2.1. Theoretical framework

Theoretically, there are two approaches in how economic growth could impact on poverty reduction and economic inequalities. The first one is the absolute approach promoted by the World Bank, based on ^{Ravallion and Chen (2001)} and ^{Kraay (2004)} works in which they point out that economic growth is the most important thing to encourage poverty reduction and the fall of inequality is inherent to it. On the other hand, the relative approach promoted by the works of ^{Kakwani and Son (1990)} and ^{Kakwani and Pernia (2000)}, focuses its objectives on postulating a growth model that, in addition to increasing the income of the poor proportionally more than that of the non-poor, generates long-term distributive changes that manifest themselves in sustained reductions in the levels of inequality.

Hence, the absolute approach ensures that there will be a trickle-down effect due to economic growth that would impact on poverty reduction. This effect is based on ^{Kuznets’s (1955)} hypothesis of U inverted shaped economic development, whereas poverty and inequalities will reduce as the economy grows. Nevertheless, ^{Chenery et al. (1974)} critiques the trickle-down effect of the economic growth on grounds that it matters, but also public policies focused on pro-distribution in favor to the poorest members of society. Later on, the relative approach based on the pro-distribution concept, sets up the growth pro-poor concept to refer to the economic growth accompanied by public policies that mitigate income inequalities and facilitate employment generation for the poorest.

Both approaches seem to agree on the role of economic growth in poverty reduction; however, the difference between them stems in the role that public policy could have in poverty reduction. While the absolute approach focuses on economic growth; the relative approach also takes into consideration the role of public policy intervention.

2.2. Empirical evidence review

There is few empirical evidence using panel data of how economic growth and public expenditure could reduce poverty rates in Latin America and the Caribbean (LAC) countries; and its results are still not conclusive in terms of the effectiveness of public policies. Some authors argued that public policies focused matters on poverty reduction, some others argued otherwise. However, it is more likely to find evidence in a cross-section data, and its results are conclusive: Economic growth is the principal variable that explains poverty reduction.

^{Gasparini, Gutierrez, and Tornarolli (2005)} were the pioneers to get empirical evidence on how growth impacts on poverty reduction in lac countries from 1989 to 2004. They used microdata from household surveys of 18 LAC countries to evidence that economies have experienced heterogenous patterns of pro-poor economic growth and poverty changes. For instance, there are seven countries which have experienced significant episodes of pro-poor income growth, and nine which have not; therefore, episodes of significant pro-poor income growth over this period have been rare in LAC. On average, economies that have grown (in terms of Gross Domestic Produtc, GDP, per capita) at more than 1% annually have been able to reduce poverty.

However, ^{Azevedo, Inchaust, and Sanfelice (2013)} extended the sample until 2010 and used a PVAR model and its variance decomposition of poverty reduction and they concluded that for LAC countries, 34% of poverty reduction is attributable to public policies, in comparison with 66% which is attributable to per capita economic growth. In the same way, ^{Ferreira et al. (2013)}, the ^{ECLAC (2015)}, ^{Crespo, Klasen, and Wacker (2016)}, and ^{Breunig and Majeed (2020)} found out that economic growth is the most important variable to explain poverty reduction; followed by public policies intervention. A common feature among these three papers is that they used Panel data estimation to get their results.

It is more likely to find empirical evidence when cross-section data is used. For example, ^{Henoch and Ramón (2015)} used the Datt-Ravallion methodology to decompose changes of poverty rate in three components: The role of economic growth, the role of redistributed public policy and an error term. They use data from Chile for the period from 1990 to 2013, and they concluded that economic growth is the principal variable which explains around 67% of the poverty reduction; and although the role of distributed public policies has helped out on poverty reduction, it was not significant.

In the same way, it is possible to find evidence from every LAC country. For instance, in 2012 the director of the Konrad Adenauer Foundation, Olaf Jacob, financed the Regional Program of Social Policies in Latin America to find evidence about the role of social public policies and economic growth in 13 Latin American countries. The studied countries were Argentina, Bolivia, Brazil, Chile, Colombia, Costa Rica, Ecuador, Guatemala, Mexico, Paraguay, Peru, Uruguay and Venezuela. Although each country has its own characteristics, a common feature among them all is that economic growth has played the most important role on poverty reduction, followed by social public policies.

