Sub-Saharan Africa’s manufacturing trade: trade costs, zeroes, and export orientation
Maarten Bosker* Utrecht University
Abstract Sub-Saharan Africa (SSA) is only a marginal player on the world’s export and import markets. Moreover, and in contrast to other developing countries SSA trade is still largely dominated by trade in primary commodities. Diversifying SSA trade by developing an exporting manufacturing sector is viewed as a vital ingredient for Africa’s future economic success (IMF, 2007; World Bank, 2007). By focusing on bilateral manufacturing exports and imports of 44 SSA countries over the period 1993-2002 and by explicitly addressing the occurrence of the many ‘zero trade flows’, the results in this paper show that the high level of trade costs faced by many SSA countries constitutes a significant barrier to its potential of developing a manufacturing sector that is able to compete on world markets. It also reveals interesting differences in the way these trade costs affect the intensive and extensive margin of SSA’s manufacturing trade as well as in the importance of trade costs for intra-SSA trade and SSA trade with the rest of the world respectively.
*
I would like to thank Harry Garretsen, Rob Alessie, Joppe de Ree and Marc Schramm for stimulating discussions and useful comments and suggestions.
1.
INTRODUCTION
Sub-Saharan Africa (SSA) is only a marginal player on the world’s export and import markets. Since 1970, the region’s share in global trade (exports plus imports) has declined from about 4% to a mere 2% in 2005 (IMF, 2007), with a similar decline when looking at imports and exports separately (Subramanian and Tamirisa, 2003). Moreover, and in contrast to other developing countries (see Collier, 2002), SSA trade is still largely dominated by trade in primary commodities. Manufacturing exports constitute about 30% of total SSA exports, with natural resource related manufacturing activity constituting the bulk of manufacturing exports, and there is little sign of an increase in export diversification over the last 20 years (IMF, 2007). SSA’s little diversified export pattern, with its heavy dependence on primary commodities, is viewed as a cause of concern limiting its economic development. Primary commodity dependence is associated with poor economic perforance (Sachs and Warner, 1997; Collier, 2002). It makes countries much more vulnerable to price shocks on the volatile world markets for primary commodities. Also, wealth effects associated with the abundance of natural resources pulls resources out of the manufacturing sector, promoting de-industrialization (Dutch disease), and hereby severely hampering the prospects for manufacturing export-led growth (Sachs and Warner, 2001). Arguably even worse, primary commodity dependence tends to produce poor governance and may even increase the probability of civil war as it provides financial means for rebel groups (Collier, 2002). Developing the manufacturing (export) sector is therefore viewed as a vital ingredient for Africa’s future economic success (IMF, 2007; World Bank, 2007). Empirical evidence showing the importance of the manufacturing sector in economic growth is abundant. Hausmann, Pritchett and Rodrik (2005) for example show the positive role of exports in growth accelerations; in fact Asia’s economic success is largely ascribed to the development of an exporting manufacturing sector (Radelet et al., 1997). Manufacturing exports boost productivity as it increases firms’ efficiency through a process of learning by doing (Bigsten et al, 2004; Menigstae and Pattillo, 2004; Bigsten and Söderbom, 2006; van Biesebroeck, 2005) and by fostering technological progress through knowledge spillovers (Radelet et al., 1997). Also, by entering export markets firms can overcome the constraint of small domestic market size faced by most African countries (Collier and Venables, 2007).
1
Several reasons for SSA’s inability to develop a manufacturing sector that is able to compete on world markets have been put forward. Recent microeconomic studies at the individual firm level (see o.a. Bigsten et al., 2006; Van Biesebroeck, 2005; Mengistae and Pattillo, 2004; Clarke, 2005; Söderbom and Teal, 2003) show that high levels of entry barriers, firm size, foreign ownership and firm’s cost efficiency1 positively affect a firm’s export potential. This is complemented by the macro-economic evidence on SSA’s poor trade performance that has identified high trade costs, potentially worsened by poor governance, as one of the main causes of the region’s marginal role on world markets (see o.a. Collier, 2002; Foroutan and Pritchett, 1993; Coe and Hoffmaister, 1999; Limao and Venables, 2001; Amjadi and Yeats, 1995; Redding and Venables, 2004; Francois and Manchin, 2007). High trade costs are detrimental for firms’ competitiveness on global markets, both by increasing the price of intermediate goods used in production and by decreasing the amount of revenue earned for the final product. For the small SSA economies, that do not significantly affect world prices, this means that firms, in order to be able to compete in global markets, will have to make up for these high trade costs by reducing costs elsewhere in the production process (usually by offering lower wages). This especially hurts labor-intensive manufacturing export activity, where profit margins are thin (Radelet and Sachs, 1998) and that makes intensive use of imported intermediates (see Collier, 2002) in the production process. Trade costs can tip the balance between losing or earning money by exporting. Poor government policy has the potential to harm trade in two ways. First they may prevent firms from starting production, let alone undertaking import and export activity, in the first place by creating a highly uncertain investment climate, with both higher costs and risks of doing business (Collier, 2002). Second, they exacerbate the problem of high trade costs. High tariffs, costs of delay, higher costs of insurance, little regional integration (or even civil conflict) and poor public service delivery that can result from poor government policies all have a detrimental effect on the amount of trade costs faced by importers and exporters alike (Amjadi and Yeats, 1995). This paper aims to contribute to this latter macroeconomic strand of literature, by making three contributions. First, given the alleged importance of developing an exporting manufacturing sector in stimulating economic growth, I focus on the 1
Although in case of firm’s cost efficiency the results suggest that causality runs from exporting to efficiency (Bigsten et al., 2004; Mengistae and Pattillo, 2004).
2
determinants, and the role of trade costs in particular, of manufacturing imports and exports only instead of looking at total SSA imports and exports. Second, instead of focusing on SSA’s trade performance compared to the rest of the world (Limao and Venables, 2001; Subramanian and Tamirisa, 2003) or to other developing countries (Coe and Hoffmaister, 1999; Foroutan and Pritchett, 1993), I look at trade of 44 SSA countries only. Hereby this paper is the first to focus on SSA trade only, looking for the determinants of the large differences in manufacturing trade performance between SSA countries (IMF 2007; Subramanian and Tamirisa, 20032) instead of trying to answer the question why SSA is (or is not) different from other developing countries or the rest of the world (ROW). Also I distinguish explicitly between SSA trade with other SSA countries and with countries in the ROW, as it has been shown that different variables may affect intra-SSA trade differently than SSA trade with the ROW (Limao and Venables, 2001; Mengistae and Pattillo, 2004; Foroutan and Pritchett, 1993). And third, following a recent paper by Helpman et al. (2007), I argue in favor of using a two-step estimation procedure stressing the need to model the probability to and the amount of trade separately (hereby more adequately addressing the issue of zero trade flows)3. Besides being preferred from an econometric point of view, this offers interesting insights in the (possibly different) ways that trade costs affect both the intensive and extensive margin of SSA manufacturing trade. The next section presents the gravity model that, as in most papers in the empirical trade literature, is the theoretical basis of my empirical analysis. It also discusses some of the difficulties faced when taking this model to the data, arguing in favor of an empirical strategy that does not model the probability to and the amount of trade simultaneously. Section 3 briefly discusses the data set used, focusing on the large number of zero trade flows and showing the substantial variation between SSA countries in terms of export and import performance. Section 4 presents the empirical results and section 5 concludes.
2
EXPLAINING BILATERAL TRADE FLOWS – THE GRAVITY MODEL
2
Subramanian and Tamirisa (2003) do not look at variation at the country level however, instead they look at the variation between the two broad groups of SSA countries, namely Anglophone and Francophone countries. 3 Where part of the argument rests on the introduction of a test for the appropriateness of assuming the same process for both the probability to and the amount of trade that is based on the widely used (see e.g. Limao and Venables, 2001; Foroutan and Pritchett, 1993; Redding and Venables, 2004) Tobit estimation method (see Appendix B).
