less meaningful. What a country produces says less about how advanced it
is, and more where its location is in the global value chain. With the ICT revolution,
multinational firms can combine their sophisticated know-how with developing countries’
labor. However, as managing a value chain requires some face-to-face interactions, distance
still matters. This could explain why most production networks concentrated in low-wage
nations are near large headquarter nations such as the USA, Germany and Japan. Lack of a
regional headquarter nation can pose limitations for future manufacturing growth in Sub-
Saharan Africa.
A key implication of global value chains is that protectionism through import substitution
policies is less likely to be successful. While in the 1960s and 1970s, advanced countries
technology advantages could be partly offset by labor costs, competition is harder now that
international supply chains allow the combination of advanced technology and cheap labor
(Baldwin 2011).
2.4 Empirical work
The empirical literature identifying the determinants of growth-enhancing structural change
is not very large. One reason is that employment data for developing countries, especially in
Sub-Saharan Africa, has not been available or reliable.
Wacziarg & Wallack (2004) investigate the effect of trade liberalization on sectoral
reallocation of labor in 25 developing countries. Measuring reallocation across nine broad
economic sectors, they find a decrease in the pace of labor reallocation after trade
liberalization. They further disaggregate the manufacturing sector into 28 sub-sectors and find
that trade liberalization has a small positive effect on reallocation within the manufacturing
sector. The effect is strongest over longer periods. Moreover, the analysis shows that
liberalization is followed by reduced employment growth in manufacturing. The analysis does
therefore not support the proposition that trade liberalization leads to large reallocation of
labor across sector (Wacziarg & Wallack 2004: 422-425).
Dabla-Norris et al. (2013) analyze the determinants of structural change by creating a
benchmark model and comparing the actual and predicted real value added share of each
16
sector. They find that country characteristics account for a large proportion of the cross-
country variation. More specifically, land area is positively related to agriculture shares;
population size is negatively associated to the agricultural share and positively related to
manufacturing share; and having a large proportion of non-working-age people is positively
associated to the service share of the economy. The capital stock is negatively associated with
agricultural and service shares, while being positively related to the manufacturing share. The
effect of capital stock is particularly strong at low levels of manufacturing, suggesting the
importance of infrastructure in early manufacturing development. As expected GDP per capita
is negatively related to the share of agriculture and has a non-linear relationship with
manufacturing, increasing at low levels of GDP and decreasing at higher levels (Dabla-Norris
2013: 9-18).
Looking at policy variables Dabla-Norris et al. find that trade openness has a strong positive
relationship with the share of manufacturing, while being negatively associated with the
agricultural share. The effect of trade openness on the service share is negative for countries
with a low share of services, and positive for countries with a high share of services. Dividing
the sample to before and after the modern era of globalization, they find that trade openness
had a stronger effect on the manufacturing share after 1992 than before. Curiously, when
using average tariff rates as an indicator of trade liberalization, liberalization is positively
associated with agriculture, and negatively correlated with manufacturing for countries with
low share of industry. A depreciation of the real exchange rate is associated with a higher
manufacturing share for countries with low industrial base (Dabla-Norris et al. 2013: 14-19).
Looking at the 1990-2005 period, McMillan et al. (2014) find that comparative advantage in
primary products, proxied by the share of raw materials of total exports4
, has a significant and
negative effect on growth-enhancing structural change. The initial employment share in
agriculture has a positive and significant effect, but only when the comparative advantage
indicator is included. The authors see this as evidence of conditional convergence, as a country
with a large rural labor share has a potential for growth-enhancing structural change, given
that it does not have a strong comparative advantage in primary products. Moreover,
including the comparative advantage indicator makes the regional dummies for Africa and
4
Comparative advantage indicator
17
Latin America insignificant, indicating that comparative advantage and the initial agricultural
share can jointly explain the average differences between the regions. Furthermore, McMillan
et al. find that an undervalued exchange rate promotes growth-enhancing structural change,
while employment rigidity discourages it. Other potential determinants such as income levels,
demography, institutional quality, and tariff levels did not turn out to be consistently
significant in their analysis (McMillan et al. 2014: 26-27).
Analyzing the period 2000 to 2010, McMillan & Harttgen (2014) find that the main
contribution to growth-enhancing structural change is the decline in the labor share engaged
in agriculture. A small part of those workers ended up in manufacturing (around 20 %), but
around 80 % moved to the service sectors. They find that the agricultural employment share
is falling faster in countries that started with a high share of the labor force in agriculture;
experience high population growth rates; benefit from higher quality of governance; and have
undertaken deeper agricultural reforms (proxy for agricultural productivity growth). McMillan
& Harttgen argue that high population growth is negatively associated with the agriculture
employment share because it reduces farm size and makes farming less attractive for the
young. The negative relationship between agricultural productivity growth and the
agricultural labor share is consistent with theories discussed above, as higher productivity
growth reduces the labor share due to either declining relative prices or changing expenditure
patterns as income increases (McMillan & Harttgen 2014: 5, 31-32).
18
3. Theoretical framework and empirical hypotheses
3.1 Growth effect of sector reallocation
Although the models discussed in the previous chapter predict trade liberalization’s effect on
sector shares, it is unclear how changes in sector shares translate into aggregate productivity
growth. Reallocation away from agriculture to either services or manufacturing will have a
positive effect on growth due to the low productivity in agriculture However, with stagnating
manufacturing productivity in Sub-Saharan Africa, the growth effect of reallocating labor from
services to manufacturing is uncertain. The effect depends on the type of service sector, and
country-specific relative productivities.
There are nevertheless strong arguments as to why manufacturing is essential to rapid and
sustained economic development. Baumol (1967) argued that unlike manufacturing, most
service activities do not allow for constant and cumulative increases in productivity through
capital accumulation, innovation, or economies of scale (Baumol 1967: 420). Felipe et al.
(2015) find that all non-oil economies with per capita income above $12 000 (2005 dollars)
today, reached at least 18 % employment share in manufacturing at their peak (Felipe 2015:
9). Rodrik (2015) sees three advantages of manufacturing compared to services. First, formal
manufacturing exhibits unconditional productivity convergence with the technological
frontier. Second, unlike other high-productive sectors such as mining, manufacturing has
traditionally absorbed a large number of unskilled labor. Third, manufacturing is a tradable
sector meaning that it does not face demand constraints from a home market populated by
low-income consumers (Rodrik 2015: 3).
While the models discussed in Section 2.3 treat services as non-tradable, services are in fact
increasingly becoming a larger part of international trade. Rodrik nevertheless argues that
highly productive, tradable services such as IT and finance are typically skill-intensive and do
not have the capacity to absorb the type of labor that is abundant in low-income countries.
Other types of services typically lack technological dynamism or are non-tradable; meaning
their ability to expand rapidly is constrained by productivity in the rest of the economy (Rodrik
2015: 24). This pessimism is challenged by Ghani & O’Connell (2014), who note that
technology improvements have enabled services to be traded like goods, and that global trade
19
in services is growing much faster than trade in goods. The increasing proliferation and
importance of global value chains can also provide opportunities for service-based growth and
technology diffusion, as provision of services is often sub-contracted nationally and globally.
The globalization of services provides African countries more opportunities to find niches
beyond manufacturing, where they can specialize and scale up. Concerning dynamism, Ghani
& O’Connell find that productivity convergence in services has actually been faster than in
manufacturing between 1990 and 2010. Unlike manufacturing, income elasticity for services
is above unity, making it unlikely that the future expansion of the services sector will be limited
by demand factors (Ghani & O’Connell 2014: 2-16). Moreover, although the service sector in
Sub-Saharan Africa is dominated by informality, so is manufacturing, as an increasing number
of manufacturing firms are informal and low-productive. Unconditional manufacturing
convergence concerns only formal firms, and this convergence is not found when including all
types of manufacturing firms (De Vries et al. 2013: 5).
As the scope for broad-based manufacturing development seems to be shrinking, reallocation
into services might therefore be equally important. With the growth effect of reallocation
across sectors being uncertain, the theoretical framework needs specific propositions on the
relationship between openness and structural change.
3.2 Hypotheses
Considering the finding by McMillan & Harttgen (2014) that the declining labor share in
agriculture accounts for most of the structural change growth in SSA, it is plausible that trade
liberalization can affect structural change through either changes in income or agricultural
productivity growth. There is a broad spectrum of literature showing the positive relationship
between trade openness and growth (e.g. Feyrer 2009). However, the literature on trade and
agricultural productivity growth is less clear, and some studies have found mixed results (Yu
& Nin-Pratt 2011).
In their paper on structural transformation, McMillan et al. (2014) note that while
globalization has facilitated technology transfer and increased efficiency in production, its
effect on structural change has been highly uneven across regions. A number of firm-level
studies that have shown that intensified import competition has made manufacturing
20
industries become more efficient by rationalizing their operations. The least productive firms
exit the industry, while the remaining firms shed excess labor. This leads to an increased
within-sector productivity, and the top tier firms close the gap with the technological frontier
in Western countries. Unlike in the Melitz model, however, the most productive firms do not
absorb the labor from exiting firms. The effect on overall productivity growth therefore
depends on what happens with the workers who are displaced. In countries that have high
levels of unemployment or large inter-sectoral productivity gaps, there is a large risk that
displaced workers end up in low-productive sectors such as low-skilled services or informality.
If this happens, the effect of increased import-competition on aggregate productivity is
unknown, and can even be negative (McMillan et al.2014: 11-12). This is an especially big risk
for Sub-Saharan Africa with its large productivity differentials between sectors, and where
almost 80 % of the labor force work in vulnerable employment, either as unpaid family
workers or self-employed, predominately in the informal sector (ILO 2014).
However, openness to investments and intermediate goods should have a positive effect on
integration into global value chains. Rather than seeing import as competition, imports can be
seen as inputs and complementary to local production. By easing access to imports, trade
liberalization can have a positive effect on structural change through value chain integration.
In fact, there is some evidence that structural change is linked to the share of foreign value
added in a country’s export, a common measure of participation in global value chains5
(AEO
2014: 31, 72).
Finally, an undervalued exchange rate can be beneficial for structural change. McMillan et al.
(2014) argue that one reason why Asian countries have experienced stronger structural
change is because they adopted competitive exchange rate in order to promote tradable
industries, while SSA and Latin American countries often globalized with an overvalued
exchange rate (McMillan et al. 2014: 25-26). Furthermore, there is evidence that undervalued
exchange rates increases growth in developing countries by boosting the size of the tradable
sector (Rodrik 2008: 388-391).
Based on the theoretical and empirical literature, I propose three distinct hypotheses:
5
Backward integration
21
1) Trade liberalization and import competition have a negative effect on structural
change in Sub-Saharan Africa, because its comparative advantage does not lie in high-
productive sectors. Specialization will therefore lead to reallocation of labor to less
productive sectors.
2) Trade liberalization has a positive effect on structural change as it leads to a
reallocation of labor from agriculture to the service sector by either:
a. Increasing income, which leads to relative lower expenditure on agricultural
goods (income effect). This will have a negative effect on agriculture’s
employment, nominal value added share, and real value added share.
b. Increasing agricultural productivity, thereby reducing relative prices of
agricultural goods (productivity effect). This will have a negative effect on
agriculture’s employment and nominal value added share, and an ambiguous
effect on its real value added share.
3) Trade openness has a positive effect on structural change because it promotes
integration into international value chains. This effect should be clear after 1980 and
be associated with a reallocation of labor into manufacturing.
22
4. Empirical strategy
Most empirical literature that has tried to identify the causes of structural transformation has
focused on the determinants of sector shares, not on whether the reallocations across sectors
were growth-enhancing. The few studies focusing on growth-enhancing structural change
have been mostly cross-country analyses. By using panel data, this paper will attempt to
identify the determinants of growth-enhancing structural change in a more rigorous way.
This chapter will first describe the structural change variable that is used as dependent
variable in the empirical analysis. I will then explain the methodology behind difference-in-
differences (DID), the main econometric technique used in the analysis. After discussing some
methodological challenges with DID, I introduce the methodology of instrumental variables
(IV) and outline how the instruments were constructed using bilateral trade data.
I estimate the following baseline regression:
π‘†π‘‘π‘Ÿπ‘’π‘π‘‘π‘’π‘Ÿπ‘Žπ‘™ π‘β„Žπ‘Žπ‘›π‘”π‘’π‘–π‘‘ = 𝛼 + 𝛿𝑑 + πœ‡π‘– + 𝛽𝑇𝐿𝑖𝑑 + πœ™π‘‹π‘–π‘‘ + πœ€π‘–π‘‘ (1)
where 𝑆𝐢𝑖𝑑 is the structural change dependent variable, 𝛼, 𝛿𝑑 and πœ‡π‘– are the constant, time and
country fixed effects respectively. Time dummies controls for common macroeconomic
shocks, while country fixed effects control for country-specific, time-invariant factors. 𝑇𝐿𝑖𝑑 is
a dummy for whether country i has experienced trade liberalization at or before time t, 𝑋𝑖𝑑 are
relevant control variables and πœ€π‘–π‘‘ is the error term. The estimated impact of the trade
liberalization episode is the estimate 𝛽̂.
4.1 Baseline regression
4.1.1 The dependent variable
The paper’s measure of structural change follows the methodology used in De Vries et al.
(2013), who divide productivity growth into within-sector growth and between-sector growth:
βˆ†π‘ƒ = βˆ‘ (𝑖 𝑃𝑖
𝑇
-𝑃𝑖
0
)𝑆𝑖
𝑇
+ βˆ‘ (𝑖 𝑆𝑖
𝑇
βˆ’ 𝑆𝑖
0
)𝑃𝑖
0
(2)
where βˆ†π‘ƒ is economy-wide productivity growth. 𝑃𝑖
𝑇
stands for productivity in sector i at time
T, while 𝑆𝑖
𝑇
describes the share of employment in sector i at time T. The first term measures
23
the within-sector productivity growth, weighted by the sector share. The second term
measures the change in sector shares, weighted by the productivity of the sector. This is the
structural change term (De Vries et al. 2013: 15-16). If the expanding sectors have higher
average productivity than the shrinking sectors, the term is positive; if the expanding sectors
have lower average productivity than the shrinking sectors, the term is negative. Labor
productivity is measured in constant prices.
The dependent variable is measured over five-year periods. As structural change is a long-
term process, it is not meaningful to measure its year-by-year fluctuations. Measuring five-
year periods is consistent with the economic growth literature, for example in Caselli et al.
(1996). In order to measure structural change as a growth rate, the term is divided by the
average productivity at the base year for each 5-year period. Finally, the term is divided by the
number of years to get the average annual growth rate of structural change for each 5-year
period. The dependent variable Structural change is thus the following decomposition:
Structural change =
βˆ‘ (𝑖 𝑆𝑖
𝑇
βˆ’π‘†π‘–
0
)𝑃𝑖
0
5βˆ—π‘ƒ0 (3)
where 𝑃0
is the economy-average productivity at the first year of each five-year period.
There are other ways to construct the structural change term. Instead of using baseline
productivity, 𝑃𝑖
0
as a weight, one could use final productivity 𝑃𝑖
𝑇
or period productivity
averages 𝑃̅𝑖, as weights. Using final productivity will typically result in a relatively smaller
contribution from structural change to growth compared to the composition using base-
period productivity levels (De Vries et al. 2013: 15-16). For these reasons, the regressions will
be rerun as a robustness checks using final productivity weights in constructing the dependent
variable:
SC with final productivity =
βˆ‘ (𝑖 𝑆𝑖
𝑇
βˆ’π‘†π‘–
0
)𝑃𝑖
𝑇
5βˆ—π‘ƒ 𝑇 (4)
The productivity equation (2) can be expanded by adding a dynamic third term:
βˆ†π‘ƒ = βˆ‘ (𝑖 𝑃𝑖
𝑇
-𝑃𝑖
0
)𝑆𝑖
𝑇
+ βˆ‘ (𝑖 𝑆𝑖
𝑇
βˆ’ 𝑆𝑖
0
)𝑃𝑖
0
+ βˆ‘ (𝑃𝑖
𝑇
βˆ’ 𝑃𝑖
0
) βˆ— (𝑆𝑖
𝑇
βˆ’ 𝑆𝑖
0
)𝑖 (5)
This new term is called dynamic structural change. The term is positive if labor reallocates to
sectors with above-average productivity growth, and negative if labor reallocates to sectors
24
with below-average productivity growth. Production equation (5) therefore includes two
reallocation effects: reallocation of workers to sectors with above average productivity levels
(static reallocation effect), and reallocation to sectors with above average productivity growth
(dynamic reallocation effect).
Although it is important to consider the dynamic effects of reallocations, this paper choose to
not use dynamic structural change as a dependent variable in the regression analysis. The term
is often negative and can be very difficult to interpret. For instance, agricultural productivity
growth can play an important part in inducing structural change, as surplus food production
allows more workers to migrate to the cities to work in β€œmodern” sectors. However, if
productivity growth is strong and the labor share in agriculture is falling, the dynamic
structural change term will be negative, even if labor is shifting to more productive sectors
(McMillan & Harttgen 2014: 15). As the paper is using panel data, the static effect of
reallocation to sectors with low productivity growth will decrease over time. Nevertheless, I
construct a combined static and dynamic structural change variable to serve as a dependent
variable for robustness checks:
Static + Dynamic SC =
βˆ‘ (𝑖 𝑆𝑖
𝑇
βˆ’π‘†π‘–
0
)𝑃𝑖
0
5βˆ—π‘ƒ0 +
βˆ‘ (𝑃𝑖
𝑇
βˆ’π‘ƒπ‘–
0
)βˆ—(𝑆𝑖
𝑇
βˆ’π‘†π‘–
0
)𝑖
5βˆ—π‘ƒ0 (6)
As the dependent variable is in five-year periods, I measure all variables in five-year periods.
