Do international remittances promote poverty alleviation? Evidence from low- and middle-income countries

This study uses annual macro-level data for 65 low- and middle-income countries (LMICs) worldwide between 2002 and 2016 to estimate the impacts of international remittances on poverty (see Table A1 in Appendix for the country list understudy).

For the current 2021 fiscal year, low- and middle-income countries consist of low-income countries (≤1.045 USD in 2020), upper-middle-income countries (1.046 – 4.095 USD), and middle-income countries and above (4,096 – 12,695 USD). This classification is based on the 2020 gross national income (GNI) per capita calculated using the World Bank Atlas Method (World Bank, 2021b).

As proposed by Foster et al. (1984), poverty headcount (P₀) and poverty gap (P₁) can be decomposed by a single equation as follows: (1) Pα=1n∑i=1np(Giz)α {P_\alpha} = {1 \over n}\sum\nolimits_{i = 1}^{{n_p}} {{{\left({{{{G_i}} \over z}} \right)}^\alpha}} where P stands for poverty and α represents the level of sensitivity of the index in measuring poverty. G_i is the income shortfall of household i in relation to the poverty line (z). Finally, n and n_p are the total number of households and the number of the poor, respectively. Eq. (1) measures the percentage of people living under the poverty line when α = 0.

P₀ = n_p/n (poverty headcount).

For the remittance, we follow Gupta et al. (2009) and Akpan et al. (2014) to use the proportion that international remittances contribute to gross domestic product (GDP) (R, %). This variable of interest consists of two components that are respectively recorded as primary and secondary income in the balance of payment, namely, personal transfers (workers’ remittance) and compensation of employees (IMF, 2009). Chami et al. (2009) commented on such aggregation. They argued that measuring international remittances by totaling those two components is inappropriate because each type of transfer has different characteristics and responds differently to economic turmoil. According to the IMF (2009), the difference between the two types of transfer lies in how long migrants who send money to households/relatives stay in their respective host countries. However, Martin (2015) argued that most nations do not know precisely how long migrants who remit money have been abroad, and thus most researchers use that aggregation when studying remittances. We prefer this second argument as a precaution against obtaining invalid remittance measurements.

The controlled causal hypothesis is crucial to ensure our analysis leads to the proper conclusion. All variables except those investigated (in this case, it is remittances) must remain constant (or “controlled”) in all experimental conditions (Schwichow et al., 2016). Hence, we include in the analysis several exogenous regressors that may affect poverty, as suggested by Anyanwu and Erhijakpor (2010): per capita GDP (IC), Gini ratio (GN), and inflation (INF). Per capita income is used to measure the level of a country's economy. The Gini index is used to portray a country's income gap. Inflation is included to account for the purchasing power of the individual country's currency. Table A2 in Appendix provides the definitions of proxies, alongside each one's unit of measurement, expected sign, data source, and summary statistics of all variables.

To achieve the objective of this study, we regress all poverty indicators mentioned earlier against international remittance and all selected control variables. Under the assumption that the unobservable individual country effects are correlated with the explanatory variable of interest R_it, cov(α_i, R_it) ≠ 0, we posit that the fixed-effects (FE hereafter) model can yield more consistent coefficient estimates than the ordinary least squares (OLS hereafter). The first model of this study can be written as follows: (2) Pαit=di+ϕRit+βXit+εit=δZit+di+εit(i=1,…, N;t=1,…, Ti) \matrix{{{P_{\alpha it}} = {d_i} + {\boldsymbol{\gamma}}{{\boldsymbol P}_{\alpha it - 1}} + \phi {\boldsymbol{R}_{it}} + {\boldsymbol{\theta}}{{\boldsymbol{R}}_{it - 1}} + \beta {{\rm{X}}_{it}} + {\varepsilon _{it}}} \hfill \cr {\left({i = 1, \ldots,\,\,N;t = 1, \ldots,\,{T_i}} \right)} \hfill \cr} where P_αit is a vector of all poverty measurements for country i at year t, R_it is the remittance received by the country as a share of the GDP, X_it is a vector of the control variables, d_i is the country FE, which considers unobserved time-invariant (country-specific) characteristics in a flexible manner.