3
The gravity model has become the standard analytical framework for the modelling of bilateral trade flows. It has been the main vehicle for empirical research in international trade since the 1960s, and is by now also firmly grounded in economic theory; in fact its theoretical foundations are multiple (see Anderson and van Wincoop, 2004; Coe and Hoffmaister, 1999 and Santos Silva and Tenreyro, 2006 for a good overview). The simplest version of the gravity equation relates the amount of exports from country i to country j at time t positively to the two countries’ economic mass at time t (usually measured by GDP), Yit and Yjt, and negatively to the level of trade costs, Tijt, involved in shipping goods from country i to country j at time t. In empirical studies an error factor is added, to take account of deviations from theory, resulting in the overall gravity equation to take the following form: α 0 + ε ijt
EX ijt = e
Yitα1Y jtα 2 Tijtα3
(1)
where a0 is included as a constant, a1, a2 and a3 are parameters to be estimated, with a1 and a2 expected to be positive and a3 to be negative, and εijt is a well behaved error term. Empirical work using the gravity equation is abundant (see e.g. Anderson and van Wincoop, 2004 for a discussion of several influential papers). Papers that look explicitly at SSA are Subramanian and Tamirisa (2003), Coe and Hoffmaister (1999), and Foroutan and Pritchett (1993). As already briefly mentioned in the introduction these papers look at total SSA imports and exports and use the gravity equation to answer the question whether or not SSA, given its economic mass and level of trade costs, over- or undertrades compared to the ROW or compared to other developing countries. In a nutshell, and leaving aside many of the more subtle conclusions in these papers, they generally find that SSA countries overtrade compared to other developing countries, but that the amount of overtrading has dropped over time (Coe and Hoffmaister, 1999; Foroutan and Pritchett, 1993). Compared to the rest of the world Africa seems to undertrade4, which, as Limao and Venables (2001) find, can to a large extent be attributed to the poor quality of infrastructure in SSA.
4
Subramanian and Tamirisa (2003) show that this conclusion does not hold for Anglophone Africa that seems to be an average trader.
4
The empirical analysis in this paper differs from the above-mentioned papers in several ways. First I focus on SSA manufacturing trade only. Second, I do not address the question whether or not SSA over- or undertrades compared to other (developing) countries; instead I stay somewhat closer to Limao and Venables (2001) and provide evidence on the hampering effect of trade costs on intra-SSA trade and SSA trade with the ROW. Focusing on the determinants of SSA (manufacturing) trade data only, this paper is the first to provide insights into the question why some countries in SSA are showing strong manufacturing trade performance (e.g. Mauritius, Ghana, and South Africa), whereas others are not able to effectively enter world markets (e.g. Gabon and Mozambique), see also Ng and Yeats (2003). This paper furthermore differs from the earlier contributions looking at SSA trade performance by employing a two-step selection based estimation procedure to estimate the coefficients in (1), modeling the probability to and the amount of trade separately. All previous papers looking at SSA trade estimate (1) directly, either in its non-linear form or after log-linearizing. The presence of zero trade flows complicates matters however. Estimating (1) in log-linear form (Limao and Venables, 2001; Foroutan and Pritchett, 1993) involves either discarding these zero trade flows altogether or using censored regression techniques (Tobit). Discarding zero trade flows may still provide consistent estimates, but only so under the questionable assumption of exogenous sample selection. Tobit estimation requires the definition of a censoring value, which is usually chosen arbitrarily5 shedding doubt on the consistency of the obtained estimates (Santos Silva and Tenreyro, 2006). Arguably more restrictive, it also implicitly assumes that the exact same model explains the probability to and the amount of trade. Given the problem with estimating (1) in loglinear form, others have proposed to estimate (1) in its non-linear form using NLS techniques (Coe and Hoffmaister, 1999; Subramanian and Tamirisa, 2003) or the recently proposed PPML technique that controls more adequately for the heteroscedasticity present in trade data (Santos Silva and Tenreyro, 2006). Similar to Tobit estimation, estimating (1) directly implictly makes the (questionable) assumption that the same process (same variables, same parameters) drives both the probability to and the amount of trade. This amounts 5
Common choices are the minimum value of observed trade, see e.g. Limao and Venables (2001), or 1.000 dollars (by adding 1 to zero trade flows, which amount to a censoring values of 1.000 dollars as trade flows are usually measured in thousands of dollars), see e.g. Foroutan and Pritchett (1993).
5
to saying that the gravity equation provides an accurate explanation for the absence of trade between two countries. Standard gravity theory predicts that each country trades with all other countries however: it does not provide an explanation for zero trade flows to occur6. Estimating (1) in non-linear form does not exclude the zero trade flows (as it avoids taking the log of zero), however it does so in a way that these zero trade flows are totally ascribed to the error term (relying on arguments of measurement error or reporting errors, see Santos Silva and Tenreyro, 2006, p.643). The occurrence of trade flows can by definition not be ascribed to the explicitly modeled part of the regression model (i.e. the gravity model). Also estimating (1) in non-linear form, employing the recently popular PPML method (Santos Silva and Tenreyro, 2006), faces the so-called ‘excess zeros’ problem that can result in biased estimates of the parameters of interest (see e.g. Greene (2000, ch.19), Greene (1994) or Mullahy, 1997). Especially when the number of zero observations is substantial (about 50% in case of SSA manufacturing trade), standard Poisson estimation severely underpredicts these zero observations7; particularly if, contrary to the assumption when using standard Poisson regression, these zeros are the result of a different underlying process than the non-zero observations8 (Appendix C shows that the number of zeroes is indeed substantially underpredicted when using Poisson estimation). Recent advances in international trade theory do provide a rationale for zero trade flows based on e.g. fixed entry costs (Hallak, 2006; Helpman et al., 2007). These papers call for the separate modeling of the probability to and the amount of trade, employing two-step selection estimation techniques. Here I follow the latter paper and employ a two-step Heckman procedure (also used in a recent paper by Francois and Manchin, 2007) to estimate the parameters of the gravity equation (1). Furthermore I propose a simple procedure to test for the appropriateness of modeling 6
See Hummels and Haveman (2004). Eaton and Kortum (2002) introduce a model that accounts for zero trade flows at the disaggregated level, at the aggregate level however their model does still imply positive trade flows between all country pairs. 7 Poisson regression techniques do exist that take account of the excess zeros by using a 2step approach (zero inflated Poisson (ZIP) for example). An arguably large disadvantage of these models is that they typically, contrary to the two-step Heckman procedure, have to assume independence of the processes describing the zero and the non-zero observations respectively (most importantly this rules out the possibility of shocks affecting both the probability to and the amount of trade). This independence assumption, in effect makes using ZIP very similar to using Poisson on the nonzero observations only (see also Appendix C). 8 See for a very useful overview of the various methods and their assumptions, the following document available on the PoveryNet website of the World Bank: http://www1.worldbank.org/prem/poverty/health/wbact/health_eq_tn11.pdf
6
the probability to and the amount of trade using the same underlying model, based on estimating the gravity model using Tobit estimation (see Appendix B for the details).
3.
THE DATA – A FIRST LOOK AT SSA MANUFACTURING TRADE
The data on SSA manufacturing trade are collected from CEPII’s Trade and Production Database, which contains information on bilateral manufacturing trade flows from 1976-20029. Within this dataset I focus on bilateral trade flows involving at least one SSA country (exporter or importer). Given poor data availability before 1993 (over this period SSA manufacturing import data are only given for 6 countries10), I further narrow the time dimension of the data to encompass the 10 year period from 1993-2002. This leaves a data set containing information on bilateral manufacturing trade flows for 44 SSA countries and 148 countries in the ROW. If there were no missing values, this would give a dataset with information on 149160 bilateral trade (i.e. 18920 intra SSA and 130240 SSA-ROW) relationships. Missing values reduce the total sample to 78657 observations (8574 intra-SSA and 70083 SSA-ROW). Table 1. Bilateral SSA manufacturing trade flows year
non-missing intra SSA
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 1993-2002
363 572 1017 1031 967 1012 1118 1032 860 602 8574
non-missing SSA-ROW 4198 5721 7272 7482 7553 7867 8528 8104 7296 6062 70083
% non zero intra SSA 0.48 0.47 0.44 0.46 0.49 0.48 0.43 0.44 0.42 0.45 0.45
% non zero SSA-ROW 0.51 0.49 0.46 0.48 0.48 0.47 0.45 0.40 0.41 0.43 0.46
Of these 78657 almost half, i.e. 35824 (or 46%) are non-zero trade flows. Table 1 gives additional details on the issue by showing the number of (non-zero) observation on a yearly basis. It can be seen that, whereas the number of missing observations decreases over time, the % if zero trade flows does not change substantially. Also this number does not differ substantially between intra-SSA trade and SSA trade with the 9
http://www.cepii.fr/anglaisgraph/bdd/TradeProd.htm. An explanation of the dataset is given at http://www.cepii.fr/tradeprod/TradeProd_cepii.xls. 10 South Africa, Kenya, Ethiopia, the Comoros, Malawi and Madagascar.
7
ROW. The presence of a substantial amount of zero observations highlights the importance of the choice of dealing with zero trade flows when estimating the gravity equation (see also Helpman et al. (2007), Coe and Hoffmaister (1999)). Given the aim of this paper to explain the variation between SSA countries in trade performance, Figure 1 below shows that there indeed exist large differences in manufacturing trade performance between SSA countries.
0
0
10
5
20
10
30
Figure 1. Manufacturing imports and exports as a share of GDP
0
.1
.2 share imports in GDP
.3
0
.4
.1
.2 share exports in GDP
.3
Trade performance, measured by the share of imports and exports in GDP respectively, differs substantially among SSA countries. In several countries’ (e.g. Mauritius, South Africa, the Seychelles) the manufacturing exports share in GDP is much higher than that of poor performing countries such as Ethiopia, Rwanda and Sudan11 and a similar pattern is observed when looking at imports.