4.1.2 Difference in differences
Difference-in-differences (DID) estimations entail identifying a specific intervention or
treatment, in this case a trade liberalization episode. One then compares the difference in
outcomes (structural change) before and after the intervention for countries that underwent
trade liberalization to the same difference for countries who did not undergo trade
liberalization (Bertrand et al. 2003: 2). The baseline multiple periods DID regression to be
estimated is thus equation (1). For the impact measured in the DID analysis to be causally
valid, one has to make the assumption of parallel trends. The parallel trends assumption states
that changes in the outcome variable over time would have been the same in both treatment
and control groups in the absence of an intervention, controlling for relevant time-varying
factors (Bertrand et al. 2003: 2-3).
25
If other factors besides trade liberalization affect the difference in trends between the groups,
the estimation will be biased. In addition to fixed effects that take care of time-invariant
country specific factors, I will control for observable time-varying variables that could affect
the outcome variable. The control variables chosen are recognized in the literature as
important determinants and conditions for structural change.
The advantage with this estimation technique compared to the cross-country analysis used in
McMillan et al. (2014) is that the fixed effects control for time-invariant, country-specific
factors that might influence the outcome variable.
4.1.3 Methodological challenges
Bertrand et al. (2003) argue that standard difference in differences estimation is potentially
subject to severe serial correlation problems. They argue that three factors make serial
correlation an important issue when estimating DID. First, DID estimations usually rely on long
time series. Second, the commonly used dependent variables in DID estimation are typically
highly positively serial correlated. Third, an intrinsic aspect of the DID model is that the
treatment variable changes very little within a country over time. These three factors can
reinforce each other so that the standard error for 𝛽̂ could severely understate the standard
deviation of 𝛽̂. Bertrand et al. test how DID perform under placebo laws, where treated states
and the years the laws passed are randomly selected. They find that the null hypothesis of no
effect is rejected at T > 1.96, far more often than the expected 5 % of the time (Bertrand et al.
2003: 6-11).
By using five-year periods, the time series in this paper’s analysis is not particularly long.
Moreover, the dependent variable (structural change) is unlikely to be serial correlated
because it is a growth rate. However, given that the treatment variable (liberalization dummy)
changes very little over time, there may be a problem with serial correlation. One solution for
small datasets like mine is to collapse the data into pre- and post-periods to produces
consistent standard errors. As the liberalization dates happen in different years, I must slightly
modify the technique. First, I regress structural change on country-fixed effects, year dummies
and relevant control variables. Then I divide the residuals into two groups: one before the
liberalization and one after the liberalization. I obtain the estimate of liberalization’s effect
26
and its standard errors from an OLS-regression of the two-period panel (Bertrand et al. 2003:
14-15). However, the power of this test declines fast, making it hard to detect an effect. With
20 states, this technique has a rejection rate of 9.5 %, almost twice as large as it should be
(Bertrand et al. 2003: 18-19).
The usual regression assumption is that the error term (πœ€π‘–π‘‘ ) is independently and identically
distributed (i.i.d.), but this can be violated in panel data. There could be β€œclustered errors”,
which means that observations within country i are correlated in some unknown way, inducing
correlation in πœ€π‘–π‘‘ within i. In the presence of clustered errors, standard errors can be wrong,
leading to incorrect inference. To correct for this, I use clustered standard errors by country
in all my regressions, which also corrects for heteroskedasticity.
Moreover, the assumption of parallel trends could be violated. To test for this, I will redo the
analysis with β€œplacebo” trade liberalization dates. In one test, I will move all the liberalization
dates one period ahead, and in the other test, I will move the liberalization dates one period
behind. If the trends were parallel, there should be no effect of the β€œplacebo” trade
liberalizations.
Finally, there are some limitations with the structural change term itself. While it decomposes
GDP per worker, it does not decompose GDP per capita. It therefore does not take into
consideration changes in labor force participation (due to demographics, changes in gender
roles etc.) or changes in unemployment (McMillan et al. 2014: 20). Reallocation from and to
unemployment would have been particularly interesting as this is the least productive activity
possible, but data limitations leaves this question for future research. However, for most
countries in Sub-Saharan Africa the option of being unemployed is not viable, and low-
productive work in the informal sector is the most likely alternative for someone losing their
job. Employment in the informal sector is included in the analysis as part of the different
economic sectors.
4.1.4 Sector contribution and sector shares
In order to understand the mechanisms behind structural change it is worth looking into
contributions from different sectors. By modifying decomposition (3), we can measure the
contribution to structural change from each sector. In decomposition (3), each expanding
27
sector contributes positively to aggregate productivity levels, even when it has below-average
productivity levels. In the modified model, I therefore adjust the structural change term to
take account of its relative productivity level. For each period, I divide the sectors into
expanding and shrinking sectors based on changes in employment shares. I calculate the
structural change effect of an expanding sector as its productivity level relative to the average
productivity level of the shrinking sectors. The modified decomposition is thus:
π‘†π‘’π‘π‘‘π‘œπ‘Ÿ π‘π‘œπ‘›π‘‘π‘Ÿπ‘–π‘π‘’π‘‘π‘–π‘œπ‘›π‘—π‘‘ = βˆ‘ (𝑆𝑗
𝑇
βˆ’ 𝑆𝑗
0
)
𝐽
𝑗 (𝑃𝑗
0
βˆ’ 𝑃0βˆ—
) (7)
where J is a set of expanding sectors and 𝑃0βˆ—
is the weighted average productivity of the
shrinking sectors. If the labor productivity(𝑃𝑗
0
) of expanding sector j is higher than the
weighted average productivity of the shrinking sectors (𝑃0βˆ—
), sector j made a positive
contribution to structural change (Timmer et al. 2014: 16-17). By using each sector’s individual
contribution to structural change as the dependent variable, I can identify through which
sectors the explanatory variables contribute to structural change.
As employment data by sector only coverseleven countries in Sub-Saharan Africa, generalizing
the results obtained in the DID analysis can be difficult. However, I expand my dataset using
changes in the real value added sector shares as my dependent variable. Although this does
not correspond perfectly to growth-enhancing structural change, it is possible to make some
inferences based on the results. First, although I cannot observe productivity per sector
directly without employment data, it is a reasonable assumption that the agricultural sector
has below average productivity in Sub-Saharan Africa. Thus, a reduction of agriculture’s value
added share in the economy is likely to be positive for growth-enhancing structural change.
The analysis can then serve as an additional robustness test. Furthermore, a larger sample
makes it possible to infer whether the results from the DID analysis are generalizable to the
larger region of Sub-Saharan Africa. A larger sample also gives the two-period robustness test
more power. Finally, by using sector shares as dependent variable I can identify which sectors
expands following trade openness. I can then test more directly some of the theories of
structural transformation and sector shares presented in Chapter 2. The dependent variable
will be the average five-year change in each sector’s value added share:
βˆ†π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘ =
π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘βˆ’ π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘βˆ’5
5
(8)
28
where π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘ is the real value added share of sector i at time t.
4.2 Instrumental variables regression
While the DID analysis described in the previous chapter takes steps to increase internal
validity, finding the true causal effect remains a challenge. Trade liberalization often coincides
with the adoption of other domestic market reforms, as well as stable fiscal and monetary
policies. Since these policies can affect structural change, trade liberalization could be
correlated with factors omitted from the regression, and thus be endogenous (Frankel &
Romer 1999: 379). Moreover, trade policies might change following economic conditions and
prospects.
Identifying a solid instrument for trade liberalization is not an easy task. Indeed, an exogenous
instrument for changes in domestic trade policy has not yet been identified in the literature.
It is more realistic to identify exogenous changes in trade costs that vary across time and
countries.
4.2.1 The gravity equation
I build on the established method of constructing instruments for trade flows loosely based
on the gravity model (e.g. Feyrer 2009). The gravity model predicts that bilateral trade is a
function of exporter characteristics, importer characteristics, and resistance factors such as
distance. Anderson and van Wincoop (2003) derived the theoretically consistent gravity
model:
π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘ =
𝑦𝑖𝑑 𝑦 𝑗𝑑
𝑦 𝑀𝑑
(
𝜏 𝑖𝑗𝑑
𝑃 𝑖𝑑 𝑃 𝑗𝑑
)1βˆ’πœŽ
, (9)
where π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘ denotes exports from country i to country j in year 𝑑, while 𝑦𝑖𝑑, 𝑦𝑗𝑑 and 𝑦 𝑀𝑑
are real GDP in country i, country j, and the world respectively. πœπ‘–π‘—π‘‘ is the bilateral resistance
term, and 𝑃𝑖𝑑 and 𝑃𝑗𝑑 are multilateral resistance terms in country i and j. 𝜎 > 1 is the elasticity
of substitution between goods (Blanchard & Onley 2015: 18-20). The bilateral resistance term
includes distance, but also other bilateral factors that determine trade such as common
border, colonial relationship, etc. Multilateral resistance terms represent the relative price
levels of the importing and exporting country. This can be captured by using country fixed
effects. After taking logs, the standard gravity model is estimated:
29
ln(π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘) = 𝛾𝑑 + πœ“π‘– + πœ™π‘— βˆ’ 𝛿 ln(𝑑𝑖𝑠𝑑𝑖𝑗) + πœ‚ ln(𝐺𝐷𝑃𝑖𝑑) + πœ‡ ln(𝐺𝐷𝑃𝑗𝑑) + πœ‘πΆπ‘œπ‘›π‘‘π‘–π‘”π‘–π‘— +
πœŒπΆπ‘œπ‘™π‘–π‘— + πœ€π‘–π‘—π‘‘ (10)
where 𝛾𝑑 are time dummies, and πœ“π‘– and πœ™π‘— are importer and exporter fixed effects
respectively. πΆπ‘ˆπ‘–π‘—π‘‘ is a dummy for both countries being in the same currency union; πΆπ‘œπ‘›π‘‘π‘–π‘”π‘–π‘—
is a dummy for shared borders; and πΆπ‘œπ‘™π‘–π‘— is a dummy for colonial relationship.
More rigorously, one can add a set of bilateral pair fixed effects that controls for all time-
invariant bilateral fixed effects. The equation above becomes:
ln(π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘) = 𝛾𝑑 + 𝛿𝑖𝑗 + πœ‚ ln(𝐺𝐷𝑃𝑖𝑑) + πœ‡ ln(𝐺𝐷𝑃𝑗𝑑) + πœ€π‘–π‘—π‘‘ (11)
where 𝛿𝑖𝑗 are country pair fixed effects.
4.2.2 Constructing the instruments
As variation in bilateral trade due to Home country characteristics could be correlated with
structural change, the instruments should therefore only rely on the variations of trade due
to plausibly exogenous factors such as geography, transportation costs and/or conditions in a
country’s Foreign trading partners. Such a variation could be caused by Foreign joining
GATT/WTO. When a Foreign joins the WTO, bilateral trade costs should fall, leading to a boost
in exports from Home that decreases with distance. The idea is that Foreign’s choice to join
the GATT/WTO is exogenous for Home country. Moreover, the distance between the two
countries should be exogenous. I create two variables, one that weighs for only distance, and
one that weighs for both distance and the Foreign GDP:
π‘Šπ‘‡π‘‚1𝑖𝑗𝑑 =
π‘Šπ‘‡π‘‚πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘› 𝑖𝑗𝑑
ln(π‘‘π‘–π‘ π‘‘π‘Žπ‘›π‘π‘’ 𝑖𝑗)
(12)
π‘Šπ‘‡π‘‚2𝑖𝑗𝑑 =
π‘Šπ‘‡π‘‚πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘› π‘–π‘—π‘‘βˆ—πΊπ·π‘ƒ_πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘› 𝑖𝑗𝑑
ln(π‘‘π‘–π‘ π‘‘π‘Žπ‘›π‘π‘’ 𝑖𝑗)
(13)
where π‘Šπ‘‡π‘‚πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘›π‘–π‘—π‘‘ is a dummy variable for Foreign country being member of GATT/WTO
at or before time t.
The next variable I construct is a variable of trade costs based on the Baltic Dry Index (BDI).
The BDI is a general indicator for the cost of transporting dry bulk cargoes, consisting mainly
of raw commodities. The BDI is thus especially relevant for the trade of low-income countries:
30
π΅π·πΌπ‘π‘œπ‘ π‘‘π‘–π‘‘ = πœƒπ‘–π‘‘ βˆ— ln(𝐡𝐷𝐼𝑑) (14)
where πœƒπ‘–π‘‘ is the share of primary commodities in exports for country i at time t. The idea is
that changes in the BDI matters more for the exports of countries with a larger share of
primary exports (Sim & Lin 2012: 4-5).
To construct each instrument, I regress bilateral exports on either π‘Šπ‘‡π‘‚1𝑖𝑗𝑑, π‘Šπ‘‡π‘‚2𝑖𝑗𝑑 or
π΅π·πΌπ‘π‘œπ‘ π‘‘π‘–π‘‘ in the bilateral equation (11). Since I am not interested in the coefficients for the
bilateral resistance factors, I use bilateral pair fixed effects (𝛿𝑖𝑗) to capture time-invariant
geographical factors:
ln(𝑒π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘—π‘‘) = 𝛾𝑑 + 𝛿𝑖𝑗 + π›½πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘π‘–π‘—π‘‘ + πœ€π‘–π‘—π‘‘ (15)
where ln(𝑒π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘—π‘‘) is the log of exports from Home to Foreign, while πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘π‘–π‘—π‘‘ is either
π‘Šπ‘‡π‘‚1𝑖𝑗𝑑, π‘Šπ‘‡π‘‚2𝑖𝑗𝑑 or π΅π·πΌπ‘π‘œπ‘ π‘‘π‘–π‘‘.
Following standard literature (Frankel & Romer 1999, Feyrer 2009), I aggregate the unlogged
bilateral relationships to arrive at total predicted exports for Home. I similarly sum the actual
exports to arrive at a value for Home’s total actual exports.
π‘π‘Ÿπ‘’π‘‘π‘–π‘π‘‘π‘’π‘‘ 𝑒π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘‘ = βˆ‘ 𝑒 𝛾̂𝑑+ 𝛿 𝑖𝑗
Μ‚ + 𝛽̂ πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘ 𝑖𝑗𝑑
𝑖≠𝑗 (16)
I then take the logs of the aggregated predicted and actual exports. To implement the
instrumental variable regression, I follow the standard two-stage least square regression
(2SLS), using the predicted export values as an instrument for the actual exports in the first
stage regression.
First stage: ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )𝑖𝑑
= πœ‹0 + πœ“Μ‚ln(βˆ‘ 𝑒 𝛾̂ 𝑑+ 𝛿𝑖𝑗
Μ‚+ 𝛽̂ πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘π‘–π‘—π‘‘)𝑖≠𝑗 + πœ’π‘–π‘‘ + πœπ‘–π‘‘ (17)
where πœ‹0 is the intercept, πœ’π‘–π‘‘ is a vector of control variables, and πœπ‘– is the residual. The first
stage decomposes the endogenous variable (ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘‘
)) into two components: a
component (πœπ‘–π‘‘) that can be correlated with the regression error in the first stage, and a
problem-free component that is uncorrelated with the second-stage regression error. The
idea is to use the problem-free component of ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )𝑖𝑑
and disregard πœπ‘– in the second
stage regression. Then I use the predicted value from the first stage regression:
31
ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )Μ‚
𝑖𝑑
= πœ‹Μ‚0 + πœ“Μ‚ln(π‘ƒπ‘Ÿπ‘’π‘‘π‘–π‘π‘‘π‘’π‘‘πΈπ‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )𝑖𝑑
(18)
where πœ‹Μ‚0 and πœ“Μ‚ are OLS estimates. The second stage is then regressing π‘†π‘‘π‘Ÿπ‘’π‘π‘‘π‘’π‘Ÿπ‘Žπ‘™ π‘β„Žπ‘Žπ‘›π‘”π‘’π‘–π‘‘
on ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )Μ‚
𝑖𝑑
Second stage: π‘†π‘‘π‘Ÿπ‘’π‘π‘‘π‘’π‘Ÿπ‘Žπ‘™ π‘β„Žπ‘Žπ‘›π‘”π‘’π‘–π‘‘ = 𝛾0 + 𝛿𝑑 + 𝛼𝑖 + 𝛽ln(𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ Μ‚ )𝑖𝑑 + πœ’π‘–π‘‘ + πœ€π‘–π‘‘ (19)
In this strategy, the instrumented endogenous variable is a country’s export. When
transportation costs go down or a foreign country joins the WTO, trade costs between two
countries decrease in a way that is unrelated to developments in the Home country. I look at
exports rather than total trade because changes in the BDI mainly affect the exports of Sub-
Saharan African countries. Previous studies have shown that trade liberalization has a positive
effect on both export and import growth (Kassim 2013: 12). Thus, the IV analysis measuring
the effect of exogenous changes to exports is directly related to the DID analysis using trade
liberalization episodes. Nevertheless, because domestic trade liberalization mainly involves
reducing the barriers to imports, the results using exports as instruments must be interpreted
differently from the DID analysis using trade liberalization dummy.
In addition to using predicted exports as instruments, I look at the effect of an exogenous
change in imports. Instead of considering what happens to exports from Home to Foreign, I
see what happens to Home’s imports from Foreign when Foreign joins the WTO/GATT. To
construct the instrument I follow the same procedure as above, weighing for distance or
distance and Foreign GDP. I regress Home’s bilateral imports from Foreign on the WTO-
variable, and country pair and time fixed effects:
ln(π‘–π‘šπ‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘—π‘‘) = 𝛾𝑑 + 𝛿𝑖𝑗 + π›½π‘Šπ‘‡π‘‚π‘–π‘—π‘‘ + πœ€π‘–π‘—π‘‘ (20)
Here, π‘Šπ‘‡π‘‚π‘–π‘—π‘‘ is a dummy for Foreign joining WTO/GATT, weighted by either distance, or by
distance and Foreign GDP. I then aggregate the actual and predicted imports. Predicted
imports is then used as an instrument for the actual imports in the first stage regression. By
looking at imports, I can better test the effect of import competition on structural change,
which would help answer hypothesis 1 in Chapter 3.
