Coastal Proximity and Individual Living Standards: Econometric Evidence from Geo-Referenced Household Surveys in Sub-Saharan Africa *

We investigate geo-referenced household-level data consisting of up to 128,609 individuals living in 11,261 localities across 17 coastal sub-Saharan African countries over 20 years. We analyze the relevance of coastal proximity, measured by geographic distance to harbors, as a predictor of individual economic living standards. Our setting allows us to account for country-time fixed effects as well as individual-specific controls. Results reveal that individuals living further away from the coast are significantly poorer measured along an array of welfare indicators. Our findings are robust to the inclusion of other geographic covariates of development such as climate (e.g. temperature, precipitation) or terrain conditions (e.g. ruggedness, land suitability). We also explore mechanisms through which coastal proximity may matter for individual welfare and decompose the estimated effect of coastal proximity via formal mediation analysis. Our results highlight the role of human capital, urbanization as well as infrastructural endowments in explaining within-country differences in individual economic welfare.


I. INTRODUCTION
Cross-country studies investigating the link between physical geography and economic development consistently provide evidence of a positive and statistically significant association between coastal proximity and national income (e.g. Bloom et al. 1998;Radelet and Sachs 1998;Gallup et al. 1999;Redding and Venables 2004;Putterman and Weil 2010;UN-OHRLLS 2013;Carmignani 2015). More recent literature analyzing subnational variation in economic activity also suggests coastal proximity as a relevant indicator of within-country income differences (e.g. Rappaport and Sachs 2003;Gennaioli et al. 2013;Motamed et al. 2014;Mitton 2016;Flückiger and Ludwig 2018;Henderson et al. 2018;Jetter et al. 2019).
To systematically complement the literature that focused on outcomes at the national or regional level, this paper analyzes the relevance of coastal proximity on individual economic welfare. We employ a repeated cross-sectional dataset from the Afrobarometer spanning almost 20 years and consisting of up to 128,609 individuals living in 11,261 geo-referenced localities across 17 coastal sub-Saharan African countries. Particularly in Africa, countries and regions with coastal access have had higher levels of economic development compared to more remote areas, which has been attributed to factors such as lower costs of trade as well as the amplifying forces of urbanization and agglomeration (Bloom et al. 1998;Gallup et al. 1999;Limão and Venables 2001;Rappaport and Sachs 2003;Atkin and Donaldson 2015;Storeygard 2016;Henderson et al. 2018). Spatial inequalities such as these have been shown to persist even when initial advantages of coastal areas may have declined in relevance (Bleakley and Lin 2012;Jedwab et al. 2017).
Our results confirm coastal proximity as a robust indicator of individual economic welfare across African countries: Living further away from the coast is associated with a significant and meaningful reduction in the likelihood of having cash employment, increases in the occurrence of cash-, food-, water-and medicinal droughts (deprivation), as well as lower 3 overall household wealth (possessions). Our results are robust to the inclusion of relevant individual-level covariates, country-time specific influences via fixed-effects, as well as an extensive set of further geographic variables related to development such as latitude, elevation, climatic factors (e.g. temperature, precipitation) and features of the terrain (e.g. ruggedness, land suitability).
We also explore potential mechanisms on how coastal proximity may matter for individual living standards and investigate several candidate factors shown to contribute to spatial disparities in the literature (for an overview see Breinlich et al. 2014). In particular, we The remainder of this paper is structured as follows: Section II presents the data and the estimation strategy. Our results are given in Section III, where we also present the insights from our mediation analysis. Concluding remarks are offered in Section IV.

