Remittances and Household Investment Decisions: Evidence from sub-Saharan Africa

Globally, one in nine people receive remittances from a migrant family member, and these transfers make up about 60% of the receiving household's income (United Nations, 2019). The United Nations estimate that about three-quarters of remittances are spent on necessities, such as food and housing, while the rest is saved or invested in income-generating activities and coping with shocks (i.e., crop failure or family emergencies) (United Nations, 2019). Earlier studies on the uses of remittances focused on household consumption expenditure (i.e., durable and non-durable goods). However, more recently, an increasing number of studies are investigating households’ use of remittances for investment purposes.

In principle, remittances can help boost the longer-term prospects of remittance-receiving households by facilitating investment in productive assets. It can also help to smoothen consumption for households affected by adverse economic shocks. However, despite the potential of remittances to stimulate capital accumulation and investment, the earning capacity of receiving households often stays unchanged even after years of receiving remittances. This observation suggests that remittance-receiving households often fail to accumulate capital and invest in income-generating activities, instead allocating remittances to immediate and conspicuous consumption (Chami et al., 2005; Kakhkharov and Ahunov, 2020; Simiyu, 2013). Additionally, dependence on remittances may reduce involvement in income-generating activities by the left-behind household members.

It is difficult to theoretically determine the net effect of remittances on investment decisions. Several empirical studies have investigated this question, but the findings are inconclusive (Démurger, 2015). Studies have found positive (Amuedo-Dorantes and Pozo, 2010; Jena, 2018), null (Acosta, 2011), and adverse effects (Simiyu, 2013) of remittances on household investment decisions. The bigger concern here is the reliability of the existing empirical studies, as these studies often suffer from selection bias and other endogeneity issues (Adams, 2011).

This study aims to examine the effect of remittances on households’ investment decisions using data from five sub-Saharan Africa (SSA) countries. More specifically, we investigate two questions. First, are remittance-receiving households more likely to make investment expenditures than non–remittance-receiving households? Second, do the household investment decisions vary by the type of investment expenditure: human capital (i.e., education and health), physical capital, and social capital? In addition, we also explore the heterogenous effect of remittance sources – internal, within-Africa, and out-of-Africa remittances. Finally, we examine three different channels – income effect, substitution effect, and migration expectations – through which remittances can affect household investment decisions.

SSA is an excellent context to study these questions as little is known about the relationship between remittances and household investment decisions in the region. To our knowledge, only a handful of studies have examined this question in the region. Moreover, most empirical studies on the subject are based on Latin American countries, with some focus on Asia (Acharya and Leon-Gonzalez, 2014) but largely ignoring SSA. The results from these regions may not be generalizable to SSA primarily due to differences in migration patterns. SSA migrants typically migrate to other African countries or outside the continent with no intention of returning, while Latin American and Asian migrants are typically temporary migrants who return to their country of origin (Ratha et al., 2011).

Our study utilizes the Migration and Remittances Household Surveys conducted by the World Bank between 2009 and 2010 as part of the Africa Migration Project (AMP). These cross-sectional household surveys provide comprehensive information about migration, remittances, housing conditions, household assets and expenditures, and other socioeconomic and demographic characteristics. The dataset also provides the opportunity to analyze remittance flows by sources, namely, internal remittances, within-Africa remittances, and out-of-Africa remittances. We use data on five predominantly remittance-receiving countries in the AMP: Uganda, Kenya, Nigeria, Burkina Faso, and Senegal. An important characteristic of the AMP surveys is that they are standardized across countries, which allows for easy aggregation and comparison. The AMP surveys enable us to provide country-specific results and explore the effect heterogeneity across countries in SSA.

We define investment expenditure as an outlay for which the household expects financial returns in the future. Human capital investments are broadly defined to have two components – expenditure on education and health. Education investments are households’ expenditure on tuition payment, purchase of school uniforms, books, and other related expenditures. Health investments are households’ expenditure on doctor fees, hospital fees, and cost of diagnostics tests and medicine. Physical capital investments are households’ expenditure on setting up a business, opening a store, purchasing farming equipment such as tractors, and purchasing other productive assets. Finally, social capital investments are households’ expenditure on festivals, weddings, and funerals. Households that spend on festivals, either as contributions to the village or in private celebrations, receive tangible returns in the form of higher social status, access to larger social networks to protect against adverse economic shocks, and access to credit markets (Rao, 2001).

