The aim of this study is to investigate empirically the effect of agricultural aid on poverty levels in countries in sub-Saharan Africa. We test the hypothesis that aid improves agriculture productivity, which in turn contributes to poverty reduction. Two econometrics models are estimated. The first model is a linear panel model. The model is specified as follows:

$povert{y}_{i,t}={\alpha }_0+{\alpha }_{1}laidpw{o}_{i,t-1}+{\alpha }_{2}lagripro{d}_{i,t}+{\alpha }_{3}lrurpop{}_{i,t}+{\alpha }_{4}lgovex{p}_{i,t}+{\alpha }_{5}polsta{b}_{i,t}+{\mu }_{i,t}$ (3.1)

where poverty is the dependent variable, measured by the headcount ratio obtained from the Povcalnet database. The explanatory variables are:

The second model is a simultaneous equations model, expressed as follows:

$\{\begin{array}{l}povert{y}_{it}={\alpha }_{1}laidpw{o}_{i,t-1}+{\alpha }_{2}lagripro{d}_{i,t}+{\theta }^{\prime }{X}_{it}+{\epsilon }_{it}\left(3.2\right)\\ laidpw{o}_{it}={\gamma }_{1}lagripro{d}_{i,t}+{\gamma }^{\prime }{X}_{it}+{\delta }_{it}\left(3.3\right)\\ lagripro{d}_{it}={\beta }_{1}laidpw{o}_{i,t-1}+{\gamma }^{\prime }{Z}_{it}+{\mu }_{it}\left(3.4\right)\end{array}$

The first equation (3.2) explains poverty levels. The main explanatory variables in this equation are agricultural productivity, measured by the value added per worker in agriculture, and agriculture aid per worker, one-year lagged. Data are in constant 2010 U.S. dollars. The control variables ( $X_{it})$ are the real GDP per capita, to control for the level of economic development among countries in our sample, one-year lagged; the rural population as a percentage of total population, given that the majority of the poor are in rural areas; and also as a proxy of employment levels., government expenditures and political stability, as an indicator of governance. The error term of the first equation is ${\delta }_{it}.$

The second equation (3.3) attempts to explain the determinant of aid to African countries. The explanatory variables are the real GDP per capita; the indicator of agricultural productivity, infant mortality as a measure of human development levels, political stability. The error term is ${\delta }_{it}.$ The third equation (3.4) explains agricultural productivity. We are interested in assessing the effect of agricultural aid per worker. $Z_{it}$ is a set of control variables, including the real GDP per capita (as in the first equation), government expenditures, rainfall, to capture the effect of climate change on productivity, arable land as a percentage of territory, political stability, as an indicator of governance. The error term in this equation is $u_{it}$.

The econometric analysis comprises three steps. A first specification is made through a linear panel model estimated by fixed effects techniques. In this model, an interactive variable between aid and agricultural productivity is introduced to capture their combined effect on poverty levels.

A second estimation is made from a system of equations. In this specification, agricultural productivity, foreign aid, and poverty are considered as endogenous. As a consistent estimation of the parameters requires an estimation method that can deal with the endogeneity problem, we use the three-stage least squares (3SLS) method, which is more efficient than a two-stage least squares (2SLS) estimation ( Wooldridge, 2010 ). The 3SLS estimator decomposes reverse causality, controls for endogeneity, takes the disturbance between residuals in different equations into account and provides the possibility of incorporating other transmission channels within a simultaneous framework. The first two stages of the 3SLS estimation, which are equivalent to a 2SLS estimation, correct the bias in coefficients arising from reverse causality. The third stage improves the estimated standard errors of the coefficients by controlling for the correlation of errors across equations ( Kaya et al., 2013 ). Before considering the method of estimation, the identifiability of the model was checked because estimation methods that are used for SEM are functions of identification criteria. For an equation in a system of equations to be identified, the number of excluded exogenous variables in that equation must be at least as great as the number of included endogenous variables, less one. In our case, each equation is over-identified. In estimating the equations, we control for unobserved time-invariant variables and unobserved time effects by including N − 1 country dummies and T − 1 time dummies.