The role of social protection expenditure on poverty reduction, in both short and long run, is supported by the evidence found by ^{Abramo, Cecchini, and Morales (2019)} in which the Conditional Cash Transfer (CCT) programs helped with two simultaneous objectives: 1) reduce poverty in the short run by consumption via monetary transfers, and 2) reduce poverty in the long run by human capital formation via conditionalities. They affirm that the effect of the CCT in poverty reduction is greater in the long-term rather than in the short-term.

3.1. Variables and data

The sample used in this paper consists of an annually frequency panel dataset of 8 Latin American countries which are: Bolivia, Colombia, Dominican Republic, El Salvador, Honduras, Panama, Paraguay and Uruguay over 2000-2019. The panel data estimated is considered as a macro panel given that the number of periods (20) is greater than the number of units (8), and it is a strongly balanced panel given that there are not missing values¹. In total, the panel contains 160 observations and the data was obtained from the statistics released by the ECLAC.

A controversy arises when it comes to work with Latin American data. First, we only work with 8 Latin American countries. Adding more economies to the analysis, will make the model lose some time variables since not all countries start measuring their poverty rate since 2000. Working with more cross-sectional data and fewer times periods have disadvantages of capturing endogeneity in our variables, whereby our results will be biased. On the other hand, not considering important economies in the region such as Brazil or Chile make us lose representativeness in Latin America. We have decided to work with the largest amount of temporal data, thus sacrificing cross-sectional units in order to have unbiased results across the paper.

As we mentioned, the aim of this paper is to determine and quantify the impact that economic growth and social protection expenditure has had on poverty reduction in Latin American countries for the period of 2000-2019. The presence of endogeneity in our variables is corrected by the estimation of a PVAR model. Not considering endogeneity would have made our results to be biased; that is why a PVAR is used. One advantage of estimating a PVAR model is that it estimates the variance decomposition which it allows to quantify the impact of how one shock affects another variable.

The paper uses 3 endogenous variables: The logarithm of the real GDP (constant price of 2010), the percentage of the population living in poverty (measure as a monetary poverty rate) and the logarithm of the central government expenditure in social protection, which could be defined as the range of public policies and social programs needed to reduce the lifelong consequences of poverty exclusion (^{UNICEF, sf}). The paper also considers one exogenous variable, which is a dummy variable that stands for 1 for the global crisis of 2008 and 0 otherwise.

3.2. Analysis procedure

When it comes to working with times series data, we must ensure that all our variable are stationary, otherwise, estimation results are going to be spurious. In order to check the stationarity properties of the variables, the paper uses the Levin, Lin and Chu (LLC) unit root test given that this test is appropriate when the panel cross-section units (8) divided by the size of the panel temporal dimension (20) tends to zero.

Then, if all endogenous variables are integrated of order one (I∼1), a test to determine cointegration must be done. If there is a long-term relationship, a Panel Vector Error Correction (VEC) will be estimated, otherwise it would be a PVAR model. To estimate the optimal lag in our model (see Appendix 1), we use the Akaike Information Criteria (AIC), the Schwarz Criteria (SC) and the Hannan-Quinn Criteria (HQ). All lag length criteria suggest to use 1 optimal lag in our model. The Cholesky’s decomposition was used to estimate our model, where variables are treated from the least to the most endogenous. The logarithm of the GDP (lGDP) is treated as the least endogenous, followed by the central government expenditure on social protection (lSPE) and the most endogenous being the poverty rate (POV). The stability of the PVAR is shown in Appendix 2, where no root lies outside the unit circle, so the specification of our model satisfies the stability condition. We also present, in Appendix 3, that the residuals of the models do not have any serial correlation problem.

After the estimation results, we can draw the impulse response functions (IRFS) to show the response of a particular variable to one standard deviation shock on each and every other variable in the system. In addition, we can display the variance decomposition (VD) in a table to show the proportion of the forecast error variance for each variable that is attributable to its own shocks and shocks in other variables in the system. Finally, a Granger causality test is used to determine whether changes in one variable cause changes in another variable.

3.3. Econometric methodology

The interdependence of relationship between variables could be solved by using a general equilibrium model. These models are useful to provide answers to economic policy issues, however, they impose certain constraints that are not always compatible with the statistical properties of the data (^{Muinelo-Gallo, Miranda, and Mordecki, 2020}). Another way used to solve interdependence, is achieved by constructing a VAR model, where all variables are treated as endogenous variables.