To identify the effect of remittances on household investment decisions, we use a recursive bivariate probit model and instrumental variables (IV) approach. The recursive bivariate probit model simultaneously estimates remittance receipt and investment decisions while incorporating the remittance-receipt variable in the investment decision equation. The identification of the recursive bivariate probit model parameters requires at least one variable (i.e., instrumental variable) in the remittance-receipt equation that is excluded from the investment decision equation (Bhattacharya et al., 2006; Horrace and Oaxaca, 2006; Jena, 2018; Wooldridge, 2002). Our primary outcome variables are binary indicators that equal one if the household made an investment expenditure in the previous 6 months before the interview and zero otherwise. Similarly, our treatment variable is a binary indicator, which equals one if the household has received remittances in the previous 12 months before the interview and zero otherwise. Since the investment decision and remittance receipts are binary variables and remittance receipt is potentially endogenous, the regression analysis employs a recursive bivariate probit model. However, we also implement a two-stage least squares (2SLS) approach as a robustness check. In addition, we conduct intensive margin analysis using the actual amounts of investment expenditures and remittances.

We account for the potential endogeneity of remittances by using historical migration networks. This instrument has been used previously in the migration literature (see Acosta, 2011; Coon, 2016; Mckenzie and Rapoport, 2011). We define historical migration networks as district-level historical migration rates. District-level historical migration rates are obtained from Population and Housing censuses – Burkina Faso 1996, Kenya 1999, Nigeria 2006, Senegal 1988, and Uganda 2002. The identifying assumption is that historical migration networks predict current migration rates and the subsequent inflow of remittances but do not directly affect a household's current investment decisions except through migration and remittances.

We find that remittances positively affect human, physical, and social capital investment in most sample countries. Our findings are consistent with past studies that find positive effects of remittances on human capital (Amuedo-Dorantes and Pozo, 2010), physical capital (Jena, 2018), and social capital (Fransen, 2015). However, we find that remittances reduce health and social capital investment in Nigeria and physical capital investment in Burkina Faso.

We check the robustness of our main results by using different specifications, different definitions of our key explanatory and outcome variables, and relaxing the exclusion restriction assumption. First, we implement Nevo and Rosen (2012)'s imperfect instrumental variables (IIV) approach that relaxes the exogeneity assumption by allowing the instrument to be correlated with the regression error term. The key assumption here is that the correlation between the instrument and error term is weaker than the correlation between the instrument and endogenous variable. Our results are largely robust to relaxing the exclusion restriction. Next, we consider a continuous treatment variable – the amount of cash remittances received by the household – and our main result persists. Finally, we used a continuous treatment (i.e., the cash amount of remittances) and a continuous outcome variable (i.e., the cash amount of investment expenditure), and the results are qualitatively similar to our main results. This suggests that our results are robust to relaxing the exclusion restriction, using different model specifications, and different definitions of the treatment and outcome variables.

We also explore the heterogeneity of our results across three different dimensions. First, we investigated how households with migrants and households without migrants differ in their investment decisions. Our results show that there is some variation in investment decisions between the two groups. However, our main results are driven by households with migrants. Next, we consider the effect of remittance sources on a household's investment decisions. Heterogeneity by remittance sources is important because the remittance literature points out that remittance sources contain critical information such as the relative size of remittances, migrant's control over the household's use of remittances, and transfer of values and norms. We find interesting patterns in investment by remittance sources: internal remittances matter more for education investment, within-Africa remittances are more likely to increase health investment, and out-of-Africa remittances are more likely to increase physical and social capital investment.