32
Finally, I construct a remoteness instrument. It is based on the instrument used in Martin et
al. (2008: 890), but different as they construct it as a dyadic variable. I construct the
remoteness variable for each country i the following way:
π‘Ÿπ‘’π‘šπ‘œπ‘‘π‘’π‘›π‘’π‘ π‘ π‘–π‘‘ = βˆ’ ln(βˆ‘
𝐺𝐷𝑃 𝑗𝑑
π‘‘π‘–π‘ π‘‘π‘Žπ‘›π‘π‘’ 𝑖𝑗
𝑅
𝑗≠𝑖 ) (21)
where 𝐺𝐷𝑃𝑗𝑑 is the GDP of i’s trading partner j. The idea is that distance to trade partners (j)
affect the cost of trading, and countries with longer distance to major trading partners trade
less on average. The variable varies over time and across countries since the trade partner’s
GDP changes over time and it affects countries differently depending on the distance to the
trade partner. As this variable is country-specific, it is applied directly into the first-stage
regression, using openness (total trade as share of GDP) as the endogenous variable.
4.3 Identification assumptions
Further scrutiny is required to determine whether the results from the 2SLS regressions show
the causal effect of trade openness on structural change. For an instrument (𝑍𝑖𝑑,) to be valid
it needs to be both relevant and exogenous.
1. Instrument relevance: π‘π‘œπ‘Ÿπ‘Ÿ(𝑍𝑖𝑑, 𝑋𝑖𝑑 β‰  0) (22)
2. Instrument exogeneity: π‘π‘œπ‘Ÿπ‘Ÿ(𝑍𝑖𝑑, πœ€π‘–π‘‘ = 0) (23)
The relevance criteria is met if the first stage F-statistic is above 10, which is the case for all
the instruments used.6
The exogenous criteria states that the instrument can only affect the
outcome variable through its effect on the endogenous variable (𝑋𝑖𝑑). The instrument must
6 I also tried to use the Generalized System of Preferences (GSP) as an instrument. As most African countries qualify for GSP,
introduction of the scheme gives an exogenous boost to trade, weighted by the distance to the provider (Martin et al. 2008:
889). However, the first stage F-statistic turned out to be too low for my analysis. Moreover, Sim and Lin (2012) applied the
BDI instrument directly into the first stage regression, without using the bilateral trade equation. However, applying the
instrument directly in the first stage does not give sufficiently high F-statistic for my sample. One problem is that the BDI is
quite volatile, and the instrument is good at capturing year-by-year changes in the cost of exports. However, averaging the
BDI over 5 years might not give the same effect on 5-year average exports.
33
thus be uncorrelated with the error term (πœ€π‘–π‘‘) in the regression. This is a challenging criterion,
and mostly depends on whether the instrument exogeneity is plausible.
BDI-cost is the instrument that has the strongest claim to exogeneity. Since all the countries
in my sample play a small part in international trade, it is highly unlikely that economic
development in these countries would affect the BDI-index (Lin & Sim 2013: 5). Whether
Foreign joins the WTO/GATT should in theory only affect the Home through its imports and
exports as trade costs go down. However, as many countries simultaneously joined
WTO/GATT, there is a chance that the instrument is correlated to domestic trade reform, and
thus partly endogenous. Moreover, when a Foreign joins the WTO/GATT, it could also affect
its outward flow of FDI given that the TRIMs-agreement provides some protection of FDI. The
Home country could thus be affected through exports, imports and inward FDI. Ideally, I
should then also instrument for inward FDI flows in my IV analysis. However, bilateral FDI
flows are not readily available for my sample, making it difficult to perform this analysis.
Although part of the IV-effect of exports or imports could be due to inward FDI, it would not
have a strong impact on the conclusion, as they are all measures of economic openness.
Finally, the remoteness index is exogenous as long as it only affects exports. Because the
remoteness index is highly dependent on geography, there is a chance that it also captures
regional trends such as development aid, democratization and security that could affect
structural change. However, since the analysis only covers Sub-Saharan Africa, it is unlikely
that there is much difference in regional trends. Nevertheless, like the other instruments it is
possible that part of the effect is due to inward investments as trading partners experience
economic growth.
34
5. Data
5.1 Structural change
The GGDC 10-Sector Database provides the data for my main dependent variable, and
includes information on ten sectors from Asian, Latin American and SSA countries. While
previous studies of structural change and sectorial productivity growth lacked reliable and
detailed data, this database presents an important step forward. The variables on structural
change, agricultural employment share and real mining value added share come from this
dataset.
The dataset includes 11 African, 11 Asian, and 9 Latin-American countries. The eleven SSA
countries7
in the database account for more than half of the total GDP and population in SSA.
The database contains annual data (1960-2010) on the value added by and employment in the
10 main sectors of the economy (see Appendix 1), making it possible to calculate labor
productivity per sector. The definition of employment is broad, including self-employed,
family workers, and other informal workers over 15 years old. To obtain employment data,
the researchers conducted an in-depth study of the available statistical resources on a
country-by-country basis. The database uses the population census to indicate absolute level
of employment, and uses labor force and business surveys to indicate trends in between
censuses (Timmer et al. 2014: 4-7). It is well known that datasets from African countries are
often unreliable due to weak capacity to collect, manage and disseminate data. However,
most of the African countries in the database have considerable experience in collecting
national accounts data and in conducting labor and household surveys. Moreover, growth rate
comparisons, which this paper uses, are more reliable than comparisons of absolute level (De
Vries et al. 2013: 9-10). Nevertheless, despite solid data collection, all results must be treated
with a certain caution.
In addition, there are some issues of missing data. In the African sub-sample, Zambia is missing
data on employment in the Government sector. This is also the case for many other developing
countries, particularly in Latin America. Moreover, some countries have missing data for the
Other sector. To ensure compatibility, the structural change index is thus first created without
7
Botswana, Ethiopia, Ghana, Kenya, Mauritius, Malawi, Nigeria, Senegal, South Africa, Tanzania and Zambia
35
the Government and the Other sector. For robustness checks, I create additional structural
change variables including the Government and the Other sectors. Excluding sectors is
potentially problematic. However, it can be argued that trade liberalization has the most direct
impact on the structure of the private sector, while the government sector is more influenced
by political decisions.
For additional robustness checks, I obtain data on sector value added measured in constant
2005 prices, from UN National Accounts Statistics. This covers almost all countries in Sub-
Saharan Africa from 1970 to 2013. To maximize the number of observations, the five-year
periods start with 1970-1974, and end with 2010-2013. List of sectors and countries are found
in Appendix 2.
5.2 Trade liberalization episodes
A trade liberalization episode is often a continuous process rather than one major reform. In
identifying trade liberalization episodes, I follow the methodology set by Wacziarg & Welch
(2008). In their paper, they update the Sachs-Warner classification of openness that defined
a country as closed if it had one of the following characteristics:
1) Average tariff rates of 40 percent or more
2) Non-tariff barriers covering 40 percent or more of trade
3) A black market exchange rate at least 20 percent lower than the official exchange rate.
This can work as a trade cost, because exporters might have to purchase foreign inputs
using foreign currency obtained in the black market, but remit their foreign exchange
receipts from exports to the government at the official exchange rate.
4) A state monopoly on major exports.
5) A socialist economic system.
In principle, the liberalization date is the date when all of the Sachs-Warner openness criteria
are fulfilled (and not reversed later). However, data limitations often made it necessary to rely
on country case studies of trade policy. A prominent critique against the Sachs-Warner criteria
by Rodrigues and Rodrik (2000) argues that the indicator for the β€œblack market premium”
played an excessive role in classifying a country as closed. Moreover, they argue that this
criterion affected only African countries, and therefore amounted to an Africa dummy.
36
Warcziarg and Welch acknowledge this as a valid critique of the dummy variable in cross-
section analyses, but claim that it is a less valid criticism of the liberalization dates. Policy
changes that reduced the β€œblack market premium” were generally accompanied by other
outward oriented policies, such as reduction of tariff and non-tariff barriers (Warcziarg &
Welch 2008: 190-196). Moreover, for the purpose of this paper, the Africa dummy critique is
irrelevant as I compare only countries within Sub-Saharan Africa. Finally, while the black
market premium is often associated with an overvalued official exchange rate, my analysis
includes controls for overvaluation of the currency.
A challenge for my analysis is to classify which five-year period the liberalization episode
started. The classification follows a simple rule; the trade liberalization dummy is assigned to
the first five-year period where the trade liberalization has lasted at least three years. The
liberalization dummy is then assigned to all periods after. Table 1 shows the liberalization
dates classified by Wacziarg-Welch, my updates, and the five-year period to which I assign the
liberalization start dates.
Table 1. Liberalization dates in SSA
Country Wacziarg-Welch (2003) Updated 5-year periods where
liberalization starts
Botswana 1979 1979 1981-1985
Ethiopia 1996 1996 1996-2000
Ghana 1985 1985 1986-1990
Kenya 1993 1993 1991-1995
Malawi n.l -
Mauritius 1969/1968 1969/1968 1971-1975
Nigeria n.l 2000 2001-2005
Senegal n.l 2003 2001-2005
South Africa 1991 1991 1991-1995
Tanzania 1995 1995 1996-2000
Zambia 1993 1993 1991-1995
Three countries were classified as closed in the Wacziarg-Welch dataset. Senegal because it
continued to maintain a monopoly of exports in the cotton industry (Wacziarg & Welch 2003:
46), Malawi because the black market exchange rate was 28.83 % lower than the official
37
exchange rate, and Nigeria because the black market exchange rate was 151.38 % lower than
the official exchange rate (Wacziarg & Welch 2008: 215). Since the sample in this paper goes
up to 2010, it is necessary to update the dataset.
Senegal phased out is marketing board in the early 2000s, leading Kireyev and Mansoor (2013:
15) to note that Senegal would today have a trade openness index of one. I set the
liberalization date in 2003, when the Senegalese textile company, SODEFITEX, privatized and
ended the State’s monopoly on cotton and textile exports.
Unfortunately, datasets on black market premiums have not been updated since 1998. Black
market premiums (BMP) are very difficult to measure due to the unavailability of public price
data on market exchanges. There is evidence that the BMP has fallen considerably in Nigeria
since 1999. According to Aluko (2007), while it was 275 % higher in 1998, it was 10 % higher
in 1999 and 9 % higher in 2007. The change in BMP coincided with a large devaluation of the
Nigerian Naira. Since Nigeria experienced low BMP uninterrupted between 1999 and 2010, I
will classify Nigeria as open from the 2001-2005 period.
While Malawi began experimenting with a floating exchange rate in 1998, they reverted to a
managed float during 2004-2006 and have had a de facto adjustable currency regime since
2006. Throughout 2006-2011, the government refused to devalue despite significant import
demand and foreign exchange shortages. This led to an overvalued exchange rate and the
development of a vibrant parallel foreign exchange market. At its peak, the foreign currency
was traded at around double the official exchange rate, before Malawi returned to a floating
regime in 2012 (Pauw et al. 2013: 1-5). Malawi has therefore not experienced uninterrupted
trade liberalization since 2000.
I also update the liberalization dates for my expanded dataset (sector share regressions).
Appendix 2 includes a list of countries with their respective liberalization dates.
5.3 Control variables
As the literature review showed, initial conditions matter for the potential for growth-
enhancing structural change. Countries with a larger share of the labor force in agriculture
have a stronger potential for growth-enhancing structural change because labor reallocating
38
away from agriculture will increase average productivity. For each five-year period, the
agricultural labor share measures the employment share of agriculture in the first year. As
natural resource dependency can also influence structural change, the Mining share of VA
measures the real value added share of mining the base year of each five-year period. Both
variables are constructed using data from GGDC. As richer countries generally have less scope
for growth-enhancing structural change than poor countries, GDP per capita in the first year
of each period is included as a control using Penn World Tables 8.1.
Whether the currency is under- or overvalued can affect the rate of structural change
(McMillan et al. 2014). An index of overvaluation is created following Rodrik (2008). To
determine whether the exchange rate is over- or undervalued, one must account for the
Balassa-Samuelson effect, which states that non-tradable goods are cheaper in poorer
countries. The first step is to construct the log of the real exchange rate. Following the Penn
World Tables 8.1, the real exchange rate is calculated by dividing the purchasing power parity
conversion factors (PPP) by the nominal exchange rate (XR).
𝑙𝑛𝑅𝐸𝑅𝑖𝑑 = ln(
𝑃𝑃𝑃 𝑖𝑑
𝑋𝑅𝑖𝑑
) (24)
If 𝑙𝑛𝑅𝐸𝑅𝑖𝑑 is above zero, this indicates that the price level is higher than the price level in the
US in 2005, and vice versa if it is below zero.
The second step is to regress the real exchange rate on GDP per capita and time dummies.
𝑙𝑛𝑅𝐸𝑅𝑖𝑑 = 𝛼 + π›½π‘™π‘›πΊπ·π‘ƒπ‘π‘Žπ‘π‘–π‘‘π‘Ž + 𝛾𝑑 + πœ€π‘–π‘‘ (25)
Finally, to arrive at the undervaluation index, I take the difference between the actual real
exchange rate and the predicted Balassa-Samuelson adjusted rate:
π‘™π‘›π‘‚π‘£π‘’π‘Ÿπ‘£π‘Žπ‘™π‘–π‘‘ = 𝑙𝑛𝑅𝐸𝑅𝑖𝑑 βˆ’ ln 𝑅𝐸𝑅̂𝑖𝑑 (26)
where ln 𝑅𝐸𝑅̂𝑖𝑑 is the predicted value from equation (26) (Rodrik 2008: 371-372). If π‘™π‘›π‘‚π‘£π‘’π‘Ÿπ‘£π‘Žπ‘™
is above zero, it means that the real exchange rate is higher than the real exchange rate
predicted by the country’s GDP per capita, and the real exchange rate is classified as
overvalued. If π‘™π‘›π‘‚π‘£π‘’π‘Ÿπ‘£π‘Žπ‘™ is below zero, the real exchange rate is below what is predicted based
39
on its income level, and the exchange rate is classified undervalued. I obtain data on the real
exchange rate from the Penn World Tables 8.1.
In addition to trade barriers and the exchange rate, the cost of a country’s exports and imports
depends on the Terms of Trade (TOT). I retrieve TOT data from DataMarket, a dataset based
on World Bank staff’s calculations using data from Thomson Reuter’s Datastream. I construct
the terms of trade indicator as an annual percentage change. For the other control variables,
I use Penn World Tables 8.1 to construct the variables on openness, population, and capital
stock per capita. World Development Indicators provide data on the age-dependency ratio,
primary enrollment, and domestic credit to the private sector. I include a variable on the age-
dependency ratio because it can affect labor supply as well as savings and consumption
behavior (Dabla-Norris et al. 2013:10). The financial openness variable is collected from the
Chinn-Ito index. Finally, I get a tariff indicator from Prati et al. (2010). Further descriptions of
the control variables can be found in Appendix 3.
To create the instruments, I use bilateral trade data from various sources. For standard gravity
variables, I use the dataset collected by CEPII. For importer and exporter GDP, I use Penn World
Tables 8.1. For bilateral trade data, I use the World Trade Flows (WTF) collected by Robert
Feenstra. WTO/GATT dummies come from Rose (2004), and were subsequently updated by
the author. Data on the Baltic Dry Index comes from Lin & Sim (2014). Primary share variables
were constructed using data from both UNCTAD (1995-2010) and NBER (1971-1994). I
construct two different primary shares. The first consists of all primary goods in exports
including fuels.8
The second excludes all fuels except for coal.9
A problem with the bilateral
trade and export data is that they do not cover Botswana before 1991. The panels will
therefore not be fully balanced when using these variables.
8
SITC Rev. 3 codes: 0 + 1 + 2 + 3 + 4 + + 67 + 68
9
SITC Rev. 3 codes: 0 + 1 + 2 + 32 + 4 + + 67 + 68
40
6. Results
6.1Descriptive statistics
Figure 1 below shows the relationship between structural change and changes in liberalization
in SSA over time (1971-2010). The solid blue line is the non-weighted five-year average
contribution of structural change in SSA (left y-axis), while the dashed red line shows the share
of countries undertaking trade liberalization within each five-year period (right y-axis).
Structural change was strong in the early 1970s, before falling steadily until its low-point in
the mid-1990s when structural change had a negative effect on growth. From the mid-1990s,
structural change again made a positive contribution to growth. SSA countries started slowly
to liberalize in the early 1980s, before there was a strong liberalization surge in the late 1980s
and early 1990s, following structural adjustment programs. From the graph, we can see that
a period of stronger structural change followed the large number of liberalization episodes in
the 1990s. By 2006, only Malawi in my sample had not liberalized.
Figure 1. Relationship between structural change and liberalization episodes
In Figure 2, the dashed blue line shows the average contribution of structural change to
growth for liberalized countries, while the solid red line shows the structural change
contribution for non-liberalized countries. Structural change was higher on average in
liberalized countries from the mid-1980s to the early 2000s, while it was higher in non-
liberalized countries at the beginning and the end of the period. Structural change seems to
be more stable and on average higher in liberalized countries. Between 1971 and 2010, the
41
average structural change contribution to growth was 0.91 % for non-liberalized countries,
and 1.34 % for liberalized countries.
Figure 2. Structural change for liberalized and non-liberalized countries
Interestingly, structural change is stable in countries with an undervalued exchange rate, while
it is highly volatile in in countries with overvalued exchange rate (see Appendix 5). Summary
statistics for the most important variables by their value on the liberalization dummy are
found in Appendix 4. Graphs of structural change in individual countries are also available in
Appendix 6.
Looking at specific sectors, Figure 3 below shows the evolution of productivity in the different
sectors. On average, mining has been the most productive sector by far, but utilities and
finance has caught up during the past decade. Manufacturing has had a fairly stable and
mediocre productivity over time. The low productivity in manufacturing can be explained by
the inclusion of informal firms in the sector. The agriculture and government sectors are by
far the least productive sectors.
