Poverty convergence with stagnation (and redistribution)∗ Luis F. Lopez-Calva† Eduardo Ortiz-Juarez‡ Carlos Rodr´ an§ ıguez-Castel´ April 8th, 2019 Abstract This paper exploits a novel dataset to provide unambiguous evidence of convergence in income, poverty, and inequality across Mexico’s mu- nicipalities over 1992-2014 in a context of stagnant overall economic growth. The findings suggest that these convergence processes stem from a combination of considerable positive performance among the poorest municipalities, and stagnant and deteriorating performance among richer counterparts. The expansion of social assistance seems to have played a crucial part in these results, with the double effect of bolstering high income growth rates among the poorest while exerting progressive changes in the distribution of income. ∗ The authors are grateful to Ted Enamorado for his comments and research assistance. avalos and Gerardo Esquivel for significant The authors would also like to thank Maria E. D´ contributions to this work, as well as Paloma Anos-Casero, Louise Cord, Joost Draaisma, Norbert Fiess, Thania de la Garza Navarrete, Rodrigo Garc´ u, Gonzalo Hern´ ıa-Verd´ andez- Licona, Sandra Mart´ ınez-Aguilar, Edgar Medina, Kinnon Scoot, Miguel Sz´ ekely, Gaston Yalonetzky, officials from CONEVAL and Mexico’s Ministry of Social Development, and participants at the EADI Nordic Conference 2017 held in Bergen, Norway, for helpful comments and suggestions. The datasets and codes necessary to replicate the exercises in this paper are available from the authors upon request. The findings, interpretations, and conclusions in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank Group, its Executive Directors, or the countries they represent. † United Nations Development Programme; e-mail: luis.lopez-calva@undp.org ‡ King’s College London; e-mail: eduardo.ortiz@kcl.ac.uk § World Bank; e-mail: crodriguezc@worldbank.org 1 1 Introduction Over 1992-2014, Mexico’s gross domestic product (GDP) expanded at an annual average rate of 2.5 percent, and of 0.9 percent in per capita terms. Comparatively, these figures placed Mexico as the second worst performer during those years among the rest of economies in continental Latin America. This moderate-to-low performance has been mirrored in terms of income- based poverty reduction. In 1992, the poverty headcount rates for the official extreme and total poverty lines reached, respectively, 21.4 and 53.1 percent of the country’s population, and by 2014 such proportions remained virtually unchanged at 20.6 and 53.2 percent, respectively. The incidence of poverty remains high not only by local standards, but also relative to countries with a similar level of development where significant poverty reductions occurred over the same period. In 2014, for instance, the share of the population living with less than $5.5 a day (2011 PPP) in Argentina and Chile was 9.1 and 10.1 percent, respectively, whereas in Mexico that share reached 33.6 percent. A closer look at subnational figures reveals also dramatic disparities within Mexico. For example, according to official data, in 2014 the total poverty headcount rate averaged 71 percent in the four poorest states, all located in the south of the country, which is twice the average of 34 percent in the four less poor states, which are mainly located in the north. Moreover, in 1,276 municipalities, i.e., more than half of total municipalities, total poverty recorded headcount rates at or above 70 percent, while only 69 out of total municipalities had headcount rates at or less than 30 percent. In various episodes during 1992-2014, Mexico has been hard hit by non-trivial economic contractions. In 1995, for instance, the internally generated Tequila Crisis plummeted GDP per capita by 8 percent, while the global financial crisis did so by almost 7 percent in 2009. Following the former crisis, poverty rates experienced a sharp increase: extreme poverty went from around 21 per- cent in 1992-4 to slightly more than 37 percent in 1996, while total poverty 2 rose from around 52 percent to 69 percent. Since then, extreme and total poverty rates experienced a decade of sustained decline, recording in 2006 the lowest levels: 14 and 43 percent, respectively. Though economic growth was disappointing in those years, progress in reducing poverty was accompanied by aggressive social reforms, including the introduction of conditional cash transfers targeted to the poorest and the expansion of federal transfers allo- cated to municipalities. Poverty bounced back after 2006, partly motivated by the global financial crisis and other shocks —e.g., the world food price crisis of 2007 and the swine flu outbreak in 2009—, and since then it has been increasing steadily up to the levels observed in 1992. This context of poor performance in income growth, recurrent economic shocks, long-run stagnation of overall poverty rates, and sizeable within- country disparities, leaves the impression that little has happened with the living standards of the population, in particular those of the poorest. Al- though social spending has been expanding1 and the income distribution has slightly improved —e.g., the Gini coefficient declined from around 0.53 in the 1990s to 0.49 in 2014—, the question of whether some groups or regions have been persistently lagging behind in pockets of poverty, remains.2 In an attempt to explore this, we perform, first, an analysis on convergence in municipalities’ mean per capita income over the period 1992-2014 follow- ing the framework used by Barro and Sala-i Martin (1991). The aim is to understand whether poorer municipalities have been capturing income gains resulting from above mentioned modest growth and social spending, and 1 Public social spending as a share of the GDP increased twofold between 1992 and 2014, from 3.9 to 7.6 percent, whereas the amount per capita did so by an average annual rate of almost 7 percent. Despite these changes, in 2014 per capita social spending in Mexico was the lowest among the OECD countries (OECD Social Expenditure Database), and across Latin America, it stood behind the levels observed in Argentina, Brazil, Chile, Costa Rica, and Uruguay (ECLAC’s statistics). 2 All figures on GDP growth rates come from the World Bank’s World Development Indicators database; data on poverty rates at $5.5 a day are from the World Bank’s LAC Equity Lab; official income-based poverty rates are from Mexico’s National Council for the Evaluation of Social Development Policy (CONEVAL); and, data on income inequality comes from the Socio-Economic Database for Latin America and the Caribbean (CEDLAS and World Bank). 3 whether there has been a reduction in regional disparities over time. Then, we analyse whether convergence in mean per capita incomes has translated into convergence in poverty headcount rates, and apply the poverty con- vergence decomposition by Ravallion (2012) to assess the effects of initial poverty on both the income growth process and the sensitivity of poverty reduction to income growth. We also shed some light into the role that ini- tial inequality and its intertemporal changes has exerted on such processes. In the analysis of convergence and the sensitivity of poverty to growth an emphasis is on subgroups of municipalities with sizeable disparities between them —e.g., urban versus rural, and those located in states along the U.S. border versus the rest—, as well as on subperiods to contextualise the various changes Mexico has experienced over the years —i.e., economic crises, ups and downs in overall poverty rates, and the expansion of public spending. This analysis exploits a five-wave municipality-level dataset spanning 1992- 2014, which contains data on mean household per capita incomes, poverty headcount rates, and levels of inequality, all of them being comparable both over time and across 2,361 municipalities, i.e., 96 percent of the total, where it was possible to compute reliable estimates in each wave. We constructed this dataset by applying the small-area estimation technique on household surveys and population censuses, which is a novel exercise as the collection of robust monetary data at that geographical level, and for such a prolonged period, is a complex and resource-consuming task. Moreover, our paper adds to the empirical literature on income growth, and to that on within-country poverty convergence, which is quite scarce. In par- ticular, it also complements past related studies in Mexico that have either used states as the unit of analysis, instead of municipalities, or considered the latter but for a significantly shorter time-span. For our analysis is disag- gregated and anchored to household-level data, it is a step forward to better understand the structural relationships between the parameters of the initial distribution of income, namely initial poverty and inequality, and subsequent growth and poverty reduction rates —at least better than in existing studies 4 where country-level figures may vanish the nonlinearities involved at a more disaggregated level, or where they do not necessarily reflect average living standards as measured by microdata. The rest of the paper is organized as follows. Section 2 reviews the relevant literature. Section 3 presents the methodology to construct the municipality- level dataset. Sections 4 and 5, respectively, test for convergence in munic- ipalities’ mean per capita and poverty headcount rates. Section 6 analyses, on one hand, the influence of initial poverty and inequality on both income growth and the efficiency of growth to reduce poverty and, on the other, de- composes the estimated magnitude of poverty convergence. Finally, Section 7 brings the main messages of each precedent section together to conclude. 