Development at the Border: Policies and National Integration in Cote D'Ivoire and its Neighbors

Regression discontinuity designs applied to a set of household surveys from the 1980-90s allow to examine whether Cote d'Ivoire's aggregate wealth was translated at the borders of neighboring countries. At the border of Ghana and at the end of the 1980s, large discontinuities are detected for consumption, child stunting, and access to electricity and safe water. Border discontinuities in consumption can be explained by differences in cash crop policies (cocoa and coffee). When these policies converged in the 1990s, the only differences that persisted were those in rural facilities. In the North, cash crop (cotton) income again made a difference for consumption and nutrition (the case of Mali). On the one hand, large differences in welfare can hold at the borders dividing African countries despite their assumed porosity. On the other hand, border discontinuities seem to reflect the impact of reversible public policies rather than intangible institutional traits.

Côte d'Ivoire is an interesting case study; at least in the 1980s and 1990s, it was much wealthier than all of its neighbors. The evidence provided by border discontinuity estimates is more mixed.
We document the hazards that commanded the alignment of boundaries during the colonial era and show that predetermined geographical and historical conditions should not account for border discontinuities in welfare.
At the eastern border with Ghana and at the end of the 1980s, large border discontinuities existed for the four outcomes. However, because the 1990s brought crisis to Côte d'Ivoire and recovery in Ghana, border differences in income were very much attenuated, and the discontinuity in nutrition vanished. Discontinuities in access to electricity and water were preserved. In contrast, at the northern border with poorer and landlocked Mali and in the mid-1990s, Côte d'Ivoire performed better in terms of income and nutrition, but not in access to utilities. The border with Guinea provides a case in which the Côte d'Ivoire advantage is canceled out along all dimensions. A more detailed analysis shows that income derived from cash crops almost fully accounts for the large differences in consumption at the borders of Ghana and Mali. Because the cocoa frontier had not yet reached the extreme west in the 1990s, the same factor explains why no discontinuity is found at the border with Guinea.
Border discontinuities reveal the role of two types of national policies: policies affecting cash crop production and public investment in utilities. Although differences in such policies have had large and visible impacts at borders, they are not irreversible, and some of them were changed in the years that followed. We conclude that large border discontinuities can be observed between African countries in the short run, but they do not necessarily reflect divergent trajectories linked to long-lasting institutional features.
Section I presents the analytical methodology and explains the econometrics.
Section II documents the historical and geographical backgrounds of borders. Section III presents survey data and border discontinuities in development outcomes, first for the eastern border with Ghana and then for the northern borders. Section IV provides further discussion of the cash crop channel and national policies.
Section V concludes.

I. Analytical methodology
Here, we discuss the conditions under which the borders we study can be considered historical "natural experiments". Consider a person born somewhere in the area now named Côte d'Ivoire. What would her welfare be if Côte d'Ivoire had been colonized by the British instead of the French, like Ghana, and then exposed Let Y be some outcome variable (income, connection to electricity, etc.) observed over a sample of people living in two countries at the same date. Let C = 0, 1 be the dummy variable indicating the country of residence. Let Y i (0) be the outcome if and when the individual (or household) i lives in the country C = 0 and Y i (1) in the country C = 1. The observed outcome reads thus: (1) The identification of the average treatment effect, E[Y (1) − Y (0)], is probably out of reach, but a regression discontinuity (RD) design based on distance to the border should correctly approach its local version (LATE) in the vicinity of the border.

Required assumptions for a border RD
Let D i stand for the distance to the border of the locality of residence, positively signed for country 1 and negatively signed for country 0, so that C i = 1{D i ≥ 0}.
Under the assumption that E[Y (0)|D = d] and E[Y (1)|D = d] are continuous provides an estimation of the average treatment effect at the border (Hahn, Todd, Van Der Klauw 2001). It is the so-called "sharp" RD estimator. As Lee (2008) argues in another context, this continuity assumption is difficult to assess and impossible to test. Lee's reformulation elucidates the conditions under which an RD replicates a random assignment around the threshold D = 0. Assume Y is generated by a partially unobservable random variable W : Y (0) = y 0 (W ) and Y (1) = y 1 (W ). W represents the "type" of individuals, households, or localities with respect to Y . Finally, let F (d|w) stand for the cdf of D conditional on W . Lee's conditions are as follows (Lee 2008, 679): Z is the predictable component of D and e is an exogenous random chance component, so that the probability of receiving treatment is somewhere between 0 and 1 for each type. Condition (ii) of unconfoundedness implies that conditional density f (d|w) is continuous in d at d = 0. In our case, this implies that within each type w and very near to the border, the probability of being allocated to one side or another is the same. However, in the overall population, D can be arbitrarily correlated with Y (0) or Y (1); Y may also be directly generated by D in addition to W (Lee 2008, 680). Under these conditions, the RD estimate is a weighted average of the difference y 1 (w) − y 0 (w) for each type w, with weights equal to the probability of being close to the border: f (0|w)/f (0).
At the locality level, Lee's conditions require that border localities are not sorted by "types" w between the two countries. This randomness of distance to the border should stem from the historical hazards of boundary alignment during the colonial period, which we document below.
At the individual level, the same conditions require that people do not "manipulate" their distance to the border through migration. This is typically an issue for embodied outcomes such as human capital. International migration based on y 1 (w) − y 0 (w), or more generally on w, is obviously the worst case. However, even 8 internal migration flows based on w are a source of bias because the center of one country may be more attractive than the other for a given type w (for instance, a larger number of good schools or good jobs in Abidjan than in Accra).
Because we do not observe types w, we cannot test directly for the validity of these assumptions, namely that the distribution of the "types" w is the same on both sides near the border.