Specifically, the country and time fixed effects can be defined as a_i = f(U_i), where U_i represents this country-specific unobserved confounder, a common cause of the outcome and treatment variables (Imai and Kim, 2019).

With respect to the empirical point of view revealed by previous studies, we expect that international remittances mitigate poverty in remittance-receiving countries. Following the results of previous studies, such as by Michálek and Výbošťok (2018) and Sehrawat and Giri (2018), we expect that per capita GDP has a negative effect on poverty, while the Gini index has a positive effect. We argue that increases in the revenue of the community imply narrowing shortfall from the poverty line, and it is associated with alleviating poverty. On the other hand, the income gap exacerbates poverty because only a few people enjoy economic growth. At the same time, the lowest layer of society remains poor or even becomes worse. By adapting mixed empirical results from de Hoyos and Medvedev (2009) and Fujii (2013), inflation is estimated to have a positive or negative effect on poverty. We believe that if price increases can stimulate people's income, inflation can improve poverty alleviation. However, inflation can exacerbate poverty if price increases only worsen people's purchasing power rather than increase income.

It should be noted that Eq. (2) assumes that the impact of remittances as a share of GDP against all poverty measurements is a causal relationship. However, the relationship between remittances and poverty may be bidirectional, which causes remittance variables to be endogenous. According to Azizi (2019), the endogeneity of remittances in relation to poverty can be explained via micro and macro perspectives and also measurement error terms as follows:

In the micro perspective, the household members who live in poverty may decide to migrate and remit their money to help their relatives cope with financial constraints.

These three situations lead to the biasness of the parameters in Eq. (2). To address these issues, we apply the 2SLS method. As the first stage, we perform regression of the auxiliary regression of remittances against the exogenous regressors in Eq. (2) and the instrumental variables (IVs). We use trade openness and exchange rate of remittance-receiving countries as instruments for the remittance variables (see Table A2 in Appendix for detailed description).

As with Gupta et al. (2009), trade openness is measured as a share of total trade (exports and imports) to GDP. We argue that economic liberalization describes the freedom of economic activity in the output and input markets, which involves exchanging labor between countries. Thus, we expect that trade openness has a positive effect on remittances. In contrast to Azizi (2019), we use remittance-receiving countries’ currency against US dollars to express the exchange rate variable. An appreciation of the domestic currency can reduce the remittance ratio because it represents a cost for the remitter (Abdih et al., 2012). Thus, this variable is expected to have a positive effect on remittance.

On the other hand, this study argues that these instruments do not affect poverty in ways other than through the remittance variable. For example, the exchange rate will not affect the income of the poor unless one of their income sources relates to the exchange rate volatility. This income must come from abroad, such as from remittances. This argument is reasonable considering that poverty used in this study is measured by income level rather than consumption level. Therefore, concerns about the influence of the exchange rate on poverty through the consumption of imported goods can be ignored. On the other hand, trade openness will not affect the income of the poor unless it is associated with the easing of immigration policies, corresponding to the migrants’ remittances sent to the needy in society. To ensure that the instruments are strong and valid, this study undertakes several tests such as under- and overidentification.

The second set of models explores the dynamic relationship between remittances and poverty. Given that the level of poverty tends to be highly persistent over time, it is reasonable to assume that current poverty depends on past poverty levels and remittances. Therefore, we estimate the variance of the poverty level with a dynamic panel model by introducing the lagged levels of poverty and remittance on the right side of the equation. The following equation captures that dynamic: (3) Pαit=di+γPαit−1+ϕRit+θRit−1+βXit+εit(i=1,…, N;t=1,…, Ti) {P_1} = {P_0} \cdot \left({\sum\limits_{i = 1}^{{n_p}} {{G_i}/z}} \right)\left({{\rm{poverty}}\,{\rm{gap}}} \right).

The main difference between Eqs. (2) and (3) is that the latter captures both the contemporary effect and the lagged effect of remittances on poverty. As mentioned earlier, estimating this dynamic model allows us to perform tests of different causal mechanisms: an increase in the income of recipient households and nonrecipients in the local community. However, it should be noted that, Eq. (3) can still produce biased and inconsistent parameters because heteroskedasticity in residuals and autocorrelation within panels (countries) always appear in data involving many units across places. In addition, the bias parameter is also attributed to the potential endogeneity of the lagged poverty level (P_αit−1) in a dynamic panel model – when this variable correlates with the random error term of the equation. That there is potential for this problem to arise is very reasonable, considering that we cannot fully capture the determinants of poverty.