2 Theory and past evidence in Mexico Both theoretical and empirical literature on economic growth offer some stylised facts on which the analysis of economic development paths at the municipality-level can be anchored. A first, widely studied stylised fact stems from the influential works by Baumol (1986) and Barro and Sala-i Martin (1991, 1992, 1995) on the convergence hypothesis, often labelled either as the catch-up effect or as the advantage of backwardness, whereby poorer coun- tries will tend to experience faster subsequent economic growth rates than in richer but slowing growing counterparts, in effect catching-up to the latter. Though what is meant by convergence is not circumscribed to those influ- ential works (e.g., see Quah 1993), they emphasise two prominent concepts that we exploit in this paper: σ -convergence, which focuses on the reduc- tion of income’s dispersion across units of analysis (e.g., see Sala-i Martin 1996), commonly assessed through a standard measure of statistical disper- sion, and, β -convergence, which assesses the negative relationship between initial levels of income and its subsequent growth rates, usually through parametric approaches; chiefly, log-linear and nonlinear growth regressions 5 —though non-parametric methods, including discrete Markov chains, are also common practice. The latter concept distinguishes at least two forms of convergence in the long-run: absolute β -convergence, whereby poorer coun- tries’ incomes converge toward a common steady-state, and, conditional β - convergence, whereby income convergence, not necessarily towards a common steady-state, is conditional on economies’ structural characteristics —a third form, related to the latter, is that of club convergence, whereby conditional convergence may cluster countries around different steady-state equilibria (e.g., see Durlauf and Johnson 1995; Quah 1996, 1997; Su 2003). A sum- mary of theoretical implications and empirical support (or lack thereof) for each of these forms is presented in Galor (1996). While most of this literature has mainly focused on incomes, a strand of re- search has opened the debate on whether countries’ income distribution con- verge towards a common invariant state, i.e., income inequality convergence enabou 1996; Ravallion 2003; Lin and Huang 2011), or whether in- (e.g., see B´ come convergence is also accompanied by poverty convergence (e.g., see Sala-i Martin 2006; Ravallion 2012; Cuaresma et al. 2017) —for instance, Ravallion (2012) demonstrates a proposition that in standard log-linear growth mod- els, with parameters independent of the initial distribution, the existence of income convergence should reveal that of poverty convergence. This latter implication, that income growth is a necessary condition for poverty reduction —and often it appears as the main driver (e.g., see ev- idence on this for Latin America in Lustig et al. 2016)—, has been widely studied in the literature. In general, the consensus is that higher growth rates tend to reduce poverty headcounts at a faster pace, in particular when using absolute poverty measures (e.g., see Ravallion 1995, 2001; Dollar and ekely 2008; Ferreira and Kraay 2002; Kraay 2006; Grimm 2007; Foster and Sz´ Ravallion 2011; Dollar et al. 2016; Fosu 2017). This advantage of economic growth, and how fast it reduces poverty, however, usually depends on both the initial income distribution and the changes it experiences with economic expansion. 6 This conditionality takes us to a second stylised fact: that the initial param- eters of the income distribution matter to both growth and the efficiency this growth has on poverty reduction. A well-established theoretical argument by which initial conditions dull economic growth and its impact is when market failures translate into credit constraints that trigger diminished investments in physical and human capital or, worse, leave investment opportunities un- exploited at all, in particular when credit rationing combines with investment indivisibilities —a combination especially harmful for the poor (e.g., see Ga- enabou 1996; Durlauf 1996; Hoff 1996; lor and Zeira 1993; Ljungqvist 1993; B´ Aghion and Bolton 1997; Piketty 1997; Banerjee and Duflo 2003). Built on similar arguments, an array of empirical studies on the constraints and determinants of growth have thus tested the role of initial parameters of the distribution in growth models and confirmed that either higher initial poverty (e.g., see Ravallion 2012) or higher initial inequality (e.g., see Alesina and Rodrik 1994; Persson and Tabellini 1994; Clarke 1995; Deininger and Squire 1998; Ravallion 1998; Knowles 2005) are significant constraints to future growth rates. Moreover, some studies have also demonstrated that such unfavourable initial parameters tend to curb the impact a given growth rate can exert on the proportionate rate of poverty reduction, as revealed by diminished elasticities of poverty to growth (e.g., see Bourguignon 2003; en 2006). Ravallion 1997, 2004, 2007, 2012; Lopez and Serv´ Despite this evidence, most of the empirical literature on (income) conver- gence do not explicitly addressees the influence the initial distribution exerts on subsequent growth and poverty reduction. In Ravallion’s (2012) sample of almost 90 countries, which recorded noticeable rates of growth and poverty reduction and for which there are unambiguous signs of income convergence, there is no significant evidence that countries starting out poorer have expe- rienced higher proportionate rates of poverty reduction in the future. This counterintuitive result is attributed to initial poverty which, as revealed by a decomposition of the speed of poverty convergence, counteracted the advan- tage of higher growth rates among poorer countries, i.e., income convergence 7 and the growth elasticity of poverty reduction. With the advantage of having constructed a municipality-level dataset on income, inequality, and poverty indicators, which covers five data points spanning almost a quarter century, we test most of the above-mentioned conclusions to provide a more disaggregated, longer-term perspective on the convergence paths and changes in those wellbeing indicators within Mex- ico. Previous studies for this country have mainly focused on income growth paths at the level of states and, while some have found a process of conver- gence in the years before the end of the import substitution model, most have consistently reported a process of divergence afterwards. For instance, Es- quivel (1999) shows that the pace of convergence across states was relatively fast over 1940-1960, but that it halted and started to reverse in the next 35 years. This divergence was confirmed by subsequent studies focused on ıa-Verd´ the 1985-2000 period (e.g., see Chiquiar 2005; Garc´ ıguez- u 2005; Rodr´ anchez-Reaza 2005; Rodr´ Pose and S´ ıguez-Oreggia 2007). In general, regional divergence in these years was linked to trade liberalisation and the entry into force of NAFTA, which bolstered the emergence of convergence clubs in those states most benefited from such reforms as they were initially endowed with relatively high skilled labour and better public infrastructure. Empirical evidence on convergence at a higher level of geographical disag- gregation, namely municipalities, has been scarcer in Mexico due to the lack of a sample with robust income information and statistical power at that level. A couple of previous studies have reported dramatic disparities be- ekely et al. tween municipalities in terms of income and poverty in 2000 (Sz´ opez-Calva et al. 2008) by applying a small-area estimation technique 2007; L´ to impute incomes from the main household income survey to the population census. Using this technique and logistic regressions, the Mexico’s National Council for the Evaluation of Social Development Policy (CONEVAL) has computed levels and changes in income-based poverty between 2000 and 2005 and in multidimensional poverty between 2010 and 2015 at the municipality- level. 8 Despite these valuable efforts, no long-run assessment of regional dispar- ities and paths in income, poverty, and inequality, based on comparable municipality-level data, has been available until the evidence shown in the ıa following sections. The closer exception is the study by Villalobos Barr´ et al. (2016) who tested a preliminary version of our dataset to analyse, through Gaussian mixture modelling, the univariate and joint distribution of human development indicators —namely, income, infant mortality, and years of schooling— and the conformation of development clusters between 1990, 2000, and 2010. Also, using the same preliminary dataset, Enamorado et al. (2016) study on the causal effect of inequality on drug-related homicides re- ported a sizeable decline in inequality in the majority of municipalities over 1990-2010. 3 Mapping income, poverty, and inequality at the municipality-level Capturing the long-run trends in income, poverty, and inequality at the municipality-level requires a dataset with intertemporally comparable indi- cators of wellbeing and that is statistically representative of the population in each municipality. The availability of such a dataset, however, often en- tails a trade-off between relatively high precision in the measurement of, say, household income, and a high level of geographical detail. One can exploit typical household-level surveys designed to capture all sources of income and thus retrieve household income with a high degree of precision. However, as any restricted sample, these surveys are usually representative at the national level only —and whenever possible, at the level of provinces, or states. Universal coverage of the population, on the other hand, can be gained from population censuses, which typically provide relevant inputs to mea- sure several dimensions of wellbeing at high levels of disaggregation —e.g., CONEVAL’s Social Backwardness Index or CONAPO’s Marginalization In- 9 dex in Mexico.3 This higher geographical detail, however, comes at the cost of lacking robust information on household incomes. While censuses are not designed to comprehensively collect income, they provide an incomplete pic- ture of households’ monetary circumstances, which, at least for the purpose of our analysis, represents its main weakness. To address the dilemma between precision and geographical detail, and hence, to make the relevant dataset available for our empirical analysis, we employ the small-area estimation technique proposed by Elbers et al. (2003) to im- pute households’ per capita income from surveys to corresponding households in censuses. Basically, this is done by predicting, from an income model in the survey, the parameters and distribution of errors that are then used to simulate the income distribution in the census dataset from which poverty and inequality indicators can be computed. There are two critical steps for this model to work properly. For a given year, the first step consists of taking the household survey to be a random sample of the total population found in the census’ sample frame. The second step is to identify a set of potential explanatory variables that are common between survey and census, and that satisfy a conceptual and statistical equality crite- rion. This means, respectively, that such variables should measure the same phenomenon in both datasets, and that their distributions are statistically indistinguishable —i.e., the sample mean is statistically equal to the popu- lation mean. Those variables that satisfy this criterion are then candidate regressors to model household per capita income in the survey dataset. Formally, the model takes a generalized least squares form 3 These indices, computed at the state, municipality, and village levels, summarise the degree of deprivation in 12 and 9 indicators, respectively, comprising the dimensions of education, access to basic services in the dwelling, and quality and spaces of dwelling, plus health (Social Backwardness Index) and labour income (Marginalization Index). The two indices are computed through the principal component analysis technique and then they are stratified into five groups according to their degree of backwardness or marginalization: very low, low, medium, high, and very high. 10 ln (yhm ) = βXhm + γZm + µhm (1) to estimate the joint distribution of per capita income y in the household h located in municipality m, conditional on two sets of covariates: Xhm , which includes household and individual characteristics, and Zm , which includes fixed characteristics of the municipality of residence. The parameter α is a household-specific effect; β and γ are the correlation parameters between the corresponding sets of covariates and ln (yhm ); and, µhm = ηm + hm represents an error term, where ηm is the component that is common to all households located in the same municipality —assumed to be homoscedastic and i.i.d.