Data
We employ the complete set of the geo-referenced Afrobarometer survey rounds, spanning a timeframe of 20 years (from 1999 to 2018) across seven survey waves (Afrobarometer 2019). 1 Afrobarometer surveys are representative at the national level and respondents are adults of the sampled households. They carry individual-and household level information on basic characteristics, including living conditions and household assets, and additionally, provide information on individuals' sentiments as well as opinions towards the economy, democracy, governance and society. Afrobarometer fits geo-coordinates (latitude and longitude) to respondents at the level of their respective enumeration area (BenYishay et al. 2017). The sampling procedure aims for eight individuals/households per EA. Our main (extended) sample of countries consists of up to 128,609 (212,037) individuals living in 11,261 (17,319) georeferenced localities across 17 (28) coastal sub-Saharan African countries (see Figure 1). We chose to restrict the main sample to coastal countries, so as to separate the distance effect from a more general "landlockedness" effect which potentially confounds distance with other influences (UN-OHRLLS 2013;Carmignani 2015). We investigate the extended sample including individuals living in landlocked countries in our robustness tests. 1 Surveys were sampled in 1999-2001, 2002-2004, 2005-2006, 2008-2009, 2011-2013, 2014-2015 and 2016-2018, respectively. 5 as well as the dummy variable Urban (0/1). We also analyze individuals' opinions towards supranational organizations aimed at increasing political as well as economic integration, Helps your Country: [AU or ECOWAS/SADC/EAC/IGAD…] (0-3) to directly relate coastal proximity with regional trade considerations. 7 To further explore a potential trade channel, we use information on individuals' occupation and test for a differential effect of distance using 2 The four different questions read: "Over the past year, how often, if ever, have you or anyone in your family gone without: Enough clean water for home use" / " […]: Enough food to eat" / " […]: A cash income" / " […]: Medicines or medical treatment?". These questions are consistently available in all Afrobarometer survey rounds. Using each question separately does not affect our main insights as shown in the Appendix. 3 The questions read "Which of these things do you personally own? Radio" / " […]? Television" / " […]?
Motor Vehicle-Car-Motorcycle". Wealth possessions were surveyed from Round 3 and onwards. Using each question separately does not affect our main insights as shown in the Appendix. 4 The question reads: "In your opinion, what are the most important problems facing this country that government should address?". 5 To measure the quality of institutions, we construct an "Institutions Score" similar to Mitton (2016), which is based on an array of questions regarding local-authorities, processes and government. The score is constituted of 21 questions measuring individuals' trust in (local) courts, police and government, their experience with the procedures of local authorities, especially regarding bribery (corruption), the enforcement of crime, and the ease of handling matters. Higher values indicate fewer experiences/better judgments of (local) institutions. 6 Nunn and Wantchekon (2011) use an identical measure für the provision of public goods, excluding roads. 7 Regional Economic Communities have the proclaimed aim to foster the movement of goods and people, and to improve living standards. The question reads: "In your opinion, how much do each of the following do to help your country, or haven't you heard enough to say".
Commercial Farmer (0/1), a dichotomous indicator for individuals working as farmers who grow their produce mainly for sale. 8

b) Main Independent Variables
To construct our main explanatory variable of interest, log(Distance to Harbor), we measure the shortest geodesic (ellipsoidal) within-country distance from each respondent's enumeration area to the respective country's major harbor(s). 9 As in Rappaport and Sachs (2003), we define all large and medium sized ports listed in the World Port Index (WPI) as "major harbors" (NGA 2019).
We also employ alternative conceptions of coastal proximity for robustness checks, namely shortest within-country distance to the coastline, log(Distance to Coastline), distance to major harbors using beelines (as the crow flies), log(Beeline Distance to Harbor), as well as distance to the coastline using beelines, log(Beeline Distance to Coast). 10 Shapefile data for country administrative areas, the boundaries of which we use to calculate within country distancesand also from which we construct the coastlinecome from the Center for Spatial Sciences at the University of California (GADM 2020).