Furthermore, the Dumitrescu and Hurlin (2011) panel causality analysis is used to analyze the causal relationship between the three main variables, namely the poverty headcount ratio, foreign aid to agriculture and agricultural productivity. This method is well-suited for our panel data structure as it accounts for heterogeneity in causal relationships across countries, improves statistical power by pooling cross-sectional information, and allows us to explore bidirectional causality. Indeed, countries differ in how aid is used, how productive agriculture is, and how poverty responds. The test accounts for these heterogeneous dynamics. The panel fixed effects and simultaneous equations models do not formally test the direction of causality. To complement these approaches, the Dumitrescu and Hurlin (2011) panel causality test is employed to determine whether past values of one variable help predict another. This multi-method strategy ensures robust and nuanced inference on the dynamics between aid, productivity, and poverty reduction. It helps confirm or challenge the assumed direction of causality in the simultaneous system. The simple model (3.5) with two variables constitutes the basic framework for studying Granger causality in a panel data context

$y_{i,t}={\alpha }_i+\underset{k=1}{\overset{K}{{\displaystyle \sum }}}{\gamma }_i^{\left(k\right)}{y}_{i,t-k}+\underset{k=1}{\overset{K}{{\displaystyle \sum }}}{\beta }_i^{\left(k\right)}{x}_{i,t-k}+{\epsilon }_{i,t}$ (3.5)

With $K\in N^{*}$and ${\beta }_i={\left({\beta }_i^1,\cdots ,{\beta }_i^K\right)}^{\prime }$. For simplicity, the individual effects ${\alpha }_i$ are supposed to be fixed in the time dimension. Initial conditions ( $y_{i,-k},\cdots ,y_{i,0}$) and ( $x_{i,-k},\cdots ,x_{i,0}$) of both individual processes $y_{i,t}$ and $x_{i,t}$ are given and observable. We assume that lag orders $K$ are identical for all cross-section units of the panel and the panel is balanced. Besides, Dumitrescu and Hurlin (2011) allow the autoregressive parameters ${\gamma }_i^{\left(k\right)}$ and the regression coefficients slopes ${\beta }_i^{\left(k\right)}$ to differ across groups.

The DH test considers the HNC¹ null hypothesis, where no Granger-causal relationships are assumed to exist for any member $i$ of the panel. The DH test is based on an aggregated Wald statistic of individual Granger causality tests defined as:

$W_{N,T}^{Hnc}=\frac{1}{N}\underset{i=1}{\overset{N}{{\displaystyle \sum }}}{W}_{i,T}$,

Where $W_{i,T}$ denotes the individual Wald statistics for the $i^{th}$cross-section unit corresponding to the individual test $H_0:{\beta }_i=0$.

Using the required stationarity tests, properties like the presence of a unit root in the panel data were verified. Since the panel data contains a large number of cross-sections that are clustered together, a test of cross-sectional dependence was then carried out. When using first-generation unit root tests, an extreme assumption of cross-sectional independence is made. Consequently, a cross-dependency test was conducted using Pesaran’s CD test, which is the most often used test. If a cross-section dependence is revealed, second generation unit root tests should then be used to ascertain the stationarity levels of the variables before conducting the causality test.

The data for this study are drawn from various sources and cover the period from 2002 to 2019. The sample comprises 34 countries in sub-Saharan Africa. Although the choice of countries is governed by the availability of data, the included countries broadly cover the whole region. Table 3 shows summary statistics of the variables. The correlation matrix is found in the appendix (Table A1). The definition and measurement units can be found in the appendix (Table A2-A3).

Source: Author’s computation.

Figure 2 illustrates the relationship between agricultural aid and poverty levels in the countries included in the study. An overall negative association is evident across these variables, with significant variances observed among the selected nations. Countries like the Democratic Republic of Congo, Burundi, and Malawi

have high poverty rates and receive low levels of aid per person, which may be due to their large populations. Conversely, Mauritius, Cape Verde, and Seychelles exhibit lower poverty rates. In Seychelles, there is a noticeable correlation between decreased poverty rates and increased agricultural assistance. Figure 3 shows a negative relationship between agricultural value added per worker, used as a proxy for agricultural productivity, and poverty rates. Prominent examples include the Democratic Republic of Congo and the Central African Republic, known for high poverty rates and reduced agricultural productivity. And, Figure 4 shows a positive correlation between aid disbursement and agriculture total factor productivity index which can be also used to measure productivity in agriculture. However, these observed associations remain descriptive and do not provide a basis for causal inferences. We will be able to ascertain whether there is a causal relationship between these variables through the econometric study that follows.