0
0
2
1
4
2
6
3
8
4
10
Figure 2. % manufacturing imports from and % manufacturing exports to other SSA countries
0
.2
.4 % imports from ssa
.6
.8
0
11
.2
.4 % exports to ssa
.6
.8
Compared with e.g. some of the emerging countries in Asia, this share is still quite low however (see Elbadawi, 1999).
8
Figure 2 furthermore shows that there also exist large differences in the % of total manufacturing imports and exports that SSA countries trade with other SSA countries. On average an SSA country imports (exports) about 12% (6%) of its total manufacturing imports (exports) from (to) other countries in SSA. However, substantial between SSA-country variation exist, with countries like South Africa and Cape Verde importing less than 5% of total manufacturing imports from other SSA countries, and countries such as Zambia, Malawi, Uganda and Zimbabwe importing more than 30% of manufacturing imports from within SSA. The same holds for exports, with countries such as Equatorial Guinea, Burkina Faso, Mauritius and the Comoros exporting less than 5% of total manufacturing exports to other SSA countries, whereas Djibouti, Kenya and Mauritania are exporting 40% or more of manufacturing exports within SSA.
4.
THE DETERMINANTS OF SSA MANUFACTURING TRADE
In order to shed some light on the question why we observe such differences in manufacturing trade performance between SSA countries, I now turn to the estimation of the gravity equation (1) using manufacturing import and export data only. Note that estimating a gravity equation using manufacturing trade data only, can also be justified theoretically. For example, trade models exhibiting increasing returns to scale and monopolistic competition derive the gravity equation explicitly for the manufacturing sector (see e.g. Redding and Venables, 2004 that also estimate a gravity equation based on this type of model). Rewriting (1) in loglinear form, and allowing for a year-specific intercept to take account of the growing amount of world trade gives: ln EX ijt = α 0 + α t + α1 ln Yit + α 2 ln Y jt + α 3 ln Tijt + ε ijt
(2)
I further specify trade costs, Tijt, to be a multiplicative12 function of the following variables that are commonly used in the literature (see Appendix A for more details on each of the variables): distance (D), sharing a common border (B), common language (CL), common colonial heritage (distinguishing between sharing a common colonizer (CC) and having had a colony-colonizer relationship (CR)), being 12
This is the usual choice in the gravity literature (see e.g. Limao and Venables, 2001; Subramanian and Tamarisa, 2003). See Hummels (2001) for a critique on this, argueing in favor of an additive specification instead.
9
landlocked (ll), being an island (isl), an index measuring the quality of infrastructure (inf) and membership of the same regional or free trade agreement (RFTA). After taking logs this trade costs function looks as follows:
ln Tijt = χ1 ln Dijt + χ 2 ln Bijt + χ 3 ln CLijt + χ 4 ln CCijt + χ 5 ln CRijt + χ 6llit + χ 7 ll jt + χ8islit + χ 9isl jt + χ10infit + χ11inf jt + χ12 RFTAijt
(3)
, which is substituted in (2). Besides including variables related to the level of trade costs faced by importers and exports alike, I also include – apart from the standard GDP variables – the following variables measuring a country’s economic mass, taking account of some of the peculiarities of SSA (again see Appendix A for the details). First I include two dummy variables that indicate whether a country experienced outbreaks of civil unrest in a specific year: civil conflict and civil war; where civil war indicates more intense fighting than civil conflict. Outbreaks of (domestic) violence or even outright war are likely to affect trade both directly and through their detrimental effect on trade costs (see Simmons, 2005). Given the fact that SSA has been the most conflict-ridden continent over the last few decades, a closer look at the effect of civil unrest on trade flows, seems warranted. Second, I include two variables that also can be argued to be of particular influence for SSA’s ability (and developing countries more generally) to trade in manufacturing goods. The first is the share of people living in rural areas. Manufacturing activity is usually located in or near urban centers (it is sometimes referred to as the ‘modern urban sector’ in models of urban-rural migration, see e.g. Harris and Todaro, 1970). Subsequently, higher urbanization rates are associated with a larger manufacturing sector, making it more likely that a country engages in trade in manufactures (both importing intermediate inputs and exporting manufacturing goods). Also, Ancharaz (2003) shows that higher urbanization shares increase the likelihood of trade policy reform, facilitating imports and exports of manufacturing goods. Finally the welfare level of the urban population is generally higher than that
of the rural population in SSA (see Sahn and Stifel, 2003), resulting in higher demand for imported manufacturing products13. Third, I include the share of the labor force employed in agriculture. The more people that are needed in agricultural production, the less there are available to produce manufacturing goods. The share of the labor force in agriculture has hardly changed in SSA over the last decades (Temple and Wößmann, 2006). Karshenas (2003) shows that, in contrast to Asia, where technological improvements in agricultural production resulted in large increases in agricultural labor productivity and making a large part of the labor force available for labour-intensive manufacturing production, agricultural labor productivity has hardly improved in SSA. At the same time urbanization14 and the general increase in population have increased demand for food in SSA increasing labor demand by the (inefficient) agricultural sector. Tiffen (2003) for example shows that in some regions in Africa the urban grain need per rural person increased by a factor 10 over the period 1951-1991. As a result agriculture still employs the largest part of the population in many SSA countries. Also, and maybe even more importantly, a lower share of the work force employed in agriculture has been shown to increase aggregate total factor productivity (Temple and Wößmann, 2006; Poirson, 2001). Poirson (2001) even shows that the stable, large share of the labor force employed in agriculture accounts for the systematic TFP growth gap between SSA and other regions. Based on these results, I argue that the percentage of the labor force employed in agriculture can be used as a proxy for aggregate productivity levels. This is particularly appealing, as recent insights from the theoretical international trade literature (e.g. Melitz, 2003) show that only the more productive firms will undertake export activity. With the use of the percentage of the labor force employed in agriculture as an additional variable in the gravity equation, I hope to capture these effects. Including these additional variables to the gravity equation, results in the following model that I estimate:
13
Other authors have included GDP per capita as a measure of welfare to the gravity equation. I choose to use % urban population instead. % urban population is highly correlated with GDP per capita, and results are very similar when using GDP per capita instead. 14 One may be worried that the urbanization rate and the share of the labor force in agriculture are almost perfectly correlated. In fact, they are not, the correlation between the two variables is about 0.65 [p-value: 0.00].
ln EX ijt = α 0 + α t + α1 ln Yit + α 2 ln Y jt + ln Tijt + + α 4 %ruralit + α 5 %rural jt + α 6 %agricit + α 7 %agric jt
(4)
+ α 8 %cconflit + α 9 %cconfl jt + α10 %cwarit + α11 %cwarjt + ε ijt
, with lnTijt as in (3). Also I interact all trade cost variables (and GDP) with a dummy that takes the value of 1 when considering intra-SSA trade, i.e. when both the importing and exporting country are located in SSA. Hereby I allow the different components of trade costs to have a different effect on intra-SSA manufacturing trade and SSA manufacturing trade with the ROW15.
4.1
Estimation results – choice of estimation method
Table 1 shows the estimation results when estimating (4) either using OLS on the non-zero observations only, using Tobit, adding 1 to the zero flows before taking logs, using Probit with the dependent variable being one (zero) if a positive (zero) trade flow is observed, and using the Heckman 2step estimation procedure using the share of labor in agriculture only in the selection equation16. Adding “ssa” to a certain variable denotes that variable interacted with an intra-SSA dummy; significance of such a “ssa”-variable indicates a different effect of that particular variable on intraSSA trade compared to SSA trade with the ROW. Before focusing on the actual parameter estimates, I first draw attention to the preferred choice of estimation strategy, arguing in favor of the 2step estimates, which I will then describe in more detail. First I consider the appropriateness of using the same model for both the probability to and the amount of trade, as e.g. Tobit or Poisson estimation do. As I discuss in Appendix B a way to test for this is to compare the results obtained using Tobit estimation to those obtained using Probit; these should (after dividing the Tobit estimates by the estimated variance) not differ too much when correctly assuming the same model for the probability to and the amount of trade (see p.573 in Wooldridge, 2002). When comparing the Tobit (column 2) with the Probit (column 3) results, differences immediately show up. Several variables, o.a. 15
Limao and Venables (2001), Mengistae and Pattillo (2004) and Foroutan and Pritchett (1993) all show that trade costs may have different effects on intra-SSA and SSA-ROW trade. 16 I will come back to this choice in more detail. Note that the selection equation is estimated using Probit, and the results are thus shown in column 3 of Table 2. The results when using Poisson and/or zero inflated Poisson regression, shown in Appendix C, are not explicitly discussed here. Note that the ‘excess zero’ problem is likely to render the Poisson results substantially biased, whereas using the zero-inflated poisson (ZIP) makes the unrealistic assumption that the probability to and the amount of trade are two independent processes (hereby not allowing for shocks that affect both processes).