We also explore the potential mechanisms through which remittances affect households’ investment decisions. In particular, we examine the income effect, substitution effect, and migration expectations effect. We use consumption expenditure and asset ownership as proxies to measure the income effect. We find that the income effect of remittances mainly drives the positive effect on capital investment. We use the labor supply response of adult household members to capture the labor substitution effect, and we find evidence of lower labor supply only in Kenya and Senegal. Finally, we examine the migration expectations effect using the children's (i.e., aged 6–15 years) labor force participation and school attendance. Our results show that children in remittance-receiving households do not disproportionately drop out of school to join the labor force in most sample countries.

Our study contributes to the existing literature in the following ways. First, we contribute to the limited but growing literature on the impact of remittances on household investment decisions by providing empirical evidence for the understudied sub-Saharan Africa region. Second, our study uses data for five SSA countries and disaggregates investment expenditure into three categories – human capital, physical capital, and social capital investment. Our analysis allows us to compare the effect of remittances across countries using a standardized dataset. Past studies only study one type of capital investment and one country at a time. For instance, Jena (2018) and Ajefu (2018) studied only physical capital investment in Kenya and Nigeria, respectively. In addition, addressing the multiplicity of the investment alternatives allows us to explore the heterogeneity among the investment types and check for substitutability. Third, we identify the heterogeneous effect of remittances from domestic, within-Africa, and out-of-Africa sources on receiving households’ investment decisions; past studies mostly focus on only international remittances or only internal remittances, and a few explore internal and international remittances. Finally, we explore several mechanisms through which remittances affect investment decisions. Existing studies assume that the income effect is the main mechanism; however, they do not demonstrate this empirically.

Our study has important policy implications. First, we provide further evidence that remittances can contribute to economic development through productive investments. Thus policymakers in SSA can design policies aimed at reducing remittance transfer costs to harness remittances and foster local economic development. Our study is also relevant for the local and international organizations designing business models and financial instruments to maximize the impact of remittances on economic development. Understanding the heterogeneous effect of remittance sources will help these organizations design effective financial instruments to boost capital formation and income generation in the remittance-receiving communities. For instance, policymakers can imitate Kenya's M-PESA – a mobile banking service – to facilitate the transfer of internal remittances. This is important since internal remittances matter most for education investment in SSA.

The rest of the paper is organized as follows. In Section 2, we provide the conceptual framework that explains the linkage between remittances and capital investment. We describe the data and methodology in Sections 3 and 4, respectively. In Section 5, we discuss our main results. Next, we present the robustness checks in Section 6, heterogeneity analysis in Section 7, and effect mechanisms in Section 8. Finally, we conclude the paper in Section 9.

In the literature, economic migration decisions have been explained by the role of remittances. In these decisions, households send migrants to urban centers, or out of the country, with a desire to increase household income level and to diversify income sources (Adams, 1998; Clemens and Ogden, 2020; Rosenzweig and Stark, 1989; Stark, 1991). Theoretical models present different motives for sending remittances: altruism, insurance contract, loan contract, and investment or inheritance motive (Lucas and Stark, 1985). The altruism model posits that remittances are sent because migrants care about their left-behind family members (Lucas and Stark, 1985; Stark, 2009). The insurance contract model suggests that remittances result from an implicit contract between the households and migrants to protect the household against shocks (Cox, et al., 1998; Rosenzweig and Stark, 1989). The loan contract model argues that remittances are repayments for an informal loan taken out by the migrants from their families to enhance their human capital and finance the cost of migration (Poirine, 1997). The first three models – altruism, insurance contract, and loan contract – are silent about the investment use of remittances or assume that remittances are not invested. The investment or inheritance motive suggests that migrants send remittances because they aspire to inherit family property, intend to return home, and consider that left-behind family members are trustworthy agents to maintain assets on their behalf (Lucas and Stark, 1985; Taylor and Wyatt, 1996).

The migration literature further points out that the four remittance motives may not be mutually exclusive. It may be the case that remittances are sent for all the motives at the same time, with each motive comprising a share of it (Poirine, 1997). It could also be the case that one of these motives becomes dominant at different stages of migration. For instance, at the early stage of migration, remittances sent back are typically for loan repayments. However, regardless of the motive, remittances are expected to positively affect household income at home if migrants earn a substantially higher income in the destination country (Stark and Bloom, 1985).