—, and hm is the component that is specific to each household —assumed to be heteroscedastic as it depends on the characteristics of the household and the municipality. The estimates of β , γ , and µhm are then applied to the corresponding sets of covariates Xhm and Zm in the census to simulate, using the bootstrap method, the distribution of household per capita income. Empirical support from extensive applications has proven this methodol- ogy to be robust (e.g., see Alderman et al. 2002; Bedi et al. 2007) and has become a common practice to improve targeting in developing countries fac- ing the dilemma between precision and geographical detail. In Mexico, the small-area estimation technique has been already applied to officially map income-based poverty at the municipal level, but only in 2000 and 2005.4 In this paper, we cover five data points over a 25-year period by pairing avail- able rounds of the Household Income and Expenditure Survey (ENIGH) and censuses collected in or around the same years (1990-92, 2000, 2005, 2010 and 2014-15).5 In each data point, the survey is a random sample of the 4 These exercises were conducted by CONEVAL by simulating income in the population censuses of 2000 and 2005. Though new rounds of census data were available in 2010 and 2015, the estimates of poverty followed a multidimensional approach based on a differ- ent methodology, and hence, a longer series of comparable income-based poverty at the municipality-level is not officially available. 5 Regarding census data, it corresponds to the general census of population and housing for the years ending in zero; for 2005, the data comes from the population and housing count ; and, for 2015 it comes from the intercensal survey. Unless otherwise stated, from 11 corresponding census’ sample frame,6 thus allowing for strict comparability of the distributions of a given variable between both data sources —even in pairings where gaps exist, i.e., 1990-92 and 2014-15, it is possible to identify common sets of covariates Xhm that satisfy the equality criterion, mainly because they are capturing virtually the same context as some households’ characteristics change slowly over time. These sets Xhm mainly include characteristics of individuals, households, and dwellings. The set Zm considers some of these variables aggregated at the municipality-level, plus data on coverage and availability of public services and infrastructure,7 and help us to increase the precision of the estimates via minimising the ratio of the variance of the error ηm relative to the variance of the total error µhm —i.e., the share of the variance of errors that results from unexplained differences across municipalities. We estimated the parameters of equation (1) using the whole sample in each round of the ENIGH, and then simulated the income distribution with 200 repetitions in the corresponding census dataset, each covering Mexico’s total population —the only exception was the census data in 2015, which is a sample of 5.9 million households; yet, it has enough statistical power to provide reliable statistics at the municipality-level. Based on the simulated income distribution, poverty and inequality indica- tors were computed at the municipality-level and validated through several tests.8 The income concept we used is that of household net per capita in- come, which includes labour income, income from businesses owned by the household, non-labour income such as public and private transfers, and an estimate of the imputed rent of owner-occupied dwellings, self-consumption, here onwards the term census will refer indistinguishable to all these three data sources. 6 Notice that in 2014-15 both the ENIGH and the intercensal survey are a random sample of the 2010 general census’ sample frame. 7 Some of these variables are derived from the census datasets, while others come from administrative records concentrated in the National System of Municipal Information (SNIM) and from the National Institute of Statistics and Geography (INEGI). 8 A detailed description of the small-area estimation methodology used in each data point is available from the authors upon request. 12 and in-kind perceptions and gifts received. The measurement of poverty was based on the Foster et al. (1984) family of indices by comparing this income concept with three poverty lines: food poverty, defined as the inability to acquire a basic food basket; capabilities poverty, defined as the inability to cover the value of the food basket plus expenditures on health and educa- tion; and, assets poverty, defined as the inability to acquire the latter plus expenditures on clothing, housing and transportation. Finally, municipali- ties’ inequality levels were computed through an array of well-known indices such as the Gini coefficient. This exercise yielded a robust, novel municipality-level dataset with income- based indicators that are comparable both over time and across 2,361 munic- ipalities for which it was possible to compute reliable estimates in each data point —these municipalities represent 96 percent of Mexico’s current mu- nicipalities and cover approximately 98 percent of the country’s population. Summary statistics of this dataset suggest that mean per capita income in Mexico has remained virtually stagnated in most of the period under study and has experienced a slight increase just after 2010. Indeed, the annualised growth rate reveals that per capita income expanded by just 0.8 percent in real terms between 1992 and 2014 —consistently with the GDP per capita performance described in the introduction. Accordingly, poverty headcount rates have not experienced a significant improvement when seen between the initial and final year, though there were important changes in the first lustrum of the 2000s (see the panel a in the Annex). In this context of relative stagnation of income growth and overall poverty rates in the long-run, the next section focuses on the growth trajectories of municipalities’ mean per capita income (constant MXN$ at August 2014 prices) with the aim of answering two key initial questions: Have poorer municipalities been persistently lagging behind in pockets of poverty, or have they captured income gains thus catching up richer municipalities? How income disparities between municipalities have evolved over time? 13 4 Convergence in municipalities’ mean per capita income As discussed in Section 2, a well-established hypothesis in the economic growth literature is that of convergence, whereby incomes in poorer areas tend to grow faster than in richer ones. To analyse the income growth paths across Mexican municipalities, we apply Barro and Sala-i-Martin’s (1991) framework on β -convergence and σ -convergence over the period 1992-2014, with a particular focus on the 2000s. Starting with β -convergence, for each time-span of length τ , the annualised growth rate in mean per capita income (y ) in municipality i between the most recent time (t) and the initial year (t − τ ) is given by gi (yit ) = ln (yit /yit−τ ) /τ (2) Hence, the empirical specification to analyse the growth process in munici- palities’ mean per capita income can be written as gi (yit ) = α + βln (yit−τ ) + µit (3) where ln (yit−τ ) is the log initial per capita income; the parameter α is a municipality-specific effect; β is a parameter indicative of the speed of abso- lute convergence; and, µit is a stochastic term. Estimates of this model, summarised in the panel a of Table 1, reveal signs of absolute β -convergence across municipalities during 1992-2014, as indicated by a significant coefficient of –0.007 meaning that per capita income in poorer municipalities grew faster than in their richer counterparts, at an annual con- vergence rate of 0.7 percent. A closer look at subperiods, however, show that the catch-up effect took place during 2000-2014 only, with a coefficient of –0.019, whereas in the 1990s no evidence of convergence was found. These 14 opposed results are also illustrated in Figure 1. Further exploring the 2000s, the speed of convergence was faster in the first five years with an annual rate of 4.3 percent, consistent with the marked reduction in overall poverty headcount rates from the high levels they reached after the Tequila Crisis, as discussed in the introduction. Income convergence was still evident after 2005, though it occurred at a slower pace —potentially slowed down by the various economic shocks that resulted in recession and non-trivial contrac- tions of Mexico’s economy. Figure 1: Municipalities’ mean per capita income converged after 2000 Source : Author’s calculations. Notes : The area of symbols is proportional to municipali- ties’ population. The regression line has a slope of 0.001 in panel a, and –0.019 in panel b (significant at the 1 percent level). Mean per capita incomes are in real terms at August 2014 prices. A breakdown by municipality’s population size yields also remarkable results. Over 1992-2014, the catch-up effect in rural municipalities, defined as those with less than 15 thousand inhabitants, was at least twice as large than that observed across urban counterparts —in fact, relative to the latter areas the speed of convergence across rural municipalities was consistently faster, and statistically significant, in each subperiod (Table 1, panels b and c ). Inter- estingly, while no evidence of convergence across urban municipalities was found in the 1990s, it actually occurred in rural ones at an annual rate of 2.7 15 percent. Moreover, although the speed of convergence halved during 2005- 2010 in both groups, relative to the previous lustrum, by 2010-2014 the pace was recovered across rural municipalities, whereas in urban ones it was even further slowed (columns 