To address these potential issues, we explore Model (3) using the dynamic SGMM estimators proposed by Arellano and Bover (1995) and Blundell and Bond (1998). In contrast to the Difference-GMM (DGMM hereafter) proposed by Arellano and Bond (1991), the SGMM estimation corrects the endogeneity problem by introducing more instruments, thereby dramatically increasing the estimator's efficiency. Another advantage of using this model is that it allows us to minimize data loss better than the DGMM. This is because, instead of subtracting the previous observation from a contemporaneous one, the SGMM subtracts the average of all future available observations of a variable. It implies that no matter how many gaps we have in our unbalanced panel data set, such a model is computable for all observations except the last for each individual (country). In addition, the Monte Carlo simulation also suggests that when the time span is short and the dependent variable (DV) is persistent, there are gains in precision and the small-sample bias is reduced when the SGMM is applied (Blundell and Bond, 1998).

To test the validity of the instrument, we run the Hansen test and the difference-in- Hansen test. The null hypothesis for the first test states that all instruments used are exogenous (orthogonal to DVs). The null hypothesis for the second test confirms the exogeneity of external instruments (consists of key explanatories and control variables) in the SGMM estimation. However, Roodman (2009) stated that the p-value of those tests might be bloated, primarily when the instruments used to overcome the endogeneity problems outnumber the country panels. Due to the relatively large number of periods under study (t = 15), the SGMM model that we have built is likely to face the instrument proliferation problem, especially when all lags are exploited as instruments. Therefore, the GMM-style instrument lag needs to be restricted to two to prevent overuse of the instruments in the model, as Roodman (2009) suggested. We treat P_αit−1, R_it, and R_it−1 as endogenous and generate the GMM-style instruments for the corresponding endogenous variables.

Table 1 displays the initial results of this study. Generally, we find that “international remittances per GDP” negatively affects all three poverty indicators. Models 1–3 use the pooled OLS regression, while Models 4–6 include country FEs. The Hausman test results suggest that the null thesis of no FE is rejected even at the 1% significance level. It indicates that the pooled OLS models likely produce inconsistent coefficient estimates for our baseline equation. Hence, Eq. (2) should be estimated by FE (within) models. Therefore, in further discussions of our regression analysis, we only dwell on the FE estimation results. In Models 7–9, we use trade openness and exchange rate in remittance-receiving countries to instrument international remittances. We find that on average, a 10-percentage-point increase in international remittances per GDP will lead to a similar decrease in poverty headcount at USD 1.90 a day, a 4.8-percentage-point decline in the poverty gap at USD 1.90 a day, and a 6.7-percentage-point decrease in the poverty gap at USD 3.20 a day. In addition, all control variables appear with the expected sign across all models.

In Models 7–9, the F-statistics of excluded instruments (TOP and log_EX) emerge with the p-values lower than the 1% significance level, indicating that they jointly have substantial impacts on international remittances in LMICs. Trade openness and exchange rate are particularly strong predictors of international remittances (see Table A3 in Appendix). In addition, the results of the underidentification test also reject the null hypothesis of the weak joint instruments (p-value of Kleibergen–Paap lagrange multiplier (LM) statistics < 0.01). Statistically, the test results provide strong evidence that the two variables are relevant instruments for international remittances. The overidentification fails to reject the null hypothesis of the joint exogeneity of the instruments (p-value of Hansen J statistics >0.1). It implies that our instruments are valid because they are orthogonal to the outcome of interest (P_αit). It also complements the underidentification test results; thus, the two attributes of “good instruments” (relevant and valid) are met. Finally, the endogeneity test rejects the null hypothesis of the exogeneity of international remittances (p-value of chi²-statistics <0.05). It indicates that the troublesome regressor of this study (R) should be treated as endogenous; hence, incorporating instruments for this variable is essential.