common border, distance ssa, contiguity, landlocked importer and % in agriculture exporter, are only significant in one of the two cases. To formally test for the validity of assuming the same model for the probability to trade and the amount of trade however, I employ a Hausman test, as proposed in Appendix B. This results in an overwhelming rejection of the null hypothesis (the test statistic is 995.9, which is much larger than 62.83 or 71.20, the 5% and 1% critical value of the χ(46)distribution respectively), indicating that Tobit, by wrongly assuming the same model for the probability to and the amount of trade, produces inconsistent estimates17. This leaves column 1 and column 4, showing the results when using OLS on the non-zero observations only or Heckman 2step estimation respectively. Both methods do not assume the same model for the probability to and the amount of trade. Recall, however, that for OLS to produce consistent estimates, the strong assumption of exogenous sample selection is required. If OLS on the non-zero observations would be consistent, the estimates obtained using 2step estimation would not differ a lot from those obtained using OLS. Comparing column 1 and 4 immediately shows that this does not seem to be the case. Instead, the estimated coefficients in column 4 are usually larger (in absolute value) than those reported in column 1. This difference can be explained by the fact that most variables affect the probability to and the amount of trade in a similar direction (see column 3). Using OLS on the selected sample only, ignores the effect of each variable on the probability to trade, resulting in lower overall estimates. Moreover the estimated coefficient on the inverse Mills ratio is very significant, indicating that the assumption of exogenous sample selection has to be rejected. Based on these results I strongly prefer the 2step estimation results to both Tobit and OLS on the non-zero sample only. However note that, as already briefly mentioned, I exclude the percentage of the labor force employed in agriculture from the conditional mean equation when using 2step estimation. Hereby assuming that this variable (after correcting for the other included variables) does not affect the amount of trade but only the probability to trade; see the assumption in (A4). The reason for doing this is that (see p.589 Wooldridge, 2003) the results using the Heckman 2step procedure are much more convincing when using such an exclusion restriction. In 17
This finding does not depend on the censoring value chosen (which is 0 here as I have added 1 to all non-zero trade flows). Results when applying Tobit after adding the minimum value of all the non-zero manufacturing trade flows to the zero observations (as e.g. Limao and Venables (2001) do) also strongly reject the consistency of the Tobit estimates. Results are available upon request.
principle this is not necessary, but it becomes quite difficult to distinguish sample selection from misspecifying the functional form of the conditional mean model when not making use of an exclusion restriction18. The choice of using the percentage of the labor force in agriculture as a determinant of the probability to trade only is made for the following reasons. First, after correcting for the other variables included, it is not significantly correlated with the amount traded when (wrongly) assuming exogenous sample selection (see column 1 in Table 2). Moreover, the results of the Probit estimation show that it does significantly affect the probability to trade (see column 3). Besides these more econometric reasons, using the percentage of the labor force in the selection equation only can also be justified from an economic point of view. Recent insights from the theoretical international trade literature (e.g. Melitz, 2003) show that only the more productive firms will undertake export activity, as only these firms are producing efficiently enough to be able to compete on world markets. In case of SSA, microeconomic studies at the individual manufacturing firm level, e.g. Van Biesebroeck (2005) and Mengistae and Pattillo (2004), seem to confirm this hypothesis19. As already mentioned earlier in this section, Temple and Wößmann (2006) and Poirson (2001) have shown that a lower share of the labor force employed in agriculture increases aggregate total factor productivity. When viewed as a proxy of the economy’s aggregate productivity, the fact that the percentage of the labor force in agriculture only significantly affects the probability that a SSA economy exports or imports manufacturing goods, can be viewed as being consistent with trade theories of the Melitz (2003) type emphasizing productivity as a positive determinant of the probability to trade. Having established 2step estimation as the preferred way to estimate the effect of each of the included variables on the amount of trade, what do the results indicate as important determinants of SSA manufacturing trade? More in particular what do they tell about the alleged trade-hampering effect of trade costs when focusing exclusively on SSA? Moreover, are there significant differences in the way that trade
18
The inverse Mills ratio is essentially a non-linear function of the variables included in the selection equation. When no additional variables are added to the selection equation, significance of the inverse Mills ratio may indicate endogenous sample selection but it may as well indicate that some (or all) variables should enter the conditional mean equation in nonlinear form. 19 Although it is in general hard to distinguish self-selection of productive firms into exporting acitivty from becoming more productive through learning-by-exporting (see Bigsten and Söderbom, 2006).
Table 2. Estimation results dependent variable
ln trade
estimation method
OLS nonzeros 1993-2002
time period
coefficients ln distance
ln trade + 1
p-values
trade? 0/1
ln trade
Tobit
Probit
Heckman 2 step
1993-2002
1993-2002
coefficients
p-values
coefficients
p-values
1993-2002 coefficients
p-values 0.000
-0.780
0.000
-1.898
0.000
-0.395
0.000
-1.514
ln distance ssa
-0.048ns
0.494
-0.199 ns
0.078
-0.140
0.000
-0.025 ns
0.773
ln gdp importer
0.753
0.000
1.903
0.000
0.464
0.000
1.416
0.000
ln gdp importer ssa
-0.371
0.000
-0.421
0.000
-0.060
0.000
-0.433
0.000
ln gdp exporter
0.937
0.000
1.947
0.000
0.436
0.000
1.559
0.000
ln gdp exporter ssa
-0.225
0.000
-0.262
0.000
-0.031
0.028
-0.254
0.000
colony - colonizer
2.104
0.000
2.577
0.000
1.702
0.000
2.479
0.000
common colonizer
0.634
0.000
1.442
0.000
0.309
0.000
1.116
0.000
-0.202 ns
0.059
-0.061 ns
0.702
-0.011 ns
0.829
-0.076 ns
0.548
ns
0.219
-1.246
0.007
0.000
2.671
0.000
common colonizer ssa contiguity
-1.056
0.001
-1.378
0.026
-0.272
contiguity ssa
1.958
0.000
3.063
0.000
0.821
common off language
0.557
0.000
1.129
0.000
0.289
0.000
0.870
0.000
common off language ssa
-0.325
0.001
-0.700
0.000
-0.164
0.000
-0.561
0.000
landlocked exporter
0.023 ns
0.556
-0.657
0.000
-0.229
0.000
-0.243
0.000
landlocked exporter ssa
-0.697
0.000
-0.545
0.000
-0.066 ns
0.106
-0.760
0.000
landlocked importer
-0.436
0.000
-0.087 ns
0.150
0.039
0.027
-0.293
0.000
ns
landlocked importer ssa
0.336
0.000
0.082
-0.006
0.888
0.340
0.002
island exporter
0.355
0.000
0.753
0.000
0.166
0.000
0.564
0.000
island exporter ssa
-1.245
0.000
-2.933
0.000
-0.690
0.000
-2.218
0.000
ns
0.245
ns
island importer
0.061
0.118
0.687
0.000
0.228
0.000
0.407
0.000
island importer ssa
-0.394
0.016
-2.637
0.000
-0.748
0.000
-1.552
0.000
ln infrastructure exporter
0.317
0.000
0.572
0.000
0.126
0.000
0.478
0.000
ln infrastructure exporter ssa
0.646
0.000
1.243
0.000
0.277
0.000
1.058
0.000
ln infrastructure importer
0.181
0.000
0.096
0.088
-0.013
0.449
0.181
0.000
ln infrastructure importer ssa
0.489
0.000
0.913
0.000
0.223
0.000
0.835
0.000
RTA or FTA
-0.601
0.007
-0.136 ns
0.732
0.013 ns
0.917
-0.467 ns
0.121
RTA or FTA ssa
1.358
0.000
1.079
0.010
0.085 ns
0.529
1.496
0.000
civil conflict importer
-0.292
0.000
-0.732
0.000
-0.196
0.000
-0.552
0.000
civil conflict exporter
-0.521
0.000
-0.499
0.000
-0.062
0.000
-0.605
0.000
civil war importer
-0.457
0.000
-1.214
0.000
-0.322
0.000
-0.924
0.000
civil war exporter
-0.978
0.000
-1.706
0.000
-0.368
0.000
-1.467
0.000
-0.063 ns
0.474
-0.387
0.002
-0.030 ns
0.431
-0.359
0.000
% rural population exporter
-1.978
0.000
-3.526
0.000
-0.651
0.000
-3.033
0.000
% in agriculture importer
0.007 ns
0.909
-0.257
0.011
-0.159
0.000
-
-
% in agriculture exporter
0.022 ns
0.753
-0.114 ns
0.231
-0.119
0.000
-
-
% rural population importer
nr observations
34524
(pseudo) R2
74401
74401
74401
0.4245
0.1775
0.3498
-
Tobit variance
-
4.210 [0.018]
-
-
Hausman test Tobit
-
995.9 [0.000]
-
-
Mills' ratio
-
-
-
3.055 [0.000]
Notes: p-values based robust standard errors in case of OLS, else standard. ns means not significant at the 5% level. In each of the estimations I have included time dummies and I allow for a separate intercept for intra-SSA trade. p-value of the Hausman test is calculated on the basis of a χ(46) distribution.
costs affect the probability to trade compared to the amount of trade? And similarly is trade with other SSA countries differently affected than trade with the ROW? The results, in column 3 and 4 of Table 2, show some interesting findings.