Table 4 presents the results of the estimation of a poverty model using the fixed effects techniques on a panel of 32 in sub-Saharan Africa².

Notes: p-values in parentheses; ***p < 0.01, **p < 0.05, *p < 0.1.

All variables except for the political stability index are in the logarithm form.

In the first column, variables are all introduced into the model. The coefficients are all significant with expected signs, showing the reducing effect that aid to agriculture, agricultural productivity, government spending and the quality of institutions could have on poverty rates. In the second column, we remove the governance variable (political stability). This variable has been eliminated from the model due to the possibility that international aid could impact poverty by bolstering recipient nations' institutional frameworks. Therefore, we eliminate it to see if there is a change in the coefficient of international aid in order to quantify the ceteris paribus effect of aid on poverty.

Having done this, we also observe no change in the coefficients of the variables. Similarly, the results obtained in column 4 are without the government spending and political stability variables. Along with a small rise in the aid coefficient, we also see that the coefficients of the variables that were first introduced maintain their sign and significance. This suggests that development aid has little or no effect on poverty through budgetary contributions and institutional strengthening. This may be explained by the fact that we are using data for aid that has been earmarked for agriculture particularly rather than the entire amount of aid that is distributed across all sectors.

In general, we discover that agricultural aid slightly lowers the poverty rate. According to the results, the effect of aid to agriculture on the poverty rate varies between 0.006% and 0.008%. Specifically, a 1% increase in aid per worker is associated with an approximate 0.006 (column I and II), and 0.008 (column III) percentage point decrease in the poverty headcount ratio, holding other factors constant. An explanation to this small effect could be that agriculture has not been a top priority for ODA spending ( Eber et al., 2020 ). According to FAO data, since 2015, agricultural ODA has consistently comprised the smallest share of total ODA. In 2018, the relative share of ODA allocated to agriculture was 4.3%, the lowest share since 2006. Disbursements for humanitarian aid and health each amounted to more than three times the disbursements for agriculture in 2018, representing 13.9% and 13.3% of total ODA disbursements, respectively ( Eber et al., 2020 )

Continuing our analysis, we address potential endogeneity concerns stemming from the reciprocal relationship between poverty levels and aid allocation within a nation. Table 5 presents the simultaneous equation model (SEM) estimates using the Three-Stage Least Squares (3SLS) approach. The expected signs are observed for the coefficients of the majority of control variables. According to the poverty equation, agricultural assistance lowers poverty levels; this relationship is statistically significant and has a tiny but negative coefficient. A 1% increase in aid per worker leads to a decline in the poverty headcount ratio by 0.007%. Agricultural productivity measured by the value added per worker in the agricultural sector has a negative and significant coefficient. Also, per capita income is associated with diminished poverty levels. Conversely, a rise in the rural population exacerbates poverty levels across the countries.

In the aid equation, GDP per capita is always significant, indicating a negative relationship with development aid. Aid is generally allocated to developing countries with high levels of poverty and lower GDP per capita ( Mahembe & Odhiambo, 2019 ). Thus, as incomes appear to rise, aid volumes will have to fall. The productivity variable is positively signed and significant indicating that the more value added per worker increases, the more aid will go to the agricultural sector.

In relation to the productivity equation, it is notable that agricultural aid emerges as a contributing factor to the enhancement of productivity. Additionally, there exists a positive correlation between GDP per capita and agricultural productivity showing that when the wealth generated in the economy is equitably distributed, this fosters an improvement in producers’ incomes. Consequently, these producers are more inclined to invest substantially in their endeavors, thereby augmenting their agricultural yield. Furthermore, a positive relationship is observed between the proportion of irrigated area and agricultural productivity, indicating that an increased allocation of land for irrigation is associated with higher levels of productivity in the agricultural sector.