Implementation of border RD estimates
To implement the border RD estimator just described, the regression functions for −h ≤ D ≤ h, and with h = 50, 75. We call this estimator "border RD".
Second and alternatively, we disregard distance to the border D and estimate a polynomial of degree three in latitude and longitude, like Dell (2010): for −h ≤ D ≤ h, again h = 50, 75, and a(h) ∈ R 9 . In this case, we assume that this cubic polynomial in latitude and longitude adequately describes the space of "types" w on both sides of the border. 3 We call this estimator "polynomial RD".
2 To correct for differences in sampling rates between countries, PSUs' sample weights are re-scaled by countries' total population. Of course, those "population weights" are still treated as probabilistic weights for statistical inference.
3 P (a, LAT, LON ) = a 10 LAT +a 01 LON +a 20 LAT 2 +a 02 LON 2 +a 11 LAT.LON +a 30 LAT 3 + Third and last, we implement a matching estimator on geographical distance that includes controls for distance to the border (i.e., combines matching and RD features). We call this "matching RD". This is inspired by Gibbons, Machin, and Silva (2009). We match each PSU j with its nearest neighbor ν(j) on the other side of the border. We sign the differences in outcome between matches ) so that positive differences designate a Côte d'Ivoire (C = 1) advantage. We then regress the signed difference on the distance to the border of the matched PSU, again with locally linear regressions on each side: for −h ≤ D ≤ h and h = 50, 75. Because different localities j in the same country can share the same nearest neighbor ν, we cluster the standard error η by ν. 4 In the case of the Mali border, where sample sizes do not allow the implementation of the narrowest 50 km bandwidth, we instead produce 100 km bandwidth estimates, where, for "border RD" and "matching RD", we add the square of the distance to the border (D 2 , interacted with the C dummy as well).
When examining border discontinuities in development outcomes, we add to all specifications a few geographical controls: we use latitude in the case of the Ghana and Guinea borders and longitude in the case of Mali as well as rainfalls, elevation, and distance to the nearest river. In specification (4), we use the difference between matched neighbors for all of these variables (again signed properly).
We also attempted a richer model that included the distance to the border of the matches. The estimates were not significantly different from those of the simpler model, although they were sometimes more imprecise. For the 75 km bandwidth sample of PSUs, the average distance to the border of the nearest neighbor matches varies between 10 and 13 km.
In the remainder of this paper, we use the acronym "BD" for border discontinuity.

II. Historical and geographical background
We first document the historical alignment of the boundaries around Côte d'Ivoire, drawing from the literature as well as from dedicated research in French colonial archives. This approach allows us to address the overlap assumption (random chance component of borders, see above). 5 Then, we document the geographical features of the boundaries. A few statistical tests assess the unconfoundedness assumption.

History
The drawing of boundaries in West Africa was arbitrary, to a large extent (e.g., Hargreaves 1985), and very often divided pre-colonial political entities. Even structured kingdoms drew no maps, and they could be composed of groups that spoke different languages, such as the Gyaman kingdom across the border from Ghana (Terray 1982). Ethnic groups are historical objects that were at least influenced, if not constructed, by pre-colonial, colonial, and post-colonial politics (e.g., Amselle andM'Bokolo 1985, Posner 2005). Furthermore, the classification of ethnic names is not independent from the national political economy. Despite these caveats, we verified in available mappings (Murdock 1959; language maps from Ethnologue: Lewis 2009) that the international boundaries we consider are not confounded by hard delimitations between ethno-linguistic areas.
During the 19th century, the largest part of the border between Côte d'Ivoire and Ghana was under the domination of the Ashanti Empire, whose capital city, Kumasi, was located in central present-day Ghana. In localities that lie no further than 75 km from this border and in the years 1986-8, surveys indicate that more than 50% of household heads belonged to the Akan ethno-linguistic grouping, which includes the Ashanti people: 56% on the Côte d'Ivoire side and 59% on the Ghana side. At the end of the 19th century, the French and British began to extend their domination from trade posts located on the coast toward the North by signing protectorate treaties with local kingdoms. Negotiations between the two colonial powers finally resulted in partitions of pre-colonial political entities in the middle part of the border (Gyaman, Indenie, Sefwi). In its southern part (Sanwi), a rebellion unsuccessfully challenged the border alignment after independence. The layout of the last demarcation on the field, with teak trees, beacons, and pillars, was achieved in 1988.
The two other borders of Côte d'Ivoire that we examine are less clearly demarcated on the field (Brownlie 1979). In surveys, the great majority of household heads belong to the Mande-Voltaic ethno-linguistic grouping. In particular, the boundary between Côte d'Ivoire and Mali lies across the Senoufo (Gur/Voltaic) area in its eastern part and the Malinke (Mande) area in its western part. The hazards of French conquest and of the wars against the Almami Samori Toure at the end of the 19th century reflected the boundaries' alignment within the French Empire and resulted in partitions of some former political entities (Kenedougou and Kong kingdoms). These borders were only stabilized after World War I.
We also use French data for the colonial period (Huillery 2009) and explore differences in initial conditions between the border areas lying inside the French 13 empire (Mali and Guinea). Pairwise comparisons do not reveal significant differences in terms of European settlement, tax revenue, or public expenditures (see table S1.1 in the supplemental appendix). One exception is perhaps that the Mali border districts exhibited higher population density. As we shall see in the following subsection, this feature has been reversed since then.