4-6 of panels b and c ). Table 1: Absolute β -convergence across municipalities, 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −0.007*** 0.001 −0.019*** −0.043*** −0.020*** −0.013*** ln (yit−τ ) (0.001) (0.003) (0.001) (0.003) (0.003) (0.003) Obs. 2,361 2,361 2,361 2,361 2,361 2,361 R2 0.102 0.000 0.342 0.313 0.076 0.022 b. Urban municipalities −0.008*** −0.003 −0.019*** −0.045*** −0.020*** −0.009*** ln (yit−τ ) (0.001) (0.004) (0.001) (0.003) (0.004) (0.004) Obs. 944 944 1,017 1,017 1,022 1,022 R2 0.138 0.002 0.334 0.323 0.076 0.012 c. Rural municipalities −0.018*** −0.027*** −0.031*** −0.077*** −0.035*** −0.068*** ln (yit−τ ) (0.001) (0.004) (0.002) (0.003) (0.004) (0.008) Obs. 1,417 1,417 1,344 1,344 1,339 1,339 R2 0.235 0.050 0.415 0.395 0.062 0.188 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of the parameter β in equation (3), weighted by municipal population at the initial year. The dependent variable is the annualised growth rate in municipalities’ mean per capita income. ln (yit−τ ) represents municipalities’ initial per capita income (log-scale, at August 2014 prices). Urban (rural) municipalities are defined as those with more (less) than 15 thousand inhabitants. The intercepts are shown in Table 1 in the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 Now, we turn to the conditional β -convergence hypothesis where paths of mean per capita income growth are conditional on factors such as municipal- ities’ initial conditions and structural characteristics. To examine this, the specification in equation (3) is rewritten as gi (yit ) = α + βln (yit−τ ) + γXit−τ + µit (4) 16 to allow for the inclusion of a set of municipality-level characteristics Xit−τ that are presumed to exert an influence on mean per capita income growth. This set includes components of municipalities’ public spending and revenues at the initial year of each period under study, which are relevant in light of the reforms to the federal transfers system undertaken in the 1990s. In particular, the 1998 reform that introduced Ramo 33, aimed at redistributing additional fiscal revenues to subnational governments for social development purposes, has allowed municipalities to benefit from larger volumes of federal transfers —e.g., average per capita conditional and unconditional (Ramo 33 ) federal transfers, respectively, have increased twofold and threefold in real terms between 2000 and 2014 (see the panel b in the Annex). Making equation (4) conditional on, for instance, per capita total public spending at the initial year, reveals that the speed of convergence over 1992- 2014 jumps from 0.7 percent found in the absolute setting to 1.2 percent, and that the pace of conditional convergence was, again, particularly faster in the first lustrum of the 2000s. Interestingly, after finding no evidence of absolute convergence in the 1990s, conditional convergence recorded a rate of 1.6 percent in those years and it was significant at the 1 percent level (Table 2, panel a ).9 Panels b and c show, respectively, the estimates in urban and rural municipalities, with two particular results. Firstly, convergence in rural municipalities occurred, again, at a faster pace than in urban ones in all peri- ods under study. Secondly, and consistently with the whole sample, there are signs of conditional convergence in urban municipalities in the 1990s, with 9 The focus is on total public spending only because no sizeable differences on the rates of convergence appear when using particular components of public spending or revenues instead —which reduce the sample significantly as no disaggregated public finance data is available for all municipalities (see Tables 2-11 and 17-26 in the Online Appendix ). Moreover, in order to exploit our panel dataset of municipalities and control for time- invariant factors, we estimate conditional convergence using fixed-effects models, which consistently confirm convergence, as reported by the standard OLS model. Random effects specifications also produce coefficients with the same signs. As extra robustness checks, we use 5-year and 10-year averages for the public spending variables and GMM techniques. The results are again consistent; i.e. poor municipalities converge at a faster rate when compared to rich municipalities. 17 at an annual rate of 2 percent. Table 2: Tests of β -convergence conditional on total public spending, 1992- 2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −0.012*** −0.016*** −0.020*** −0.047*** −0.020*** −0.015*** ln (yit−τ ) (0.001) (0.005) (0.001) (0.003) (0.004) (0.003) 0.003*** 0.012*** 0.004*** 0.008** −0.008* 0.024*** Public spending (0.001) (0.004) (0.001) (0.003) (0.004) (0.005) Obs. 2,234 2,234 2,193 2,193 2,116 2,045 R2 0.166 0.056 0.342 0.318 0.089 0.061 b. Urban municipalities −0.013*** −0.020*** −0.021*** −0.049*** −0.020*** −0.014*** ln (yit−τ ) (0.002) (0.006) (0.002) (0.003) (0.004) (0.004) 0.003** 0.012*** 0.006*** 0.011*** −0.006 0.028*** Public spending (0.001) (0.005) (0.002) (0.004) (0.005) (0.006) Obs. 923 923 971 971 985 937 R2 0.216 0.067 0.345 0.333 0.086 0.066 c. Rural municipalities −0.020*** −0.044*** −0.031*** −0.083*** −0.035*** −0.077*** ln (yit−τ ) (0.001) (0.005) (0.002) (0.003) (0.005) (0.009) 0.002*** 0.016*** 0.001 0.009*** −0.013*** 0.012** Public spending (0.001) (0.002) (0.001) (0.003) (0.004) (0.005) Obs. 1,311 1,311 1,222 1,222 1,131 1,108 R2 0.253 0.112 0.417 0.405 0.095 0.220 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of parameters β and γ in equation (4), weighted by municipal population at the initial year. The dependent variable is the annualised growth rate in municipalities’ mean per capita income. ln (yit−τ ) and public spending are for the initial year and in per capita terms (log-scale, at August 2014 prices). Urban (rural) municipalities are defined as those with more (less) than 15 thousand inhabitants. The intercepts are shown in Tables 2-11 in the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 In general, the conditional model moves the rates of convergence upwards in most cases in comparison with the absolute model. The only exception is 2005-2010, when the magnitude of convergence remained virtually un- changed. A plausible explanation is that the coefficient for initial per capita public spending was found negative during those years, when the economy 18 was hard hit by various adverse shocks. The expectation, confirmed in the remaining cases, is that the point estimate of that variable is positive and significant, meaning that the initial level of public spending exerts a positive influence on income growth through, for instance, the allocation of resources to public investment or transfers and subsidies. If the model in equation (4) controls for the latter components instead of total public spending, it can be verified that both public investment and transfers and subsidies exhibit a negative and significant sign during 2005-2010.10 Hence, it seems that the initial level of per capita public spending in 2005 was not enough to promote income growth through those channels in an environment of economic and fiscal contraction towards the end of the 2000s, and hence to accelerate the pace of convergence. In another variation of equation (4), we also control for the annualised growth rate in the number of beneficiary families of Mexico’s flagship conditional cash transfers (CCT) program in order to capture the influence of its expansion on the speed of convergence since its launch as Progresa in 1997. By 2000, this program benefited around 2.4 million families living in extreme poverty, and five years later that number reached 4.9 million, which is equivalent to an annual growth rate of 20 percent. While the expansion continued after 2005, it was at a significant lower rate of 2.4 percent annually, reaching 5.7 and 6 million families in 2010 and 2014, respectively. Table 3 summarises the estimates of this conditional model suggesting that, in general, the speed of convergence increased relative to the corresponding coefficients in Table 2. The point estimate for the CCT variable exhibits a positive and significant effect in both 2000-2014 and 2000-2005, but it is par- ticularly high in the former coinciding with the dramatic expansion of CCT’s coverage. This expansion seems to have boosted the rate of convergence in the first years of the decade through the increase of per capita income in mu- nicipalities with the poorest populations (column 2). After 2005, the sign of that variable became negative and had no apparent influence on the pace of 10 See Tables 2-11 in the Online Appendix. 19 convergence, suggesting that the subsequent growth in CCT’s coverage was low enough to otherwise exert a substantial effect on municipalities’ mean per capita income. A noticeable finding throughout all previous specifications is that the conver- gence process continued after 2010. Though it occurred at a slower pace than in the previous two lustrums when assessing the whole sample, it was particu- larly high across rural, poorer municipalities during 2010-2014. What could explain this result when, as suggested before, the expansion of the CCT’s coverage should not have had much effect in the last part of the period under study? More and better federal transfers allocated to municipalities could hold the answer. A recent redistributive assessment of the Social Infrastruc- ture Contributions Fund (FAIS), which is a crucial component of Ramo 33, suggests that the identification of priority attention zones within the country improved the targeting and implementation of federal transfers for munic- ipality’s social infrastructure, and that the latter had a positive effect, yet modest, in both the level and growth of household income across all munici- ıguez-Castel´ palities over 2000 and 2014 (Rodr´ an et al., 2017). Furthermore, the study highlights that such transfers were crucial to improve a number of socioeconomic indicators within municipalities, in particular over 2010-2014, which could reflect a better targeting on the less advantaged groups. A critical aspect of all previous results is that the convergence process took place in a context of overall low growth in mean per capita income, which averaged 0.8 percent over 1992-201411 . A closer look at the data reveals a relatively higher growth rate in this period among the poorest municipali- 11 The literature focusing on growth at the level of states have consistently reported patterns of regional divergence after the entry into force of NAFTA (see Section 2 and World-Bank 2018). Notice that these patterns can coexist with the documented process of income convergence across municipalities over 1992-2014. There are at least two ex- planations behind this coexistence. The first source of the discrepancy is that state-level analyses typically use state’s GDP, a metrics that while measuring the value of production it often fails to reflect average living standards as measured by microdata, as in this paper. A second source is the unit of analysis; while results from state-level studies tend to be biased by the weight exerted by large urban agglomerations concentrating a number of municipalities, in municipality-level analysis that issue can be separated. 