42
Figure 3. Sectoral productivity over time
Finally, Figure 4 below shows the evolution of sector labor shares. Agriculture is by far the
largest sector. To facilitate the identification of changes in the other sectors, agriculture has
its own y-axis (left). The most evident development is the fall of the agriculture labor share
(blue line). Wholesale and retail trade (red line) is the sector that has absorbed the largest
part of the workers from agriculture. The manufacturing share (dark brown line) grew a little
in the 1970s before dipping in the 1980s, rebounding slightly, and then stagnating. The most
productive sector, mining (dark green), has a very low labor share that has declined further
since the 1990s.
Figure 4. Sectoral labor share over time
43
6.2Difference in differences
Although some SSA countries have structural change data going back to 1961, I look at the
period 1971-2010 in order to have a balanced panel. I also look at the 1981-2010 sub-sample.
The results of the difference-in-differences regressions are shown in Table 2.
Table 2. Difference in differences
(1) (2) (3) (4) (5) (6)
71-10 71-10 71-10 81-10 81-10 81-10
Dependent variable Structural change contribution to growth
Liberalization dummy 0.0229 0.0295* 0.0260* 0.0312** 0.0266** 0.0220*
(0.0205) (0.0133) (0.0127) (0.0136) (0.0110) (0.00996)
GDP per capita 0.0488** 0.0342 0.0299* 0.0160 0.0157
(0.0155) (0.0189) (0.0140) (0.0118) (0.00937)
Agriculture labor share 0.176*** 0.127*** 0.220*** 0.152*** 0.165***
(0.0235) (0.0266) (0.0262) (0.0272) (0.0253)
Exports primary share 0.0254 -0.00391 -0.0674**
(0.0275) (0.0310) (0.0272)
Overvalued XR index -0.0348* -0.0436*** -0.0491***
(0.0182) (0.0126) (0.0119)
Observations 88 84 84 66 66 64
R-squared 0.121 0.358 0.387 0.474 0.540 0.575
Number of countries
Country & Time FE
Cluster SE by country
11
Yes
Yes
11
Yes
Yes
11
Yes
Yes
11
Yes
Yes
11
Yes
Yes
11
Yes
Yes
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
In column (1), only the liberalization dummy is included as an explanatory variable. Although
positive, it is not significant. As seen by the R-square of 0.121, this regression explains little of
the variation in structural change. However, the liberalization dummy is significant at the 10
% level when I add control variables in column (2). Specifically, the liberalization dummy is
only significant when I include the primary share of exports as a control. A limitation of this
control variable is that it does not cover Botswana before 1991, leading to an unbalanced
panel. There is thus the possibility of a sampling effect biasing the results.
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid
Thesis Eivind Breidlid

More Related Content

Similar to Thesis Eivind Breidlid

AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS
AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS
AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS
Dr Lendy Spires
Β 
Final Thesis - Catherine Mahony (11377841)
Final Thesis - Catherine Mahony (11377841)Final Thesis - Catherine Mahony (11377841)
Final Thesis - Catherine Mahony (11377841)
Katie Mahony
Β 
Ellen mac arthur-foundation-towards-the-circular-economy-vol.1
Ellen mac arthur-foundation-towards-the-circular-economy-vol.1Ellen mac arthur-foundation-towards-the-circular-economy-vol.1
Ellen mac arthur-foundation-towards-the-circular-economy-vol.1
Amalia Minguzzi
Β 
Ba-2130106-R.S.Hunter
Ba-2130106-R.S.HunterBa-2130106-R.S.Hunter
Ba-2130106-R.S.Hunter
RenΓ©e Hunter
Β 
Master's Thesis Ivan Sabato -- LinkedIn
Master's Thesis Ivan Sabato -- LinkedInMaster's Thesis Ivan Sabato -- LinkedIn
Master's Thesis Ivan Sabato -- LinkedIn
Ivan Sabato
Β 
Transforming the-apec-outcome-on-eg-into-seti ictsd
Transforming the-apec-outcome-on-eg-into-seti ictsdTransforming the-apec-outcome-on-eg-into-seti ictsd
Transforming the-apec-outcome-on-eg-into-seti ictsd
Dr Lendy Spires
Β 
WEF_FinancialDevelopmentReport_2010
WEF_FinancialDevelopmentReport_2010WEF_FinancialDevelopmentReport_2010
WEF_FinancialDevelopmentReport_2010
Eva Gustavsson
Β 
Cluster and Entrepreneurship in Emerging Industries
Cluster and Entrepreneurship in Emerging IndustriesCluster and Entrepreneurship in Emerging Industries
Cluster and Entrepreneurship in Emerging Industries
Gerd Meier zu Koecker
Β 
Ellen mac arthur foundation towards the circular economy vol.1
Ellen mac arthur foundation towards the circular economy vol.1Ellen mac arthur foundation towards the circular economy vol.1
Ellen mac arthur foundation towards the circular economy vol.1
Glenn Klith Andersen
Β 
Jules Arntz-Gray MA Thesis
Jules Arntz-Gray MA ThesisJules Arntz-Gray MA Thesis
Jules Arntz-Gray MA Thesis
Jules Arntz-Gray
Β 
drivers-low-carbon-development-china-industrial-zones-en
drivers-low-carbon-development-china-industrial-zones-endrivers-low-carbon-development-china-industrial-zones-en
drivers-low-carbon-development-china-industrial-zones-en
Lisa Muirhead
Β 
Balkans’s Agriculture Value Chain. Current Point of View
Balkans’s Agriculture Value Chain.  Current Point of ViewBalkans’s Agriculture Value Chain.  Current Point of View
Balkans’s Agriculture Value Chain. Current Point of View
International Journal of Economics and Financial Research
Β 
Interconnected Economies: Benefitting from Global Value Chains
Interconnected Economies: Benefitting from Global Value ChainsInterconnected Economies: Benefitting from Global Value Chains
Interconnected Economies: Benefitting from Global Value Chains
Dr Lendy Spires
Β 
363-wcms_366005.pdf
363-wcms_366005.pdf363-wcms_366005.pdf
363-wcms_366005.pdf
Mideksa1
Β 

Similar to Thesis Eivind Breidlid (20)

AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS
AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS
AIDFORTRADE AT A GLANCE 2013 CONNECTING TO VALUE CHAINS
Β 
Final Thesis - Catherine Mahony (11377841)
Final Thesis - Catherine Mahony (11377841)Final Thesis - Catherine Mahony (11377841)
Final Thesis - Catherine Mahony (11377841)
Β 
Ellen mac arthur-foundation-towards-the-circular-economy-vol.1
Ellen mac arthur-foundation-towards-the-circular-economy-vol.1Ellen mac arthur-foundation-towards-the-circular-economy-vol.1
Ellen mac arthur-foundation-towards-the-circular-economy-vol.1
Β 
Driving Transformative Change: The Role of the Private Sector
Driving Transformative Change: The Role of the Private SectorDriving Transformative Change: The Role of the Private Sector
Driving Transformative Change: The Role of the Private Sector
Β 
Ba-2130106-R.S.Hunter
Ba-2130106-R.S.HunterBa-2130106-R.S.Hunter
Ba-2130106-R.S.Hunter
Β 
Master's Thesis Ivan Sabato -- LinkedIn
Master's Thesis Ivan Sabato -- LinkedInMaster's Thesis Ivan Sabato -- LinkedIn
Master's Thesis Ivan Sabato -- LinkedIn
Β 
Transforming the-apec-outcome-on-eg-into-seti ictsd
Transforming the-apec-outcome-on-eg-into-seti ictsdTransforming the-apec-outcome-on-eg-into-seti ictsd
Transforming the-apec-outcome-on-eg-into-seti ictsd
Β 
WEF_FinancialDevelopmentReport_2010
WEF_FinancialDevelopmentReport_2010WEF_FinancialDevelopmentReport_2010
WEF_FinancialDevelopmentReport_2010
Β 
Informe de desarrollo financiero 2010 wef
Informe de desarrollo financiero 2010  wefInforme de desarrollo financiero 2010  wef
Informe de desarrollo financiero 2010 wef
Β 
Cluster and Entrepreneurship in Emerging Industries
Cluster and Entrepreneurship in Emerging IndustriesCluster and Entrepreneurship in Emerging Industries
Cluster and Entrepreneurship in Emerging Industries
Β 
Ellen mac arthur foundation towards the circular economy vol.1
Ellen mac arthur foundation towards the circular economy vol.1Ellen mac arthur foundation towards the circular economy vol.1
Ellen mac arthur foundation towards the circular economy vol.1
Β 
World FZO Global Value Chain Study
World FZO Global Value Chain StudyWorld FZO Global Value Chain Study
World FZO Global Value Chain Study
Β 
World FZO Global Value Chain Study
World FZO Global Value Chain StudyWorld FZO Global Value Chain Study
World FZO Global Value Chain Study
Β 
Jules Arntz-Gray MA Thesis
Jules Arntz-Gray MA ThesisJules Arntz-Gray MA Thesis
Jules Arntz-Gray MA Thesis
Β 
drivers-low-carbon-development-china-industrial-zones-en
drivers-low-carbon-development-china-industrial-zones-endrivers-low-carbon-development-china-industrial-zones-en
drivers-low-carbon-development-china-industrial-zones-en
Β 
Service innovationyearbook 2009-2010
Service innovationyearbook 2009-2010Service innovationyearbook 2009-2010
Service innovationyearbook 2009-2010
Β 
Balkans’s Agriculture Value Chain. Current Point of View
Balkans’s Agriculture Value Chain.  Current Point of ViewBalkans’s Agriculture Value Chain.  Current Point of View
Balkans’s Agriculture Value Chain. Current Point of View
Β 
Cash managment thesis
Cash managment thesisCash managment thesis
Cash managment thesis
Β 
Interconnected Economies: Benefitting from Global Value Chains
Interconnected Economies: Benefitting from Global Value ChainsInterconnected Economies: Benefitting from Global Value Chains
Interconnected Economies: Benefitting from Global Value Chains
Β 
363-wcms_366005.pdf
363-wcms_366005.pdf363-wcms_366005.pdf
363-wcms_366005.pdf
Β 

Thesis Eivind Breidlid

  • 1. Sciences Po PSIA – Paris School of International Affairs Master in International Economic Policy Trading up or trading down? The effect of trade openness on structural change in Sub-Saharan Africa Eivind Breidlid Master’s thesis supervised by Sergei Guriev, Professor of Economics Academic Year 2015/2016
  • 2. 1 Acknowledgements I would first sincerely thank my supervisor Sergei Guriev for his guidance, quick feedback, and very helpful suggestions. His input and advice has been invaluable in this process. Furthermore, I would like to thank Dr. Nicholas Sim who graciously shared his data on the Baltic Dry Index. I would also like to thank Bendik Elstad for all our many discussions over a cup of coffee outside Bibliotheque de l’École doctorale. Finally, my deepest gratitude goes to Elyse, who not only applied her brilliant mind to proofread this paper, but also gave me tremendous support throughout the process.
  • 3. 2 Abstract The purpose of this paper is to investigate the effect of trade openness on structural change in Sub-Saharan Africa. Structural change happens when labor reallocates from low-productive to high-productive sectors. The effect of trade openness on structural change is uncertain in theory. By definition, trade prompts the reallocation of resources across sectors. On the one hand, trade liberalization allows for the expansion of the most productive firms at the expense of the least productive firms, which raises aggregate productivity. On the other hand, trade liberalization leads to specialization in comparative advantage sectors, which for Sub-Saharan Africa is mainly in low- productive agriculture. To identify the effect of trade openness empirically, I use a panel dataset on sector employment and value added to conduct difference-in-differences (DID) and instrumental variable (IV) analyses. I find that trade openness has a positive effect on structural change in Sub-Saharan Africa, and that this result is quite robust to different tests. Moreover, an overvalued exchange rate was found to have a negative effect on structural change. The most plausible transmission channel for trade openness is the income effect, where trade liberalization leads to increased income, which again leads to more expenditure on services instead of agricultural products. Therefore, structural change in Sub-Saharan Africa is driven by an expansion of services rather than manufacturing. The findings of this paper lead to two big questions; first, whether services can be a dynamic growth sector in Sub-Saharan Africa like manufacturing has been for other developing countries in the past; and second, whether manufacturing can play a bigger part in the future as opportunities for integration into global value chains emerge.
  • 4. 3 Table of contents Acknowledgements................................................................................................................................1 Abstract..................................................................................................................................................2 Table of contents....................................................................................................................................3 List of tables ...........................................................................................................................................4 List of figures..........................................................................................................................................5 List of tables and figures in the appendix...............................................................................................5 1. Introduction ...................................................................................................................................6 2. Literature review................................................................................................................................8 2.1 Structural transformation.............................................................................................................8 2.2 Stylized facts about structural transformation.............................................................................9 2.2.1 Patterns of structural transformation ...................................................................................9 2.2.2 Growth-enhancing structural change..................................................................................10 2.3 Trade and other determinants of structural transformation......................................................11 2.4 Empirical work............................................................................................................................15 3. Theoretical framework and empirical hypotheses ...........................................................................18 3.1 Growth effect of sector reallocation ..........................................................................................18 3.2 Hypotheses.................................................................................................................................19 4. Empirical strategy.............................................................................................................................22 4.1 Baseline regression.....................................................................................................................22 4.1.1 The dependent variable.......................................................................................................22 4.1.2 Difference in differences .....................................................................................................24 4.1.3 Methodological challenges..................................................................................................25 4.1.4 Sector contribution and sector shares.................................................................................26 4.2 Instrumental variables regression ..............................................................................................28 4.2.1 The gravity equation............................................................................................................28 4.2.2 Constructing the instruments..............................................................................................29 4.3 Identification assumptions .........................................................................................................32 5. Data..................................................................................................................................................34 5.1 Structural change .......................................................................................................................34 5.2 Trade liberalization episodes......................................................................................................35 5.3 Control variables ........................................................................................................................37 6. Results..........................................................................................................................................40 6.1 Descriptive statistics.............................................................................................................40 6.2 Difference in differences ......................................................................................................43 6.3 Robustness checks................................................................................................................47
  • 5. 4 6.3.1 Alternative dependent variables .........................................................................................47 6.3.2 Alternative trade openness variables ..................................................................................48 6.3.3 Sensitivity analysis...............................................................................................................52 6.3.4 Two-period regression.........................................................................................................56 6.4 Sector contribution...............................................................................................................57 6.5 Sector shares as dependent variables ..................................................................................61 6.6 Instrumental variables regression ........................................................................................63 7. Discussion.........................................................................................................................................67 7.1 Does trade openness promote structural change? ....................................................................67 7.2 Why does trade liberalization lead to structural change? ..........................................................68 7.2.1 The agricultural channel ......................................................................................................68 7.2.2 The value chain channel ......................................................................................................69 7.2.3 The role of the mining sector ..............................................................................................70 7.3 Role of the exchange rate...........................................................................................................72 8. Concluding remarks..........................................................................................................................75 Bibliography .........................................................................................................................................77 A. Articles, books, journals ...........................................................................................................77 B. Datasets....................................................................................................................................80 Appendices...........................................................................................................................................81 List of tables Table 1. Liberalization dates in SSA.....................................................................................................36 Table 2. Difference in differences.........................................................................................................43 Table 3. Difference in differences with interactions.............................................................................45 Table 4. Tariff indicator as explanatory variable.................................................................................49 Table 5. Openness as explanatory variable (1) ....................................................................................50 Table 6. Openness as explanatory variable (2) ....................................................................................51 Table 7. Placebo liberalization dates ...................................................................................................52 Table 8. Analysis excluding Nigeria.....................................................................................................54 Table 9. Two-period regression ...........................................................................................................57 Table 10. Sector contribution 1971-2010.............................................................................................57 Table 11. Sector contributions as dependent variable..........................................................................58 Table 12. Mining contribution as dependent variable..........................................................................60 Table 13. Agriculture value added share as dependent variable..........................................................61 Table 14. Two-period, Agriculture value added share .........................................................................62 Table 15. Bilateral trade equation .......................................................................................................64 Table 16. Second stage regression .......................................................................................................65
  • 6. 5 List of figures Figure 1. Relationship between structural change and liberalization episodes ....................................40 Figure 2. Structural change for liberalized and non-liberalized countries.............................................41 Figure 3. Sectoral productivity over time..............................................................................................42 Figure 4. Sectoral labor share over time...............................................................................................42 Figure 5. lvr2plot SSA 1971-2010..........................................................................................................53 Figure 6. lvr2plot SSA 1981-2010..........................................................................................................53 Figure 7. Added variable plot SSA 1971-2010 without Nigeria.............................................................56 Figure 8. lvr2plot SSA 1971-2010 without Nigeria................................................................................56 List of tables and figures in the appendix Appendix 1. Sectors in the GGDC database.........................................................................................81 Appendix 2: List of sectors and countries in the UN National Account Statistics.................................82 Appendix 3. Description of control variables.......................................................................................84 Appendix 4. Summary statistics............................................................................................................85 Appendix 5. Structural change growth with over- and undervalued currencies ...................................85 Appendix 6. Structural change and liberalization over time in individual SSA countries.....................86 Appendix 7. Alternative measures of structural change .......................................................................88 Appendix 8. Sector contribution to SC growth over time......................................................................90 Appendix 9. Industry and Services contribution to SC .........................................................................90 Appendix 10. Change in Manufacturing and Mining value added share as dependent variable..........91 Appendix 11. Service and Industry value added as dependent variable ...............................................92 Appendix 12. First stage regressions....................................................................................................93 Appendix 13. Discussion chapter regressions ......................................................................................94
  • 7. 6 1. Introduction Africa needs structural transformation not structural adjustment Carlos Lopez, Executive Director UNECA1 Economic growth is essential to reduce poverty and improve living standards in developing countries. While high economic growth in East and South Asia has led to an unprecedented reduction in poverty, growth in Sub-Saharan Africa (SSA) has been disappointing, with a high incidence of negative growth rates in the 1980s and 1990s. However, since 2000 the average per capita annual growth in SSA has been close to 3 percent. It is however, argued that the high growth rates the past decade are due to a favorable external environment leading to high demand for the region’s natural resources. There are therefore concerns that SSA countries will not be able to sustain high growth rates when the benign external environment shifts. For the growth to be sustainable, it is essential that the recent boom translated into structural transformation and diversified economic activities (Rodrik 2014: 1-2, 5-7). Aggregate labor productivity growth happens either because of productivity growth within sectors, or because high-productive sectors employ a larger share of the labor force. Labor reallocation between sectors with different productivity levels is known as structural transformation or structural change. This paper defines growth-enhancing structural change as the reallocation of labor from low-productive to high-productive sectors.2 Large productivity gaps between sectors are characteristic of low-income countries. The agricultural sector is simultaneously the least productive sector and the largest employer in the economy. Countries that start with a large share of their labor force in low-productive agriculture therefore have a large potential for economic growth through structural transformation. Most Sub-Saharan African countries followed import-substitution policies in the 1960s and 1970s, aiming to nurture a domestic industry by protecting it from foreign competition. However, following the debt crisis in the early 1980s, SSA countries accepted financial aid from the World Bank and the IMF on the condition that they liberalize their trade regimes. Hence, 1 Lopez, Carlos (2013). β€œAfrica needs structural transformation not structural adjustment”, The Executive Secretary’s Blog 2 If not specified otherwise, the term β€œstructural change” means growth-enhancing structural change.