Geography
First, we check the continuity of the density of the distance to the border. The usual RD tests for sorting (Lee 2008, McCrary 2008) must be adapted to our context. We cannot only count the number of primary sample units because the sample stratification and sample rates differ between country surveys. For each bandwidth and each border size, we compute standardized "relative PSU weights" by dividing the original weights by their mean. The test then detects whether relatively more people are found closer to the border on one side compared to the other. The first row in Table 1 shows the result of this test in the case of the Ghana border. No border discontinuity is detected; one minor exception is the polynomial method at a bandwidth of 50 km. The same is found for Mali and Guinea at 75 km distance (see appendix table A.1), which will be our preferred bandwidth everywhere. We then characterize the PSUs of household surveys by five geographical variables: rainfall, elevation, distance to the nearest river, population density, distance to the closest regional city, and distance to the capital city (for more details on variable sources and construction, see supplemental appendix S2). The geographical locations of cities with 5,000 inhabitants or more in the years 1960 and 1990 are drawn from the Africapolis database. 6 Even natural geography should not be too quickly considered to be predetermined before the drawing of the boundaries.
Deforestation can influence rainfall, and even elevation or watercourses can, to some extent, be reshaped by human activity. Further, geographical discontinuities are measured on a sample of localities whose settlement is not random. However, we expect to find no or few discontinuities in rainfalls, altitude, and hydrography, and we regard this result as corroborating the quasi-randomness of the boundaries' alignment. Conversely, constructed geography, such as population density or city distribution, is far from being independent from boundaries. Nevertheless, we want to explore the extent to which discontinuities in welfare should be linked to discontinuities in urban structures. Finally, due to differences in countries' shape and spatial organization, distance to the capital or main city does not vary smoothly at borders. Any border effect includes a change in the capital city and a shift in the distance to it.
The major part of the border of Côte d'Ivoire with Ghana does not follow a natural line, except the lagoon in the extreme South and the Black Volta river in the extreme North. Given our sample distribution, these two parts contribute little to our estimates, and withdrawing them changes nothing. The border with Ghana does not exhibit any discontinuity in our geographical variables, except distance to the capital city: the Ghana border is closer to Abidjan than it is to Accra by 130 to 150 km (table 1).
The same is found for Mali in the 75 km bandwidth, except that here the border is farther away from Abidjan than from Bamako by approximately 300 km (top panel of table A.1). On both sides, as seen in figure 1, very few PSUs are found on the western part, which is underpopulated, probably due to the prevalence of parasitic diseases. Again, withdrawing the two or three most western PSUs (above longitude 7 • W) is innocuous. When enlarging the bandwidth to 100 km, additional Ivorian PSUs are found to lie relatively farther away from the border in the South. Their inclusion produces a significant "matching RD" discontinuity in rainfall. This leads us to prefer the narrower 75 km bandwidth.  considering the distance to locations that were already cities in 1960, especially at the Mali border. We link this post-independence urbanization process to the rapid economic growth of Côte d'Ivoire and, more locally, to the expansion of either cocoa (Ghana) or cotton (Mali) production. Lower population density on the Mali side could also be linked to the persistence of "river blindness" (onchocerciasis), which was previously fought and eradicated in Côte d'Ivoire.
Finally, in the case of Guinea, half of the alignment is based on rivers; "parts of watercourses, from map evidence, are tortuous, indecisive and many-armed" (Brownlie 1979, 374). The southern part of this border is also rather mountainous.
A few discontinuities are found for elevation and distance to rivers that could put the BD estimates in question. We give it less weight in our comments and conclusions. However, no discontinuity in welfare will be identified at this border.

III. Data and main results
After a short presentation of the survey data, we analyze border discontinuities in development outcome variables.