Remittances affect household investment decisions through three main channels – income effect, substitution effect, and migration expectations (Amuedo-Dorantes, 2014). First, remittances help ease households’ resource constraints; the income effect of remittances reduces the need for households to send their children to join the labor force and enable households to pay tuition and other education-related expenses. Several studies in the literature have found a positive effect of remittances on education investments and the education outcomes of children left behind (Alcaraz et al., 2012; Amuedo-Dorantes and Pozo, 2010; Cox-Edwards and Ureta, 2003). Similarly, higher resource availability leads to better health outcomes of the household members through investment in improved lifestyle and spending more on health care (Ambrosius and Cuecuecha, 2013; Berloffa and Giunti, 2019; Hines and Simpson, 2018; Salas, 2014). Furthermore, the income effect of remittances positively affects physical capital investment through facilitating savings and improving access to the financial market (Amuedo-Dorantes and Pozo, 2014; Chiodi et al., 2012; Jena, 2018).

The income effect of remittances may also affect a household's spending on social events such as birthdays, wedding ceremonies, and funerals. In developing countries where social safety net programs are relatively weak and private insurance services are inaccessible, households rely on informal risk coping mechanisms to mitigate the impact of adverse economic shocks. Relying on relatives, friends, and community members is the most frequently used informal coping mechanism (Carter and Maluccio, 2003; Gerry and Li, 2010). Rao (2001) shows that spending on big social events generates tangible returns, such as paying a lower price for items in the local marketplace and achieving higher social status. A relatively stronger social network and social status signal the creditworthiness of the household and increase access to credit markets. Thus, investing in building social networks through spending on social events is a form of social capital as it not only hedges against future shocks but also generates other economic returns. However, remittances can act as a risk mitigation strategy, which may reduce the need to rely on social networks to cope with economic shocks. Consequently, the effect of remittances on social capital investment is ambiguous and can only be determined empirically.

Second, remittances can have a substitution effect because they may raise the reservation wage – the lowest wage at which a person is willing to work – of the left-behind household members and reduce the opportunity cost of leisure. Assuming leisure is a normal good, the substitution effect provides left-behind household members with incentives to lower labor supply. This phenomenon is also related to a moral hazard problem whereby left-behind household members are less inclined to engage in income-generating activities, which eventually leads to dependency on remittances (Amuedo-Dorantes, 2014; Démurger, 2015). The lowering of labor supply in response to receiving remittances is well documented in the literature (Amuedo-Dorantes and Pozo, 2006; Binzel and Assaad, 2011; Mendola and Carletto, 2012). In addition, remittance dependence due to the substitution effect of remittances may reduce the likelihood of capital investments as some physical assets (e.g., tractors) need to be combined with labor to be productive.

Finally, remittances can affect households’ investment decisions through migration expectations. Left-behind household members in a remittance-receiving household may have high expectations of migration and be reluctant to engage in income-generating activities. Moreover, migration expectations of left-behind household members can negatively affect human capital formation. Children may drop out of schools if they perceive lower returns to education in their destination country. For instance, Mckenzie and Rapoport (2011) found that boys in migrant households in Mexico are less likely to complete junior secondary school due to migration expectations and lower returns of Mexican education in the US labor market, especially in the context of illegal migration.

Theoretically, it is difficult to unambiguously determine the effect of remittances on household investment decisions. Thus, we set out to empirically investigate this phenomenon. Focusing on three investment categories – human capital, physical capital, and social capital – also helps us understand whether household investment decisions vary by type of investment expenditure.

We used data from the Migration and Remittances Household Surveys conducted by the World Bank between 2009 and 2010. These household surveys are part of the AMP and are designed to provide information about the volume, causes, and impacts of migration and remittances in sub-Saharan Africa (Plaza et al., 2011). An important feature of the surveys is that they are standardized across countries, which allows for easy aggregation and comparison. We use data on five predominantly remittance-receiving countries from the AMP, namely Burkina Faso, Kenya, Nigeria, Senegal, and Uganda. The surveys are cross-sectional and provide comprehensive information about migration, remittances, housing conditions, household assets and expenditures, and other socioeconomic and demographic characteristics. The surveys contain information about households with no migrants, internal (domestic) migrants, within-Africa migrants, and out-of-Africa migrants, which we use to create a variable representing the source of remittances. The principal respondent to the survey was the household head or their representative, who reported information about the migrant(s).