As it was mentioned before, the paper takes into consideration eight units of cross-section analysis and 20 years for each unit, so the paper performs a dynamic empirical analysis of simultaneous equation model using the PVAR methodology. According to ^{Grossmann, Love, and Orlov (2014)} a PVAR methodology allows to specify the model with few theoretical information about the relationships among variables. Also, the PVAR methodology is useful when there exists a problem of endogeneity, and, finally, the use of time series and cross section dimension allows us to make more complete use of information.

The paper estimates a PVAR model with three endogenous variables with one optimal lag. Formally:

Where the cross section of the panel is represented by i = 1,2,…,8 and it stands for each country. The time series is represented by t = 1,2,…,20 and it goes from 2000 to 2019. Y _it represents the endogenous variables which are a 1×3 vector. The Y _it vector of endogenous variables is represented by the GDP (constant price of 2010), the population living in poverty (as a percentage of total population) and the central government expenditure in social protection. Likewise, X _it represents the exogenous variable which is a dummy variable that stands for 1 for the global crisis of 2008 and 0 otherwise; while μ _i represents the country effects variable that captures the unobservable heterogeneity. Finally, ε_i is the error term of the model estimated, where is assumed that E(ε _it ) = 0, E(ε _it ,ε _is ) = 0 ∀ t > s.

4.1. Unit root test

Table 1 presents the main results of the unit root test for the lGDP, the lSPE and the POV. It is found that, both lGDP and lSPE are not stationary in levels, but they are in their first difference, I(1); while POV variable is in its level, I(0).

Cointegration exists when these two requirements are met: All variables are integrated in the same order, normally speaking it is said that variables are integrated of order one, I(1); and the error term of the model is stationary at its level, I(0). Before performing a cointegration test we must ensure that these requirements are met. In Table 1, we can see that not all variables are integrated in the same order, so the requirements to perform a cointegration test has not been met, so we proceed to estimate a PVAR.

4.2. Estimation result

Estimating a PVAR model implies that all variables are treated as endogenous, that is to say that each variable is affected by the others; however, Cholesky’s decomposition works to set up an order from the least endogenous variable to the most. According to the literature review section, Cholesky’s decomposition is as follow: The lGDP will impact on the lSPE and this will on the POV.

Table 2 presents the main results of our PVAR estimation. As we can see, all endogenous variables are statistically significant to explain the POV variable, being the most important the first lag of the lGDP and the lag of the POV, both of them at 99% level of confidence; followed by the lSPE at 90% level of confidence. These results are according with the empirical evidence previous mentioned, where the economic growth is the most important variable followed by public policy spending to explain poverty reduction.

4.3. Post estimation test

An important utility to use a PVAR estimation is that it allows the dynamic analysis of the series. In this subsection, first we report the impulse response function, then the variance decomposition and, finally, the Granger causality test outcome is presented.

4.3.1. Impulse response function (IRF)

According to ^{Hamilton (1994)} and ^{Enders (2015)}, the impulse response function (IRF) works to describe the effect of an innovation in the jth variable on future values of each of the variable in the system. In other words, the aim of the IRF is to describe how a shock in certain variable would impact on the present and future of another variable. Figure 3 graphs the POV orthogonal response due to impulse or shock from lGDP and lSPE, in one standard deviation magnitude. The solid lines represent the impulse responses, and the dashed lines represent a 95% confidence interval.

Source: Author’s estimation.

Figure 3 Impulse response function 

As we can see from Figure 3, one standard deviation shock on the change of the lGDP causes a significant decrease in the poverty rate for twenty periods after which the effect dissipates. However, one standard deviation shock to the change of the logarithm of social protection expenditure does not show any significant movement in the poverty rate.

4.4. Cross-section analysis

The results show that economic growth has the strongest influence on poverty reduction, in both short and long run, however, there is heterogeneity of economic growth in poverty reduction as evidenced in each country. For instance, in absolute values, Bolivia was the country that has experienced the greatest reduction in its poverty rate, going from 66.4% in 2000 to 43.6% in 2019.Nevertheless, in relative values, El Salvador was the country which has a better performance in reducing its poverty rate in terms of economic growth: Its elasticity was of -4.1%, it means that a 1% economic growth is associated with a poverty reduction of 4.1%, followed by Colombia (-2.9%) and Bolivia (-2.8%) [See Table 5].

The only country which has experienced an increase in its poverty rate during the same period of analysis was Honduras and this is because its private investment is limited by low economic returns to investment due to 1) inadequate education of its labor force, 2) high administrative costs and 3) crime and security (^{U.S. Agency for International Development, 2021}).