Project 9 Final Slides Aker Klein O Connell

Editor's Notes

• #2: Good afternoon, and thank you for the opportunity to present here at the NBER Africa Project. This is a joint project between Michael Klein, Steve O’Connell and myself, and one that has really I think brought together people from three different disciplines to work on a common project, and so I would like to thank Gates and the organizers for that. We are very much looking forward to your questions and feedback.The title of our presentation today is “Borders or Barriers? The Impact of Borders on Agricultural Markets in West Africa.”
• #3: As economists, we have often recognized the potential welfare-enhancing effects of international trade. There is more debate, however, about the ease with which such trade takes place across international borders. Since Engels and Rogers’ seminal work assessing the role of the international borders on price dispersion between the US and Canada, there has been a large body of literature testing for the presence of a border effect between countries. While a variety of these papers find that a border effect exists for specific products (of the same retailer, like Safeway) between high-income countries, there are wide estimates of the magnitude and economic significance of this effect. Consequently, border effects have been cited as one of the “great puzzles” in international economics.
• #4: Now understanding the border effect is more than a theoretical concern. Borders incur excessively (or maybe unnecessary) high transaction costs, which is a deadweight loss. This is of particular concern in sub-Saharan Africa, and especially the landlocked countries of West Africa, which are some of the poorest countries in the world. In addition such countries rely primarily upon trade with their neighbors to meet their food and income needs.In this paper, we look at the specific case of West Africa. The countries in this area – particularly Niger – is one of the poorest countries in the world, ranked last on the UN’s Human Development Index (HDI). Their economies are primarily based upon rainfed agriculture, and trade with its coastal neighbors (Nigeria and Benin) and Burkina Faso.
• #5: This is really the focus of this research. Our goal is to assess the extent to which borders impose “additional” costs – in effect, become barriers -- that segment markets within West Africa. To do this, we rely upon the arbitrary nature of geographic borders in Africa to measure their impact upon price dispersion for grains and cash crops, using both traditional approaches in the intenrational economics literature and a RD design. Next, we try to go beyond this estimate of the international border effect to identify alternative explanations and mechanisms that can mitigate or exacerbate the border effect, namely, ethnicity. As this paper is primarily empirical in nature, we rely upon three distinct and unique datasets. Three datasetsMarket-level time series monthly panel from 1999-2007 for Niger, northern Nigeria, Benin, Burkina Faso and MaliGIS data on market location, road distances between markets and distances to international and internal bordersTrader and farmer-level market data
• #6: Whatwefindisthatthereis an international border effect, but itseconomicsignificancedoes not detertradebetween the two countries. This effectisrelatively the same for grains and cowpeas, twodifferentcommodities.A commonethnicityacross the international border appears to reducethis border effect.Whilethiscanmitigate the border, thereappears to be an internal border effectalongethniclines. This does not appear to be due to infrastructure differences, or unwillingness to trade, but somefundamentaldifferences in the role of women in agricultural trade and the importance of your social (ethnic) network in providingtradecredit. 3
• #7: Before we begin, I would like to give you a bit of background on borders and trade in West Africa, primarily focusing on Niger-Nigeria sub-region.
• #8: The Niger/Nigeria border reached its current configuration (1,500 km) within 20 yrs of the Berlin Conference of 1884/85. Its placement reflected two simultaneous processes, which, while not random, were arbitrary with respect to African interests. For example, the French wanted access to Chad across southern Niger. This moved the border from something roughly corresponding to the desert/arable fault line (a geographical barrier perhaps) to a location within the arable area. The British insisted that the Hausa/Fulani kingdoms of northern Nigeria be kept intact. Now in the centuries before the colonial powers arrived, trade of course existed between these two countries. In particular, the Hausa in the states of Northern Nigeria were linked to the sub-region through a series of trading routes, which were based mainly upon ecological specialization. There were two major routes: The trans-Saharan trade, which primarily focused in trade in slaves, textiles, salt, grains and livestock. The second major trade route was westward, and involved trade in kola nuts between Niger, Nigeria and Ghana.
• #9: These traditional trade routes were extremely important for trade of different commodities, and many of those for grains and cowpeas still exist until today. Today, both grains, cash crops and livestock are bought and sold through an extensive system of traditional grain markets, each of which occurs on a weekly basis. The average distance between these markets is about 350 km, but distance between markets for which trade occurs ranges from 5 km to over 1000 km. These photos provide some examples of the grains trade in Niger at the weekly grain markets.