Survey data on development outcomes
We gather a database composed of 15 multi-topic household surveys: seven for Côte d'Ivoire, six for Ghana, one for Mali, and one for Guinea. These surveys were implemented between 1986 and 1998 (table A.2). We mainly use "income surveys", which correspond to the frame of Living Standard Measurement Surveys (LSMS), as designed by the World Bank in the 1980s. These surveys allow the measurement of consumption and income sources, and we code the geographical location of PSUs with precision using the locality name. We complement these income surveys with Demographic and Health Surveys (DHS), which do not measure consumption but record anthropometric data as well as housing conditions. All sample designs are two-stage and regionally stratified, and each PSU contains between 12 and 25 households.
We choose to analyze four different welfare outcomes: consumption per capita, children's height-for-age, access to electricity, and access to a man-made source of water. Despite the multi-topical nature of these surveys, very few other welfare outcomes are usable for comparison.
First, we construct a household expenditure variable that includes all current expenditures, such as food, clothing, transportation, housing, and imputed rents, and expenditures for education. We only exclude overly infrequent or badly measured expenditures such as those in health, durable goods, and transfers. Except for Mali, we also compute the value of the consumption of own food production, which we add to household expenditures to obtain a total consumption variable.
We use monthly data on the national consumer price index and express individ- and uniform producer price (Mali border; the Malian survey only records physical cotton output). We also extract data on wage incomes earned by household members.
Second, height stature is available for children from six months to four years (59 months) of age, and only for 6-35 months in some DHS. We construct heightfor-age Z-scores using the World Health Organization standards (WHO 2006) and code children as stunted when the Z-score is below -2.
Third, from both types of surveys, we construct a dummy variable indicating whether the household uses electricity as the main source of light in the house.
The surveys do not distinguish between being connected to a network or using a private generator.
Fourth and last, a second dummy variable codes whether the household has access to a man-made source of water (i.e., any source of water other than rivers, lakes, pools, or rainfalls) (Ghana border only). Column (1) reports the difference between survey means at the national level, and columns (2a) to (4b), respectively, report the BD obtained when implementing our three estimation methods with two bandwidths, 50 or 75 km, as in table 1. In 1986-8, we are able to restrict the estimates to the sub-sample of heads born in border administrative districts and in the country of residence. This allows us to test whether discontinuities hold among people sharing the same region of origin or "ethnicity" because part of the above-mentioned literature is concerned with this dimension, especially the fact that shared preferences or intense trade flows may equalize welfare among ethnic groups. We favor the district of birth variable  -12*** -38*** -25*** -44*** -35*** -37*** -29*** (1) Although we discuss the channels driving these results in more detail in section IV, we need to assess the role of differences in price levels because discontinuities in "nominal" consumption may not match discontinuities in purchasing power.

The border with Ghana
According to the World Bank (2012) 9 In a completely different context, Gopinath, Gourinchas, Hsieh, and Li (2011) find retail and wholesale markets to be segmented at the border between Canada and US. However, relative prices co-move with the nominal exchange rate, so there is little room for changes in the real exchange rate. estimates lie between 16 and 32 pp. In 1993, which is when income differences at the border must have been reduced to a minimum (see table 2 and above), BDs in early-age stunting are no longer observed. For Côte d'Ivoire, Cogneau and Jedwab (2012) found that the 1990 drastic cut in cocoa producer prices had a large impact on the height of two-to five-year-olds, less so on the youngest.
Finally, BDs for access to electricity seem large in magnitude for most estimates; they are always above 29 percentage points in 1986-8 but suffer from a lack of precision when the smallest 50 km bandwidth is used. BDs for access to safe water are positive but even less precise. With larger sample sizes on the Côte d'Ivoire side in 1992-3, both are found to lie in the same range and display higher statistical significance. 10 The results are not shown for the year 1998; despite a small sample on the Ivorian side (11 PSUs only), the results seem to confirm the persistence of the relative superiority of Côte d'Ivoire over Ghana in terms of utilities. 11

The northern borders
We now turn to the northern borders of Côte d'Ivoire with Mali and, secondarily,