Notes: p-values in parentheses; ^*** p < 0.01, ^** p < 0.05, ^* p < 0.1. Country and time fixed effects are included in the regressions.

An increase in gross domestic product (GDP) per capita, reflecting a higher quality of living is associated with a rise in agricultural production. Higher income levels are linked to a rise in the availability of upgraded agricultural implements and the ability to pay for better working conditions.

The study emphasizes how agricultural production can effectively reduce poverty, albeit in tiny proportions. Many vulnerable populations in sub-Saharan Africa heavily depend on agriculture for their main source of food, which is consistent with existing academic research. The increase in agricultural production can significantly improve living conditions by increasing food supply and reducing food prices. The findings align with Gallup et al.’s (1997) finding that increased agricultural output had a beneficial effect on the income of the poorest twenty percent of the population. Thirtle et al. (2001) found that a 1% increase in agricultural output is associated with a 0.83% decrease in the population living on less than one USD per day. Thus, it can be said that boosting agricultural productivity is essential for developing countries.

In the analysis process, a causality test is also conducted. The cross-section dependence tests show evidence of cross-dependence across countries in the sample ( Table 6 ). This is expected, given that the countries in the sample are developing countries, belonging to the same economic region, whose main characteristics are high levels of poverty. Therefore, the cross-sectionally ADF (CADF) of Pesaran (2007) which is a second-generation panel unit root test is undertaken ( Table 7 ). The headcount poverty ratio and agricultural aid are stationary at first differenced while the agricultural productivity is stationary at level.

Table 6. Cross section dependence test.

Notes: ***, **, *: significance at 1% level, 5% level and 10% level.

Source: Authors’ compilations.

Thereafter, the cross-sectionally ADF (CADF) of Pesaran (2007) which is a second-generation panel unit root test is undertaken ( Table 8 ). The headcount poverty ratio and agricultural aid are stationary at first differenced while the agricultural productivity is stationary at level.

Notes: Figures in the table are p values. ***, **, * express significance at 1% level, 5% level and 10% level.

Source: Authors’ computations.

Table 8 presents the results of homogeneous non causality hypothesis test between the poverty headcount ratio, foreign aid to agriculture and agricultural productivity in the sample based on three test statistics, namely, the average Wald statistic, $W_{HNC}$, the asymptotic standardized statistic, $Z_{HNC}$ and the approximated standardized statistic based on finite sample moments, ${\tilde{Z}}_{HNC}$. The causality tests are conducted with stationary variables, and the optimal lag criteria is the Akaike information criterion (AIC).

Notes: The numbers on parentheses are probability values related to the tests statistics. ***, **, *: significance at 1% level, 5% level and 10% level.

Source: author’s computation.

A p-value of less than 1%, 5%, or 10% suggests that there is a causal relationship for at least one country in the sample, contrary to the null hypothesis of the causality test, which states that there is no causal relationship between the variables for any individual. Based on the three causality test statistics, the analysis’s findings show that there is a unidirectional relationship between poverty and agricultural aid, with the relationship going in the direction from aid to poverty. This supports the outcomes from the earlier regressions. All three tests demonstrate that agricultural productivity granger causes poverty, and two of the three indicators ( $W_{HNC}$ et $Z_{HNC}$) likewise demonstrate the relationship between agricultural aid and agricultural productivity. Agricultural aid has an effect on labor productivity in the agricultural sector, which in turn has an impact on poverty.

Furthermore, the results suggest a bi-directional relationship between foreign agricultural aid and agricultural productivity, as well as between poverty and agricultural productivity. Aid can provide farmers with access to modern technology, better quality seeds, improved irrigation systems, etc. This can increase their productivity and, consequently, reduce poverty by increasing farm incomes. On the other hand, poverty can also negatively affect agricultural productivity. Poor farmers may have limited access to resources such as land, water and credit, which can hamper their ability to invest in improved farming practices. As a result, their productivity may remain low, maintaining their poverty level.