We define investment expenditure as an outlay for which the individual or household expects financial returns in the future. Following Jena (2018), we define physical capital investment as households’ expenditure on setting up a business, opening a store, purchasing farming equipment such as tractors, and purchasing other productive assets. Human capital investments are broadly defined to have two components – expenditure on education and health. Education expenditures include tuition payment, purchase of school uniforms, and books. In many SSA countries, the public education system is subsidized; for instance, Nigeria's Universal Basic Education (UBE) up to Junior High School. Such public programs imply that there is little need to spend on tuition fees. However, despite subsidized public education, households still incur educational expenses such as buying textbooks, uniforms, and after-school lessons. Health care expenditures include doctor fees, hospital fees, and cost of diagnostics tests and medicine. Like education, the public health system in many SSA countries is highly subsidized. Despite subsidized health systems, not all services are available in public facilities, and not all services are free. Out-of-pocket expenditures on diagnostic tests and medicine comprise a substantial share of health expenditure in most SSA countries. Finally, following Rao (2001), we define social capital investments as households’ expenditure on festivals, weddings, and funerals.

Households report the investment expenditures made during the last 6 months before the interview date. At the extensive margin, we use dummy indicators capturing whether the household made an investment expenditure or not in the preceding 6 months. Along with this extensive margin analysis, we conduct an intensive margin analysis using the actual amount of investments.

The control variables are household head characteristics such as gender of household head, whether the household head is a paid employee, whether the household head is self-employed, whether the household head has secondary education, whether the household head has above secondary education, whether the household head is aged 45–60 years old, whether the household head is above 60 years old, and socioeconomic characteristics of the household, such as number of children, number of elderly, and location of the household. We also control for the overall resource availability to the household by including per capita income. Per capita income is proxied by per capita expenditure following the standard practice in the literature as income data often suffer from measurement errors (Deaton, 2018; Jena, 2018).

Table 1 presents summary statistics of the outcome and control variables for the countries analyzed. Considerable variation exists between remittance-receiving and non–remittance-receiving households across all the countries in our sample. The first noticeable factor in Table 1 is that remittance-receiving households, on average, are more likely to be female headed than non–remittance-receiving households. Another interesting observation is that household heads in remittance-receiving households are on average less likely to hold paid employment or self-employment compared to non–remittance-receiving households in all the countries in our sample. Furthermore, for all the investment categories considered, remittance-receiving households, on average, spend more than non–remittance-receiving households. Kenya, on average, receives the highest amount of remittances, followed by Senegal and then Nigeria. Conversely, Burkina Faso receives the smallest remittances on average. Table A1 in Appendix presents correlation coefficients between the treatment and outcome variables.

We model households’ investment decisions as a function of their remittance-receipt status and a vector of other explanatory variables. Adams (2011) and many others have noted that empirical analyses of migration and remittances have failed to provide needed insights because of various econometric issues. One such issue is endogeneity, which can arise from selection bias and simultaneity. First, migrants are a self-selected group as migration and remittance transfers are not random events. Remittance-receiving households might differ systematically from non–remittance-receiving households in unobservable characteristics, such as migration aspirations, entrepreneurial ambitions, level of altruism, and household-specific norms. Given that these characteristics are unobservable, estimating a regression model without properly accounting for them may lead to the classic omitted variables bias. Next, simultaneity may arise from the reason for sending the remittances. For example, it could be the case that the migrant sends remittances to take advantage of a business opportunity in the home community. In this case, remittances did not lead to investment expenditures; instead, the migrant's desire to invest led to the transfer of remittances. Thus, researchers need to address endogeneity issues carefully to obtain unbiased estimates.

