The Water Suitcase of Migrants: Assessing Virtual Water Fluxes Associated to Human Migration

Disentangling the relations between human migrations and water resources is relevant for food security and trade policy in water-scarce countries. It is commonly believed that human migrations are beneficial to the water endowments of origin countries for reducing the pressure on local resources. We show here that such belief is over-simplistic. We reframe the problem by considering the international food trade and the corresponding virtual water fluxes, which quantify the water used for the production of traded agricultural commodities. By means of robust analytical tools, we show that migrants strengthen the commercial links between countries, triggering trade fluxes caused by food consumption habits persisting after migration. Thus migrants significantly increase the virtual water fluxes and the use of water in the countries of origin. The flux ascribable to each migrant, i.e. the “water suitcase”, is found to have increased from 321 m3/y in 1990 to 1367 m3/y in 2010. A comparison with the water footprint of individuals shows that where the water suitcase exceeds the water footprint of inhabitants, migrations turn out to be detrimental to the water endowments of origin countries, challenging the common perception that migrations tend to relieve the pressure on the local (water) resources of origin countries.


S1 Text. Additional elaborations and results.
This supplementary text provides additional information supporting some statements given in the main paper, both form a methodological and quantitative point of view. All variables employed in the following analyses are detailed in the Data section of the main paper or within the supplementary text itself. In the tables shown hereafter, "rta" stands for regional trade agreements and "tta" stands for tariff and trade agreements. Notations following the coefficient values identify the level of significance of the considered variable at a t-Student test, i.e. '*' (p < 0.05), '**' (p < 0.01), '***' (p < 0.001); "N. observations" identifies the number of links, or country pairs, considered in the model calibration, and "R 2 -adjusted" is the coefficient of determination of the model output, adjusted by the number of calibrated parameters.

A Alternative model results: the Fixed Effects formulation and other count models
The ordinary least squares (OLS) method has traditionally been used for estimating the coefficients of the baseline gravity specification in its log-linear form. While the base-line gravity equation (Eq. (1) in the main text) is commonly used in empirical analysis, multilateral trade resistances (MTR) may be taken into account [1]. According to Baier and Bergstrand [2], a simple way of treating MTR is the use of fixed effect (FE) model, which was first emphasized by Harrigan [3]. This model specification consists of including origin-specific and destination-specific countries' intercepts in the regression. This approach includes the effect of cross-country variations of explanatory variables but precludes, in cross-sectional regressions, the analysis of country-specific determinants. We thus employ the FE model dropping all country-specific variables but the dummy indicating the belonging of the country to the tariff and trade agreement. The resulting multiplicative model reads as follow: where α i and α j are the n-dimension vector of coefficients to be estimated relative to origin and destination countries' fixed effects, respectively, while a 1 , a 2 ,... a 9 are the coefficients related to migration, distance, contiguity, common language, colony, common currency, regional trade agreement, and tariff trade agreement (of origin and destination country), respectively.   zero-inflated Negative Binomial (ZINB). Poisson and NegBin are valid alternatives to OLS, which are based on considering the dependent variable as a count process, removing the assumption of normal distribution and enabling the accounting of a high share of zero flows. In particular, NegBin introduces a parameter, called θ, controlling for distribution's overdispersion. The zero inflated models consist on a two step procedure. In the first step, we estimate the probability of existence of a link between two countries (at least one m 3 of VW) with a logit regression. In the second step we estimate the value of VW with a Poisson or NegBin model, conditional to the first step. The ZIP and ZINB permits to obtain consistent estimates even when big amount of zeros are present.

B Commonality analysis
Considering a multivariate regression between a dependent variable Y and a set of regressors x 1 ,x 2 ,x 3 , ..., the variance of Y can be expressed partly by the contribution of each regressor alone (unique contribution), and partly by the join contribution of other regressors, exerted through regressor correlation. In order to identify the contribution of each regressor and disentangle the role of correlation, a commonality analysis is performed [4]. This is a statistical tool to determine the partitioning of variance into the contribution given by each regressor into a multivariate linear regression. The following table shows that the magnitude is affected by the part conjointly explained by the other explanatory variables in the model. The commonality coefficients report a common value that increases over time (from 0.202 in 1990 to 0.325 in 2010) and a unique value of 0.003 for 1990, that increases at 0.010 for the 2000 cross section and 0.025 for year 2010.