4.2
Estimation results –the amount of SSA manufacturing trade
First I discuss the results regarding the intensive margin of SSA manufacturing trade in column 4, as this (contrary to the extensive margin) is what has been looked at in detail by earlier papers looking at SSA trade performance (Foroutan and Pritchett, 1993; Limao and Venables, 2001; Coe and Hoffmaister, 1999 and Subramanian and Tamirisa, 2003). It allows me to more readily compare my results to these previous studies. The next subsection will focus in more detail on the extensive margin of SSA’s manufacturing trade, discussing some of the interesting differences with the results discussed here. First regarding the non-trade cost related variables. Importer and exporter GDP both have the expected positive sign; interestingly, the trade-stimulating effect of an increase in GDP is much lower when considering intra-SSA trade. This indicates that an increase in GDP of a SSA country will increase its trade with the ROW more than its trade with other SSA countries, suggesting that as SSA countries get richer the focus of their manufacturing trade activity shifts away from other SSA countries in favor of countries in the ROW. Civil unrest negatively affects trade, hereby confirming the finding of e.g. Simmons (2005) that violence severely hampers trade. Furthermore when comparing the effect of civil conflict with outright civil war, the results show that the latter, as expected, has a much higher negative impact on trade flows. Interestingly, civil unrest seems to hurt exports more than imports. Finally, a higher degree of urbanization results in more exports and imports of manufacturing goods. Part of this effect is likely due the fact that countries with a higher urbanization rate tend to have less restrictive trade policies (Ancharaz, 2003). The effect on manufacturing exports is however much larger than on imports. An explanation for this finding could be that, given that manufacturing in SSA is mostly located in urban areas and (unskilled) labor-intensive, increasing urbanization will suppress wages due to increased supply of unskilled labor, hereby lowering firms’ production costs, making it easier for them to be competitive on world markets. Next I turn to the results regarding the effect of trade costs on SSA manufacturing trade, offering particularly interesting insights in the different effect of
some of these trade cost variables on intra-SSA and SSA-ROW trade. First I discuss the results regarding the bilateral trade cost variables. Distance negatively affects the amount of trade between countries. As Foroutan and Pritchett (1993), but contrary to Limao and Venables (2001), I do not find evidence that the penalty on distance is higher for intra-SSA trade. For intra-SSA trade I find the usual positive effect of sharing a common border on trade flows (see e.g. Limao and Venables, 2001; Subramanian and Tamirisa, 2003 and Foroutan and Pritchett, 1993). For SSA-ROW trade however, I find a surprising negative effect. Considering that the only SSA countries that border non-SSA countries are those bordering North African countries, this indicates that these SSA countries trade less with their North African neighbors than with nonAfrican countries (see also IMF, 2007); perhaps the impregnable Sahara desert plays a role here. Sharing a colonial history has a strong positive affect on the amount of trade in manufactures. Especially SSA trade with its former colonizer(s) is much higher than trade with other countries in the world, hereby explaining to a large extent the large share of SSA trade with Europe in overall SSA trade (see also IMF, 2007). Having a common colonizer also boasts bilateral trade and I find no indication that the effect is different for intra-SSA trade compared to trade with the ROW. Corroborating earlier findings by Foroutan and Pritchett (1993) and Coe and Hoffmaister (1999) I find that sharing a common language stimulates both intra-SSA and SSA-ROW trade. The trade facilitating effect of language similarity is much larger for trade with the rest of the world however (the common border and common colonizer variable may already be capturing some of the language effect in case of intra-SSA trade). A bilateral trade cost variable that is of particular interest is the variable capturing the effect of being a member of the same regional or free trade agreement (RFTA). I focus specifically on all-African RFTAs, whose effect has been mostly contested in the literature (Foroutan and Pritchet (1993) find ambiguous effects depending on the specific SSA trade agreement). I find that intra-SSA trade in manufactures substantially benefits from having a RFTA. Hereby providing evidence in favor of those who argue for increased African integration (one of the explicit goals of e.g. the African Union). The finding that having a RFTA does not significantly affect SSA-ROW trade is not so surprising since the only non-SSA countries being
part of an all-African RFTA are some of the North African countries (see also the results regarding the common border variable). The final three trade cost related variables that I look at are all country-specific variables, i.e. being landlocked, being an island and the quality of a country’s infrastructure. I find that being landlocked depresses both SSA imports and exports of manufacturing goods to the ROW, corroborating the findings in Coe and Hoffmaister (1999). Most manufactured goods are exported over water; having no direct access to the sea makes landlocked countries dependent on their neighboring SSA countries, increasing transport costs (see Limao and Venables, 2001) and having a negative effect on trade flows. The result when looking at intra-SSA trade show however some interesting differences. Being landlocked affects intra-SSA exports even more negatively than SSA trade with the rest of the world; on the contrary, being landlocked slightly increases the amount imported from other SSA countries. This difference is quite interesting, as it indicates that landlocked countries in SSA are more dependent on imported manufacturing goods from other SSA countries compared to the SSA countries that do have direct access to the sea. SSA countries with direct access to the sea, while exporting more to other SSA countries than their landlocked counterparts, import less from other SSA countries indicating that they are looking more towards the ROW for their imported manufacturing goods. Being an island nation in SSA increases trade with the ROW, confirming findings by e.g. Limao and Venables (2001). However I find that trade with other SSA countries is much lower for these same island nations. This suggests that these island nations (e.g. Mauritius, Cape Verde and Sao Tome and Principe) are oriented away from the African mainland when it comes to trade in manufacturing goods. The findings on these two ‘location’ variables suggest that SSA countries are much more oriented towards the ROW than towards other SSA countries when it comes to manufacturing trade: island nations trade much less with the African mainland and countries with direct access to the sea import more manufacturing goods from the ROW than from other SSA countries. The final trade cost related variable that I look at is the quality of a country’s infrastructure, which arguably is the most interesting variable from a policy perspective given the large amounts of money allocated by donors to fund infrastructrure improvements in SSA ($7.7 billion by members of The Infrastructure
Consortium for Africa alone20). Confirming the results in Limao and Venables (2001) and Francois and Manchin (2007), I find that improved quality of infrastructure has large positive effects on both the amount of manufacturing goods imported and exported. Maybe even more interestingly, improving the quality of infrastructure has a much larger positive effect on intra-SSA trade than on SSA trade with the ROW. These findings show that the current focus on improving SSA infrastructure (see e.g. the aim of the Sub-Saharan African Transport Policy Progam21 and The Infrastructure Consortium for Africa20) is warranted. Reducing transport costs by improving infrastructure will have a positive effect on the export potential of SSA’s manufacturing sector by on the one hand lowering price of intermediate inputs used in the production process while at the same time making the price of final output more competitive on world markets.
4.3
Estimation results – the probability to trade
The results regarding the extensive margin of SSA manufacturing trade in column 3, obtained by virtue of the use of a 2step estimation strategy, reveal several interesting differences with the results discussed in the previous subsection. I only discuss the most salient differences here. Variables that affect the probability of trade in very much the same way as they do the amount of trade are not explicitly discussed here. First the trade hampering effect of distance. As in case of the amount of trade, distance is found to negatively affect the probability to trade. However, in contrast with the amount of trade, the distance penalty on the probability to trade is significantly larger when considering intra-SSA trade. This is similar to the finding of Limao and Venables (2001), however, contrary to their results, this extra distance penalty only shows up when considering the extensive margin of SSA trade and is no longer present when considering its intensive margin. Second, were SSA countries found to trade significantly less with their contiguous Northern African neighbors, the probability that these SSA countries trade with their neighbors across the Sahara is not significantly lower. Also the substantial higher burden of being landlocked on the amount exported to other SSA countries, 20
See http://www.icafrica.org/fileadmin/documents/AR2006/ICA_Annual_Report_-_Volume_1__FINAL_March_2007.pdf. 21 For more info see: http://web.worldbank.org/WBSITE/EXTERNAL/COUNTRIES/AFRICAEXT/ EXTAFRREGTOPTRA/EXTAFRSUBSAHTRA/0,,menuPK:1513942~pagePK:64168427~piPK:6416 8435~theSitePK:1513930,00.html?
does not show itself when considering the probability to trade. Similarly, I do not find a different effect of being a landlocked importer on the probability to trade with another SSA country compared to the probability to trade with a country in the ROW. Arguably most interesting are the different effects of the (importer) infrastructure variable and that of having a regional or free trade agreement. First, and contrary to the significant negative effect found on the amount of trade, the quality of infrastructure does not significantly influence the probability a SSA country imports manufacturing goods from another country in the ROW. This probably reflects to some extent the fact that the infrastructure index used (see Appendix A) does not explicitly take the quality of a country’s sea-/airport(s) into account. These are arguably of much more importance when considering countries’ probability to trade with other countries in the ROW. The fact that the quality of infrastructure still has the expected negative effect on the probability to import goods from another country in SSA, the transport of which is more likely to involve land-based modes of transport that are explicitly captured by the infrastructure index, seems to corroborate this explanation. Having a regional or free trade agreement does not significantly influence the probability to trade. Combining this with the fact that it does affect the amount of trade (see the previous subsection), this seems to indicate that two countries are usually already trading bilaterally when forming a regional or free trade agreement. The establishment of a regional or free trade agreement does therefore not so much influence the probability to trade; it does however boost the amount traded between countries.