  • 8. 7 from the mid-1980s to the early 1990s, SSA countries undertook trade liberalization reforms under so-called Structural Adjustment Programs (Kassim 2013). There is a large body of literature on the effect of trade liberalization on growth and incomes (e.g. Feyrer 2009). Nevertheless, the literature on the effect of trade liberalization on the structural change component of growth is limited. As trade openness induces countries to specialize in the products for which they have a comparative advantage, trade liberalization naturally leads to reallocation of production factors across sectors. However, if the comparative advantage of SSA countries lies in low-productive sectors such as agriculture, the effect on structural change can be negative. This study therefore asks the following questions: What is the effect of trade openness on structural change, and through which mechanisms does the effect work? To answer these questions, I will use a panel dataset on employment and value added for 11 SSA countries. The main identification strategies will be difference-in-differences (DID) regression and instrumental variable (IV) analysis. I structure the thesis as follows. In Chapter 2, I present some important stylized facts on structural change, before discussing the main theories on trade openness and structural transformation. The chapter concludes with some empirical findings from the literature. Chapter 3 then describes the theoretical framework including the hypotheses to be tested. In Chapter 4, I explain the empirical strategy of using difference-in-differences and instrumental variables analysis. Chapter 5 describes the datasets and variables used in the empirical analysis, before the empirical results are presented in Chapter 6. In Chapter 7, I discuss the results in the context of my theoretical framework and hypotheses, before Chapter 8 concludes.
  • 9. 8 2. Literature review This chapter first describes the relationship between structural transformation and economic development. I then present stylized facts on the evolution of structural change over time and across regions. In the second part of the chapter, I will discuss theories on the relationship between trade openness and structural change. Finally, I will show some empirical results from the literature on the determinants of structural change. 2.1 Structural transformation Economists have long recognized the importance of structural transformation in economic development. Arthur Lewis recognized structural change as key to economic development. As the capitalist (modern) sector expands, surplus labor is drawn from the mainly rural subsistence sector. This process of economic development continues until the surplus labor disappears, as wages between the modern and traditional sectors equalize (Lewis 1954: 8). Simon Kuznets noted that economic growth is characterized by a high rate of structural transformation, from agriculture to non-agricultural activities, and later from industry to services (Kuznets 1973: 248). More recently, Restuccia et al. (2008) find that low agricultural productivity and a high share of labor devoted to agriculture are the main explanations behind the large income differences between rich and poor countries (Restuccia et al. 2008: 235). Reallocating labor towards more productive activities would hence have a major impact on economic growth. Moreover, amidst concern that growth in SSA has failed to create sufficient productive employment opportunities in order to absorb the growing labor force (ILO 2014, World Bank 2014), structural change is important because it relates to the component of growth caused by increasing employment shares in high-productive sectors. Page (2015) finds evidence that growth originating from structural change is more effective at reducing poverty than productivity growth within sectors. This is because growth caused by structural change involves low-skilled workers gaining employment in sectors with higher productivity and hence higher wages (Page 2015: 230-233).
  • 10. 9 2.2 Stylized facts about structural transformation 2.2.1 Patterns of structural transformation Herrendorf et al. (2013) investigate the evolving relationship between economic development and sectoral shares of employment and value added, using historical time series for currently rich countries, and panel data for a broader set of countries. An increase in GDP per capita is associated with a continuous decrease in agriculture’s employment and value added shares, and a continuous increase in the service sector’s employment and value added shares. Manufacturing, on the other hand, follows an inverse U-shaped development. At low levels of development, income growth is accompanied by a rapid expansion of manufacturing shares, before the relationship flattens. At high levels of income, manufacturing shares decrease with income growth. The value added and employment shares of services accelerate around the same income level manufacturing shares peak, which is about US$8000 in 1990 international dollars (Herrendorf et al. 2013: 9-16). Rodrik (2015) shows that countries’ manufacturing shares are increasingly peaking at earlier stages of development. Controlling for GDP per capita and population, the average country’s manufacturing employment share is 11.7 percentage points lower post-2000 than in the 1950s, and 8.8 percentage points lower than in the 1960s. The only region without a negative time trend in manufacturing employment controlling for income and population is Asia. Moreover, countries’ manufacturing shares peak at a lower level of income. Since 1990, countries have reached peak manufacturing at incomes that are around 40 % of the levels experienced before 1990 (Rodrik 2015: 8-15). Felipe et al. (2015) find that global productivity growth in manufacturing has been similar to non-manufacturing between 1970 and 2010. Although manufacturing productivity grew faster within countries, this was offset by a reallocation of manufacturing jobs to less productive countries. Furthermore, because manufacturing shifted to countries with more labor-intensive production, the manufacturing employment share has not declined globally. Nevertheless, manufacturing productivity growth has accelerated the last decades. Since 1990, it has been 50 % higher than aggregate productivity growth, coinciding with the increased contribution from China and other East- Asian countries. As global manufacturing employment is no longer expanding, industrialization is more difficult for low-income countries today because of growing
  • 11. 10 international competition from mainly labor-abundant Asian countries (Felipe et al. 2015: 11- 23). De Vries et al. (2013) analyze the evolution of structural change in Sub-Saharan Africa. They find that rate of structural change was strong between 1960 and 1975, with a large decrease in agricultural employment share, and a corresponding strong increase in manufacturing’s share of employment and value added. Despite the strong increase in employment share, manufacturing managed to increase their relative productivity from 1960 to 1975, showing few signs of decreasing marginal productivity. The period between 1975 and 1990, however, showed negative growth and very slow reallocation of labor from agriculture. The 20 years following 1990 have been a period of rapid reallocation of labor across sectors. Agricultural employment share fell from 61.6 % in 1990 to 49.8 % in 2010. However, this reallocation of labor from agriculture did not benefit the manufacturing sector, as its share of employment decreased slightly. Market services on the other hand nearly doubled their share of employment from 24.1 % to 36.8 %. Unfortunately, the relative productivity of market services fell considerably, indicating low marginal productivity (De Vries et al. 2013: 10-12). 2.2.2 Growth-enhancing structural change The authors then analyze whether labor reallocated from low-productive to high-productive sectors (static structural change), and whether labor reallocated from low productivity growth to high productivity growth sectors (dynamic structural change). They find that Sub-Saharan Africa experienced a strong reallocation of labor from low-productive to high-productive sectors between 1960 and 1975, while the reallocation gains were small between 1975 and 1990. Between 1990 and 2010, labor moved from lower to higher productive sectors, producing a static growth effect. At the same time, labor reallocated to sectors that experienced lower productivity growth, resulting in a large negative dynamic reallocation effect. Thus, the aggregate contribution from structural change to growth is thus negligible in this period. These results point to low marginal productivity in the expanding sectors. For example, reallocation of labor to market services has a strong static effect on labor productivity (1 %), but this positive impact is offset by the negative dynamic effects (-0.96 %) (De vries et al. 2013: 15-22).
  • 12. 11 McMillan et al. (2014) find that between 1990 and 2005, structural change made a negative contribution to productivity growth in both Sub-Saharan Africa and Latin America. Meanwhile, structural change made a positive contribution to growth in Asia, explaining most of the productivity growth gap between Asia and the two other regions. This is disappointing considering the large structural change potential in Sub-Saharan Africa, with its high share of labor in low-productive agriculture. The effects are not, however, uniform across the region. Ghana, Ethiopia and Malawi all experienced growth-enhancing structural change as the employment share in agriculture fell, while manufacturing employment increased. In Nigeria and Zambia, on the other hand, manufacturing and tradable services contracted, while the employment share in agriculture increased (McMillan et al. 2014: 18-23). While structural change is negative between 1990 and 1999, structural change made a positive contribution to labor productivity after 2000. Nigeria and Zambia then had an expansion in manufacturing and a contraction in agriculture and services. Overall, half of the African countries in the sample experienced an increase in the employment share in manufacturing post-2000, although the size of the changes is not enough to transform these economies rapidly (McMillan et al. 2014: 23-25). 2.3 Trade and other determinants of structural transformation Fundamentally, international trade allows sectoral expenditure shares to deviate from sectoral production. This allows each country to run a net export surplus in its sector of comparative advantage. Thus, patterns of specialization induced by trade openness directly affect the labor share. Moreover, as trade affects relative prices, sectoral expenditure shares change, shifting sectoral production and labor shares further (Uy et al. 2013: 668). In classic models of trade, the movement of labor and capital across sectors are what allows countries to reap the benefits of openness. Countries gain from trade by moving resources to their comparative advantage sectors; which is defined by relative technological differences in the Ricardian model or by relative factor endowments according to the Heckscher-Ohlin model. Models with increasing returns to scale predict geographical agglomeration of production, which can lead to observable reallocation of labor across sectors. However, new trade models that focus on intra-industry trade do not have clear predictions on labor
  • 13. 12 reallocation across broad economic sectors, as specialization due to intra-industry trade might be highly disaggregated within economic sectors (Wacziarg & Wallack 2004: 411-412). In the Melitz model of heterogeneous firms, trade openness leads to both export opportunities and import competition. The most productive firms take advantage of the export opportunities to expand their production. As profits increase, the most productive firms demand more labor, thereby driving up wages. The least productive firms shrink or exit due to increased product market and labor market competition that reduces their profits (Melitz 2002: 23-26, Janiak 2006: 2). Openness therefore leads to reallocation of labor from the least to the most productive firms. Theoretical studies of structural transformation focus on two economic mechanisms that drive the reallocation of labor across sectors. One type of models focuses on demand factors, with income effects driving the process of structural transformation. The other main type of models focuses on supply factors, as differential productivity growth drives changes in relative prices and sectoral labor allocations (Dabla-Norris et al. 2013: 5). Herrendorf et al. (2013) present a three-sector model consisting of agriculture, manufacturing and services. In the model preferences are non-homothetic, meaning that the relative share of consumption does not stay constant as income changes. Specifically the income elasticity is below one for agriculture, constant for manufacturing and above one for services. Keeping relative prices constant, an increase in income will lead to more expenditure devoted to services at the expense of agriculture. This leads to corresponding changes in sectoral employment and value added shares. Alternatively, structural change is driven by changes in relative prices, due to differential technological progress across sectors. The sector that has the highest productivity growth will experience a relative price decrease. Assuming price elasticity of demand is below one (inelastic), the quantity increase in demand will not fully compensate the fall in price, reducing the nominal value added of that sector. As the sector can produce the same amount of goods with fewer workers, labor shifts away from this sector. Correspondingly, the sector with the lowest productivity growth will experience higher relative prices, and increase their share of employment and nominal value added share (Herrendorf 2013: 41-46).
  • 14. 13 In the absence of trade, income growth will lead to a higher share of employment in services, and the sector with the highest productivity growth will experience a declining share of employment. However, Matsuyama (2009) introduces international trade where countries’ comparative advantage follows relative sector-productivities (Ricardian specialization). If the Home country has larger productivity growth in manufacturing than in services, labor should shift to services. However, if productivity growth in manufacturing is higher in Home country than abroad, comparative advantage will shift and increase manufacturing employment at Home. Total manufacturing employment will fall, and manufacturing employment abroad will unambiguously decline. The overall effect on manufacturing employment at Home is ambiguous. While the trade effect boosts manufacturing employment, the relative price effect is negative. Allowing for openness can help explain the finding that countries with low productivity growth in manufacturing are experiencing a stronger fall in manufacturing employment than countries with high productivity growth (Matsuyama 2009: 482-484). Uy et al. (2013) include evolving trade costs in the model. In the model, the manufacturing and agriculture sectors are tradable, while the service sector is non-tradable. Opening the economy up to trade has three main impacts on structural change. First, declining trade costs will lead to net exports in the comparative advantage sector. Hence, labor will move from its comparative disadvantage sector to its comparative advantage sector. As trade costs decline over time, each country’s comparative advantage is increasingly revealed, leading to increased specialization and increased labor reallocation to the comparative advantage sector. Second, a country with relative higher productivity growth within a sector will supply an increasing share of the world demand; this counteracts the negative effect of relative prices on the labor share caused by high productivity growth. However, as productivity increases and export demand slows in the long run, the relative price effect dominates the trade effect, leading to a decline in domestic manufacturing employment. This could explain the non-linear relationship between income and manufacturing employment share previously discussed. Third, lower trade costs leads to higher income growth, which in turn reinforces the effect of non-homothetic preferences on labor reallocations, leading to a stronger reallocation of labor from agriculture to services (Uy et al. 2013: 671-673). Using a similar framework, Rodrik (2015) notes that most developing countries are pure price takers on global markets. If a small open economy has high productivity growth in the
  • 15. 14 manufacturing sector, the relative price of manufacturing goods will not fall, and the effect on manufacturing employment and value added shares is unambiguously positive. However, as price takers, developing countries also import relative price declines caused by relative high productivity growth in the manufacturing sectors of other countries. For manufacturing employment to expand in developing countries, productivity growth differentials between manufacturing and non-manufacturing, must not only be positive, but also need to exceed the decline in relative prices on the world market. Higher productivity growth in manufacturing than non-manufacturing does therefore not directly lead to increased share of manufacturing in the economy. Only the countries with the highest productivity growth globally will expand their labor force in the manufacturing sector (Rodrik 2015: 15-22). One clear prediction from these models is that the labor shares of sectors that produce tradable goods should differ across countries that have different sectoral productivities. Countries reallocate labor to the sector where they have a comparative advantage, and continue this process as long as productivity growth is sufficiently high and global demand is not saturated. However, comparative advantage is not static and labor could reallocate to new sectors with relative high productivity growth. A decline in the relative price of manufacturing experienced by small developing countries could be the result of technological progress elsewhere, large increase in supply (e.g., the rise of China), or domestic trade liberalization. Countries that lag behind global productivity growth in manufacturing could experience a reallocation of labor away from manufacturing when opening up to trade. This could explain why countries in Sub-Saharan Africa have failed to expand their manufacturing sector the last 20 years, as the region’s manufacturing productivity has fallen behind the technological frontier (De Vries et al. 2013: 13-15)3 . However, Baldwin (2011) argues that globalization has undergone radical change from the mid-1980s, giving developing countries new opportunities to industrialize. Until the 1980s, globalization mostly concerned falling trade costs, while since the 1980s globalization is driven by lower transmission costs. Developing countries can now expand manufacturing by joining global and regional supply chains. Countries are no longer required to build the whole supply chain at home, they can simply produce the components for which their technological 3 Relative productivity actually increased in the 1970s, but experienced a rapid fall in the 1980s and has continued downwards since.