Guinea. 12
Although 1993 was a very bad year for Côte d'Ivoire, the macroeconomic figures in table 2 show that this country was still far above Mali in terms of household 10 Using DHS data only and for the 75 km bw, the results are similar, although again less precise due to the smaller sample size (for electricity connection: +44**, +40*, +35; for water safety: +27*, +44***, +32**).
12 In previous versions of this work, we also looked at the border with Burkina Faso. However identification had drawbacks due to sample size and distribution. Borders between northern neighbors were also studied. See supplemental appendix section S3. consumption per capita, by around +300 USD at 1993 exchange rates and prices.
This national account figure is consistent with the survey means comparison for cash expenditures (+221 USD difference), although with some caveats already mentioned. When focusing on administrative districts along the border, the advantage of the Côte d'Ivoire side is halved but remains significant at approximately +110 USD. Even if cotton output per capita was fairly comparable at the national level, the Malian border districts produced very little cotton in 1993, providing a +43 USD per capita advantage to Ivorian districts.
Particularly because of its bauxite resources, Guinea is wealthier than Mali, and the difference from Côte d'Ivoire only reaches +200 USD in consumption per capita (not shown). Furthermore, although mines are not located in the border areas, survey mean comparisons among border districts show that the Guinean side is no poorer than the Ivorian side (not shown).
At the Mali border, discontinuities in log cash expenditures per capita are even higher than those found at the border with Ghana, ranging from +0.84 to +2.02 depending on the estimates, the bandwidth used, and the population considered (table 4 top panel). Recall, however, that the narrower 75 km bandwidth can be deemed more reliable for identification (see above). With this bandwidth, the upper bound is decreased to +1.52. The restriction to household heads who are from the Mande-Voltaic group again slightly reduces the estimates by dropping a few Ivorian southerners, such as civil servants sent to the North. 13 When we ad-13 The Mande include, in particular, the Bambara, Bobo, Diula, Malinke, and Soussou, whereas the Voltaic or Gur include the Lobi, Mossi, and Senoufo people. The Mande and Voltaic groups are close together in linguistic terms and display some mixing on the map. Ethnic codifications are not homogeneous; in particular, the Malian survey records the language of interview rather than the "ethnic group". However, district of birth is only recorded in the Côte d'Ivoire survey, so we have no alternative. ditionally withdraw international migrants (i.e., mainly relatively wealthy Malian migrants to Ivorian regional cities such as Korhogo), a slight attenuation is again observed, but the BDs still range between +0.8 and +1.2. Here, monetary welfare comparisons are facilitated by the fact that the two countries share the same currency, namely the CFA franc. We acknowledge that the BDs in nominal consumption may mix discontinuities in real terms and in price levels. However, although cattle is a traditional export of northern countries and kola nut is a traditional export of forested Côte d'Ivoire, many goods flow one way or another across the border depending on market demand (e.g., cereals such as rice or millet, textiles, spare parts, etc.), so persistent price differentials should be limited (Labazée 1993). At the national level, World Bank (2012)  In addition to income, climate and ecology strongly determine height in Africa (Moradi 2012). In West Africa, all anthropometric data confirm that people from the savannah are taller than people from the forest because of the protein intake they obtain from milk and meat (cattle breeding is constrained in the South by the presence of the tsetse fly). In the Malian survey, children from the North were taller than children born close to the Côte d'Ivoire border. Thus, when restricting the comparison to the same ecological area, we recover large BDs in nutrition: within our preferred 75 km bandwidth, children from the Côte d'Ivoire side are 13 to 21 pp less likely to be stunted. The next section will provide some evidence that the bulk of this discontinuity can be explained by parental income.
Finally, whereas the average Ivorian household is 36 pp more likely to have electricity compared to its Malian counterparts (col.1), we find no evidence of a discontinuity in electrification at the border.
In the case of the border with Guinea, no BDs in consumption per capita, stunting, or access to electricity are found (bottom panel of table 4).