c) Further Covariates
To isolate coastal proximity from other, potentially correlated, geographic influences of We thank an anonymous referee for pointing out this additional extension. 9 We measure distances using the projection of coordinates along the earth's ellipsoid (using WGS 84, EPSG 4326). We add +1 (kilometer) before logging all our distance measure prior to taking the logarithm. 10 Beeline distances disregard country boundaries, i.e. cross country borders for shorter distances.
8 Temperature and Monthly Rainfall (Fick and Hijmans 2017). We also include seven dummy variables indicating the dominant natural vegetation of the area according to Olson et al. (2001). 11 We account for individuals' access to rivers or lakes by adding two dummy variables indicating whether individuals live within 25 kilometers of a navigable river or major lake, i.e. Navigable River (0/1) and Major Lake (0/1), and thereby analyze an extended set of traderelated covariates together with our main explanatory variable, log(Distance to Harbor) (see Henderson et al. 2018). 12 We also add in individual-level covariates Age, squared Age and a dichotomous indicator of gender, Female (0/1). The importance of urbanization-agglomeration aspects, argued to be particularly relevant in African contexts (Young 2013;Motamed et al. 2014;Chauvin et al. 2017;Jedwab et al. 2017;Gollin et al. 2017;Flückiger and Ludwig 2018;Henderson et al. 2018), is encapsulated by three distinct indicators of urbanization: Primate City (0/1), a dummy indicating whether individuals live within 25 kilometers of a capital or primate city, Population Density (CIESIN 2017), a continuous measure of population density (per sq. kilometer), as well as Urban (0/1), a dichotomous indicator included in the Afrobarometer survey.
Descriptive statistics for all variables are presented in Table A1, part a) and b), of the Appendix.
, , = + log( ) , , + ′ , , + , + , , , , represents the respective welfare indicator of individual in country , surveyed at survey-sampling period . captures the influence of the logged (within-country) distance to major harbors such that the link between distance and the respective welfare indicator can be interpreted as a semi-elasticity. Standard errors are clustered at level of the survey enumeration area, i.e. at the survey cluster level. Binary dependent variables are estimated with a simple Linear Probability Model (LPM) specification. 13 represents a matrix of control variables which allows us to account for all influences potentially conflating the relationship between coastal distance and individual economic welfare. In contrast to the cross-country (crossregional) literature, our setting allows us to account for country-time fixed effects , such that we can explore a within-country estimate of distance to harbor on (individual) outcomes net of time-specific influences as well as country-specific influences at specific points in time, such as the Kenyan Post-Election Crisis of 2007-2008. , , is an idiosyncratic error term.
We explore potential mechanisms and factors affecting the link between coastal proximity and individual living standards both via a "bad control" approach as well as a formal mediation analysis, after establishing the relevance of coastal proximity for individual living standards. Numerous robustness checks for the persistence of the observed links are offered (mostly relegated to the Appendix). we report the predicted change of the respective dependent variable when moving from the minimum distance in the sample (i.e. effectively living by a major port) to living as far as 564 kilometers away from the harbor (3 rd quartile of sample) and compare the predicted change of each individual welfare indicator to the respective sample mean reported in brackets. The results show that distance to harbors is statistically significantly and negatively related to cash employment (columns 1 and 2) and positively, statistically significantly, related to deprivation (columns 3 and 4). Quantitatively, increasing individuals' distance to major ports to the 3 rd quartile in the sample translates into to a 5.5 percentage-point decrease in the probability of having part-or full-time cash-employment (column 2) and can explain 17% of the occurrence of monetary-, medicinal-as well as food-and water-related shortages compared to the mean in the sample (column 4). Coastal remoteness is significantly related to having fewer (wealth) possessions: increasing individuals' distance to the 3 rd quartile corresponds to a 12 percentagepoint decrease in the probability of owning a radio, a tv or a motor vehicle, accordingly, a 23% reduction compared to the mean in the sample (column 6). Importantly, the results for our indices of deprivation (columns 3 and 4) and possessions (columns 5 and 6) also hold when analyzing the variables that compose our indices separately (see Table A3). Notes: Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and 3rd quartile harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.