• #10: Now, West Africa has attempted to promote cross-border trade through a system of monetary and trade unions. There are two relevant ones in West Africa: 1) UEMOA, created in 1994, which covers eight Francophone countries in West Africa, and which shares a common currency – the CFA franc – and a common external tariff; and 2) ECOWAS.The actual CFA franc was created in 1945, with two devaluations in 1948 and 1994 (changing the relative value of the CFA to the French franc). UEMOA is also part of ECOWAS, created in 2001 among all West African states. In 2007, ECOWAS CET is an instrument for tariff setting and liberalisation in order to ensure common market access in the West African region.
• #11: Now, as I mentioned previously, there is a large body of research that has attempted to measure the impact of a border effect, especially in high-income countries. These can be divided into three approaches.
• #12: The first approach, used by Engel and Rogers, compares within country price dispersion to cross-country dispersion. The approach within these papers is to regress a measure of price dispersion on distance between two cities, a dummy variable for the border and city fixed effects. The idea behind this analysis is that a border is like a “treatment effect” and, but for this, price relationships would be the same in the two countries
• #13: Now, a major criticism of this approach, as noted by Gorodnichenko and Tesar (2005), is that these estimates of the border effect are unidentified unless the degree of within-country price dispersion is the same across countries. They attempt to “control” for this by including country heterogeneity in the regressions.
• #14: Now, as I mentioned previously, there is a large body of research that has attempted to measure the impact of a border effect, especially in high-income countries. These can be divided into three approaches.
• #15: The concerns raised byGorodnichenko and Tesar inform our empirical strategy. During the first part of the strategy, we use a regression-based approach using market pairs to assess the impact of international borders on price dispersion. This approach examines whether the border can increase price dispersion (extra transaction costs) when there is trade between markets.However, recognizing the endogeneity concerns inherent in this approach, we use the RD design – as proposed by Gopinath et al – to measure the border effect, effectively limiting our sample to markets located close to the border. This of course has some of its own issues, but we can discuss those later.
• #16: To measure the border effect, we rely upon three primary datasets.
• #20: To giveyou an idea of whatthese data look like for the two countries, here are somebriefsummarystatistics. I willdiscussthese in more detaillater, but I wantyou to mainly focus on the pricedifferences for the twocommodities – millet isrelatively more expensive in Niger as comparedwithNorthern Nigeria, and cowpeaisrelativelycheaper, although the differenceis not statisticallysignificant.
• #21: Now, as a first brush, we estimate the border effect by using the approach in much of the literature, which is based upon market pairs. This involves the log difference in prices as a function of the distance (or transport costs) between markets and border.ln(pij,t / pik,t) = βDdj,k+ βDBj,k+Σ βzDz+εij,k,tDistance between cities j and k: dj,kBorder between cities j and k?: Bj,kdummy variableMarket dummy variables: DzOur primary coefficient of interest is b2, the coefficient on the border effect. In this case, we do not use all of the market pairs, but limit the market pairs that are within a certain distance of each other (ie, 200 km).
• #22: Using this approach reveals a positive and statistically significant border effect between the two countries. Using the basic specification, the international border increases price dispersion by 2.1 percent, which increases slightly once other covariates are included. Mindful of the criticism by Gorodnichnenko and Tesar, we include a dummy variable for country-specific market pairs, first for Niger, then for Nigeria. In doing so, the coefficient estimates change slightly, but not of the magnitude seen in other papers. And finally, when including market-specific fixed effects, the border effect increases price dispersion by 1.8 percent. The last column includes an additional term, namely, whether the markets in the pair share a common ethnicity or not (1 = different, 0 – same). What we see here is that the international border effect is relatively larger when trade is between markets of a different ethnicity. This will be important later. But in general, the magnitude of the border effect is small -- the km-equivalent border effect is 5 km, which is in contrast to the km-equivalent in the other border effects literature. The last column includes The interaction term reveals the interaction of two barriers: 1) the presence of a border; and 2) different ethnicities between two markets.The first joint effect shows that, in the presence of different ethnicity, the international border effect is higher.The second joint effect shows that, in the presence of an international border, price dispersion is higher between two markets of different ethnicity. (These two mean the same?)Km-equivalent border effect: exp(b1/b2), where b1 is coefficient on ln(distance) and b2 is coefficient on border.The coefficients are b1=.033 and b2=.024
• #23: Nowobviously, there are numerousthreats to our identification of this border effect.