Since the investment decision and the remittance receipt are binary variables in our main estimation and the latter is likely to be endogenous, we employ a recursive bivariate probit model (Bhattacharya et al., 2006; Horrace and Oaxaca, 2006; Jena, 2018; Wooldridge, 2002). The recursive bivariate probit model accounts for endogeneity by simultaneously estimating remittance receipt and investment decisions while incorporating the remittance-receipt variable in the investment decision equation. We estimate the following recursive bivariate probit model: (1) Ri1*=Xi′β1+εi1 R_{i1}^* = X_i^\prime{\beta _1} + {\varepsilon _{i1}} (2) Yi1*=Ri1δ1+Zi′β2+εi2 Y_{i1}^* = {R_{i1}}{\delta _1} + Z_i^\prime{\beta _2} + {\varepsilon _{i2}} and E[εi1|X,Z]=E[εi2|X,Z]=0.Var[εi1|X,Z]=Var[εi2|X,Z]=1Cov[εi1,εi2|X,Z]=ρ \matrix{ {E\left[ {{\varepsilon _{i1}}|X,Z} \right] = E\left[ {{\varepsilon _{i2}}|X,Z} \right] = 0.} \hfill \cr {Var\left[ {{\varepsilon _{i1}}|X,Z} \right] = Var\left[ {{\varepsilon _{i2}}|X,Z} \right] = 1} \hfill \cr {Cov\left[ {{\varepsilon _{i1}},{\varepsilon _{i2}}|X,Z} \right] = \rho } \hfill \cr } where Ri1* R_{i1}^* and Yi1* Y_{i1}^* are latent dependent variables that determine the propensity of remittance receipt and the propensity to make an investment expenditure by the household, respectively. Xi′ X_i^\prime and Zi′ Z_i^\prime are vectors of covariates, and ɛ_i1 and ɛ_i2 are unobservable error terms and are assumed to be correlated. The correlation between the remittance-receipt equation and investment decision equation is ρ. We let two observable indicator variables represent the latent variables Ri1* R_{i1}^* and Yi1* Y_{i1}^* such that: (3) Ri1={1 if Ri1*>00 if Ri1*≤0} {R_{i1}} = \left\{ {\matrix{ {1\,if\,R_{i1}^* > 0} \cr {0\,if\,R_{i1}^* \le 0} \cr } } \right\} (4) Yi1={1 if Yi1*>00 if Yi1*≤0} {Y_{i1}} = \left\{ {\matrix{ {1\,if\,Y_{i1}^* > 0} \cr {0\,if\,Y_{i1}^* \le 0} \cr } } \right\} where R_i1 indicates the remittance-receipt status of the household, and Y_i1 captures the households’ investment decision. This study aims to empirically obtain estimates for the parameter δ₁ in Eq. (2), the parameter corresponding to the endogenous variable, R_i1.

Based on Eqs (3) and (4), the four basic probabilities of the bivariate probit model are: Prob[Ri1=1, Yi1=1]=F[Xi′β1,Zi′β2+δ1;ρ]Prob[Ri1=1, Yi1=0]=F[Xi′β1,−Zi′β2+δ1;−ρ]Prob[Ri1=0, Yi1=1]=F[−Xi′β1,Zi′β2;−ρ]Prob[Ri1=0, Yi1=0]=F[−Xi′β1,−Zi′β2;ρ] \matrix{ {{Prob}\left[ {{R_{i1}} = 1,\,{Y_{i1}} = 1} \right] = F\left[ {X_i^\prime{\beta _1},Z_i^\prime{\beta _2} + {\delta _1};\rho } \right]} \hfill \cr {{Prob}\left[ {{R_{i1}} = 1,\,{Y_{i1}} = 0} \right] = F\left[ {X_i^\prime{\beta _1}, - Z_i^\prime{\beta _2} + {\delta _1}; - \rho } \right]} \hfill \cr {{Prob}\left[ {{R_{i1}} = 0,\,{Y_{i1}} = 1} \right] = F\left[ { - X_i^\prime{\beta _1},Z_i^\prime{\beta _2}; - \rho } \right]} \hfill \cr {{Prob}\left[ {{R_{i1}} = 0,\,{Y_{i1}} = 0} \right] = F\left[ { - X_i^\prime{\beta _1}, - Z_i^\prime{\beta _2};\rho } \right]} \hfill \cr } where F[·] indicates the distribution function of the bivariate normal distribution with correlation parameter ρ.