IV. The role of public policies
We argue that two main factors determine the presence or absence of discontinuities in welfare at the borders of Côte d'Ivoire with its neighbors. The first factor lies in the public policies that regulate the cash crop sectors (here, cocoa, coffee, and cotton), either through administered producer prices or through agricultural extension and subsidies to production. The second factor is public investment, either national or local, in infrastructures and utilities such as electricity and safe water.
Because cash crop income is mainly derived from rural areas, whereas utilities initially reach cities, we introduce the urban structure into the analysis. This approach also allows us to test the robustness of our BD estimates. We use a variable already considered in the subsection dedicated to geography, the distance to cities with more than 5,000 inhabitants in 1990. Due to sample size constraints, we only distinguish two urbanization classes: 0 to 5 km from a city and 5 km or more. We then compute new "matching RD" estimates with PSUs matched with the nearest neighbor from the same class. 14 We report matching RD estimates for the whole sample ("All") and for the sample restricted to remote rural areas ("5+"); we do this for our preferred 75 km bandwidth. The results show that the "All" and "5+" estimates look very much the same. Indeed, cities or peri-urban areas ("0-5") are seldom found close to the borders, and it is mainly differences between remote rural areas that "make" the border discontinuities (table 5). In addition to this "urban/rural" breakdown, we report BD estimates for cash crop output and income.
At the border with Ghana (1986-8), BDs in log cash expenditures are very robust to the breakdown between urban/peri-urban (0-5 km) and remote rural (5+ We now turn to cash crop output and income. At the Ghana border, no discontinuity in cocoa output is detected; both sides are important cocoa production areas at that time. 15 However, the real producer price for cocoa in Côte d'Ivoire is more than twice that of Ghana: at 1988 prices, 14 For the rest, the estimation is kept the same; see equation (4). In the case of the Mali border, where only five urban or peri-urban PSUs are found in the 75 bandwidth, we match these five PSUs with the nearest neighbor, regardless of its urbanization class. See also table 5 footnote.
15 At the national level, the Ivorian superiority in cocoa output is due to the younger trees of the central and western regions, not to the eastern border area. Source: Authors' analysis based on data described in the text. Coverage: PSUs in the 75 km bandwidth window. Notes: "Matching RD" estimates with 75 km bandwidth. PSUs are matched to the nearest neighbor on the other side of the border within the same urbanization class: 0-5 km or 5+ km from the nearest city (5,000 + inhabitants) as of 1990. Matched differences are then analyzed as in equation (4) (see text). In the case of the Mali border, where only 5 PSUs are in the 0-5 km class, the "All" estimate (col.2a) does not match PSUs according to class, it is therefore exactly the same as in Table 4 col. (4a). Controls for latitude (Ghana and Guinea) or longitude (Mali), rainfalls, elevation, and distance to river are always included. a: 1988 prices and exchange rates in columns (1a-b); 1993 prices and exchange rates in col. the average real producer price for 1986-8 is 1.55 USD per kg in the former country compared with 0.65 in the latter. In contrast, in the case of coffee, a large BD in output is observed because this crop is not produced in Ghana due to historically low administered producer prices in this country. Among rural villages, we therefore find a large BD in cocoa and coffee income of +184 USD per capita (table 5, col. 1b). Additional disaggregation (not shown) shows that cocoa accounts for two-thirds and coffee for the remaining one-third. Before 1990, the cocoa producer price differential generated a strong incentive to smuggle cocoa beans across the Côte d'Ivoire border (Bulír 1998 This procedure involves assuming that (i) the same saving rate applies to cash crop income on both border sides and (ii) this saving rate is consistently estimated by OLS.
17 In 1986-8, the survey figures for the average annual wage of civil servants and public firms workers reached 5,223 USD in Côte d'Ivoire compared with 663 USD in Ghana. For private firm workers, they were 3,498 USD and 540 USD, respectively. The minimum wage in Côte d'Ivoire (1,400 USD) was also six times higher than in Ghana (240 USD). border, the BD in cotton income mainly stems from a higher output on the Côte d'Ivoire side because producer prices are only slightly higher (90 vs. 85 CFA franc per kg in Mali). Further, this Ivorian advantage is specific to the border area; both national official sales and survey figures show that the two countries produced roughly the same quantities of cotton (see table 2). In fact, the Mali side of the border was not yet producing significant quantities of cotton. In this country, cotton production historically began more northward, around the city of Koutiala, and mostly took off in the mid-1970s. As in Côte d'Ivoire, it was strongly regulated and subsidized by the State through a parastatal company (Compagnie Malienne de Développement des Textiles). We can again explain the BD in consumption among rural villages by the discontinuity in cash crop income. As in the case of Ghana (with cocoa and coffee), when controlling for cotton income, the BD in cash expenditures decreases from +117 USD to an insignificant -13. Unreported results also show that richer farmers invest in cattle. We identify BDs in the number of cows owned by the household (+1.76, s.e.= 0.32) as well as in the number of goats and sheep (+0.53, s.e.=0.10). Differences in public health policies may also be involved because the fight against parasitic diseases (river blindness or onchocerciasis, sleeping sickness or trypanosomiasis) was undertaken much earlier in Côte d'Ivoire.
At the Guinea border, although almost no cocoa and approximately half as much coffee is grown on the Guinea side, the BD in cash crops income is small but statistically significant (+16 USD). In comparison with the Ghana and Mali cases, we argue that the absence of a large difference in cash crop income may account for why no border discontinuity is found in consumption per capita or in children's stunting.
Regarding this latter variable, table 5 also suggests that a large part of the BDs can be accounted for by household consumption. In the Ghana case, controlling for cash expenditures per capita and its square reduces the BD in stunting from -32 among rural villages (-26 in the whole sample) to an insignificant -8 percent (-10 in the whole sample). In the Mali case, the same dramatic reduction is observed, from -24 in rural villages to -8 percent (-20 to -5 in the whole sample). Combined with the results obtained on the impact of cash crop income, this last result lends some support to the idea that the cash crop channel explains most of the border discontinuities in both household consumption and child nutrition.
In the Ghana case, electricity reaches border rural villages on the Côte d'Ivoire side, whereas none of their Ghanaian counterparts is connected. This simple fact translates into a significant +22-percentage-point BD (+28 pp in the whole sample, see last row of table 5). In 1993, these BDs were reinforced: +37 pp in rural areas and +30 in the whole sample (see table 5 footnote). The same features are found for access to safe water. In contrast, at the two other borders, all remote rural areas are in the dark. Although Ivorian regional cities are more connected to electricity than their northern counterparts, they are too far from the border to significantly contribute to the BDs. In explaining these features, it is difficult to disentangle a pure wealth effect from a more discretionary uneven allocation of public investment. 18 Recent history shows that price policies were reverted during structural adjustment under pressure from donors. In 1989-90, Côte d'Ivoire halved both its cocoa and coffee producer prices before raising them again after the CFA franc devaluation in 1994. In the meantime, Ghana significantly raised its cocoa producer price so that the difference between the two countries became negligible at the end of the 1990s. Changes in cash crop production were slower, but they also occurred. Between 1994 and 2001, Mali more than doubled its cotton production in the South through the extensive cultivation of new lands; however, the cultivation of cotton had not yet reached the most southern border area (Dufumier and Bainville 2006). In contrast, the Ivorian "cocoa frontier" has moved westward and today reaches the southern Guinea border. With these developments, some of the border discontinuities that we identified for the 1990s may have changed. After