Robustness Checks and Extensions
We conduct a large array of robustness checks on our main results and summarize them in Table A4 of the Appendix. All interpretations regarding the relevance of coastal proximity for individual living standards remain robust: (a) We re-estimate our main results by altering the distance specification to a simple "beeline" ("as the crow flies") measure. (b) We use a different conceptualization of coastal proximity by regressing our outcome variables on individuals' distance to the coastline instead of port locations, using log(Distance to Coastline). (c) Accordingly, we test beeline distances to the coastline with, log(Beeline Distance to Coastline).
We add dummies for living within 25 kilometers to a major harbor (d) or the coast (e), Harbor (0/1) and Coast (0/1) to separate the distance effect from a pure "coastal access" effect. 16 Table A14.
All robustness checks corroborate our general findings of a negative, independent, statistically significant relationship between coastal proximity and individual living standards.
The results reiterate the relevance of coastal proximity, in varying conceptualizations, in predicting individual economic welfare.
Next to the above-mentioned robustness tests, we extend our analysis and (a) expand our main sample to include individuals living in landlocked countries (see "Extended Sample" in 13 Figure 1) and also (b) analyze the persistence of our estimated effects over time. For (a), we include individuals living in landlocked countries, to explore a potential "placebo" group compared to individuals living in coastal countries. This allows us to compare the effect of sheer coastal distance within countries from a landlockedness-effect, the one often explored in the literature (see e.g. UN-OHRLLS 2013;Carmignani 2015). The idea is that differences in individual coastal proximity within landlocked countries should influence individual welfare to a lower degree given that national borders need to be crossed, creating further, potentially large, restrictions. 17 As expected, Table A15 suggests that the relevance of individual distance to harbors tends to be less pronounced for individuals living in landlocked countries. (b) The relative importance of trade-related factors of geography might be expected to change along a country's developmental path (see Henderson et al. 2018). Hence, we estimate differential effects using an interaction effect constituted of log(Distance to Harbor) and Young (0/1), which indicates respondents below the median age in the sample (33). The results in Table A16 of the Appendix show a clear pattern. The negative effect of distance becomes less stark for younger generations, potentially hinting at a reduction in the relevance of trade-related aspects over time (see Henderson et al. 2018). 18 17 Empirically, we add an interaction term constituted of log(Distance to Harbor) and a binary variable indicating whether the country is Landlocked. The sum of the coefficients log(Distance to Harbor) and the interaction term represents the total effect of distance to coast for individuals living in landlocked countries. 18 We thank an anonymous referee for pointing out this additional extension. Notes: Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in column (7) includes individuals from rounds 2 through 7 of the Afrobarometer, column (8) and [(9)] includes data from rounds 2, 4, [5] and 6. We also report estimated interquartile differences in the respective dependent variables between minimum-, and 3rd quartile harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01. Individual educational attainment has been linked to economic welfare at the cross- we also find a negative, statistically significant, and quantitatively large relationship between (coastal) remoteness and living in urban environments (column 3). Given the strong interconnection between coastal proximity and urbanization, we explore the differential effects of distance for individuals from urban environments in Table A17 separately, by estimating the interaction term log(Distance to Harbor) * Urban(0/1). The results show that, while less pronounced, the distance penalty remains a statistically significant predictor of individual living standards, for two of our three outcomes, for respondents living in urban surroundings.

Mechanisms Explaining the Relevance of Coastal Proximity
Regarding institutions, we proceed similar to Mitton (2016; 107) and construct an index, results at the individual level suggest that coastal proximity is negatively associated with respondents' sentiments that infrastructure needs are issues of concern (see column 6 and 7).
The actual access to basic infrastructure (as measured by our composite infrastructure index), is also negatively associated with distance to major ports. phenomenon, we find that survey respondents further away from the coast exhibit a higher tendency to report their respective Regional Economic Communities (RECs) or the African Union (AU) as helpful to their country, which is consistent with them wishing to improve trade opportunities. Moreover, Table A18 shows that, the distance penalty is significantly increased for commercial farmers, i.e. farmers who mainly grow their produce for sale. Commercial farming is likely to depend on access to markets and trade opportunities, leaving commercial farmers more vulnerable to a distance penalty.

Bad Controls and Mediation Analysis
All results highlight coastal proximity as a statistically as well as an economically meaningful indicator of individual living standards and as a relevant predictor for diverse mechanisms that systematically relate and contribute to economic development and spatial inequalities. As coastal remoteness need not be destiny (Motamed et al. 2014), we aim to gauge the empirical importance of our controls as well as our potential mechanisms on our main explanatory variable by investigating the relevance of a bad controls problem (a) and by performing a formal mediation analysis (b).
Notes: Odd columns present the coefficient (changes) of our main explanatory variable log(Distance to Harbor) when subsequently adding seven distinct (sets of) control variables to a parsimonous basline regression constituted of our main coefficient and country-time fixed effects. Even column report the corresponding changes in the total R-squared compared to the previous specification, respectively. The results in each row come from seperate regressions and observations are held constant across rows. Inclusion of mediating factors in (h), specifically variables on infrastructure, limits the sample to rounds 2 through 7 of coastal, sub-saharan African countries included in the Afrobarometer. Remaining changes in the number of observations across columns stem from differencs in the response rates of dependent variables (see notes of