• #24: The first of these is the criticism included by Gorodnichenko and Tesar, which point out that the border effect will be underidentified if the underlying price volatilities. We checked for this by including the country-specific market pair dummies. Another way to do this is to plot the residuals from a regression of the absolute value of the differences in the log of prices on a variety of covariates, assuming those residuals show deviations from the LoP.
• #25: In doing so, we see a slight difference in the distributions, but some difference, providing some support of the previous regression results. Similarity of distributions for within country market pairs suggests that different underlying price distributions is not a concern
• #26: The second major threat to our identification – althoughperhaps an alternative explanationratherthan a threat – isthat the markets are segmented (thereis no tradebetweenthem), sowewould not expectthere to be LOP. In this case, the border effect
• #27: Lookingat the trader-level data for border markets in Niger and nothern Nigeria, wefindthat traders in bothmarketstradewith cross-border marketswithin a 50-km radius – although a slightlyhigherpercentage of Nigerien traders – and that the magnitude of thistradeis about the same for traders in bothmarkets.
• #29: Now, a large criticism of this type of estimate of the border effect is the potential for bias in the results, since borders (albeit arbitrary) are not necessarily randomly assigned. One approach that has been used in the literature recently – byGopinath et al and by Ryan Bubb at Harvard – is the regression discontinuity approach. The idea here is that The intuition in applying the RD in this case is the following: The border represents a discontinuity in “treatment”, which is determined by your assignment to either side of the border. Since the level of treatment jumps discontinuously at the border, any discontinuous change in the outcome variable at the border could be attributed to the change in treatment status. Usually, people criticize RD because it only helps identify some treatment effect at the border, which limits its external validity. In our setting, the effect at the border is exactly what we need to know: why should there be a discontinuous change in price at the border
• #30: In a regression framework, this essentially means regressing the log of prices in market j at time t on a variable for Niger, the distance of the market to the border, an interaction term and other potential market-specific variables. The variable of interest here is γi, which measures the size of the border effect. This reveals a counterfactual: the price remains higher to one side of the border even if the distance shrinks to zero. As is the case in the RD literature (Imbens and Lemieux), the regression is estimated suing a local linear regression framework, whereby the regression is only estimated for prices in those markets that are within a certain distance from the border. The border effect therefore answers the question: How do prices change when crossing from a positive to a negative distance from the border, where this distance is some small number (epsilon)?*NOTES*:RD can be either sharp (level of treatment, N, goes from 0 to 1 at Z0) or fuzzy (where the level of treatment at Z0 increases discontinuously, but not from zero to one.The treatment effect is either the change in y at D0 in the sharp design, or the change in y over the change in treatment at Z0, which is the local Wald estimate of the causal impact. The log function log (x) is only defined for x>0, and it is <0 when x is inbetween 0 and 1. In this case, then, the change in the log of the price is equal to gamma times the change in N (going from treatment or not). This means as we go from Niger to Nigeria, the change in prices increases by x% (also known as a semi-elasticity).
• #31: Is thisapproachvalid? This is an oldmapthatprovides a bit of the intiution for the validity of thisapproach. Herewesee a close-up of the international border from the time of the international border, with villages on eitherside. In many cases, thesemarkets are «arbitrarily close » to the border.
• #32: Now, one of the primary assumptions in using the RD approach is whether the discontinuity in the outcome is related to discontinuities in other observable and unobservable characteristics. While we can’t directly test for the discontinuity in unobservables, we can test for the discontinuity in observable characteristics. #4 can be tested by comparing the means in small bins of Z to the left and right of the threshold, or as a regression with various powers of Z. But, since the goal is to compute the effect at Z0, most often use the local linear regression. The biggest issue is how to choose the bandwidth or kernel. The distance is uniquely identifying the villages and/or the markets in both slides.