2002, the civil war in Côte d'Ivoire and the five-year partition between North and
South may have further reduced the Ivorian advantage at the northern borders, at least until 2007. In the South, Ghana continued to catch up with its neighbor (Eberhardt and Teal 2010). Hence, although infrastructure showed greater persistence in 1993 and in 1998, border differences may have been attenuated since then. Unfortunately, the data that we gathered do not cover those more recent economic conditions.

V. Conclusion
The borders between Côte d'Ivoire and its neighbors divide fairly comparable areas in terms of geography, anthropology, and precolonial history. At the end of the 1980s and in the 1990s, Côte d'Ivoire was by far the wealthiest country, particularly because of its export crops. By applying regression discontinuity designs to a set of household surveys, we show that this higher wealth was translated at the borders with Ghana and Mali in the form of large and consistent discontinu-ities in consumption and in child stunting. We provide evidence for the fact that the bulk of these border discontinuities can be explained by cash crop production and/or pricing policies. At the border of Guinea, where no economically significant discontinuity is found in cash crop income, no discontinuity in consumption or nutrition is detected either. At the border of Ghana, we also identify persistent border discontinuities in access to electricity and safe water, even among rural villages, linked to high public investment in southern Côte d'Ivoire. At the northern borders, electricity only reaches Ivorian cities, not rural areas.
Thus, we obtain a balanced conclusion. National borders in West Africa are not highly porous, abstract lines with no impact on the welfare of the communities across them. However, our analysis suggests that border discontinuities reflect reversible public policies rather than intangible institutional traits (although political and structural factors can account for how long it takes to change policies). In rural areas, the discontinuities in consumption stem from differences in cash crop output and earnings. These differences are not necessarily permanent, and they changed during the two past decades because of policy shifts, political shocks, and medium-term agronomic developments. We hope that we have shown that border discontinuities in welfare were observed in the past and may continue to mirror the differences between national policies in the future. Appendix Source: Authors' analysis based on data described in the text. Coverage: PSUs in the bandwidth window (75 or 100 km from the corresponding border). Notes: See equations (2), (3) and (4) for each estimator. For "Border RD" and "Matching RD", the only control variable is longitude (Mali), or latitude (Guinea). For 100 km bandwidth, sample sizes allow for using a quadratic in distance to border rather than a simple linear specification. a, b: See Table 1.

The border between Côte d'Ivoire and Guinea
The northern part of this border was laid out in the context of the first war against the Almami Samori Toure whose "first empire" was centered on Kankan and extended southward in present-day Sierra Leone, eastward to Odienne, and northward to the banks of river Niger near Bamako where the French had just arrived, coming from Senegal. 23 Between 1891  There is no detailed description of the border alignment between Côte d'Ivoire and Guinea, whether in French colonial enactments, or in international agreements since independence. 24 No demarcation is known to have taken place, so that French maps still provide the best available evidence, revealing that some sectors are still indefinite.
However no dispute has ever been reported between the two countries.

The borders of Côte d'Ivoire with Mali and Burkina Faso
In this region, the French were again confronted with the Almami Samori Toure they had already defeated in 1894 and whom they had pushed to the East (cf. supra on Guinea border it was at that time a border between British Northern Territories of the Gold Coast (making part of Gold Coast administration since 1897) and French "Soudan" (Sudan).
The alignment was broadly a straight line following the 11th degree of north latitude.