C Different types of commodities
Virtual water fluxes are associated with the trade of different commodities and are summed across all commodities to obtain the total VW fluxes associated to trade (see the Data section in the main text). Commodities can also be grouped into categories to evaluate the associated VW flows; categories here considered are: crops, luxury food, animal-based and non-edible products. The list of products and corresponding categories can be found in [5]. We analyzed, by means of the OLS gravity model in Equation (1) (main text) and the FE specification in Equation (1) in the SI, the effect of migration on the VW trade associated to each category, and we found that all categories are positively affected by migrants (see the coefficients in the following table). The effect of migration is higher for VW trade associated to crops and animal-based commodities, while migrants have less importance in explaining VW trade associated to non edible commodities. Model coefficients also show an increasing temporal trend, as found in the analyses of total VW trade.

D Migration flows data
It may be argued that the effect of migrants moving in a certain period (migration flows) can be significant, in place of that of migration stocks. However, we suspect a temporal delay between the migrants settlement and trade initiation/intensification thus we expect the effect of stocks to be higher. To check for this hypothesis, we used two different sources of data. First, we constructed migration flows data for each pair of countries and each direction from stock data, by taking the differences between subsequent decades (  [6] from the same UN stock data, but accounting for births and deaths rates [7]. They estimated migration flows for the periods 1990-1995, 1995-2000, 2000-2005, 2005-2010, that we summed up to match the decades of our analysis. We re-ran the OLS and the FE models for the two decades (see the following tables), and we computed commonality coefficients on OLS method using both our differenced UN data and Abel-Sanders data (migration AS in the tables). Results highlight that also migration flows significantly affect the VW fluxes, but to a smaller extent compared to migrant stocks.  We used the instrumental variable approach to overcome the issues of potential reverse causality of migration and VW trade. Several approaches have been proposed in the literature [8,9,10] to remove the effect of spurious correlations using an instrumental variable. We adopt the two-stage least square method where, at the first stage, a regression is set between the migration flows and the time-lagged migration flows. At the second step, we use the fitted values of the first step as an exogenous migration variable in the instrumented regression. Looking at the migration coefficients in the instrumented regression (following table) compared with the same (non instrumented) coefficients (in this file, Section 4), it is possible to see that results are robust to different data and method applications, proving that the role of reverse causality in the VW trade-migrants relations is marginal.  Considering the lagged migration flow variable as instrument and the results for our migration dataset only, the diagnostic tests provide the following results. The first-stage F-test for instrument relevance shows coefficients equal to 5047.6 (OLS gravity model) and 3372.3 (FE gravity model), both with very small p-value (< 0.0005); the Wu-Hausman F-test for esogeneity shows coefficients equal to 27.5 (OLS) and 8.9 (FE), both having p-value < 0.0005.

F The role of refugees and asylum seekers
An asylum seeker is defined as someone who leaves its own country, often for political reasons or because of war, and who travels to another country hoping that the government will protect him and allow him to live there. A refugee is an asylum seeker who obtained the permission to live in the destination country. These migrants which usually leave countries undergoing a crisis are unlikely to be capable of increasing the VW trade due to the food demand. For this reason, we think that a refugee does not cause VW fluxes, thus we used refugees data as a placebo test.
To check this hypothesis we perform the OLS and the FE gravity models for the period 2005-2010, replacing total migration flows with refugees and asylum seekers flows and averaging population and per capita gdp over the period 2005-2010. We used data from UNHCR Population Statistics Database [11], which reports information on refugees flows for every year for the period 2000-2013, where asylum seekers are taken as the people applying for asylum during the year, and the number of total decisions is taken as a proxy for the number of refugees. Results (in the following table) show a smaller effect of both refugees and asylum seeker on VW trade, which turns negative in the FE specification. Therefore, it is possible to argue that VW flows changes are affected only to a limited extent by the flow of forced migrants.