• #35: Whenusing the regressionresults, theseconfirm the visual inspection. Crossing the border leads to a 25 percent jump in prices, and thisjumpisstatisticallysignifcantacross all specifications. Whenwe change the bandwidth to 50 km of the border, the effectstillremains, although the magnitude isrelativelysmaller.What is the reason behind this result? Have Niger-Nigeria done a good job of promoting free trade, or have they done a « bad » job at patrolling the border? Or are there other factors that could be explaining the magnitude of the border effect? Well, similar to the market pair results, we add in a variable for ethnicity. Once we include this we find belonging to the Hausa group - -which is common across the borders – significantly reduces the magnitude of this price jump, and the border effect falls from 25 percent to about 10 percent. The similar finding exists for cowpeas.
• #36: Whenusing the regressionresults, theseconfirm the visual inspection. Crossing the border leads to a 25 percent jump in prices, and thisjumpisstatisticallysignifcantacross all specifications. Whenwe change the bandwidth to 50 km of the border, the effectstillremains, although the magnitude isrelativelysmaller.Whatis the reasonbehindthisresult? Have Niger-Nigeria done a good job of promoting free trade, or have theydone a « bad » job atpatrolling the border? Or are thereotherfactorsthatcouldbeexplaining the magnitude of the border effect? Well, similar to the market pair results, weadd in a variable for ethnicity. Once weincludethiswefindbelonging to the Hausa group - -whichiscommonacross the borders – significantlyreduces the magnitude of thispricejump, and the border effectfallsfrom 25 percent to about 10 percent. The similarfindingexists for cowpeas.
• #38: The question is, of course, defining an ethnic border. Within Niger alone, there are 8 major ethnic groups, the major ones are the Zarma, Hausa, Kanuri, Tuareg, Puehl. The Hausa are the dominant ethnic group, representing 55% of the population, followed by the Zaram/Songhai (21%), Toureg (9.3%), Peuhl (8.5%), Kanuri (5%) and other (1.2%). They are strongly spatially differentiated throughout the country, with the Toureg in the north, with the Kanuri to the far southeast, and the Hausa to the southeast, and the Zarma/Songhai to the southwest. The Peuhl are somewhat interspersed with the Zarma in the Western region of the country, along the Burkina-Niger border. This map shows the general locations of different ethnic groups. .
• #39: Now, defining an ethnic border istrickier, but not impossible. This isactuallylessarbitrarythanyoumightthink – when I was in Niger in March, I was right along the area of the Hausa-Zarmaregion. Weleftfrom one village – SabonYayi – whereby the predominant composition was Hausa. Wethentraveledapproximately 5-10 km on a dirt road to another village – Zozo TomboKoida – whichwaspredominantlyZarma. So being able to identify a border, whilesubject to error, is not impossible.
• #40: The idea here is to use the similar type of RD design to determine the internal border effect, again by restricting our observations close to the actual Zarma/Hausa border.
• #42: The issue, of course, istwofold.
• #43: Once we focus on thosemarketswithin 50 km of the border, the effectissimilar in magnitude of the international border effect but slightlylarger, and robust to the inclusion of othercovariates. Once weexpandthis to 100 km, the effectisrelativelysmaller and lesssignificant, whichwewouldexpect. But westill have the « distance to border » issue.
• #44: Whatcouldbeexplainingtheseresults? Coulditbedifferent transport costsbetween the markets? This seemsunlikely, sincetheywere the same. Coulditbedifferentinvestments due by the colonial powers? Again, given the similarity in the market-levelcharacteristics, thisseemsunlikely. Coul
• #45: Whatcouldbeexplainingtheseresults? Coulditbedifferent transport costsbetween the markets? This seemsunlikely, sincetheywere the same. Coulditbedifferentinvestments due by the colonial powers? Again, given the similarity in the market-levelcharacteristics, thisseemsunlikely. Coul
• #53: Now, before we discuss the regression-based results, let’s look at some kernel distributions to see the intuition for where we’re going.In these distributions, we regress the absolute value of the log of price ratios in two different markets for product i at time t on a variety of factors we think might explain percentage changes in price dispersion.We then capture and plot the eit to show the measure of price dispersion. te |ln(pit/pjt)| = b0 + b1 ln(total transport cost)(i,j,t) + b2 (1 if i urban, j not urban, or conversely) + b3 (1 if i drought, j not drought, or conversely)(t) + e(i,j,t)We plot the εijt as a measure of deviation from the Law of One Price.
• #54: We see the same thing for cowpeas – similar underlying price distributions, so this would appear not to be driving the results.
• #56: RDD ResultsCowpea