20 Table 3: Tests of β -convergence conditional on total public spending and CCT data, 2000-2014 (1) (2) (3) (4) 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −0.025*** −0.059*** −0.026*** −0.015*** ln (yit−τ ) (0.002) (0.004) (0.003) (0.004) 0.003** 0.006** −0.010*** 0.025*** Public spending (0.001) (0.003) (0.003) (0.005) 0.035*** 0.067*** −0.023 −0.055** Annual growth in CCT’s families (0.010) (0.014) (0.026) (0.024) Obs. 1,957 1,957 2,106 2,035 R2 0.367 0.348 0.182 0.065 b. Urban municipalities −0.025*** −0.060*** −0.027*** −0.014*** ln (yit−τ ) (0.002) (0.005) (0.003) (0.004) 0.004*** 0.008** −0.009** 0.029*** Public spending (0.001) (0.003) (0.004) (0.006) 0.033*** 0.066*** −0.023 −0.055** Annual growth in CCT’s families (0.010) (0.015) (0.027) (0.025) Obs. 878 878 975 927 R2 0.369 0.364 0.197 0.072 c. Rural municipalities −0.033*** −0.095*** −0.035*** −0.076*** ln (yit−τ ) (0.002) (0.004) (0.005) (0.009) 0.000 0.010*** −0.013*** 0.008 Public spending (0.001) (0.003) (0.004) (0.006) 0.011 0.058*** −0.072 −0.110* Annual growth in CCT’s families (0.013) (0.011) (0.044) (0.057) Obs. 1,079 1,079 1,131 1,108 R2 0.434 0.426 0.098 0.230 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of parameters β and γ in equation (4), weighted by municipal population. The dependent variable is the annualised growth rate in municipalities’ mean per capita income. ln (yit−τ ) and public spending are for the initial year and in per capita terms (log-scale, at August 2014 prices). Urban (rural) municipalities are defined as those with more (less) than 15 thousand inhabitants. The intercepts are shown in Tables 3-6 and 8-11 in the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 21 ties —e.g., 2.5 percent annually among the poorest 10 percent—, whereas it was negative among the richest ones —e.g., –0.6 percent annually among the top 10 percent. Indeed, non-anonymous growth incidence curves (GIC) for some revealing periods (Figure 2) show that, over 1992-2000, the bottom 10 percent of municipalities distribution experienced positive growth averag- ing 2 percent annually, while the rest observed negative rates: –1.1 percent among the remaining 90 percent, and –1.9 percent among the top 10 percent. Figure 2: Positive growth among poorest municipalities, and stagnant or negative among richer ones Source : Author’s calculations. The story in 2000-2014 was, in general, more optimistic. During these years, the vast majority of municipalities experienced positive growth, though were again those at the bottom who exhibited relatively higher rates. This per- formance was mainly driven by the high rates achieved during the first lus- trum of the decade, which benefited a larger share of municipalities at the bottom: mean per capita income among the poorest half expanded by 6.8 percent annually, while among the upper half it increased by an annual rate 22 of 0.4 percent only and reduced by 1.3 percent among the top 10 percent. In 2005-2010, the economic slowdown took a toll on municipalities’ income performance, with growth rates averaging 0.6 percent annually and, with the exception of the poorest 10 percent municipalities, the rest experienced an average rate of –0.8 percent. Thus, the observed process of income convergence stems from a combination of positive and relatively high growth in mean per capita income among the first decile of municipalities, and stagnant and negative growth among those located at the middle and top of the distribution, respectively. To explore more on such process, we focus on two additional groups of municipalities characterised by dissimilar levels of development and exposure to economic shocks: those located in states along the U.S. border, which are more eco- nomically integrated with the U.S. and exhibit higher levels of mean per capita income, and the rest (NB). The estimates of equation (4) across both groups, conditional on per capita public spending, show that the speed of convergence was evident through- out all periods, and consistently higher in municipalities of the former group (Table 4, panels b and c ). A careful look at how income growth performed in each group reveals some clues to better understand that result. For instance, over 1992-2000, convergence in NB municipalities resulted, again, from rela- tively high growth rates among the poorest municipalities and negative rates among the rest. By contrast, the speed of convergence across those in border states stems from an inverted-U-shaped growth pattern. That is, while mean per capita income among both the poorest and richest 20 percent contracted, such contraction occurred at a lower annual rate in the former: –0.2 and –0.7 percent, respectively. Remarkably, the bulk of municipalities in the middle of the distribution experienced positive growth rates. It seems, then, that while the Tequila Crisis had nation-wide adverse effects, some relatively poorer mu- nicipalities in states along the U.S. border may have slightly benefited from the devaluation of the currency and the entry into force of NAFTA12 , thus 12 As reference, growth in mean per capita income over 1992-2000 was positive in mu- 23 catching-up to their richer counterparts —and relatively faster than in the rest of the country. Table 4: Tests of β -convergence across municipalities in border and non- border states, conditional on total public spending; 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −0.012*** −0.016*** −0.020*** −0.047*** −0.020*** −0.015*** ln (yit−τ ) (0.001) (0.005) (0.001) (0.003) (0.004) (0.003) Obs. 2,234 2,234 2,193 2,193 2,116 2,045 R2 0.166 0.056 0.342 0.318 0.089 0.061 b. Municipalities in states along the U.S. border −0.017*** −0.028*** −0.022*** −0.051*** −0.060*** −0.044** ln (yit−τ ) (0.004) (0.010) (0.004) (0.012) (0.009) (0.017) Obs. 267 267 262 262 267 266 R2 0.250 0.113 0.226 0.256 0.198 0.055 c. Municipalities in non-border states −0.011*** −0.020*** −0.019*** −0.044*** −0.014*** −0.017*** ln (yit−τ ) (0.001) (0.006) (0.001) (0.003) (0.004) (0.003) Obs. 1,967 1,967 1,931 1,931 1,849 1,779 R2 0.154 0.052 0.307 0.274 0.056 0.089 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of the parameter β in equation (4), weighted by municipal population at the initial year. The dependent variable is the annualised growth rate in municipalities’ mean per capita income. ln (yit−τ ) and public spending are for the initial year and in per capita terms (log-scale, at August 2014 prices). Full specifications are shown in Tables 2-6 and 12-16 of the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 As another example, the difference in the speed of convergence between NB municipalities (1.4 percent) and those in border states (6 percent) over 2005- 2010, may be explained by the following growth patterns. Income growth averaged almost 8 percent annually among the poorest 10 percent munici- palities of both groups, while it decreased among the richest top 10 percent. The difference lies in the magnitude of this loss: it averaged –1.6 percent an- nicipalities located in border states, with an annual rate of 0.3 percent, whereas it was negative among NB municipalities: –0.9 percent. 24 nually in NB municipalities, whereas it was –5.4 percent annually in border states. In this case, then, it seems that the U.S. originated housing bubble that unleashed the global financial crisis had a strong regional bias, with dis- proportionate effects on those municipalities most integrated with the U.S.13 —similar growth patterns may also explain the difference in the speed of convergence between both groups over 2010-2014. A salient outcome of the documented process of β -convergence within the country is that it was quite effective in reducing regional disparities at the municipality level, in particular after 2000, which is consistent with empirical evidence of an overall decline of income inequality in the following years (e.g., see Esquivel et al. 2010), and also confirmed with our data, as shown later. Figure 3 shows the evolution of the standard deviation of logged mean per capita income across municipalities, or σ -convergence. Starting with the whole sample, after regional disparities increased sharply in the 1990s, they experienced a steep decline during the first lustrum of the 2000s and continued declining moderately up to 2010. Regional disparities remained relatively unchanged after this year, yet it is significant that, relative to 1992, income dispersion was almost 8 percent lower by 2014. When dividing the sample, similar results in terms of trends and orders of magnitude are evident across both urban and NB municipalities, with declines in income dispersion of 8.6 and 6.1 percent, respectively, between 1992 and 2014. Two additional results are worth noticing. Firstly, income disparities in rural municipalities deteriorated slightly after the sharp decline in the first half of the 2000s and, although such differences reduced again after 2010, the level recorded in 2014 was virtually the same than in 1992. Secondly, the relatively high β -convergence coefficients across municipalities in border states, seem to have reduced income dispersion along the U.S. border at a rate of 22 percent between 1992 and 2014. 13 Indeed, growth rates in mean per capita income in municipalities located in states along the U.S. border averaged –0.1 percent annually, whereas NB counterparts recorded an annual average rate of 0.8 percent. 25 Figure 3: Disparities between municipalities reduced sharply over the 2000s Source : Author’s calculations. 5 Testing for poverty convergence Have poorer, converging municipalities been able to translate their relative income gains into poverty reduction? If per capita income follows a lognormal distribution, then any change in the poverty headcount rate is determined, in a magnitude η , by two components: one that is attributable to changes in income and one that is attributable to changes in the distribution of in- come. The relationship between each component and changes in poverty is illustrated in Figure 4 over 1992-2014. As expected, those municipalities that experienced relatively higher rates of poverty reduction, according to the food poverty line14 , were those that experienced higher growth rates in mean per capita income (panel a ), but also experienced progressive changes in the distribution of income (panel b ) —as such changes imply transferring 14 The bulk of the following analysis focuses on extreme poverty as measured by the food poverty line, defined in Section 3. This is deliberately done because the narrative is consistent, and conclusions hold, when using higher thresholds such as the capabilities and assets poverty lines. 