5.
CONCLUSIONS
Sub-Saharan Africa (SSA) is only a marginal player on the world’s export and imports markets. Moreover, and in contrast to other developing countries SSA trade is still largely dominated by trade in primary commodities. Manufacturing exports constitute about 30% of total SSA exports. Developing the manufacturing (export) sector is viewed as a vital ingredient for Africa’s future economic success (IMF, 2007; World Bank, 2007) as empirical evidence showing the importance of the manufacturing sector in economic growth is abundant; in fact Asia’s economic success is largely ascribed to the development of an exporting manufacturing sector (Radelet et al., 1997).
Given the alleged importance of developing an exporting manufacturing sector in stimulating economic growth, this papers looks for the determinants of SSA manufacturing export performance. Instead of comparing SSA’s trade performance to the rest of the world or to other developing countries, this paper focusses on SSA trade only; looking for the determinants, and the role of trade costs in particular, of the large differences in manufacturing trade performance between SSA countries, instead of trying to answer the question why SSA is (or is not) different from other developing countries or the ROW. Also I distinguish explicitly between intra-SSA trade and SSA trade with the ROW as it has been shown that different variables may constitute a larger burden to intra-SSA trade than to SSA trade with the ROW. Given the large amount of zero trade flows in my data set, I pay specific attention to the choice of estimation method, argueing in favor of using a two-step estimation procedure. Hereby I stress the need to model the probability to and the amount of trade separately, instead of assuming the same process underlying both the extensive and intensive margin of trade (as e.g. Tobit or Poisson regression do). Besides being preferred from an econometric point of view, the use of a 2step estimation method also reveals interesting differences in the way trade costs, and distance and the establishment of a regional or free trade agreement in particular, affect the intensive and extensive margin of SSA’s manufacturing trade. The results show that high trade costs severely hamper SSA’s potential to import and export manufacturing goods. Both the amount of trade and the probability to trade are negatively influenced when a SSA country faces higher trade costs. I also find interesting differences in the way trade costs affect intra-SSA and SSA trade with the ROW. Colonial ties in particular are still very important determinants of bilateral trade flows from SSA to the ROW. Furthermore, most SSA countries, in particular the island nations and countries with direct access to the sea, seem to be much more oriented towards the ROW than towards other SSA countries when it comes to manufacturing trade. Evidence of such an outward orientation of SSA countries also shows up in the finding that an increase in a SSA country’s GDP boosts its trade with the ROW more than its trade with other SSA countries. Finally, and interesting from a policy perspective, I find that investments in infrastructure and increased regional integration in the form of regional or free trade agreements, have a positive effect on the amount of manufacturing goods imported and exported by SSA countries, and on intra-SSA trade in particular. These findings show that the current focus on improving
SSA infrastructure (see e.g. the aim of the Sub-Saharan African Transport Policy Progam and The Infrastructure Consortium for Africa), or the efforts by the African Union to increase intra-SSA integration, will likely contribute to the development of a SSA manufacturing sector that is able to more effectively compete on world markets.
REFERENCES Amjadi, A. and A.J. Yeats, 1995. Have transport costs contributed to the relative decline of Sub-Saharan African exports? Some preliminary evidence. World Bank Policy Research Paper, no.1559. Ancharez, V., 2003. Determinants of trade policy reform in Sub-Saharan Africa. Journal of African Economies, 12, p.417-443. Anderson and van Wincoop, 2004.Trade Costs, Journal of Economics Literature, Vol.XLII, p.691-751. Bigsten, A., P. Collier, S. Dercon, M. Fafchamps, B. Gauthier, J.W. Gunning, A. Oduro, R. Oostendorp, C. Pattillo, M. Söderbom, F. Teal and A. Zeufack, 2004. Do African manufacturing firms learn from exporting? Journal of Development Studies, 40, p.115-141. Bigsten, A. and M. Söderbom, 2006. What have we learned from a decade of manufacturing enterprise surveys in Africa? The World Bank Research Observer, 21, p.241-265. Canning, D., 1998. A database of world infrastructure stocks, 1950-1995. World Bank Research Paper, Washington D.C. Clark, G.R.G., 2005. Beyond tariffs and quotas: why don’t African manufacturers export more? World Bank Policy Research Working Paper, no.3617. Coe, D.T. and A.W. Hoffmaister, 1999. North-South trade: is Africa unusual? Journal of African Economies, 8, p.229-256. Collier, P., 2002. Primary commodity dependence and Africa’s future. Keynote speech, World Bank’s Annual Bank Conference on Development Economics. Collier, P. and A.J. Venables, 2007. Trade preferences and manufacturing export response; lessons from theory and policy. Working paper, Oxford University. Elbadawi, I.A., 1999. Can Africa export manufactures? The role of endowment, exchange rates and transaction costs. World Bank Policy Research Working Paper, no. 2120. Foroutan, F. and L. Pritchett, 1993. Intra-Sub-Saharan African Trade: is it too Little? Journal of African Economies, 2, p.74-105. Francois, J., and M. Manchin, 2007. Institutions, infrastructure and trade. World Bank Policy Research Working Paper, no.4152. Greene, W.H., 2000. Econometric Analysis, 4th Edition. Prentice Hall, New Jersey.
Greene, W.H.,1994. Accounting for excess zeros and sample selection in Poisson and negative binomial regression models. Working paper No. EC-94-10., Stern School of Business, New York University, Dept of Economics. Hallack, J.C., 2006. Product quality and the direction of trade. Journal of International Economics, 68, p.238-265. Harris, J. and M. Todaro, 1970. Migration, unemployment & development. American Economic Review, 1, p.126-142. Hausman, J.A., 1978. Specification tests in Econometrics, Econometrica 46, p.12511271. Hausmann, R., L. Pritchett and D. Rodrik, 2005. Growth accelerations. Journal of Economic Growth, 10, p.303-329. Helpman, E., M. Melitz and Y. Rubinstein, 2007. Estimating trade flows: trading partners and trading volumes. NBER working paper No.12927. Haveman, J. and Hummels, D., 2004. Alternative hypotheses and the volume of trade: the gravity equation and the extent of specialization. Canadian Journal of Economics, 37, p.199-218. International Monetary Fund, 2007. Regional economic outlook: Sub-Saharan Africa, Washington, D.C. Karshenas, M., 2001. Agriculture and economic development in sub-Saharan Africa and Asia. Cambridge Journal of Economics, 25, p.315-342. Limao, N. and A.J. Venables, 2001. Infrastructure, geographical disadvantage, transport costs, and trade. The World Bank Economic Review, 15, p.451-479. Mullahy, J., 1997. Heterogeneity, excess zeros, and the structure of count data models. Journal of Applied Econometrics, Vol.12-3, p.337-350. Melitz, M., 2003. The impact of trade on intra-industry reallocations and aggregate industry productivity. Econometrica, 71, p.1695-1725. Mengistae, T. and C. Pattillo, 2004. Export orientation and productivity in SubSaharan Africa. IMF Staff Papers, 51, p.327-353. Ng, F and A.J. Yeats, 2000. On the recent trade performance of Sub-Saharan African countries: cause for hope or more of the same. Africa Region Working Paper Series, The World Bank, Washington D.C. Poirson, H., 2001. The impact of intersectoral labour reallocation on economic growth. Journal of African Economies, 10, p.37-63.