The identification of the recursive bivariate probit model parameters requires at least one variable (i.e., instrumental variable) in the remittance-receipt equation (i.e., Eq. (1)) that is excluded from the investment decision equation (i.e., Eq. (2)). A credible instrument should be strongly correlated with the endogenous regressor of interest (i.e., receipt of remittances in our case) but uncorrelated with the outcome of interest (i.e., investment decisions). While the first condition can be easily tested, the second condition is practically untestable. Consequently, it is difficult to find credible instruments, and only a few instruments are generally acceptable in the migration literature. We use historical migration networks as an instrument for remittances as they are one of the generally acceptable instruments that have been widely used in the migration literature (Acosta, 2011; Alcaraz et al., 2012; Calero et al., 2009; Coon, 2016; Hildebrandt et al., 2005; Mckenzie and Rapoport, 2011).

The argument for using historical migration networks as an instrument is that such networks can reduce the cost of migration and induce current migration by providing access to information and facilitating services at the destination (i.e., assistance with accommodation and employment opportunities). The identifying assumption is that historical migration networks predict current migration rates and the subsequent inflow of remittances but do not directly affect a household's current investment decisions except through migration and remittances. Households with more extensive migration networks are expected to have lower migration costs, which increases their likelihood of having a migrant member and receiving remittances (Coon, 2016).

We define historical migration networks as district-level historical migration rates. District-level historical migration rates are obtained from Population and Housing censuses – Burkina Faso 1996, Kenya 1999, Nigeria 2006, Senegal 1988, and Uganda 2002.

Data source: Minnesota Population Center. Integrated Public Use Microdata Series, International: Version 7.3 [dataset]. Minneapolis, MN: IPUMS, 2020. https://doi.org/10.18128/D020.V7.2

We define a district as a second-tier administrative unit within a country. This refers to a district in Uganda and Kenya, local government in Nigeria, and department in Burkina Faso and Senegal. Districts cover large geographic areas and populations, making it difficult for households to affect the migration networks in any significant way. The domestic migration network is about 0.1% in Kenya, Nigeria, and Burkina Faso, which suggests that, on average, one in 1,000 households have a domestic migrant (see Table 1). The domestic migration network in Uganda is negative, which suggests that, on average, districts experience more in-migration than out-migration. We have international migration network data for Kenya, Burkina Faso, and Senegal. Of these countries, Burkina Faso has the highest international migration network – about three in 1,000 households. In our estimation, we use both domestic and international migration networks in Kenya and Burkina Faso. However, because data is unavailable elsewhere, we use only domestic migration networks in Uganda and Nigeria and only international migration networks in Senegal. We argue that either historical domestic or international migration rates capture the overall migration network in the district.

The first stage regression shows that the migration network is statistically significant at less than 5% significance level (see Table 2). The F-statistic in the first stage is higher than 10 for all the countries under review except Uganda. However, it is close to 10 for Uganda. We check for overidentifying restrictions on our instruments using the Hansen's J-statistic. The joint null hypothesis states that the instruments are valid, and rejecting the null hypothesis implies that at least one of the instruments is not valid. In our case, we cannot reject the null hypothesis for any of the countries because the p-values are higher than the traditional significance levels.

A potential threat to identification using district-level historical migration networks as an instrument is that previous remittance flows, return migration, and the transfer of knowledge via migration may be correlated with district-level factors such as education facilities, health facilities, and a better investment climate. Consequently, historical migration networks could be correlated with the current level of infrastructure in a district. One way to account for this violation of the exclusion restriction is to control for district-level variation in infrastructure. However, it is difficult to find data on infrastructure at the district level in SSA. Consequently, we implement the IIV approach introduced by Nevo and Rosen (2012) as a robustness check. Nevo and Rosen (2012)'s IIV approach relaxes the exogeneity assumption by allowing the instrument to be correlated with the regression error term. The key assumption here is that the correlation between the instrument and error term is weaker than the correlation between the instrument and endogenous variable. The IIV approach produces bound estimates of the endogenous parameter of interest rather than a point estimate.