a) Bad Controls
We add in all of our baseline covariates and the explored mechanisms in step-wise fashion and report the corresponding changes to our main coefficient, log(Distance to Harbor), as well as changes in the residual variance. Results are presented in Table 3. Row (a) shows the coefficient of log(Distance to Harbor) in a regression including country-time fixed effects only. Row (b) proceeds to add in our basic controls, i.e. Age, Age squared and Female (0/1), as is done in Table 1. Row (c) adds our three urbanization controls to the specification, and so on. 19 The results show that, while the coefficient size of log(Distance to Harbor) diminishes as expected, coastal proximity remains a statistically relevant predictor of individual living standards throughout all rows and columns. The covariates contributing most to the specifications, as seen by changes in the coefficient (odd column numbers) as well as changes in the R-squared (even column numbers), are Urbanization Controls, Educational Level and Infrastructure, which are the ones we will explore as potential mediators next.

b) Mediation Analysis
To further evaluate the link between coastal proximity, its potential channels of influence as well as individual welfare, we conduct a formal mediation analysis. We empirically decompose the total effect of coastal proximity and individual welfare into indirect effects, i.e. effects which run through the proposed mediating factors, and direct effects, i.e. effects of coastal proximity that are unrelated to the proposed channels.
1 measures the direct effect of coastal proximity on our different welfare indicators , and 2 measures the effect of distance to harbor on the respective mediator (e.g. education, urbanity, infrastructure). represents the direct effect of the mediator on the outcome variable such that the indirect effect is retrieved by multiplying 2 * (Alwin and Hauser 1975; MacKinnon et al. 2007). The total effect is then given by a summation of the direct ( 1 ) and indirect effects ( 2 * ). 20 Figure A2 provides a visual representation of the mediation analysis.
As before, is a matrix including all our usual controls. We keep country-time fixed effects , to evaluate a stringent setting.  Table 2 and Table 3, Education Level, Urbanization Controls and Infrastructure and estimate their mediating effect on our three main outcome variables (results for our proxies of Institutions and Trade are relegated to Table A19 in the Appendix).
The results in Table 4 suggest that a substantial part of the total effect of distance to harbors is mediated by educational attainment. Including respondents' level of schooling in the main specification (equation 2) reduces the coefficient size of the direct effect of coastal proximity by 28% (see proportion mediated at the bottom of the is therefore picking up a substantial part of the total effect of coastal distance, in similar magnitude as do educational differences.
It is important to note that while both education and urbanization absorb variation in explaining individual living standards on their own (Table 3), as well as through their mediation of coastal proximity ( -0.009*** -0.012*** -0.009*** 0.035*** 0.055*** 0.037*** -0.019*** -0.028*** -0.019*** (0 The relevance of coastal proximity on economic development has been often ascribed to trade-related factors, especially among "late developers" (see Henderson et al. 2018). Table   A19 explores this link, estimating the direct and indirect effect of regional and supra-regional institutions fostering trade, as measured by respondents' evaluation of the African Union (AU) and their "corresponding" Regional Economic Community (REC), respectively. The results show that more positively perceived trade organizations correlate positively with individual living standards (row 4), which emphasizes a potential need of trade facilitation independent of individuals distance harbors or (global) markets.