26 resources from richer to poorer populations, thus stimulating poverty reduc- tion. Figure 4: Municipalities that reduced poverty experienced relatively higher growth rates in income, but also progressive changes in the distribution of income Source : Author’s calculations. Notes : The area of symbols is proportional to municipal- ities’ total population. The regression line has a slope of −1.42 in panel a and 0.63 in panel b (both significant at the 1 percent level). Focusing on the first component, for now, let gi (Pit ) = δ + ηgi (yit ) + νit (5) be the partial elasticity of poverty to growth in municipalities’ mean per capita income, representing the percent change in the poverty headcount rate as a result of a one percent increase in income, holding the income distribution constant. gi (Pit ) is the annualised change in poverty rates, calculated as in equation (2); η is the elasticity parameter, with the expectation that η < 0; δ is a municipality-specific effect; and, νit is a stochastic term.15 15 Similarly, gi (Pit ) = δ + ηgi (Git ) + νit can represent the partial inequality elasticity of poverty, or the percent change in the poverty headcount rate as a result of a one 27 Estimates of equation (5) confirm that higher growth rates in income tend to reduce poverty. In 1992-2014, for instance, a one percent growth rate in municipalities’ mean per capita income would lead to a 1.4 percent decline in the food poverty headcount rate (Table 5, panel a ). The results also suggest that food poverty is more responsive to growth among both urban municipalities and those located in states along the U.S. border, relative to their corresponding counterparts (panels b through e ). According to the data, such counterparts consistently exhibit higher food poverty rates over time: around 30 percent higher in rural municipalities than in urban ones, and twice the size in NB municipalities than in those located in border states. Thus, food poverty tends to be more responsive to growth in municipalities where poverty rates are relatively lower, which fits well- known evidence that under log normality, holding the income distribution constant, the growth elasticity will decrease in absolute value as the poverty rate increases (Bourguignon, 2003). In other words, poverty itself seems to act as a barrier to poverty reduction.16 Regardless of the context-specific magnitude of the growth elasticity param- eter, the fact that growth in municipalities’ mean per capita income tends to reduce food poverty rates, plus the previous evidence of income convergence, imply that those municipalities with relatively high initial poverty headcount rates, ln (Pit−τ ), should have experienced higher subsequent rates of poverty reduction over the period under study. To test this, let gi (Pit ) = α + βln (Pit−τ ) + µit (6) percent increase in inequality, holding per capita income constant, with the expectation that η > 0, and with gi (Git ) being the annualised rate of change in inequality. The growth and inequality elasticity parameters can be denoted as η y and η G , respectively, and hence, under log normality, changes in poverty rates can be expressed as gi (Pit ) ≈ η y gi (yit ) + η G gi (Git ). 16 These elasticities, in general, are also more responsive to growth the lower the value of the poverty line; for instance, relative to the food poverty line, the elasticity almost invariably reduces by half, in absolute value, when using the assets poverty line (see Table 33 in the Online Appendix ). 28 Table 5: Growth elasticities of food poverty reduction across municipalities, 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −1.425*** −1.291*** −1.671*** −1.504*** −1.472*** −1.736*** gi (yit ) (0.089) (0.072) (0.083) (0.114) (0.082) (0.077) Obs. 2,361 2,361 2,361 2,361 2,361 2,361 R2 0.535 0.411 0.517 0.432 0.427 0.549 b. Urban municipalities −1.513*** −1.348*** −1.736*** −1.620*** −1.457*** −1.751*** gi (yit ) (0.109) (0.091) (0.092) (0.133) (0.102) (0.094) Obs. 944 944 1,017 1,017 1,022 1,022 R2 0.540 0.391 0.511 0.456 0.400 0.521 c. Rural municipalities −1.142*** −1.018*** −1.210*** −0.942*** −1.482*** −1.680*** gi (yit ) (0.041) (0.052) (0.150) (0.060) (0.067) (0.111) Obs. 1,417 1,417 1,344 1,344 1,339 1,339 R2 0.599 0.488 0.580 0.312 0.540 0.686 d. Municipalities in states along the U.S. border −1.878*** −1.112** −1.837*** −1.276*** −1.551*** −1.904*** gi (yit ) (0.308) (0.531) (0.166) (0.371) (0.255) (0.175) Obs. 267 267 267 267 267 267 R2 0.501 0.127 0.600 0.201 0.339 0.660 e. Municipalities in non-border states −1.298*** −1.263*** −1.434*** −1.436*** −1.324*** −1.698*** gi (yit ) (0.065) (0.077) (0.095) (0.143) (0.074) (0.092) Obs. 2,094 2,094 2,094 2,094 2,094 2,094 R2 0.587 0.464 0.453 0.435 0.450 0.516 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of the parameter η in equation (5), weighted by municipal population at the initial year. The dependent variable is the annualised growth in the food poverty headcount rate. gi (yit ) is the annualised change in mean per capita income at the municipality-level (at August 2014 prices). The intercepts are shown in Table 32 of the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 be the empirical specification for the annualised proportionate change in poverty rates, or poverty convergence, where β is the speed of convergence parameter. 29 Indeed, estimates of equation (6) suggest that poorer municipalities reduced their headcount rates at a faster pace than richer, and poverty increasing, counterparts over 1992-2014 —in fact, food poverty rates among the 20 per- cent municipalities with the lowest incidence in 1992 recorded non-trivial increases by 2014 (Figure 5). Figure 5: Convergence in food poverty rates, 1992-2014 Source : Author’s calculations. Notes : The area of symbols in panel a is proportional to municipalities’ total population. The regression line has a slope of –0.012 in panel a (significant at the 1 percent level). A closer look at sub periods reveals a positive sign of the convergence pa- rameter in the 1990s, indicating that poorer municipalities became poorer after the Tequila Crisis, or at least that their poverty rates stagnated, while sizeable signs of convergence are found after 2000, in particular during 2000- 2005 (Table 6, panel a ). The breakdown by population size in panels b and c reveals that both urban and rural municipalities experienced poverty con- vergence, though convergence in the latter occurred even in the 1990s and, in general, at a faster pace than in the former. Sizeable poverty convergence in the 1990s also occurred across municipalities located in states along the U.S. border, whereas the opposite sign was found across NB municipalities. After 2000, though convergence unambiguously 30 occurred in both groups, interestingly those in border states exhibited a con- siderably higher coefficient during 2005-2010 (panels d and e ). The evidence in Section 4 helps to explain these results: poorer municipalities in border states were able to converge relatively faster in the 1990s and late-2000s be- cause mean per capita incomes in richer counterparts were disproportionately affected by the economic contractions that characterised those years. 6 Initial distribution and the speed of poverty convergence While poorer municipalities have experienced convergence for most of the pe- riod 1992-2014, little is known about the influence of the parameters of the initial distribution of income in shaping the speed of convergence. Focus- ing on initial poverty, we build on Ravallion’s (2012) decomposition of the so-called poverty convergence elasticity to explore how municipalities’ ini- tial poverty headcount rates might affect their advantage of starting poorer through two channels: the growth rates in mean per capita income, and the impact of that growth on poverty reduction, as revealed by the partial elasticity of poverty to mean per capita income. Starting with the first channel, we estimate three augmented versions of the β -convergence model in equation (4). In the first one, the annualised growth rates in mean per capita income depend on municipalities’ initial per capita income plus their initial food poverty headcount rates, gi (yit ) = α + βln (yit−τ ) + γln (Pit−τ ) + µit (7) Estimates of the parameter γ reveal some adverse effects of initial poverty on income growth at any given initial mean, though the coefficient is sizeable 31 Table 6: Tests of food poverty convergence across municipalities, 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −0.012*** 0.014** −0.031*** −0.055*** −0.038*** −0.048*** ln (Pit−τ ) (0.002) (0.006) (0.002) (0.006) (0.005) (0.006) Obs. 2,361 2,361 2,361 2,361 2,361 2,361 R2 0.206 0.025 0.565 0.333 0.175 0.164 b. Urban municipalities −0.013*** 0.014** −0.031*** −0.057*** −0.038*** −0.044*** ln (Pit−τ ) (0.002) (0.007) (0.002) (0.007) (0.006) (0.006) Obs. 944 944 1,017 1,017 1,022 1,022 R2 0.229 0.024 0.575 0.353 0.193 0.142 c. Rural municipalities −0.017*** −0.032*** −0.033*** −0.077*** −0.031*** −0.114*** ln (Pit−τ ) (0.001) (0.007) (0.006) (0.015) (0.007) (0.009) Obs. 1,417 1,417 1,344 1,344 1,339 1,339 R2 0.192 0.064 0.375 0.276 0.031 0.426 d. Municipalities in states along the U.S. border −0.023*** −0.044*** −0.027*** −0.058*** −0.097*** −0.038* ln (Pit−τ ) (0.004) (0.008) (0.004) (0.011) (0.013) (0.023) Obs. 267 267 267 267 267 267 R2 0.375 0.185 0.281 0.194 0.392 0.031 e. Municipalities in non-border states −0.009*** 0.019** −0.028*** −0.053*** −0.020*** −0.055*** ln (Pit−τ ) (0.002) (0.008) (0.003) (0.010) (0.003) (0.006) Obs. 2,094 2,094 2,094 2,094 2,094 2,094 R2 0.114 0.041 0.507 0.282 0.067 0.243 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of the parameter β in equation (6), weighted by municipal population at the initial year. The dependent variable is the annualised growth in the food poverty headcount rate. ln (Pit−τ ) represents municipalities’ initial poverty headcount rate. All variables are in log-scale. Urban (rural) municipalities are defined as those with more (less) than 15 thousand inhabitants. The intercepts are shown in Table 34 of the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 (–0.022) and significant at the 1 percent level in the 1990s only (Table 7, panel a ). An interesting, opposed result is shown in column 4, where the food poverty headcount rate in 2000 exerted a positive effect (0.007) on growth in the subsequent five years. While such effect is low and significant only at the 32 10 percent level, it coincided with the faster expansion of CCT across the poorest households located in