  • 16. 15 disadvantage is least marked, and export it into an international supply chain. Industrialization is both easier and less meaningful. What a country produces says less about how advanced it is, and more where its location is in the global value chain. With the ICT revolution, multinational firms can combine their sophisticated know-how with developing countries’ labor. However, as managing a value chain requires some face-to-face interactions, distance still matters. This could explain why most production networks concentrated in low-wage nations are near large headquarter nations such as the USA, Germany and Japan. Lack of a regional headquarter nation can pose limitations for future manufacturing growth in Sub- Saharan Africa. A key implication of global value chains is that protectionism through import substitution policies is less likely to be successful. While in the 1960s and 1970s, advanced countries technology advantages could be partly offset by labor costs, competition is harder now that international supply chains allow the combination of advanced technology and cheap labor (Baldwin 2011). 2.4 Empirical work The empirical literature identifying the determinants of growth-enhancing structural change is not very large. One reason is that employment data for developing countries, especially in Sub-Saharan Africa, has not been available or reliable. Wacziarg & Wallack (2004) investigate the effect of trade liberalization on sectoral reallocation of labor in 25 developing countries. Measuring reallocation across nine broad economic sectors, they find a decrease in the pace of labor reallocation after trade liberalization. They further disaggregate the manufacturing sector into 28 sub-sectors and find that trade liberalization has a small positive effect on reallocation within the manufacturing sector. The effect is strongest over longer periods. Moreover, the analysis shows that liberalization is followed by reduced employment growth in manufacturing. The analysis does therefore not support the proposition that trade liberalization leads to large reallocation of labor across sector (Wacziarg & Wallack 2004: 422-425). Dabla-Norris et al. (2013) analyze the determinants of structural change by creating a benchmark model and comparing the actual and predicted real value added share of each
  • 17. 16 sector. They find that country characteristics account for a large proportion of the cross- country variation. More specifically, land area is positively related to agriculture shares; population size is negatively associated to the agricultural share and positively related to manufacturing share; and having a large proportion of non-working-age people is positively associated to the service share of the economy. The capital stock is negatively associated with agricultural and service shares, while being positively related to the manufacturing share. The effect of capital stock is particularly strong at low levels of manufacturing, suggesting the importance of infrastructure in early manufacturing development. As expected GDP per capita is negatively related to the share of agriculture and has a non-linear relationship with manufacturing, increasing at low levels of GDP and decreasing at higher levels (Dabla-Norris 2013: 9-18). Looking at policy variables Dabla-Norris et al. find that trade openness has a strong positive relationship with the share of manufacturing, while being negatively associated with the agricultural share. The effect of trade openness on the service share is negative for countries with a low share of services, and positive for countries with a high share of services. Dividing the sample to before and after the modern era of globalization, they find that trade openness had a stronger effect on the manufacturing share after 1992 than before. Curiously, when using average tariff rates as an indicator of trade liberalization, liberalization is positively associated with agriculture, and negatively correlated with manufacturing for countries with low share of industry. A depreciation of the real exchange rate is associated with a higher manufacturing share for countries with low industrial base (Dabla-Norris et al. 2013: 14-19). Looking at the 1990-2005 period, McMillan et al. (2014) find that comparative advantage in primary products, proxied by the share of raw materials of total exports4 , has a significant and negative effect on growth-enhancing structural change. The initial employment share in agriculture has a positive and significant effect, but only when the comparative advantage indicator is included. The authors see this as evidence of conditional convergence, as a country with a large rural labor share has a potential for growth-enhancing structural change, given that it does not have a strong comparative advantage in primary products. Moreover, including the comparative advantage indicator makes the regional dummies for Africa and 4 Comparative advantage indicator
  • 18. 17 Latin America insignificant, indicating that comparative advantage and the initial agricultural share can jointly explain the average differences between the regions. Furthermore, McMillan et al. find that an undervalued exchange rate promotes growth-enhancing structural change, while employment rigidity discourages it. Other potential determinants such as income levels, demography, institutional quality, and tariff levels did not turn out to be consistently significant in their analysis (McMillan et al. 2014: 26-27). Analyzing the period 2000 to 2010, McMillan & Harttgen (2014) find that the main contribution to growth-enhancing structural change is the decline in the labor share engaged in agriculture. A small part of those workers ended up in manufacturing (around 20 %), but around 80 % moved to the service sectors. They find that the agricultural employment share is falling faster in countries that started with a high share of the labor force in agriculture; experience high population growth rates; benefit from higher quality of governance; and have undertaken deeper agricultural reforms (proxy for agricultural productivity growth). McMillan & Harttgen argue that high population growth is negatively associated with the agriculture employment share because it reduces farm size and makes farming less attractive for the young. The negative relationship between agricultural productivity growth and the agricultural labor share is consistent with theories discussed above, as higher productivity growth reduces the labor share due to either declining relative prices or changing expenditure patterns as income increases (McMillan & Harttgen 2014: 5, 31-32).
  • 19. 18 3. Theoretical framework and empirical hypotheses 3.1 Growth effect of sector reallocation Although the models discussed in the previous chapter predict trade liberalization’s effect on sector shares, it is unclear how changes in sector shares translate into aggregate productivity growth. Reallocation away from agriculture to either services or manufacturing will have a positive effect on growth due to the low productivity in agriculture However, with stagnating manufacturing productivity in Sub-Saharan Africa, the growth effect of reallocating labor from services to manufacturing is uncertain. The effect depends on the type of service sector, and country-specific relative productivities. There are nevertheless strong arguments as to why manufacturing is essential to rapid and sustained economic development. Baumol (1967) argued that unlike manufacturing, most service activities do not allow for constant and cumulative increases in productivity through capital accumulation, innovation, or economies of scale (Baumol 1967: 420). Felipe et al. (2015) find that all non-oil economies with per capita income above $12 000 (2005 dollars) today, reached at least 18 % employment share in manufacturing at their peak (Felipe 2015: 9). Rodrik (2015) sees three advantages of manufacturing compared to services. First, formal manufacturing exhibits unconditional productivity convergence with the technological frontier. Second, unlike other high-productive sectors such as mining, manufacturing has traditionally absorbed a large number of unskilled labor. Third, manufacturing is a tradable sector meaning that it does not face demand constraints from a home market populated by low-income consumers (Rodrik 2015: 3). While the models discussed in Section 2.3 treat services as non-tradable, services are in fact increasingly becoming a larger part of international trade. Rodrik nevertheless argues that highly productive, tradable services such as IT and finance are typically skill-intensive and do not have the capacity to absorb the type of labor that is abundant in low-income countries. Other types of services typically lack technological dynamism or are non-tradable; meaning their ability to expand rapidly is constrained by productivity in the rest of the economy (Rodrik 2015: 24). This pessimism is challenged by Ghani & O’Connell (2014), who note that technology improvements have enabled services to be traded like goods, and that global trade
  • 20. 19 in services is growing much faster than trade in goods. The increasing proliferation and importance of global value chains can also provide opportunities for service-based growth and technology diffusion, as provision of services is often sub-contracted nationally and globally. The globalization of services provides African countries more opportunities to find niches beyond manufacturing, where they can specialize and scale up. Concerning dynamism, Ghani & O’Connell find that productivity convergence in services has actually been faster than in manufacturing between 1990 and 2010. Unlike manufacturing, income elasticity for services is above unity, making it unlikely that the future expansion of the services sector will be limited by demand factors (Ghani & O’Connell 2014: 2-16). Moreover, although the service sector in Sub-Saharan Africa is dominated by informality, so is manufacturing, as an increasing number of manufacturing firms are informal and low-productive. Unconditional manufacturing convergence concerns only formal firms, and this convergence is not found when including all types of manufacturing firms (De Vries et al. 2013: 5). As the scope for broad-based manufacturing development seems to be shrinking, reallocation into services might therefore be equally important. With the growth effect of reallocation across sectors being uncertain, the theoretical framework needs specific propositions on the relationship between openness and structural change. 3.2 Hypotheses Considering the finding by McMillan & Harttgen (2014) that the declining labor share in agriculture accounts for most of the structural change growth in SSA, it is plausible that trade liberalization can affect structural change through either changes in income or agricultural productivity growth. There is a broad spectrum of literature showing the positive relationship between trade openness and growth (e.g. Feyrer 2009). However, the literature on trade and agricultural productivity growth is less clear, and some studies have found mixed results (Yu & Nin-Pratt 2011). In their paper on structural transformation, McMillan et al. (2014) note that while globalization has facilitated technology transfer and increased efficiency in production, its effect on structural change has been highly uneven across regions. A number of firm-level studies that have shown that intensified import competition has made manufacturing
  • 21. 20 industries become more efficient by rationalizing their operations. The least productive firms exit the industry, while the remaining firms shed excess labor. This leads to an increased within-sector productivity, and the top tier firms close the gap with the technological frontier in Western countries. Unlike in the Melitz model, however, the most productive firms do not absorb the labor from exiting firms. The effect on overall productivity growth therefore depends on what happens with the workers who are displaced. In countries that have high levels of unemployment or large inter-sectoral productivity gaps, there is a large risk that displaced workers end up in low-productive sectors such as low-skilled services or informality. If this happens, the effect of increased import-competition on aggregate productivity is unknown, and can even be negative (McMillan et al.2014: 11-12). This is an especially big risk for Sub-Saharan Africa with its large productivity differentials between sectors, and where almost 80 % of the labor force work in vulnerable employment, either as unpaid family workers or self-employed, predominately in the informal sector (ILO 2014). However, openness to investments and intermediate goods should have a positive effect on integration into global value chains. Rather than seeing import as competition, imports can be seen as inputs and complementary to local production. By easing access to imports, trade liberalization can have a positive effect on structural change through value chain integration. In fact, there is some evidence that structural change is linked to the share of foreign value added in a country’s export, a common measure of participation in global value chains5 (AEO 2014: 31, 72). Finally, an undervalued exchange rate can be beneficial for structural change. McMillan et al. (2014) argue that one reason why Asian countries have experienced stronger structural change is because they adopted competitive exchange rate in order to promote tradable industries, while SSA and Latin American countries often globalized with an overvalued exchange rate (McMillan et al. 2014: 25-26). Furthermore, there is evidence that undervalued exchange rates increases growth in developing countries by boosting the size of the tradable sector (Rodrik 2008: 388-391). Based on the theoretical and empirical literature, I propose three distinct hypotheses: 5 Backward integration
  • 22. 21 1) Trade liberalization and import competition have a negative effect on structural change in Sub-Saharan Africa, because its comparative advantage does not lie in high- productive sectors. Specialization will therefore lead to reallocation of labor to less productive sectors. 2) Trade liberalization has a positive effect on structural change as it leads to a reallocation of labor from agriculture to the service sector by either: a. Increasing income, which leads to relative lower expenditure on agricultural goods (income effect). This will have a negative effect on agriculture’s employment, nominal value added share, and real value added share. b. Increasing agricultural productivity, thereby reducing relative prices of agricultural goods (productivity effect). This will have a negative effect on agriculture’s employment and nominal value added share, and an ambiguous effect on its real value added share. 3) Trade openness has a positive effect on structural change because it promotes integration into international value chains. This effect should be clear after 1980 and be associated with a reallocation of labor into manufacturing.
  • 23. 22 4. Empirical strategy Most empirical literature that has tried to identify the causes of structural transformation has focused on the determinants of sector shares, not on whether the reallocations across sectors were growth-enhancing. The few studies focusing on growth-enhancing structural change have been mostly cross-country analyses. By using panel data, this paper will attempt to identify the determinants of growth-enhancing structural change in a more rigorous way. This chapter will first describe the structural change variable that is used as dependent variable in the empirical analysis. I will then explain the methodology behind difference-in- differences (DID), the main econometric technique used in the analysis. After discussing some methodological challenges with DID, I introduce the methodology of instrumental variables (IV) and outline how the instruments were constructed using bilateral trade data. I estimate the following baseline regression: π‘†π‘‘π‘Ÿπ‘’π‘π‘‘π‘’π‘Ÿπ‘Žπ‘™ π‘β„Žπ‘Žπ‘›π‘”π‘’π‘–π‘‘ = 𝛼 + 𝛿𝑑 + πœ‡π‘– + 𝛽𝑇𝐿𝑖𝑑 + πœ™π‘‹π‘–π‘‘ + πœ€π‘–π‘‘ (1) where 𝑆𝐢𝑖𝑑 is the structural change dependent variable, 𝛼, 𝛿𝑑 and πœ‡π‘– are the constant, time and country fixed effects respectively. Time dummies controls for common macroeconomic shocks, while country fixed effects control for country-specific, time-invariant factors. 𝑇𝐿𝑖𝑑 is a dummy for whether country i has experienced trade liberalization at or before time t, 𝑋𝑖𝑑 are relevant control variables and πœ€π‘–π‘‘ is the error term. The estimated impact of the trade liberalization episode is the estimate 𝛽̂. 4.1 Baseline regression 4.1.1 The dependent variable The paper’s measure of structural change follows the methodology used in De Vries et al. (2013), who divide productivity growth into within-sector growth and between-sector growth: βˆ†π‘ƒ = βˆ‘ (𝑖 𝑃𝑖 𝑇 -𝑃𝑖 0 )𝑆𝑖 𝑇 + βˆ‘ (𝑖 𝑆𝑖 𝑇 βˆ’ 𝑆𝑖 0 )𝑃𝑖 0 (2) where βˆ†π‘ƒ is economy-wide productivity growth. 𝑃𝑖 𝑇 stands for productivity in sector i at time T, while 𝑆𝑖 𝑇 describes the share of employment in sector i at time T. The first term measures
  • 24. 23 the within-sector productivity growth, weighted by the sector share. The second term measures the change in sector shares, weighted by the productivity of the sector. This is the structural change term (De Vries et al. 2013: 15-16). If the expanding sectors have higher average productivity than the shrinking sectors, the term is positive; if the expanding sectors have lower average productivity than the shrinking sectors, the term is negative. Labor productivity is measured in constant prices. The dependent variable is measured over five-year periods. As structural change is a long- term process, it is not meaningful to measure its year-by-year fluctuations. Measuring five- year periods is consistent with the economic growth literature, for example in Caselli et al. (1996). In order to measure structural change as a growth rate, the term is divided by the average productivity at the base year for each 5-year period. Finally, the term is divided by the number of years to get the average annual growth rate of structural change for each 5-year period. The dependent variable Structural change is thus the following decomposition: Structural change = βˆ‘ (𝑖 𝑆𝑖 𝑇 βˆ’π‘†π‘– 0 )𝑃𝑖 0 5βˆ—π‘ƒ0 (3) where 𝑃0 is the economy-average productivity at the first year of each five-year period. There are other ways to construct the structural change term. Instead of using baseline productivity, 𝑃𝑖 0 as a weight, one could use final productivity 𝑃𝑖 𝑇 or period productivity averages 𝑃̅𝑖, as weights. Using final productivity will typically result in a relatively smaller contribution from structural change to growth compared to the composition using base- period productivity levels (De Vries et al. 2013: 15-16). For these reasons, the regressions will be rerun as a robustness checks using final productivity weights in constructing the dependent variable: SC with final productivity = βˆ‘ (𝑖 𝑆𝑖 𝑇 βˆ’π‘†π‘– 0 )𝑃𝑖 𝑇 5βˆ—π‘ƒ 𝑇 (4) The productivity equation (2) can be expanded by adding a dynamic third term: βˆ†π‘ƒ = βˆ‘ (𝑖 𝑃𝑖 𝑇 -𝑃𝑖 0 )𝑆𝑖 𝑇 + βˆ‘ (𝑖 𝑆𝑖 𝑇 βˆ’ 𝑆𝑖 0 )𝑃𝑖 0 + βˆ‘ (𝑃𝑖 𝑇 βˆ’ 𝑃𝑖 0 ) βˆ— (𝑆𝑖 𝑇 βˆ’ 𝑆𝑖 0 )𝑖 (5) This new term is called dynamic structural change. The term is positive if labor reallocates to sectors with above-average productivity growth, and negative if labor reallocates to sectors
  • 25. 24 with below-average productivity growth. Production equation (5) therefore includes two reallocation effects: reallocation of workers to sectors with above average productivity levels (static reallocation effect), and reallocation to sectors with above average productivity growth (dynamic reallocation effect). Although it is important to consider the dynamic effects of reallocations, this paper choose to not use dynamic structural change as a dependent variable in the regression analysis. The term is often negative and can be very difficult to interpret. For instance, agricultural productivity growth can play an important part in inducing structural change, as surplus food production allows more workers to migrate to the cities to work in β€œmodern” sectors. However, if productivity growth is strong and the labor share in agriculture is falling, the dynamic structural change term will be negative, even if labor is shifting to more productive sectors (McMillan & Harttgen 2014: 15). As the paper is using panel data, the static effect of reallocation to sectors with low productivity growth will decrease over time. Nevertheless, I construct a combined static and dynamic structural change variable to serve as a dependent variable for robustness checks: Static + Dynamic SC = βˆ‘ (𝑖 𝑆𝑖 𝑇 βˆ’π‘†π‘– 0 )𝑃𝑖 0 5βˆ—π‘ƒ0 + βˆ‘ (𝑃𝑖 𝑇 βˆ’π‘ƒπ‘– 0 )βˆ—(𝑆𝑖 𝑇 βˆ’π‘†π‘– 0 )𝑖 5βˆ—π‘ƒ0 (6) As the dependent variable is in five-year periods, I measure all variables in five-year periods. 4.1.2 Difference in differences Difference-in-differences (DID) estimations entail identifying a specific intervention or treatment, in this case a trade liberalization episode. One then compares the difference in outcomes (structural change) before and after the intervention for countries that underwent trade liberalization to the same difference for countries who did not undergo trade liberalization (Bertrand et al. 2003: 2). The baseline multiple periods DID regression to be estimated is thus equation (1). For the impact measured in the DID analysis to be causally valid, one has to make the assumption of parallel trends. The parallel trends assumption states that changes in the outcome variable over time would have been the same in both treatment and control groups in the absence of an intervention, controlling for relevant time-varying factors (Bertrand et al. 2003: 2-3).