S2.1. Consumption
The consumption variable is the sum of four distinct components: 1. Consumption of own food production; 2. Food expenditures including meals outside the household; 3. Housing expenditures: rents paid and imputed rents; home cleaning and reparation; water, electricity and other fuels; 4. Other current expenditures, including education.
Consumption of own production other than food is disregarded. Gifts received in kind are not included for they were not collected in all countries. Gifts and transfers to other households are disregarded, as well as tax payments. Expenditures for ceremonies and for health were not included for being too infrequent. For the same reason, durable goods acquisition and reparation are not included (furniture, domestic appliances, radio and TV sets, vehicles).
No correction was considered for within countries regional price levels differences (this kind of information is only available for the early surveys in Côte d'Ivoire and Ghana). Monthly data on national consumer price index (CPI) was used to express all components in a common base-year (1988 or 1993), taking into account their specific recall period and the month and year of recording. Although very much imperfect as CPI are not disaggregated by products, this correction is however better than nothing for periods of high inflation: in Ghana whatever the date, in Burkina Faso and Mali just after the CFA franc devaluation in 1994. 34 Household consumption levels are then translated in current dollars at base year exchange rates and prices (1988,1993), using either official or parallel (in the cases of Ghana before 1989 and Guinea) exchange rates.
Regarding consumption of own food production, all surveys directly ask households about the market value for each product, with a recall period that may vary from one survey to another. In GLSS3-4 (Ghana 1992 and1998) and EIBC (Guinea), quantities consumed are also recorded: in that case, within PSUs median unit prices can be computed and a second market value constructed. Market values are then translated into monthly consumption (with a multiplier depending on the recall period) and divided by the corresponding monthly CPI, then turned into annual consumption using declarations about how often the product is consumed during the year (when available). They are finally summed across products. When two measurements have been constructed (cases of GLSS3-4 and EIBC), we check they are fairly correlated (minimum correlation coefficient of 0.5), and we take the average of the two. 35 Regarding food expenditures, the procedure is nearly the same, except that quantities are not collected. All surveys record food expenditures during the last 15 days or the last month. This allows to construct a first simple measurement of food expenditures: (i) translation into a monthly basis (double the value when the recall period is 15 days), (ii) division by monthly CPI, (iii) multiplication by 12 to obtain the annual expenditures estimate. The Côte d'Ivoire (CILSS and ENV) surveys and the early Ghana surveys (GLSS1-2) additionally record how many months each product has been consumed during the year. This allows to construct a second measurement where the last step is replaced by (iii) multiplication by number of months of consumption to obtain the annual expenditures estimate. In the CILSS and GLSS1-2 this number of months is matched with a second declaration of expenditures over a recall period of one year (rather than one month). Here again, we keep the average of the two measurements after having checked their correlation.
Regarding housing expenditures, information usually comes from specific survey modules, and is collected over or twelve months recall period. Combined with monthly or yearly CPIs, we then straightforwardly obtain an annual aggregate at base-year prices.
For house-owners or households with free accommodation, the imputed rent is the predicted value derived from a regression estimated on tenants. The regression relates the rent paid to the characteristics of the house only; one such regression is estimated for each survey, no correction for selection is made. We check that the resulting housing budget shares are sensible.
Last, for all other expenditures, the longest recall period (usually 12 months) has been preferred when two recall periods are available. Exceptions are hygiene, cigarettes and fuels. Information on education expenditures (including transportation and sometimes food) are usually derived from a specific module.
The food consumption aggregate is trimmed separately by dropping observations for which the logarithm is under or above the mean by 5 standard deviations. Households declaring no food consumption are directly withdrawn. This "clean" food consumption amount is then summed with other expenditures and the total consumption amount is trimmed again with the same ±5 standard deviations rule. In the end, in all surveys, less than 1% of households are withdrawn from the sample by this trimming procedure.

S2.2. Geographical variables
We obtain the geographical coordinates of survey clusters from combining NGA GEOnet Names Server, Falling Rain Global Gazetteer and regional maps. These coordinates are then used to construct the geographical attributes of clusters: altitude, rainfall, distance to the nearest river and distance to national borders.  Figure S2.1 thereafter provides a map giving the location of administrative districts located in the border areas (for more details, see historical appendix). The Ghanaian region of Brong-Ahafo extends far from the border with Côte d'Ivoire, but in total only one quarter of Ghanaian survey clusters are more than 100 km away from the border, the mean distance being 77 km (GLSS1-2 and GLSS4).

S3. Additional results at other borders
As alluded to in the main text, in previous versions of this work we also looked at the border of Côte d'Ivoire with Burkina Faso. However identification had drawbacks.
The western half of this border follows the Leraba and the Comoé rivers, and a few discontinuities in elevation and distance to river are detected. The main difficulty actually comes from the distribution of localities along the boundary. The middle part of this border is not very populated and hardly represented in the sample, as the Côte d'Ivoire side lies just above a large national park; further, most Ivorian localities are urban or peri-urban, lying around the large cities of Bouna (East) or Ferkéssédougou (West). Finally and most importantly, the difference in urban structure accounts for all the border discontinuities in consumption per capita or children stunting that can be identified in the first place. It is hard to know whether this feature reflects a small sample effect or a real world pattern.
We also looked at BDs between neighbors of Côte d'Ivoire that happen to be adjacent, this making a set of six additional country-period pairs: Guinea/Mali in 1994, Burkina Faso/Mali in 1994-2001, and Burkina Faso/Ghana in 1998, and even 1992 and 2003 using DHS data. The striking result is that we could not find any significant and robust BD, with one exception only: Ghanaian households seem to benefit from an increased access to electricity at the border with Burkina Faso in 2003. These results suggest that large border discontinuities are much less likely to emerge when countries display similar levels of wealth. Macroeconomic figures indeed show that it is only in the second half of the 2000s that Ghana started to overtake the three other countries, and to catch up with Côte d'Ivoire (Eberhardt and Teal 2010). Meanwhile, Burkina Faso and Mali were staying more or less at par, and Guinea had fallen down to their level.