also the most marginalised municipalities.17 As initial poverty rates are not independent from other parameters of the distribution, in the second version of the model we add municipalities’ initial inequality, ln (Git−τ ), measured by the Gini coefficient, as a third regressor. The results now reveal a positive and significant, yet moderate effect of initial poverty rates on income growth during both 1992-2014 and 1992-2000, and a more sizeable effect during 2000-2005 (Table 7, panel b ), which supports the plausible argument that initially poorer municipalities experienced higher subsequent growth in their mean per capita income as a result of expanding CCT among the poorest. In the rest of subperiods, the coefficients are is statistically indistinguishable from zero. To investigate these results further, we tested the previous augmented model by adding extra controls for concepts of either public spending or revenues, and with and without CCT data. Invariably, the story holds under different specifications: positive and significant effects of initial food poverty head- count rates on income growth are found over 2000-2014, and in particular during the expansion of CCT’s coverage in 2000-200518 . One of such spec- ifications is shown in the panel c of Table 7, in which the point estimates for the annualised growth rate in the number of beneficiary families exhibit positive and significant effects in the first years of the program’s expansion —consistently with the findings in the conditional β -convergence model in Section 4. Turning to the second channel, i.e., the growth elasticity of poverty reduction, it can be analysed through a variation of equation (5) by regressing gi (Pit ) on the growth rate in mean per capita income interacted with the initial 17 The coefficient over 2000-2005 even increases for higher values of the poverty line: 0.014 and 0.041 when using, respectively, the capabilities and assets poverty lines. In both cases, the effects are statistically significant at the 1 percent level (see Table 37 in the Online Appendix ). 18 The various specifications of this augmented model are shown in Tables 37-44 of the Online Appendix. 33 Table 7: Growth in mean per capita income conditional on initial parameters of the income distribution, 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. Conditional on initial poverty −0.009*** −0.033*** −0.022*** −0.032*** −0.036*** −0.003 ln (yit−τ ) (0.003) (0.008) (0.003) (0.007) (0.008) (0.010) −0.001 −0.022*** −0.002 0.007* −0.010* 0.006 ln (Pit−τ ) (0.002) (0.005) (0.002) (0.004) (0.005) (0.006) Obs. 2,361 2,361 2,361 2,361 2,361 2,361 R2 0.103 0.039 0.344 0.318 0.084 0.024 b. Conditional on initial poverty and inequality 0.005 0.014* −0.014*** −0.000 −0.022** 0.003 ln (yit−τ ) (0.003) (0.008) (0.004) (0.008) (0.010) (0.013) 0.008*** 0.009** 0.002 0.021*** −0.002 0.009 ln (Pit−τ ) (0.002) (0.005) (0.002) (0.004) (0.007) (0.008) −0.045*** −0.155*** −0.031*** −0.130*** −0.058*** −0.016 ln (Git−τ ) (0.005) (0.015) (0.011) (0.026) (0.017) (0.024) Obs. 2,361 2,361 2,361 2,361 2,361 2,361 R2 0.222 0.170 0.363 0.379 0.098 0.025 c. Conditional on initial poverty and inequality and extra controls − − −0.003 −0.004 −0.021** 0.003 ln (yit−τ ) − − (0.004) (0.010) (0.009) (0.014) − − 0.016*** 0.035*** 0.007 0.015 ln (Pit−τ ) − − (0.003) (0.007) (0.006) (0.010) − − −0.038*** −0.112*** −0.054*** −0.031 ln (Git−τ ) − − (0.006) (0.016) (0.014) (0.025) − − 0.007*** 0.011*** 0.007*** 0.012*** Public sector payroll − − (0.001) (0.003) (0.003) (0.004) − − −0.000 0.000 −0.003* −0.000 Public investment − − (0.000) (0.001) (0.002) (0.003) − − −0.001 0.000 −0.008*** 0.007** Public transfers/subsidies − − (0.001) (0.002) (0.001) (0.003) − − 0.033*** 0.059*** −0.019 −0.062** Growth in CCT’s families − − (0.011) (0.015) (0.025) (0.025) Obs. − − 1,793 1,793 1,910 2000 R2 − − 0.440 0.403 0.234 0.075 Source : Author’s calculations. Notes : For each period under study, the table shows esti- mates of parameters in equation (7) and its extensions, weighted by municipal population at the initial year. The dependent variable is the annualised growth rate in real per capita income. All monetary variables are for the initial year and in per capita terms (log-scale, at August 2014 prices). The hyphens in panel c indicate that models conditional on CCT data were not estimated as the latter is available from 2000 onwards. The intercepts are shown in Tables 37-40 of the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 34 poverty headcount rates. This adjusted rate is given by the growth rate in municipality’s mean per capita income times one minus municipality’s initial poverty headcount rate (Pit−τ ), which penalises more the sensitivity of food poverty to subsequent growth rates in municipalities starting out relatively poorer. The poverty-adjusted growth elasticity of poverty reduction is then defined as gi (Pit ) = η (1 − Pit−τ ) gi (yit ) + νit (8) The estimates for the whole sample of municipalities are shown in Table 8 (panel a ) —notice they increased in absolute value in all periods, rela- tive to the ordinary elasticities in Table 5. To illustrate the implications of the poverty-adjusted elasticity, consider for instance the value of –1.983 in 1992-2014. If a municipality’s initial food poverty rate is 10 percent and it experiences a 4-percent annual growth in mean per capita income, then that municipality would expect an annual poverty reduction of 7.1 percent. If ini- tial poverty stands at 70 percent instead, and the annual income growth rate is again 4 percent, then the municipality would expect a poverty reduction of only 2.4 percent annually. Then, as in previous section, poverty tends to be less responsive to growth, or the elasticity declines in absolute value, the higher the initial poverty rate. Interestingly, however, our estimates reveal that poverty-adjusted elasticities are consistently higher, in absolute value, in poorer municipalities than in richer counterparts. For instance, at an initial food poverty rate of 63 percent or more, at or above one standard deviation above the mean, a one percent increase in growth during 1992-2014 would lead to an annual decline in the poverty rate of almost 3.4 percent, whereas in municipalities with initial food poverty at 20 percent or less, at or below one standard deviation below the mean, the elasticity is almost −2 (column 1 of panels b and c ). Hence, contrary to the linear relationship by which the ordinary growth elas- 35 Table 8: Poverty-adjusted growth elasticities, 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. All municipalities −1.983*** −1.885*** −2.288*** −2.280*** −1.874*** −1.990*** (1 − Pit−τ ) gi (yit ) (0.151) (0.135) (0.163) (0.195) (0.118) (0.116) Obs. 2,361 2,361 2,361 2,361 2,361 2,361 R2 0.499 0.421 0.453 0.489 0.432 0.451 b. Municipalities with relatively low initial food poverty rates −1.984*** −1.756*** −1.870*** −2.247*** −1.337*** −2.182*** (1 − Pit−τ ) gi (yit ) (0.233) (0.233) (0.167) (0.249) (0.170) (0.155) Obs. 426 426 436 436 383 440 R2 0.486 0.293 0.365 0.423 0.277 0.601 c. Municipalities with relatively high initial food poverty rates −3.387*** −2.863*** −2.872*** −3.911*** −2.444*** −3.082*** (1 − Pit−τ ) gi (yit ) (0.124) (0.242) (0.142) (0.264) (0.106) (0.095) Obs. 433 433 458 458 425 457 R2 0.882 0.785 0.596 0.621 0.828 0.881 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of the parameter η in equation (8), weighted by municipal population at the initial year. The dependent variable is the annualised growth in the food poverty rate. (1 − Pit−τ ) gi (yit ) is the annualised change in mean per capita income at the municipality- level (at August 2014 prices), adjusted by municipalities’ initial food poverty headcount rate. The intercepts are shown in Table 45 of the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 ticity of poverty reduction decreases in absolute value as poverty rates in- crease, it can be readily verified that the poverty-adjusted growth elasticity follows a concave relationship with poverty (Figure 6). In words, those munic- ipalities with very high levels of food poverty in 1992 experienced sufficiently higher subsequent growth in mean per capita income to achieve substantial rates of poverty reduction by 2014, which unambiguously occurred (see Fig- ure 5, panel b ) —at least as substantial as in contexts of low poverty and relatively high growth in income. A salient result is observed during the first lustrum of the 2000s. Coinciding with the expansion of CCT, a one percent increase in the poverty-adjusted growth rate would lead to a 3.9 percent re- duction of food poverty headcount rates among the poorest municipalities, whereas the corresponding poverty reduction in less poor counterparts would be 2.2 percent only (Table 8, column 4 of panels b and c ). 36 Figure 6: Efficiency of growth to reduce food poverty, by initial poverty rates; 1992-2014 Source : Author’s calculations. Notes : The area of symbols is proportional to municipali- ties’ total population. For visibility purposes the elasticities in both charts are capped at −10. To better understand how the extent of poverty convergence was shaped by municipalities’ initial food poverty rates, we exploit all previous evidence computed for each channel to apply Ravallion’s (2012) decomposition of the poverty convergence elasticity. This decomposition results from the deriva- tive of equations (7) and (8) as −1 ∂gi (Pit ) ∂ln (Pit−τ ) = ηβ (1 − Pit−τ ) + ηγ (1 − Pit−τ ) − ηgi (yit ) Pit−τ ∂ln (Pit−τ ) ∂ln (yit−τ ) ∂gi (Pit ) where ∂lnPit−τ is the speed of food poverty convergence, equivalent to the parameter β in equation (6); the first element at the right-hand side of the equation is the mean convergence effect ; the second element, ηγ (1 − Pit−τ ), is the effect of initial poverty ; and, the third element, ηgi (yit ) Pit−τ , repre- sents the poverty elasticity effect. Using the estimates of η in Table 8; the parameters β and γ in Table 7; the ordinary elasticities of municipalities’ 37 initial food poverty with respect to their initial mean per capita income19 ; and, the sample means of Pit−τ and gi (yit ), the computation of this decom- position yields virtually the same food poverty convergence rates calculated previously. For instance, the convergence rate from the above decomposition is −0.011 during 1992-2014, which is very close to the coefficient of −0.012 computed with equation (6) for the same period (see the panel a of Table 6). The decomposition of that rate reveals that the convergence effect accounted for −0.007 and that poverty was actually responsive to growth, with a poverty elasticity effect of −0.005. By contrast, municipalities’ initial poverty rates exerted an adverse, yet moderate effect of 0.001. In the 1990s only, un- surprisingly, a convergence effect of −0.024 was more than counteracted by both the initial poverty (0.024) and the poverty elasticity (0.015) effects, thus confirming the significant poverty divergence of 0.014 found in those years. Moving to the period 2000-2014, both convergence and poverty elasticity ef- fects explain in similar magnitudes (−0.016 and −0.020, respectively) the totality of the speed of poverty convergence (−0.034), with only a slightly adverse effect of initial poverty of 0.002. These results confirm that the pro- cess of income convergence, and the efficiency of growth to reduce poverty, effectively translated into poverty convergence during 1992-2014 in general, and particularly after 2000. Focusing on the first lustrum of the 2000s, probably the most revealing period under study, the decomposition offers a remarkable result: the three effects moved in the same favourable direction. The convergence rate of −0.055 was mostly explained, again in similar magnitudes, by the convergence effect (−0.024) and the poverty elasticity effect (−0.022). But, saliently, munici- palities’ initial poverty also contributed with an effect of −0.009, equivalent to 16 percent of the speed of poverty convergence. This result supports the evidence in Tables 7 and 8 for this period, which suggested plausibly that 19 The computation of these elasticities through OLS yields the following values: −1.505 in 1992, −1.662 in 2000, −1.664 in 2005, and −1.553 in 2010. 