Radelet, S., J.D. Sachs and J.-W. Lee, 1997. Economic growth in Asia. Prepared as background for Asian Development Bank’s study: Emerging Asia: Changes and Challenges. HIID discussion paper No.609. Radelet, S. and J.D. Sachs, 1998. Shipping costs, manufactured exports and economic growth. Mimeo, Harvard Institute for International Development. Redding, S. and A.J. Venables, 2004.Economic Geography and International Inequality.Journal of International Economics, 62, p. 53-82. Sachs, J.D. and A.M. Warner, 2001. The curse of natural resources. European Economic Review, 45, p.827-838. Sachs, J.D. and A.M. Warner, 1995. revised 1997, 1999. Natural resource abundance and economic growth. NBER working paper No. 5398, Cambridge, MA. Sahn, D.E. and C. Stifel, 2003. Urban-rural inequality in living standards in Africa. Journal of African Economies, 12, p.564-597. Santos Silva, J.M.C. and S. Tenreyro, 2006. The log of gravity. The Review of Economics and Statistics, 88, p.641-658. Simmons, B., 2005. Rules over real estate: trade, territorial conflict, and international borders as institution. Journal of Conflict Resolution, 49, p.823-848. Söderbom, M. and F. Teal, 2003. Are manufacturing exports the key to economic success in Africa? Journal of African Economies, 12, p.1-29. Subramanian A. And N.T. Tamirisa, 2003. Is Africa integrated in the global economy? IMF Staff Papers, vol.50, no.3, Washington, D.C. Temple, J. and L. Wößmann, 2006. Dualism and cross-country growth regressions. Journal of Economic Growth, 11, p.187-228. Tiffen, M., 2003. Transition in Sub-Saharan Africa: agriculture, urbanization and income growth, World Development, 31, p.1343-1366. Van Biesebroeck, J., 2005. Exporting raises productivity in Sub-Saharan African manufacturing firms. Journal of International Economics, 67, p.373-391. Wooldridge, J.M., 2003. Introductory Econometrics – A Modern Approach. Thomson, USA. World Bank, 2007. Accelerating Development Outcomes in Africa - Progress and Change in the Africa Action Plan. Africa Region, The World Bank, Washington D.C.
APPENDIX A.
Variables included in the gravity equation
GDP Gross Domestic Product, from Penn World Tables 6.2, if not available (for Angola, Haiti, French Polynesia, New Caledonia, Azerbaijan, Armenia, Belarus in selected years) from World Bank Development Indicators 2003 or World Bank Africa Database 2006. Distance Great circle distance between main cities, from CEPII. Contiguity Dummy variable indicating if two countries share a common border, from CEPII. Common official language Dummy variable indicating if two countries share a common official language, from CEPII Common colonizer Dummy variable indicating if two countries have been colonized by the same colonizer, from CEPII. Colony – Colonizer relationship Dummy variable indicating if two countries have ever had a colony-colonizer relationship, from CEPII. Landlocked Dummy variable indicating if a country has no direct access to the sea. Island Dummy variable indicating if a country is an island. Infrastructure index Following Limao and Venables (2001), the index constructed as the unweighted average of four variables (each normalized to have a mean of 0 and standard deviation 1 over the whole sample period as well as in each year). The four components are: - Roads Km road per km2, from World Bank Development Indicators 2003, World Bank Africa Database 2006 and Canning (1998). - Paved roads Km paved road per km2, from World Bank Development Indicators 2003, World Bank Africa Database 2006 and Canning (1998). - Railways Km railways per km2, from Canning (1998). - Telephone main lines Telephone main lines per 1000 inhabitants, from World Bank Development Indicators 2003, World Bank Africa Database 2006 and Canning (1998). As Limao and Venables (2001) I ignore missing values, making the implicit assumption that the four variables are perfect substitutes to a transport services production function. Regional or Free trade agreement Dummy variable indicating if two countries are both a member of one of the following regional or free trade agreements: ECOWAS, ECCAS, COMESA, SADCC, UEMOA, CEMAC (or UDEAC), EAC, IGAD or CENSAD. Civil conflict Dummy variables indicating if a country experienced the use of armed force between two parties, of which at least one is the government of a state, that results in at least
25 and at most 999 battle-related deaths, from the International Peace Research Institute, Oslo. Civil war Dummy variables indicating if a country experienced the use of armed force between two parties, of which at least one is the government of a state, that results in at least 1000 battle-related deaths, from the International Peace Research Institute, Oslo. % rural population Share of the population living in rural areas, from World Bank Development Indicators 2003 and World Bank Africa Database 2006. % labor force in agriculture Average proportion of the total labor force recorded as working in agriculture, hunting, forestry, and fishing (ISIC major division 1) over the period 1993-2002. Labor force comprises all people who meet the International Labour Organization’s definition of the economically active population, from World Bank Development Indicators 2003 and World Bank Africa Database 2006.
APPENDIX B. Testing the appropriateness of using the same model for the probability to and the amount of trade
As discussed in section 2 above, in order to deal with the occurence of zero bilateral trade flows, one can either estimate the model using OLS and the non-zero observations only (OLS), using a censored regression model that assumes the same model for the probability to and the amount of trade if positive trade is observed zero (Tobit), or using a two-step estimation method that models these two separately (2step) 22. The difference in assumptions under which each of these methods provides consistent esimates of the parameters of interest can be shown by looking at the following framework that encompasses all three methods23:
yi = X i β + ui s = 1[ X iϕ + Z iγ + vi > 0]
(A1)
, where s = 1 if one observes a positive trade flow and s = 0 in case of a zero trade flow. y denotes a bilateral trade flow, X contains variables that determine both the
22
Here I focus explicitly on the use of Tobit estimation. The discussion does not directly apply to the non-linear estimation procedures proposed in the literature. However these non-linear methods, although being able to incorporate the zero observations, assume that standard gravity models also offer an explanation for these zero observation, which they in fact do not. 23 Here I leave aside the distributional assumptions made when employing each of these three methods, i.e. normality of the residuals in case of both Tobit and 2step estimation; referring to Wooldridge (2003) for a more detailed exposition. Also for ease of exposition I use a censoring value of zero, it is easy extend it to different censoring values.
conditional mean of y and sample selection and Z contains variables that determine selection only. The main point of interest is to obtain a consistent estimate of β. Under what assumptions do the three above-introduced methods, OLS, Tobit and 2step, produce such a consistent estimate? The assumptions needed in each of the three cases are the following: OLS:
E[u|X,s = 1] = 0
(A2)
Essentially what is needed for OLS on the non-zero observations only to deliver a consistent estimate of β, is to have exogenous sample selection. That is, selection depends on a function of X only, or, if it also depends on additional variables contained in Z or additional random terms summarized by v these need to be independent of u. Tobit:
b = ϕ, γ = 0, u = v, E[u|X] = 0
(A3)
The assumption that the error process in the selection and the conditional mean model are the same, immediately implies that OLS using the non-zero observations only would deliver an inconsistent estimate of β. However, note that the assumptions in (A3) are also in some sense more restrictive than those made when simply using OLS on the non-zero sample only. They imply that Tobit assumes the exact same model (same parameters, same variables and same error process) for selection as for the conditional mean. It does not allow a certain variable to have dissimilar effects on selection and the conditional mean (this is an assumption that I will use later to test for the appropriateness of using Tobit).
2step:
E[u|X,Z] = 0, E[v|X,Z] = 0
(A4)
Under these assumptions OLS on the non-zero sample yields an inconsistent estimate of b as it is not a priori assumed that the error term u in the conditional mean equation is uncorrelated with the error in the selection equation v, i.e. E[u|v] = 0 does not necessarily hold. If this were to hold, OLS on the non-zero sample would also yield consistent estimates of b. Compared to Tobit, 2step estimation allows for differences in both parameters and error process in the selection and conditional mean equation
and also for additional variables in the selection model, that are required to be uncorrelated with u however24. Given these assumptions, which is the appropriate method to use? The assumption made when using OLS and the non-zero observations only, seems to be somewhat restrictive a priori. The assumption of exogenous sample selection is quite demanding. For example, it can be easily imagined that a shock that negatively affects the probability to trade (e.g. a drastic drop in prices on world markets, or a exchange rate shock) also negatively affects the amount to trade25, which would immediately render b estimated using OLS on the non-zero observations only inconsistent. When using Tobit, the assumption is explicitly made that the exact same model can be used for selection and the conditional mean. Here I propose a way to test for this assumption by comparing the estimated parameters using Tobit to the estimated parameters obtained using Probit. Recall that Tobit assumes the following process: yi = X i β + ui s = 1[ X i β + ui > 0]
i .i .d
, where u ~ N ( 0, σ 2 )
(A5)
, where s = 1 if one observes a positive trade flow and s = 0 in case of a zero trade flow. Probit assumes the exact same process, but for the selection process only, i.e.: i .i .d
s = 1[ X iδ + ui > 0] , where u ~ N ( 0, σ 2 )
(A6)
, where s = 1 if one observes a positive trade flow and s = 0 in case of a zero trade flow. Furthermore, and important for the test I propose, under the above assumptions the estimated parameter using Tobit and the estimated parameter using Probit are related: δ = β/σ. When using Probit, estimating the selection equation only, one cannot seperately identify β and σ; when using Tobit instead, one can seperately identify β and σ by making the additional assumption that the same model explains selection and the conditional mean. To evaluate whether this assumption is appropriate, one can compare the Probit estimate of δˆ to the Tobit estimate of βˆ / σˆ .