  • 26. 25 If other factors besides trade liberalization affect the difference in trends between the groups, the estimation will be biased. In addition to fixed effects that take care of time-invariant country specific factors, I will control for observable time-varying variables that could affect the outcome variable. The control variables chosen are recognized in the literature as important determinants and conditions for structural change. The advantage with this estimation technique compared to the cross-country analysis used in McMillan et al. (2014) is that the fixed effects control for time-invariant, country-specific factors that might influence the outcome variable. 4.1.3 Methodological challenges Bertrand et al. (2003) argue that standard difference in differences estimation is potentially subject to severe serial correlation problems. They argue that three factors make serial correlation an important issue when estimating DID. First, DID estimations usually rely on long time series. Second, the commonly used dependent variables in DID estimation are typically highly positively serial correlated. Third, an intrinsic aspect of the DID model is that the treatment variable changes very little within a country over time. These three factors can reinforce each other so that the standard error for 𝛽̂ could severely understate the standard deviation of 𝛽̂. Bertrand et al. test how DID perform under placebo laws, where treated states and the years the laws passed are randomly selected. They find that the null hypothesis of no effect is rejected at T > 1.96, far more often than the expected 5 % of the time (Bertrand et al. 2003: 6-11). By using five-year periods, the time series in this paper’s analysis is not particularly long. Moreover, the dependent variable (structural change) is unlikely to be serial correlated because it is a growth rate. However, given that the treatment variable (liberalization dummy) changes very little over time, there may be a problem with serial correlation. One solution for small datasets like mine is to collapse the data into pre- and post-periods to produces consistent standard errors. As the liberalization dates happen in different years, I must slightly modify the technique. First, I regress structural change on country-fixed effects, year dummies and relevant control variables. Then I divide the residuals into two groups: one before the liberalization and one after the liberalization. I obtain the estimate of liberalization’s effect
  • 27. 26 and its standard errors from an OLS-regression of the two-period panel (Bertrand et al. 2003: 14-15). However, the power of this test declines fast, making it hard to detect an effect. With 20 states, this technique has a rejection rate of 9.5 %, almost twice as large as it should be (Bertrand et al. 2003: 18-19). The usual regression assumption is that the error term (πœ€π‘–π‘‘ ) is independently and identically distributed (i.i.d.), but this can be violated in panel data. There could be β€œclustered errors”, which means that observations within country i are correlated in some unknown way, inducing correlation in πœ€π‘–π‘‘ within i. In the presence of clustered errors, standard errors can be wrong, leading to incorrect inference. To correct for this, I use clustered standard errors by country in all my regressions, which also corrects for heteroskedasticity. Moreover, the assumption of parallel trends could be violated. To test for this, I will redo the analysis with β€œplacebo” trade liberalization dates. In one test, I will move all the liberalization dates one period ahead, and in the other test, I will move the liberalization dates one period behind. If the trends were parallel, there should be no effect of the β€œplacebo” trade liberalizations. Finally, there are some limitations with the structural change term itself. While it decomposes GDP per worker, it does not decompose GDP per capita. It therefore does not take into consideration changes in labor force participation (due to demographics, changes in gender roles etc.) or changes in unemployment (McMillan et al. 2014: 20). Reallocation from and to unemployment would have been particularly interesting as this is the least productive activity possible, but data limitations leaves this question for future research. However, for most countries in Sub-Saharan Africa the option of being unemployed is not viable, and low- productive work in the informal sector is the most likely alternative for someone losing their job. Employment in the informal sector is included in the analysis as part of the different economic sectors. 4.1.4 Sector contribution and sector shares In order to understand the mechanisms behind structural change it is worth looking into contributions from different sectors. By modifying decomposition (3), we can measure the contribution to structural change from each sector. In decomposition (3), each expanding
  • 28. 27 sector contributes positively to aggregate productivity levels, even when it has below-average productivity levels. In the modified model, I therefore adjust the structural change term to take account of its relative productivity level. For each period, I divide the sectors into expanding and shrinking sectors based on changes in employment shares. I calculate the structural change effect of an expanding sector as its productivity level relative to the average productivity level of the shrinking sectors. The modified decomposition is thus: π‘†π‘’π‘π‘‘π‘œπ‘Ÿ π‘π‘œπ‘›π‘‘π‘Ÿπ‘–π‘π‘’π‘‘π‘–π‘œπ‘›π‘—π‘‘ = βˆ‘ (𝑆𝑗 𝑇 βˆ’ 𝑆𝑗 0 ) 𝐽 𝑗 (𝑃𝑗 0 βˆ’ 𝑃0βˆ— ) (7) where J is a set of expanding sectors and 𝑃0βˆ— is the weighted average productivity of the shrinking sectors. If the labor productivity(𝑃𝑗 0 ) of expanding sector j is higher than the weighted average productivity of the shrinking sectors (𝑃0βˆ— ), sector j made a positive contribution to structural change (Timmer et al. 2014: 16-17). By using each sector’s individual contribution to structural change as the dependent variable, I can identify through which sectors the explanatory variables contribute to structural change. As employment data by sector only coverseleven countries in Sub-Saharan Africa, generalizing the results obtained in the DID analysis can be difficult. However, I expand my dataset using changes in the real value added sector shares as my dependent variable. Although this does not correspond perfectly to growth-enhancing structural change, it is possible to make some inferences based on the results. First, although I cannot observe productivity per sector directly without employment data, it is a reasonable assumption that the agricultural sector has below average productivity in Sub-Saharan Africa. Thus, a reduction of agriculture’s value added share in the economy is likely to be positive for growth-enhancing structural change. The analysis can then serve as an additional robustness test. Furthermore, a larger sample makes it possible to infer whether the results from the DID analysis are generalizable to the larger region of Sub-Saharan Africa. A larger sample also gives the two-period robustness test more power. Finally, by using sector shares as dependent variable I can identify which sectors expands following trade openness. I can then test more directly some of the theories of structural transformation and sector shares presented in Chapter 2. The dependent variable will be the average five-year change in each sector’s value added share: βˆ†π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘ = π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘βˆ’ π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘βˆ’5 5 (8)
  • 29. 28 where π‘†β„Žπ‘Žπ‘Ÿπ‘’π‘–π‘‘ is the real value added share of sector i at time t. 4.2 Instrumental variables regression While the DID analysis described in the previous chapter takes steps to increase internal validity, finding the true causal effect remains a challenge. Trade liberalization often coincides with the adoption of other domestic market reforms, as well as stable fiscal and monetary policies. Since these policies can affect structural change, trade liberalization could be correlated with factors omitted from the regression, and thus be endogenous (Frankel & Romer 1999: 379). Moreover, trade policies might change following economic conditions and prospects. Identifying a solid instrument for trade liberalization is not an easy task. Indeed, an exogenous instrument for changes in domestic trade policy has not yet been identified in the literature. It is more realistic to identify exogenous changes in trade costs that vary across time and countries. 4.2.1 The gravity equation I build on the established method of constructing instruments for trade flows loosely based on the gravity model (e.g. Feyrer 2009). The gravity model predicts that bilateral trade is a function of exporter characteristics, importer characteristics, and resistance factors such as distance. Anderson and van Wincoop (2003) derived the theoretically consistent gravity model: π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘ = 𝑦𝑖𝑑 𝑦 𝑗𝑑 𝑦 𝑀𝑑 ( 𝜏 𝑖𝑗𝑑 𝑃 𝑖𝑑 𝑃 𝑗𝑑 )1βˆ’πœŽ , (9) where π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘ denotes exports from country i to country j in year 𝑑, while 𝑦𝑖𝑑, 𝑦𝑗𝑑 and 𝑦 𝑀𝑑 are real GDP in country i, country j, and the world respectively. πœπ‘–π‘—π‘‘ is the bilateral resistance term, and 𝑃𝑖𝑑 and 𝑃𝑗𝑑 are multilateral resistance terms in country i and j. 𝜎 > 1 is the elasticity of substitution between goods (Blanchard & Onley 2015: 18-20). The bilateral resistance term includes distance, but also other bilateral factors that determine trade such as common border, colonial relationship, etc. Multilateral resistance terms represent the relative price levels of the importing and exporting country. This can be captured by using country fixed effects. After taking logs, the standard gravity model is estimated:
  • 30. 29 ln(π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘) = 𝛾𝑑 + πœ“π‘– + πœ™π‘— βˆ’ 𝛿 ln(𝑑𝑖𝑠𝑑𝑖𝑗) + πœ‚ ln(𝐺𝐷𝑃𝑖𝑑) + πœ‡ ln(𝐺𝐷𝑃𝑗𝑑) + πœ‘πΆπ‘œπ‘›π‘‘π‘–π‘”π‘–π‘— + πœŒπΆπ‘œπ‘™π‘–π‘— + πœ€π‘–π‘—π‘‘ (10) where 𝛾𝑑 are time dummies, and πœ“π‘– and πœ™π‘— are importer and exporter fixed effects respectively. πΆπ‘ˆπ‘–π‘—π‘‘ is a dummy for both countries being in the same currency union; πΆπ‘œπ‘›π‘‘π‘–π‘”π‘–π‘— is a dummy for shared borders; and πΆπ‘œπ‘™π‘–π‘— is a dummy for colonial relationship. More rigorously, one can add a set of bilateral pair fixed effects that controls for all time- invariant bilateral fixed effects. The equation above becomes: ln(π‘‘π‘Ÿπ‘Žπ‘‘π‘’π‘–π‘—π‘‘) = 𝛾𝑑 + 𝛿𝑖𝑗 + πœ‚ ln(𝐺𝐷𝑃𝑖𝑑) + πœ‡ ln(𝐺𝐷𝑃𝑗𝑑) + πœ€π‘–π‘—π‘‘ (11) where 𝛿𝑖𝑗 are country pair fixed effects. 4.2.2 Constructing the instruments As variation in bilateral trade due to Home country characteristics could be correlated with structural change, the instruments should therefore only rely on the variations of trade due to plausibly exogenous factors such as geography, transportation costs and/or conditions in a country’s Foreign trading partners. Such a variation could be caused by Foreign joining GATT/WTO. When a Foreign joins the WTO, bilateral trade costs should fall, leading to a boost in exports from Home that decreases with distance. The idea is that Foreign’s choice to join the GATT/WTO is exogenous for Home country. Moreover, the distance between the two countries should be exogenous. I create two variables, one that weighs for only distance, and one that weighs for both distance and the Foreign GDP: π‘Šπ‘‡π‘‚1𝑖𝑗𝑑 = π‘Šπ‘‡π‘‚πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘› 𝑖𝑗𝑑 ln(π‘‘π‘–π‘ π‘‘π‘Žπ‘›π‘π‘’ 𝑖𝑗) (12) π‘Šπ‘‡π‘‚2𝑖𝑗𝑑 = π‘Šπ‘‡π‘‚πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘› π‘–π‘—π‘‘βˆ—πΊπ·π‘ƒ_πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘› 𝑖𝑗𝑑 ln(π‘‘π‘–π‘ π‘‘π‘Žπ‘›π‘π‘’ 𝑖𝑗) (13) where π‘Šπ‘‡π‘‚πΉπ‘œπ‘Ÿπ‘’π‘–π‘”π‘›π‘–π‘—π‘‘ is a dummy variable for Foreign country being member of GATT/WTO at or before time t. The next variable I construct is a variable of trade costs based on the Baltic Dry Index (BDI). The BDI is a general indicator for the cost of transporting dry bulk cargoes, consisting mainly of raw commodities. The BDI is thus especially relevant for the trade of low-income countries:
  • 31. 30 π΅π·πΌπ‘π‘œπ‘ π‘‘π‘–π‘‘ = πœƒπ‘–π‘‘ βˆ— ln(𝐡𝐷𝐼𝑑) (14) where πœƒπ‘–π‘‘ is the share of primary commodities in exports for country i at time t. The idea is that changes in the BDI matters more for the exports of countries with a larger share of primary exports (Sim & Lin 2012: 4-5). To construct each instrument, I regress bilateral exports on either π‘Šπ‘‡π‘‚1𝑖𝑗𝑑, π‘Šπ‘‡π‘‚2𝑖𝑗𝑑 or π΅π·πΌπ‘π‘œπ‘ π‘‘π‘–π‘‘ in the bilateral equation (11). Since I am not interested in the coefficients for the bilateral resistance factors, I use bilateral pair fixed effects (𝛿𝑖𝑗) to capture time-invariant geographical factors: ln(𝑒π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘—π‘‘) = 𝛾𝑑 + 𝛿𝑖𝑗 + π›½πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘π‘–π‘—π‘‘ + πœ€π‘–π‘—π‘‘ (15) where ln(𝑒π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘—π‘‘) is the log of exports from Home to Foreign, while πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘π‘–π‘—π‘‘ is either π‘Šπ‘‡π‘‚1𝑖𝑗𝑑, π‘Šπ‘‡π‘‚2𝑖𝑗𝑑 or π΅π·πΌπ‘π‘œπ‘ π‘‘π‘–π‘‘. Following standard literature (Frankel & Romer 1999, Feyrer 2009), I aggregate the unlogged bilateral relationships to arrive at total predicted exports for Home. I similarly sum the actual exports to arrive at a value for Home’s total actual exports. π‘π‘Ÿπ‘’π‘‘π‘–π‘π‘‘π‘’π‘‘ 𝑒π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘‘ = βˆ‘ 𝑒 𝛾̂𝑑+ 𝛿 𝑖𝑗 Μ‚ + 𝛽̂ πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘ 𝑖𝑗𝑑 𝑖≠𝑗 (16) I then take the logs of the aggregated predicted and actual exports. To implement the instrumental variable regression, I follow the standard two-stage least square regression (2SLS), using the predicted export values as an instrument for the actual exports in the first stage regression. First stage: ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )𝑖𝑑 = πœ‹0 + πœ“Μ‚ln(βˆ‘ 𝑒 𝛾̂ 𝑑+ 𝛿𝑖𝑗 Μ‚+ 𝛽̂ πΌπ‘›π‘ π‘‘π‘Ÿπ‘’π‘šπ‘’π‘›π‘‘π‘–π‘—π‘‘)𝑖≠𝑗 + πœ’π‘–π‘‘ + πœπ‘–π‘‘ (17) where πœ‹0 is the intercept, πœ’π‘–π‘‘ is a vector of control variables, and πœπ‘– is the residual. The first stage decomposes the endogenous variable (ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘‘ )) into two components: a component (πœπ‘–π‘‘) that can be correlated with the regression error in the first stage, and a problem-free component that is uncorrelated with the second-stage regression error. The idea is to use the problem-free component of ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )𝑖𝑑 and disregard πœπ‘– in the second stage regression. Then I use the predicted value from the first stage regression:
  • 32. 31 ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )Μ‚ 𝑖𝑑 = πœ‹Μ‚0 + πœ“Μ‚ln(π‘ƒπ‘Ÿπ‘’π‘‘π‘–π‘π‘‘π‘’π‘‘πΈπ‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )𝑖𝑑 (18) where πœ‹Μ‚0 and πœ“Μ‚ are OLS estimates. The second stage is then regressing π‘†π‘‘π‘Ÿπ‘’π‘π‘‘π‘’π‘Ÿπ‘Žπ‘™ π‘β„Žπ‘Žπ‘›π‘”π‘’π‘–π‘‘ on ln( 𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ )Μ‚ 𝑖𝑑 Second stage: π‘†π‘‘π‘Ÿπ‘’π‘π‘‘π‘’π‘Ÿπ‘Žπ‘™ π‘β„Žπ‘Žπ‘›π‘”π‘’π‘–π‘‘ = 𝛾0 + 𝛿𝑑 + 𝛼𝑖 + 𝛽ln(𝐸π‘₯π‘π‘œπ‘Ÿπ‘‘π‘ Μ‚ )𝑖𝑑 + πœ’π‘–π‘‘ + πœ€π‘–π‘‘ (19) In this strategy, the instrumented endogenous variable is a country’s export. When transportation costs go down or a foreign country joins the WTO, trade costs between two countries decrease in a way that is unrelated to developments in the Home country. I look at exports rather than total trade because changes in the BDI mainly affect the exports of Sub- Saharan African countries. Previous studies have shown that trade liberalization has a positive effect on both export and import growth (Kassim 2013: 12). Thus, the IV analysis measuring the effect of exogenous changes to exports is directly related to the DID analysis using trade liberalization episodes. Nevertheless, because domestic trade liberalization mainly involves reducing the barriers to imports, the results using exports as instruments must be interpreted differently from the DID analysis using trade liberalization dummy. In addition to using predicted exports as instruments, I look at the effect of an exogenous change in imports. Instead of considering what happens to exports from Home to Foreign, I see what happens to Home’s imports from Foreign when Foreign joins the WTO/GATT. To construct the instrument I follow the same procedure as above, weighing for distance or distance and Foreign GDP. I regress Home’s bilateral imports from Foreign on the WTO- variable, and country pair and time fixed effects: ln(π‘–π‘šπ‘π‘œπ‘Ÿπ‘‘π‘ π‘–π‘—π‘‘) = 𝛾𝑑 + 𝛿𝑖𝑗 + π›½π‘Šπ‘‡π‘‚π‘–π‘—π‘‘ + πœ€π‘–π‘—π‘‘ (20) Here, π‘Šπ‘‡π‘‚π‘–π‘—π‘‘ is a dummy for Foreign joining WTO/GATT, weighted by either distance, or by distance and Foreign GDP. I then aggregate the actual and predicted imports. Predicted imports is then used as an instrument for the actual imports in the first stage regression. By looking at imports, I can better test the effect of import competition on structural change, which would help answer hypothesis 1 in Chapter 3.
  • 33. 32 Finally, I construct a remoteness instrument. It is based on the instrument used in Martin et al. (2008: 890), but different as they construct it as a dyadic variable. I construct the remoteness variable for each country i the following way: π‘Ÿπ‘’π‘šπ‘œπ‘‘π‘’π‘›π‘’π‘ π‘ π‘–π‘‘ = βˆ’ ln(βˆ‘ 𝐺𝐷𝑃 𝑗𝑑 π‘‘π‘–π‘ π‘‘π‘Žπ‘›π‘π‘’ 𝑖𝑗 𝑅 𝑗≠𝑖 ) (21) where 𝐺𝐷𝑃𝑗𝑑 is the GDP of i’s trading partner j. The idea is that distance to trade partners (j) affect the cost of trading, and countries with longer distance to major trading partners trade less on average. The variable varies over time and across countries since the trade partner’s GDP changes over time and it affects countries differently depending on the distance to the trade partner. As this variable is country-specific, it is applied directly into the first-stage regression, using openness (total trade as share of GDP) as the endogenous variable. 4.3 Identification assumptions Further scrutiny is required to determine whether the results from the 2SLS regressions show the causal effect of trade openness on structural change. For an instrument (𝑍𝑖𝑑,) to be valid it needs to be both relevant and exogenous. 1. Instrument relevance: π‘π‘œπ‘Ÿπ‘Ÿ(𝑍𝑖𝑑, 𝑋𝑖𝑑 β‰  0) (22) 2. Instrument exogeneity: π‘π‘œπ‘Ÿπ‘Ÿ(𝑍𝑖𝑑, πœ€π‘–π‘‘ = 0) (23) The relevance criteria is met if the first stage F-statistic is above 10, which is the case for all the instruments used.6 The exogenous criteria states that the instrument can only affect the outcome variable through its effect on the endogenous variable (𝑋𝑖𝑑). The instrument must 6 I also tried to use the Generalized System of Preferences (GSP) as an instrument. As most African countries qualify for GSP, introduction of the scheme gives an exogenous boost to trade, weighted by the distance to the provider (Martin et al. 2008: 889). However, the first stage F-statistic turned out to be too low for my analysis. Moreover, Sim and Lin (2012) applied the BDI instrument directly into the first stage regression, without using the bilateral trade equation. However, applying the instrument directly in the first stage does not give sufficiently high F-statistic for my sample. One problem is that the BDI is quite volatile, and the instrument is good at capturing year-by-year changes in the cost of exports. However, averaging the BDI over 5 years might not give the same effect on 5-year average exports.