To check the robustness of the regression results shown in Models 7–9 in Table 1, we add one political variable as an exogenous regressor for poverty, obtained from the World Governance Indicators (WGI) provided by the World Bank, namely, political stability and absence of violence/terrorism (POL) (see Table 2). This variable measures perceptions related to the possibility of political instability and politically motivated violence, including terrorism (Kaufmann et al., 2010). This political factor has a negative influence on all poverty measurements in this study. Qualitatively, the presence of political variables in the FE- IV model does not change the significance and sign of the coefficient of remittances to poverty. However, quantitatively, the magnitude of the parameter estimate is slightly smaller; they are 1.005, 0.471, and 0.661, respectively. Compare this with remittances in FE-IV that do not involve political factors, where the sizes of the effects are 1.012, 0.475, and 0.665, respectively (see Models 7–9 in Table 1). To put it differently, our FE-IV models remain robust with the inclusion of political factors into the analysis.

Table 3 summarizes the results from the SGMM estimations, which now estimate the contemporaneous and lagged effects of international remittances on the poverty level in a dynamic setting. The Hansen J statistics fail to reject the null of the joint exogeneity of all instruments across specifications 1–6 (p-value > 0.1), which brings some confidence that our instruments are jointly valid (p-value >0.1). The difference-in-Hansen test also indicates that the GMM-type internal instruments are also valid (p-value > 0.1). Moreover, the Arellano–Bond test for autoregressive (AR) (2) across all models failed to reject the null hypothesis of no second-order serial correlation (p-value > 0.1). This implies that original disturbances are serially uncorrelated, and the moment conditions are correctly specified, making our estimations safe from bias.

Other key findings of this study emerge in Table 3. The contemporaneous impact of international remittances is still consistent in promoting poverty alleviation (see Models 1–3). It is indicated by the negative and significant effects of remittance on poverty headcount ratio at USD 1.90 a day (p-value < 0.1) and poverty gap ratio at USD 3.20 a day (p-value < 0.1), but with insignificant effect on poverty gap ratio at USD 1.90 a day (p-value > 0.1). In addition, the effect sizes of remittances became larger and significant (p-value < 0.05) when political variables were included in the estimation models (see Models 4–6), but the effect remained insignificant on the poverty gap ratio at USD 1.90 a day (p-value > 0.1). These contemporaneous effects of international remittances are much larger than their lagged effects, which were found to be insignificant across poverty measures and specifications (p-value > 0.1).

As discussed earlier, the contemporaneous effects of remittances mainly capture their direct and immediate impact on increasing migrant families’ income, which is then configured to reduce poverty among migrant families. By contrast, the lagged effects of remittances capture the indirect effects of remittances on the income of other members of the community of origin through local market linkages (spillover mechanism). After some temporal lag, remittances may promote local development and alleviate poverty. Our findings seem to provide some support that a greater proportion of poverty reduction could be explained by the direct effect of international remittances on increasing the wealth index among recipient households.

In addition, the insignificant impact of the first lag of remittance variable on poverty is confusing. This indirect effect should be substantially greater than the direct effects on which researchers and policymakers usually focus (Taylor and Dyer, 2009). However, these results can still be justified at the practical level. As Barham and Boucher (1998) and Brown and Jimenez (2008) note, the impact of migration on household poverty is seen in the net effect of a reduced labor supply in local communities and the incentive effect of receiving remittances. For example, Adaku (2013) has found in northern Ghana that labor loss due to migration tends to keep households poor. In addition, a net detrimental effect has been found by Tuladhar et al. (2014) in Nepal, where migration causes labor shortages while agricultural households receiving remittances do not invest in increasing agricultural productivity. On the other hand, the Food and Agriculture Organization (FAO) (2018) asserts that the main channel through which the effects of migration can spread to other members of the community of origin are the dynamic effects resulting from investment and the response in terms of labor supply and demand.

An insignificant spillover effect of remittance was found in this study probably because those main channels do not work optimally. Many local community workers are lost, and recipient families have no incentive to invest. This argument is reinforced by the results of the Migrating out of Poverty (MOOP) Consortium survey, which revealed that investment spending in agriculture by migrant households in most developing countries only represents a small part of the total remittance use, while the largest share of remittances (30%–40%) is dedicated to daily consumption (Poggi, 2018). However, the negative effect of outmigration on labor availability can be offset by reinvestment of remittances (Huy and Nonneman, 2016). Perhaps this is one of the reasons why the effect of the lagged remittance variable coefficient is left negative even though the impact is not significant.