The two should be ‘close’ when the Tobit model holds, if not Tobit is inappropriate suggesting that selection and conditional mean should be modelled allowing for 24
In principle one does not necessarilty have to include all the variables in X in the selection equation, however wrongly excluding a variable in X immediately leads to inconsistency, whereas wrongly including a variable in X comes at little cost. 25 When using the 2step method this assumptions can also be tested using the so-called inverse Mills ratio, see below.
different parameters on the X variables and possibly for additional variables, Z, in the selection model. Instead of informally evaluating the ‘closeness’ of the Probit and the Tobit estimate (see p.573 in Wooldridge, 2003), I propose to test whether the two estimates are close using a standard Hausman (1978) test. The logic behind using this test is the following: H0:
Tobit and Probit are both consistent. Tobit is efficient however as it allows the estimation of the variance of the error process by assuming the same process for selection as well as for the conditional mean.
H1:
Probit consistent but Tobit inconsistent. Tobit is inconsistent by wrongly assuming the same process for selection and the conditional mean.
The resulting Hausmann (1978) test statistic is:
(δˆ
probit
− ( βˆ / σˆ )tobit
'
) (V (δˆ)
probit
− V ( βˆ / σˆ )tobit
−1
) (δˆ
probit
)
− ( βˆ / σˆ )tobit ~ χ 2 (h)
(A7)
where V denotes the covariance matrix of the corresponding estimate and h is the number of variables in X. Returning to the subject of this paper, not rejecting H0 indicates that one can indeed make the assumption that one can use the same model for the probability to and the amount of trade and use Tobit to obtain a consistent estimate of β. When instead rejecting H0 this suggests it is more appropriate to separately model the probability to trade and the amount of trade. This is what is done when estimating β using the 2step estimation strategy. In the first step the selection equation is estimated using Probit. Using the estimated parameters of the selection equation, one can construct the so-called inverse Mills ratio:
λ ( X ϕˆ + Z γˆ ) = φ ( X ϕˆ + Z γˆ ) / Φ ( X ϕˆ + Z γˆ )
(A8)
, where φ and Ф denote the pdf and cdf of the normal distribution respectively. Next in the second step, one estimates the equation modelling the conditional mean including this inverse Mills ratio as an additional regressor. The inclusion of the inverse Mills ratio, controls for the biased resulting for endogenous sample selection and ensures that β is estimated consistently (see Wooldridge, 2003). The coefficient on the inverse Mills ratio also provides a test regarding the exogeneity of the sample selection process. If the coefficient on the inverse Mills ratio is not significant, exogenous sample selection cannot be rejected and OLS on the non-zero observations
only would also yield consistent estimates; else only using the 2step method yields consistent estimates of β.
APPENDIX C.
(Zero-inflated) Poisson results
The results of estimating (2) using either Poisson or Zero-Inflated Poisson (ZIP) are shown in Table C1 below. Note (see also p.6 and footnote 7) that the Poisson results are likely biased on the basis of the ‘excess-zeros’ problem faced when considering SSA trade data (see Table 1). Indeed the expected number of zero observations, given the Poisson estimates, is only 181726 compared to 39977 ‘true’ zero observations. Some of the results are indeed quite puzzling, e.g. the positive effect of civil conflict or of being landlocked; also the insignificant effect of language similarity and of sharing a common border contradict some of the (established) findings in the empirical trade literature. ZIP estimation does take this into account by allowing a different process ‘explaining’ these zero observations (now the number of predicted zeros, 4011327, is much closer to the true number of zeros); note also that the Vuong statistic (see Greene, 2000 p.891) indicates that ZIP is preferred over standard Poisson estimation. ZIP estimation however requires the questionable assumption that the probability to and the amount of trade are two independent processes: e.g. random shocks are not allowed to affect both and also there are no omitted variables (uncorrelated with the included variables) affecting both the probability to and the amount of trade. Note that, given this independence assumption, ZIP estimation amounts to the same results as using Poisson on the non-zero observations only (compare column 2 and 3).
26
Calculated as
27
Calculated as
∑
N P ( y = 0 | X ) = exp(− exp( X i βˆ )) , see Greene (1994). ∑ i i i =1 i =1 N
∑ [ϕ ( X ) + (1 − ϕ ( X )) P( y N
i =1
, where st
i
i
i
i
i
N = 0 | X i ) ] = ∑ i =1 ϕi ( X i ) + (1 − ϕi ( X i )) exp(− exp( X i βˆ ))
ϕi ( X i ) is the predicted probability of observing a zero observation given Xi obtained from the
1 stage probit (see Table 2, column 3).
Table C1. Poisson and Zero-Inflated Poisson dependent variable
trade
trade
trade
estimation method
Poisson
ZIP
Poisson nonzeros
time period
1993-2002 coefficients
p-values
1993-2002 coefficients
p-values
1993-2002 coefficients
p-values
ln distance
-0.641
0.000
-0.585
0.000
-0.585
0.000
ln distance ssa
0.222 ns
0.094
0.251 ns
0.083
0.251 ns
0.083
ln gdp importer
0.786
0.000
0.751
0.000
0.751
0.000
ln gdp importer ssa
-0.343
0.000
-0.389
0.000
-0.389
0.000
ln gdp exporter
0.983
0.000
0.932
0.000
0.932
0.000
-0.082 ns
0.216
-0.133 ns
0.079
-0.133 ns
0.079
colony - colonizer
1.152
0.000
1.126
0.000
1.126
0.000
common colonizer
0.610
0.000
0.524
0.005
0.524
0.005
ln gdp exporter ssa
common colonizer ssa
-0.253
ns
0.237
-0.374
ns
0.086
-0.374
ns
contiguity
0.244 ns
0.473
0.141 ns
0.086
0.678
0.141 ns
contiguity ssa
0.400 ns
0.281
0.494 ns
0.678
0.193
0.494 ns
common off language
-0.257
ns
0.193
0.536
-0.249
ns
0.546
-0.249
ns
common off language ssa
1.125 ns
0.546
0.072
1.199 ns
0.076
1.199 ns
landlocked exporter
0.076
0.363
0.000
0.411
0.000
0.411
0.000
landlocked exporter ssa
-0.505
0.001
-0.562
0.001
-0.562
0.001
landlocked importer
-0.990
0.046
-1.020
0.041
-1.020
0.041
landlocked importer ssa island exporter island exporter ssa
1.057
0.000
1.088
0.000
1.088
0.000
-0.180 ns
0.494
-0.165 ns
0.493
-0.165 ns
0.493
-2.595
0.000
-2.557
0.000
-2.557
0.000
island importer
-0.492
ns
island importer ssa
-0.595 ns 1.369
ln infrastructure exporter ln infrastructure exporter ssa ln infrastructure importer
0.114
ns
-0.514
0.144
-0.540 ns
0.002
1.334
0.623
0.571
0.000
ln infrastructure importer ssa
-0.045
ns
RTA or FTA
0.182 ns 1.916
RTA or FTA ssa
0.153
ns
civil conflict importer
1.919
ns
civil conflict exporter
-0.341 ns
civil war importer
0.079
ns
0.747
-0.061
0.482
0.132 ns
0.000
1.938
0.094
1.948
ns
0.060
-0.327 ns
-0.159 ns
0.403
civil war exporter
-0.906
% rural population importer
-1.212
% rural population exporter
-0.514
0.134
0.238
-0.540 ns
0.238
0.001
1.334
0.001
0.745
0.570 ns
0.134
ns
0.000
0.079
ns
0.570
0.745 0.000
0.652
-0.061
ns
0.652
0.591
0.132 ns
0.591
0.000
1.938
0.000
0.092
1.948
ns
0.092
0.074
-0.327 ns
0.074
-0.115 ns
0.534
-0.115 ns
0.534
0.000
-0.834
0.000
-0.834
0.000
0.015
-1.158
0.029
-1.158
0.029
-2.378
0.000
-2.331
0.000
-2.331
0.000
% in agriculture importer
-1.198 ns
0.160
-1.229 ns
0.160
-1.229 ns
0.160
% in agriculture exporter
0.400
0.014
0.428
0.007
0.428
0.007
nr observations
74401
74401
34524
(pseudo) R2
0.715
-
0.668
-
-
126.2 [0.000]
Vuong statistic
Notes: bold italic numbers denote differences in sign and/or significance (at the 5% level) between the (zero-inflated) Poisson and the 2step estimation results. ns means not significant at the 5% level. Zeroinflated poisson (ZIP) uses the same probit results as shown in Table 2 to explain the occurrence of zero trade flows. The Vuong statistic tests the null of standard Poisson against ZIP under the alternative.