IV. CONCLUSION
We systematically investigate the role of coastal proximity in explaining intra-national differences in individual living standards across sub-Saharan Africa economies using an extensive dataset covering up to 128,609 observations distributed across 11,261 localities over 20 years. We employ geo-referenced individual data to complement the existing literature that focused on outcomes at the national or regional level. Analyzing individuals' distance to harbors and their corresponding living standards allows us to test whether the insights of the cross-country and cross-regional contexts also apply at the individual level. Moreover, we can utilize the comprehensiveness of our dataset to explore a large set of indicators and potential channels of influence to gauge the relevance of coastal proximity and to investigate the mechanism through which it may matter for individual living standards.
Our results show that coastal proximity, as measured by geographical distance to harbors, Exploring potential channels, we find that human capital, urbanization, as well as access to infrastructure mediate relevant parts of the link between coastal proximity and economic development. This highlights that even though coastal proximity is a relevant indicator for individual living standards across Africa, coastal proximity need not be "destiny". Fostering education as well as infrastructural outlays might help in mitigating problems associated with coastal remoteness. Nevertheless, the systematic robustness of coastal proximity as a predictor for individual living standards even in stringent settings suggests that there are relevant development costs of remoteness alone that need to be addressed (see also UN-OHRLLS 2013).   1999-2001, 2002-2004, 2005-2006, 2008-2009, 2011-2013, 2014-2015 and 2016-2018 (i.e. Round 1 through Round 7) geo-referenced Afrobarometer survey rounds. Variation in the number of observations size stem from differences in response rates of variables as well as changes in questions asked across surveys. Geographic covariates come from an array of sources described in the data section of the paper.  1999-2001, 2002-2004, 2005-2006, 2008-2009, 2011-2013, 2014-2015 and 2016-2018 (i.e. Round 1 through Round 7) geo-referenced Afrobarometer survey rounds. Variation in the number of observations size stem from differences in response rates of variables as well as changes in questions asked across surveys. Geographic covariates come from an array of sources described in the data section of the paper. † Not asked in survey rounds 1 and 2. ‡ Not asked in survey round 1. ' Only asked in survey rounds 2, 4 and 6. * Only asked in survey rounds 2, 4, 5 and 6. ** Only asked in survey rounds 1,2, and 3. Notes: Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and 3rd quartile harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.   Table 1. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (9) through (14) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.

Coefficient on log(Distance to Harbor) -if not indicated otherwise
(see Table A5)  Table A11) (see Table A13) (see Table A7) Notes: This table summarizes the robustness checks on our main findings. Results in each row and column come from separate regressions with changes to the baseline estimates as indicated. a) Replaces the main explanatory variable of (logged) within-country distance used in Table 1 with a simple beeline (as the crow flies) distance to the nearest harbor. b) Replaces the main explanatory variable with a within-country distance to the country's coastline. (c) Replaces the main explanatory variable with a beeline (as the crow flies) distance to the coastline. (d) Adds a dummy indicator of living within 25 kilometers to a major harbor to our baseline specification. (e) Includes a dummy indicator of living within 25 kilometers to the coastline to our baseline specification. (f) Estimates the basline specification but holds the number of observations constant across all columns, using the maximum amount of individuals for which all inpendent and dependent variables are available. Note that this drops rounds 1 and 2 of the Afrobarometer entirely, given that questions on ownership of household items were not asked in these rounds. (g) Estimates the baseline specification but excludes distances larger than the 80th percentile. (h) Estimates the basline specification but excludes localities marked with a precision code of 2 or larger (scale 1-8). (i) Estimates the baseline specification but uses the included Afrobarometer survey weights. (j) Estimates the baseline specification but implements standard error clustering at the country-sample level, instead of the survey enumeration area level. We use the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer in all columns and rows. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.
(see Table A12)  Table 1 with a beeline (as the crow flies) distance to the nearest harbor. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.

Urbanization Controls
Urban (

Trade-related Controls
Navigable River (0/1) -0.011 -0.130*** 0.032*** (0  This table is equivalent to Table 1, but exchanges the within-country distance to harbors used in Table 1 with within-country distance to the coastline. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.

Urbanization Controls
Urban (

Trade-related Controls
Navigable River (0/1) -0.  This table is equivalent to Table 1, but exchanges the within-country distance to harbors used in Table 1 with a beeline (as the crow flies) distance to the coastline. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.