  • 34. 33 thus be uncorrelated with the error term (πœ€π‘–π‘‘) in the regression. This is a challenging criterion, and mostly depends on whether the instrument exogeneity is plausible. BDI-cost is the instrument that has the strongest claim to exogeneity. Since all the countries in my sample play a small part in international trade, it is highly unlikely that economic development in these countries would affect the BDI-index (Lin & Sim 2013: 5). Whether Foreign joins the WTO/GATT should in theory only affect the Home through its imports and exports as trade costs go down. However, as many countries simultaneously joined WTO/GATT, there is a chance that the instrument is correlated to domestic trade reform, and thus partly endogenous. Moreover, when a Foreign joins the WTO/GATT, it could also affect its outward flow of FDI given that the TRIMs-agreement provides some protection of FDI. The Home country could thus be affected through exports, imports and inward FDI. Ideally, I should then also instrument for inward FDI flows in my IV analysis. However, bilateral FDI flows are not readily available for my sample, making it difficult to perform this analysis. Although part of the IV-effect of exports or imports could be due to inward FDI, it would not have a strong impact on the conclusion, as they are all measures of economic openness. Finally, the remoteness index is exogenous as long as it only affects exports. Because the remoteness index is highly dependent on geography, there is a chance that it also captures regional trends such as development aid, democratization and security that could affect structural change. However, since the analysis only covers Sub-Saharan Africa, it is unlikely that there is much difference in regional trends. Nevertheless, like the other instruments it is possible that part of the effect is due to inward investments as trading partners experience economic growth.
  • 35. 34 5. Data 5.1 Structural change The GGDC 10-Sector Database provides the data for my main dependent variable, and includes information on ten sectors from Asian, Latin American and SSA countries. While previous studies of structural change and sectorial productivity growth lacked reliable and detailed data, this database presents an important step forward. The variables on structural change, agricultural employment share and real mining value added share come from this dataset. The dataset includes 11 African, 11 Asian, and 9 Latin-American countries. The eleven SSA countries7 in the database account for more than half of the total GDP and population in SSA. The database contains annual data (1960-2010) on the value added by and employment in the 10 main sectors of the economy (see Appendix 1), making it possible to calculate labor productivity per sector. The definition of employment is broad, including self-employed, family workers, and other informal workers over 15 years old. To obtain employment data, the researchers conducted an in-depth study of the available statistical resources on a country-by-country basis. The database uses the population census to indicate absolute level of employment, and uses labor force and business surveys to indicate trends in between censuses (Timmer et al. 2014: 4-7). It is well known that datasets from African countries are often unreliable due to weak capacity to collect, manage and disseminate data. However, most of the African countries in the database have considerable experience in collecting national accounts data and in conducting labor and household surveys. Moreover, growth rate comparisons, which this paper uses, are more reliable than comparisons of absolute level (De Vries et al. 2013: 9-10). Nevertheless, despite solid data collection, all results must be treated with a certain caution. In addition, there are some issues of missing data. In the African sub-sample, Zambia is missing data on employment in the Government sector. This is also the case for many other developing countries, particularly in Latin America. Moreover, some countries have missing data for the Other sector. To ensure compatibility, the structural change index is thus first created without 7 Botswana, Ethiopia, Ghana, Kenya, Mauritius, Malawi, Nigeria, Senegal, South Africa, Tanzania and Zambia
  • 36. 35 the Government and the Other sector. For robustness checks, I create additional structural change variables including the Government and the Other sectors. Excluding sectors is potentially problematic. However, it can be argued that trade liberalization has the most direct impact on the structure of the private sector, while the government sector is more influenced by political decisions. For additional robustness checks, I obtain data on sector value added measured in constant 2005 prices, from UN National Accounts Statistics. This covers almost all countries in Sub- Saharan Africa from 1970 to 2013. To maximize the number of observations, the five-year periods start with 1970-1974, and end with 2010-2013. List of sectors and countries are found in Appendix 2. 5.2 Trade liberalization episodes A trade liberalization episode is often a continuous process rather than one major reform. In identifying trade liberalization episodes, I follow the methodology set by Wacziarg & Welch (2008). In their paper, they update the Sachs-Warner classification of openness that defined a country as closed if it had one of the following characteristics: 1) Average tariff rates of 40 percent or more 2) Non-tariff barriers covering 40 percent or more of trade 3) A black market exchange rate at least 20 percent lower than the official exchange rate. This can work as a trade cost, because exporters might have to purchase foreign inputs using foreign currency obtained in the black market, but remit their foreign exchange receipts from exports to the government at the official exchange rate. 4) A state monopoly on major exports. 5) A socialist economic system. In principle, the liberalization date is the date when all of the Sachs-Warner openness criteria are fulfilled (and not reversed later). However, data limitations often made it necessary to rely on country case studies of trade policy. A prominent critique against the Sachs-Warner criteria by Rodrigues and Rodrik (2000) argues that the indicator for the β€œblack market premium” played an excessive role in classifying a country as closed. Moreover, they argue that this criterion affected only African countries, and therefore amounted to an Africa dummy.
  • 37. 36 Warcziarg and Welch acknowledge this as a valid critique of the dummy variable in cross- section analyses, but claim that it is a less valid criticism of the liberalization dates. Policy changes that reduced the β€œblack market premium” were generally accompanied by other outward oriented policies, such as reduction of tariff and non-tariff barriers (Warcziarg & Welch 2008: 190-196). Moreover, for the purpose of this paper, the Africa dummy critique is irrelevant as I compare only countries within Sub-Saharan Africa. Finally, while the black market premium is often associated with an overvalued official exchange rate, my analysis includes controls for overvaluation of the currency. A challenge for my analysis is to classify which five-year period the liberalization episode started. The classification follows a simple rule; the trade liberalization dummy is assigned to the first five-year period where the trade liberalization has lasted at least three years. The liberalization dummy is then assigned to all periods after. Table 1 shows the liberalization dates classified by Wacziarg-Welch, my updates, and the five-year period to which I assign the liberalization start dates. Table 1. Liberalization dates in SSA Country Wacziarg-Welch (2003) Updated 5-year periods where liberalization starts Botswana 1979 1979 1981-1985 Ethiopia 1996 1996 1996-2000 Ghana 1985 1985 1986-1990 Kenya 1993 1993 1991-1995 Malawi n.l - Mauritius 1969/1968 1969/1968 1971-1975 Nigeria n.l 2000 2001-2005 Senegal n.l 2003 2001-2005 South Africa 1991 1991 1991-1995 Tanzania 1995 1995 1996-2000 Zambia 1993 1993 1991-1995 Three countries were classified as closed in the Wacziarg-Welch dataset. Senegal because it continued to maintain a monopoly of exports in the cotton industry (Wacziarg & Welch 2003: 46), Malawi because the black market exchange rate was 28.83 % lower than the official
  • 38. 37 exchange rate, and Nigeria because the black market exchange rate was 151.38 % lower than the official exchange rate (Wacziarg & Welch 2008: 215). Since the sample in this paper goes up to 2010, it is necessary to update the dataset. Senegal phased out is marketing board in the early 2000s, leading Kireyev and Mansoor (2013: 15) to note that Senegal would today have a trade openness index of one. I set the liberalization date in 2003, when the Senegalese textile company, SODEFITEX, privatized and ended the State’s monopoly on cotton and textile exports. Unfortunately, datasets on black market premiums have not been updated since 1998. Black market premiums (BMP) are very difficult to measure due to the unavailability of public price data on market exchanges. There is evidence that the BMP has fallen considerably in Nigeria since 1999. According to Aluko (2007), while it was 275 % higher in 1998, it was 10 % higher in 1999 and 9 % higher in 2007. The change in BMP coincided with a large devaluation of the Nigerian Naira. Since Nigeria experienced low BMP uninterrupted between 1999 and 2010, I will classify Nigeria as open from the 2001-2005 period. While Malawi began experimenting with a floating exchange rate in 1998, they reverted to a managed float during 2004-2006 and have had a de facto adjustable currency regime since 2006. Throughout 2006-2011, the government refused to devalue despite significant import demand and foreign exchange shortages. This led to an overvalued exchange rate and the development of a vibrant parallel foreign exchange market. At its peak, the foreign currency was traded at around double the official exchange rate, before Malawi returned to a floating regime in 2012 (Pauw et al. 2013: 1-5). Malawi has therefore not experienced uninterrupted trade liberalization since 2000. I also update the liberalization dates for my expanded dataset (sector share regressions). Appendix 2 includes a list of countries with their respective liberalization dates. 5.3 Control variables As the literature review showed, initial conditions matter for the potential for growth- enhancing structural change. Countries with a larger share of the labor force in agriculture have a stronger potential for growth-enhancing structural change because labor reallocating
  • 39. 38 away from agriculture will increase average productivity. For each five-year period, the agricultural labor share measures the employment share of agriculture in the first year. As natural resource dependency can also influence structural change, the Mining share of VA measures the real value added share of mining the base year of each five-year period. Both variables are constructed using data from GGDC. As richer countries generally have less scope for growth-enhancing structural change than poor countries, GDP per capita in the first year of each period is included as a control using Penn World Tables 8.1. Whether the currency is under- or overvalued can affect the rate of structural change (McMillan et al. 2014). An index of overvaluation is created following Rodrik (2008). To determine whether the exchange rate is over- or undervalued, one must account for the Balassa-Samuelson effect, which states that non-tradable goods are cheaper in poorer countries. The first step is to construct the log of the real exchange rate. Following the Penn World Tables 8.1, the real exchange rate is calculated by dividing the purchasing power parity conversion factors (PPP) by the nominal exchange rate (XR). 𝑙𝑛𝑅𝐸𝑅𝑖𝑑 = ln( 𝑃𝑃𝑃 𝑖𝑑 𝑋𝑅𝑖𝑑 ) (24) If 𝑙𝑛𝑅𝐸𝑅𝑖𝑑 is above zero, this indicates that the price level is higher than the price level in the US in 2005, and vice versa if it is below zero. The second step is to regress the real exchange rate on GDP per capita and time dummies. 𝑙𝑛𝑅𝐸𝑅𝑖𝑑 = 𝛼 + π›½π‘™π‘›πΊπ·π‘ƒπ‘π‘Žπ‘π‘–π‘‘π‘Ž + 𝛾𝑑 + πœ€π‘–π‘‘ (25) Finally, to arrive at the undervaluation index, I take the difference between the actual real exchange rate and the predicted Balassa-Samuelson adjusted rate: π‘™π‘›π‘‚π‘£π‘’π‘Ÿπ‘£π‘Žπ‘™π‘–π‘‘ = 𝑙𝑛𝑅𝐸𝑅𝑖𝑑 βˆ’ ln 𝑅𝐸𝑅̂𝑖𝑑 (26) where ln 𝑅𝐸𝑅̂𝑖𝑑 is the predicted value from equation (26) (Rodrik 2008: 371-372). If π‘™π‘›π‘‚π‘£π‘’π‘Ÿπ‘£π‘Žπ‘™ is above zero, it means that the real exchange rate is higher than the real exchange rate predicted by the country’s GDP per capita, and the real exchange rate is classified as overvalued. If π‘™π‘›π‘‚π‘£π‘’π‘Ÿπ‘£π‘Žπ‘™ is below zero, the real exchange rate is below what is predicted based
  • 40. 39 on its income level, and the exchange rate is classified undervalued. I obtain data on the real exchange rate from the Penn World Tables 8.1. In addition to trade barriers and the exchange rate, the cost of a country’s exports and imports depends on the Terms of Trade (TOT). I retrieve TOT data from DataMarket, a dataset based on World Bank staff’s calculations using data from Thomson Reuter’s Datastream. I construct the terms of trade indicator as an annual percentage change. For the other control variables, I use Penn World Tables 8.1 to construct the variables on openness, population, and capital stock per capita. World Development Indicators provide data on the age-dependency ratio, primary enrollment, and domestic credit to the private sector. I include a variable on the age- dependency ratio because it can affect labor supply as well as savings and consumption behavior (Dabla-Norris et al. 2013:10). The financial openness variable is collected from the Chinn-Ito index. Finally, I get a tariff indicator from Prati et al. (2010). Further descriptions of the control variables can be found in Appendix 3. To create the instruments, I use bilateral trade data from various sources. For standard gravity variables, I use the dataset collected by CEPII. For importer and exporter GDP, I use Penn World Tables 8.1. For bilateral trade data, I use the World Trade Flows (WTF) collected by Robert Feenstra. WTO/GATT dummies come from Rose (2004), and were subsequently updated by the author. Data on the Baltic Dry Index comes from Lin & Sim (2014). Primary share variables were constructed using data from both UNCTAD (1995-2010) and NBER (1971-1994). I construct two different primary shares. The first consists of all primary goods in exports including fuels.8 The second excludes all fuels except for coal.9 A problem with the bilateral trade and export data is that they do not cover Botswana before 1991. The panels will therefore not be fully balanced when using these variables. 8 SITC Rev. 3 codes: 0 + 1 + 2 + 3 + 4 + + 67 + 68 9 SITC Rev. 3 codes: 0 + 1 + 2 + 32 + 4 + + 67 + 68
  • 41. 40 6. Results 6.1Descriptive statistics Figure 1 below shows the relationship between structural change and changes in liberalization in SSA over time (1971-2010). The solid blue line is the non-weighted five-year average contribution of structural change in SSA (left y-axis), while the dashed red line shows the share of countries undertaking trade liberalization within each five-year period (right y-axis). Structural change was strong in the early 1970s, before falling steadily until its low-point in the mid-1990s when structural change had a negative effect on growth. From the mid-1990s, structural change again made a positive contribution to growth. SSA countries started slowly to liberalize in the early 1980s, before there was a strong liberalization surge in the late 1980s and early 1990s, following structural adjustment programs. From the graph, we can see that a period of stronger structural change followed the large number of liberalization episodes in the 1990s. By 2006, only Malawi in my sample had not liberalized. Figure 1. Relationship between structural change and liberalization episodes In Figure 2, the dashed blue line shows the average contribution of structural change to growth for liberalized countries, while the solid red line shows the structural change contribution for non-liberalized countries. Structural change was higher on average in liberalized countries from the mid-1980s to the early 2000s, while it was higher in non- liberalized countries at the beginning and the end of the period. Structural change seems to be more stable and on average higher in liberalized countries. Between 1971 and 2010, the
  • 42. 41 average structural change contribution to growth was 0.91 % for non-liberalized countries, and 1.34 % for liberalized countries. Figure 2. Structural change for liberalized and non-liberalized countries Interestingly, structural change is stable in countries with an undervalued exchange rate, while it is highly volatile in in countries with overvalued exchange rate (see Appendix 5). Summary statistics for the most important variables by their value on the liberalization dummy are found in Appendix 4. Graphs of structural change in individual countries are also available in Appendix 6. Looking at specific sectors, Figure 3 below shows the evolution of productivity in the different sectors. On average, mining has been the most productive sector by far, but utilities and finance has caught up during the past decade. Manufacturing has had a fairly stable and mediocre productivity over time. The low productivity in manufacturing can be explained by the inclusion of informal firms in the sector. The agriculture and government sectors are by far the least productive sectors.
  • 43. 42 Figure 3. Sectoral productivity over time Finally, Figure 4 below shows the evolution of sector labor shares. Agriculture is by far the largest sector. To facilitate the identification of changes in the other sectors, agriculture has its own y-axis (left). The most evident development is the fall of the agriculture labor share (blue line). Wholesale and retail trade (red line) is the sector that has absorbed the largest part of the workers from agriculture. The manufacturing share (dark brown line) grew a little in the 1970s before dipping in the 1980s, rebounding slightly, and then stagnating. The most productive sector, mining (dark green), has a very low labor share that has declined further since the 1990s. Figure 4. Sectoral labor share over time
  • 44. 43 6.2Difference in differences Although some SSA countries have structural change data going back to 1961, I look at the period 1971-2010 in order to have a balanced panel. I also look at the 1981-2010 sub-sample. The results of the difference-in-differences regressions are shown in Table 2. Table 2. Difference in differences (1) (2) (3) (4) (5) (6) 71-10 71-10 71-10 81-10 81-10 81-10 Dependent variable Structural change contribution to growth Liberalization dummy 0.0229 0.0295* 0.0260* 0.0312** 0.0266** 0.0220* (0.0205) (0.0133) (0.0127) (0.0136) (0.0110) (0.00996) GDP per capita 0.0488** 0.0342 0.0299* 0.0160 0.0157 (0.0155) (0.0189) (0.0140) (0.0118) (0.00937) Agriculture labor share 0.176*** 0.127*** 0.220*** 0.152*** 0.165*** (0.0235) (0.0266) (0.0262) (0.0272) (0.0253) Exports primary share 0.0254 -0.00391 -0.0674** (0.0275) (0.0310) (0.0272) Overvalued XR index -0.0348* -0.0436*** -0.0491*** (0.0182) (0.0126) (0.0119) Observations 88 84 84 66 66 64 R-squared 0.121 0.358 0.387 0.474 0.540 0.575 Number of countries Country & Time FE Cluster SE by country 11 Yes Yes 11 Yes Yes 11 Yes Yes 11 Yes Yes 11 Yes Yes 11 Yes Yes Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 In column (1), only the liberalization dummy is included as an explanatory variable. Although positive, it is not significant. As seen by the R-square of 0.121, this regression explains little of the variation in structural change. However, the liberalization dummy is significant at the 10 % level when I add control variables in column (2). Specifically, the liberalization dummy is only significant when I include the primary share of exports as a control. A limitation of this control variable is that it does not cover Botswana before 1991, leading to an unbalanced panel. There is thus the possibility of a sampling effect biasing the results.