We present our main estimation results in Table 3. The treatment variable is “received remittances,” which is an indicator variable that takes one if a household received remittances in the 12 months before the survey, and zero otherwise. The outcome variables are investment decision indicators that equal one if a household made a capital investment in the 6 months before the survey and zero otherwise. Columns 1–4 show the naïve probit estimates, while columns 5–8 show the recursive bivariate probit estimates. Panels A–E present the results for Uganda, Kenya, Nigeria, Burkina Faso, and Senegal, respectively. All columns of Table 3 include control variables.

We present the average marginal effects from the probit and the recursive bivariate probit models in Table 3 and the estimated coefficients in Table A3 in Appendix. In addition, we present the robust standard errors in parentheses. Abadie et al. (2017) argue that when the treatment assignment is at the participant level, there is no need to cluster standard errors. From an experimental design perspective, our treatment assignment and our unit of analysis are at the household level, and so, we do not cluster our standard errors; rather, we present robust standard errors. Comparing the results of naïve probit and recursive bivariate probit estimates, we see that the naïve probit estimates are biased downwards. Downward bias in the probit estimation implies the presence of endogeneity due to reverse causality. Consequently, we focus on interpreting the recursive bivariate probit estimates.

Remittances can stimulate investment in income-generating activities by relaxing liquidity constraints in receiving households. However, remittance dependence and other unintended consequences can reduce investment in income-generating activities. In the context of SSA, we study whether the remittance-receiving households make any investment expenditures, and if they do, what kind of investments they make. With a few exceptions, we find that remittances increase investment in human, physical, and social capital in our sample countries. The income effect of remittances mainly drives this positive effect on investment. We do not find evidence of a strong income effect in Burkina Faso, and it is only in this country that we find remittances reduce the likelihood and amount of physical capital investment. This finding is not surprising as the average remittances received by the households in Burkina Faso are the lowest in our sample countries. Furthermore, we find evidence of substitution effect by the left-behind household members in Kenya and Senegal, but the effect is strong enough to influence investment decisions only in Senegal. Similarly, we find evidence of migration expectations only in Senegal, and remittances do not increase human capital expenditure there.

We also explore the heterogeneous effect of remittance sources on households’ investment decisions. We find some interesting patterns: internal remittances matter more for education investment, within-Africa remittances are more likely to increase health investment, and out-of-Africa remittances are more likely to increase physical and social capital investment. We argue that internal remittances are more likely to increase education investment because they create relatively lower migration expectations than within-Africa and out-of-Africa remittances. Similarly, out-of-Africa remittances are more likely to increase physical and social capital investment due to the relatively strong income effect generated from the larger remittances.

Our study has important policy implications for SSA's economic development. First, we provide further evidence that remittances can contribute to economic development through productive investments. Given that migrants send about 15% of their total income as remittances, there is great potential to harness remittances by devising policies to reduce remittance transfer costs. This also coincides with the Sustainable Development Goal (SDG) 10.7.C (United Nations, 2015), which aims to reduce the cost of sending remittances to SSA to less than 3% by 2030 from the current 9% (World Bank, 2018). Our study is also relevant for the local and international organizations designing business models and financial instruments to maximize the impact of remittances on economic development. Understanding the heterogeneous effect of remittance sources will help these organizations to design effective financial instruments to boost capital formation and income generation in the remittance-receiving communities.

Furthermore, our study highlights the importance of social capital investment, which suggests that researchers and policymakers should devote more attention to this investment type. Policymakers seeking to boost human and physical capital investments should also focus on social capital investment decisions.

Although highly complementary to the existing literature, our findings must be evaluated against the fact that our analysis is not free from limitations. We use cross-sectional data, which makes us unable to follow the same household over time. Given the rising importance of remittances, a multi-country longitudinal study is required to generate deeper knowledge for policy action.