Trade-related Controls
Harbor ( Age 2 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Female (0/1) -0.106*** -0.106*** 0.020*** 0.020*** -0.083*** -0.083*** (0  This table is equivalent to Table 1, but adds dummies indicating individuals living within 25km of the harbor or coast, in turn. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.   This table is equivalent to Table 1, but holds the number of observations constant across all columns, using the maximum amount of individuals for which all inpendent and dependent variables are available. Note that this drops rounds 1 and 2 of the Afrobarometer entirely, given that questions on ownership of household items were not asked in these rounds. Therefore, the sample used is comprised of coastal, sub-saharan African countries included in survey rounds 3 through 7 of the Afrobarometer. Results in each column come from seperate regressions. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the country-sample level. * p < .1, ** p < .05, *** p < .01.   This table is equivalent to Table 1, but excludes distances larger than the 80th percentile. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the country-sample level. * p < .1, ** p < .05, *** p < .01.   This table is equivalent to Table 1, but excludes localities marked with a precision code of 2 or larger (scale 1-8). Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the country-sample level. * p < .1, ** p < .05, *** p < .01.   This table is equivalent to Table 1, but uses the included Afrobarometer survey weights to produce the estimates. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.   This table is equivalent to Table 1, but implements standard error clustering at the country-sample level, instead of the survey enumeration area level. Results in each column come from seperate regressions and are estimated using the main sample of coastal, sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective dependent variables between minimum-, and the 3rd quartile of harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the country-sample level. * p < .1, ** p < .05, *** p < .01.   (2019), analyzing changes in the estimate of our main explanatory variable "log(Distance to Harbor)" when adding the full set of controls as well as fixed-effects, using our three main outcome variables. Columns (1) and (2) present the uncontrolled β°, as well as the controlled β' and columns (3) and (4) depict their respective regression's R-Squared. Column (5) shows the lower-and upper bound estimate of the identified set assuming R max = 1 and δ = 0.5. The bias-adjusted upper bound is calculated using β * = β' -δ((β°-β')*(R max -R' 2 ))/(R' 2 -R°2l)), the lower bound is given by β'. Results are estimated using the main sample of coastal, subsaharan African countries included in survey rounds 1 through 7 of the Afrobarometer. The sample used row 3 do not include individuals surveyed in Rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification.The standard errors reported are clustered at the survey-enumeration area level. * p < .1, ** p < .05, *** p < .01. Notes: We analyze the differential effect of our main explanatory variable for individuals living in landlocked countries. Row one shows the uninteracted effect of log(Distance to Harbor), i.e. the distance effect of individuals in coastal countries, row two shows the differential effect for being landlocked. Row three depicts the combined effect of the two constituent terms, i.e. the effect of log(Distance to Harbor) for individuals living in landlocked countires. Results in each column come from seperate regressions and are estimated using the sample of coastal and landlocked sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/ independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.  Notes: We analyze the differential effect of our main explanatory variable for individuals above and below the median age within the sample. Row one shows the uninteracted effect of log(Distance to Harbor), i.e. the distance effect of individuals at and above the median age (33), row two shows the differential effect for being in the younger strata. Row three depicts the combined effect of the two constituent terms, i.e. the effect of log(Distance to Harbor) for individuals younger than the median age (33). Results in each column come from seperate regressions and are estimated using the sample of coastal and landlocked sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/ independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.  Notes: We analyze the differential effect of our main explanatory variable for individuals living in urban environments. Row one shows the uninteracted effect of log(Distance to Harbor), i.e. the distance effect of individuals in rural settings, row two shows the differential effect for urban sample participants. Row three depicts the combined effect of the two constituent terms, i.e. the effect of log(Distance to Harbor) for individuals living in urban environments together with the corresponding p-value in brackets. Results in each column come from seperate regressions and are estimated using the sample of coastal and landlocked sub-saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differencs in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.  Notes: We analyze the differential effect of our main explanatory variable for individuals working as commercial farmers, i.e. farmers who produce mainly for sale. Row one shows the uninteracted effect of log(Distance to Harbor), i.e. the distance effect of individuals not working as commercial farmers, row two shows the differential effect for commercial farmers. Row three depicts the combined effect of the two constituent terms, i.e. the effect of log(Distance to Harbor) for individuals working as commercial farmers together with the corresponding p-value in brackets. Results in each column come from seperate regressions and are estimated using the sample of coastal and landlocked sub-saharan African countries included in survey rounds 1, 2, and 3 of the Afrobarometer for which this detailed occupational data is available. Remaining changes in the number of observations across columns stem from differencs in the response rates of dependent/ independent variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level. * p < .1, ** p < .05, *** p < .01.