38 starting out (very) poor in 2000 was associated with high growth rates in mean per capita income in the next five years. In a context of disappointing economic growth, such high rates could have been the result of the explosive expansion of cash transfers among the extreme poor, and of social spending in general, potentially having the double effect of bolstering per capita incomes enough to have reduced food poverty while exerting progressive changes in the distribution —which, in turn, may promote poverty reduction (see Figure 4, panel b ). To shed some light on the latter, we now explore the role of inequality. For starters, municipalities’ initial inequality tends to exert sizeable and signifi- cant adverse effects on subsequent growth rates in mean per capita income (Table 7), consistently with a large body of empirical literature on growth. Moreover, our data also reveals that initial inequality tends to curb the im- pact growth in mean per capita income has on food poverty reduction, thus aligning with cross-country empirical evidence that high inequality makes the poor to accrue a lower share of the gains from growth in income. For instance, in those municipalities with a Gini coefficient at or below one stan- dard deviation below the mean in 1992, equivalent to 0.37 or less, a one percent growth in mean per capita income over 1992-2014 would lead to a decline of food poverty of roughly 2 percent annually, whereas in those with an initial Gini of 0.48 or more, at or above one standard deviation above the mean, the poverty reduction would occur at 1.07 percent per year (Table 9, column 1 of panel a ). This tendency of food poverty to be less responsive to growth in more unequal municipalities is confirmed in all subperiods, and it generally holds when growth rates in mean per capita income are adjusted by initial poverty (panel b ), or even by initial inequality (panel c ) as gi (Pit ) = η (1 − Git−τ ) gi (yit ) + νit (9) 39 Table 9: Growth elasticities of poverty in low and high inequality contexts, 1992-2014 (1) (2) (3) (4) (5) (6) 1992-2014 1992-2000 2000-2014 2000-2005 2005-2010 2010-2014 a. Ordinary growth elasticities Municipalities with relatively low initial inequality −2.015*** −1.569*** −1.779*** −1.738*** −1.673*** −2.261*** gi (yit ) (0.143) (0.158) (0.273) (0.417) (0.096) (0.292) Obs. 370 370 313 313 371 336 R2 0.734 0.414 0.693 0.400 0.541 0.659 Municipalities with relatively high initial inequality −1.069*** −0.924*** −1.241*** −1.184*** −1.662*** −1.613*** gi (yit ) (0.050) (0.073) (0.239) (0.273) (0.214) (0.141) Obs. 364 364 344 344 342 364 R2 0.769 0.508 0.359 0.329 0.496 0.584 b. Poverty-adjusted growth elasticities Municipalities with relatively low initial inequality −2.588*** −2.357*** −1.561*** −3.904*** −3.019*** −2.561*** (1 − Pit−τ ) gi (yit ) (0.191) (0.335) (0.879) (1.265) (0.190) (0.397) Obs. 370 370 313 313 371 336 R2 0.636 0.434 0.305 0.264 0.584 0.533 Municipalities with relatively high initial inequality −1.676*** −1.468*** −1.829*** −1.875*** −2.611*** −2.045*** (1 − Pit−τ ) gi (yit ) (0.062) (0.104) (0.420) (0.352) (0.339) (0.253) Obs. 364 364 344 344 342 364 R2 0.785 0.497 0.400 0.396 0.525 0.519 c. Distribution-corrected growth elasticities Municipalities with relatively low initial inequality −3.042*** −2.336*** −2.383*** −2.418*** −2.403*** −3.149*** (1 − Git−τ ) gi (yit ) (0.221) (0.249) (0.375) (0.590) (0.137) (0.415) Obs. 370 370 313 313 371 336 R2 0.717 0.401 0.678 0.395 0.542 0.656 Municipalities with relatively high initial inequality −2.191*** −1.905*** −2.362*** −2.301*** −3.051*** −2.709*** (1 − Git−τ ) gi (yit ) (0.106) (0.154) (0.452) (0.518) (0.385) (0.236) Obs. 364 364 344 344 342 364 R2 0.769 0.515 0.367 0.339 0.479 0.579 Source : Author’s calculations. Notes : For each period under study, the table shows estimates of the parameter η in equations (5) (8) and (9), weighted by municipal population at the initial year. The dependent variable is the annualised growth in the food poverty rate. The growth rates in mean per capita income are the annualised changes at the municipality-level (at August 2014 prices). The intercepts are shown in Table 48 of the Online Appendix. Robust s.e. in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1 40 which yields a distribution-corrected growth elasticity of poverty, as proposed by Ravallion (1997), where Git−τ is the initial Gini coefficient. A closer examination of the data, however, suggest that the relationship be- tween initial inequality and the efficiency of growth to reduce poverty in a country with dramatic regional disparities is far from being linear. The non-linearity is confirmed in Figure 7. Even when growth elasticities are computed using the ordinary growth rate, there is a dim indication that food poverty rates over 1992-2014 were more responsive to growth in some highly unequal municipalities than in low-inequality counterparts (panel a ). This indication becomes clearer after penalising more the income growth rates in municipalities with relatively higher Gini coefficient in 1992 (panel b ). There- fore, sizeable changes in mean per capita income and (hence) in food poverty rates occurred not only among the poorest municipalities, as documented previously, but also among those with relatively high initial inequality. Figure 7: Efficiency of growth to reduce food poverty, by initial inequality levels; 1992-2014 Source : Author’s calculations. Notes : The area of symbols is proportional to municipali- ties’ total population. For visibility purposes the elasticities in both charts are capped at −10. Indeed, the distribution of municipalities according to their Gini coefficient 41 in 1992 reveals that poverty reduction tended to be slightly higher among the top 40 percent more unequal municipalities (Figure 8, panel a ). Inter- estingly, inequality among the latter also declined markedly over 1992-2014, which suggests that the magnitude of food poverty reduction observed among poorer municipalities was not the result of income gains only, but also that of progressive changes. It also suggests a plausible process of inequality con- vergence across municipalities, which is confirmed with a coefficient of −0.04 during 1992-2014 (significant at the 1 percent level) that results from the standard model for the annualised proportionate change in inequality, as in equations (3) or (6).20 In addition, it can be confirmed that in the majority of initially poorest municipalities where subsequent food poverty reduction took place, the latter was accompanied by a decline in the Gini coefficient (panel b ). Figure 8: Annualised rates of change in food poverty and inequality, 1992- 2014 Source : Author’s calculations. These results are suggestive, in general, of a decline in inequality in the coun- 20 The magnitude and significance of the inequality convergence parameter is robust to the specification that regress the annualised absolute difference in inequality levels on the enabou (1996). initial Gini coefficient, as in B´ 42 try over the period under study, which is confirmed by a population-weighted average reduction of −0.8 Gini points for the whole sample of municipalities during 1992-2014. This reduction, however, was far from being generalised across municipalities. About 71 percent of total municipalities, which con- centrate almost half of the country’s population, experienced a decline in in- equality above the national average, reaching −5.3 Gini points, and slightly above 4 percent of municipalities also improved their inequality level, though at a lower rate than the national figure, reaching only −0.4 Gini points. The remaining 25 percent of municipalities, which home the other half of the country’s population, experienced a deterioration in inequality of around 3.4 Gini points, on average. Despite this latter result, which is basically a reflection of the rebound of inequality in the country after 2010, it highlights that a vast majority of municipalities experienced, in general, progressive changes in their income distribution and that this occurred for most of the last quarter century: the population-weighted national average shows a de- cline of −1.2 and −4.1 Gini points in the 1990s and in the first decade of the 2000s, respectively. 7 Summing up Between 1992 and 2014, Mexico experienced relative stagnation in both growth in per capita income and poverty reduction and, by the latter year, subnational disparities were sizeable. These traits leave the impression that little has happened in the long run with the living standards of the popu- lation. In between those years, however, several changes did actually occur. On one hand, Mexico’s economy was hard hit by economic crises in both the 1990s and 2000s and, while they adversely affected overall poverty rates, their impacts seem to have had a regional bias. On the other hand, impor- tant social policy reforms were undertaken in those years, highlighting the introduction and aggressive expansion of CCT among the poorest. Coincid- ing with these reforms, overall poverty rates recorded a considerable reduc- 43 tion over the decade 1996-2006 and, since the early 2000s income inequality started to decline. How living standards at a high level of geographical disaggregation evolved in this context? The findings of this paper reveal the following answers. Firstly, mean per capita income in the poorest municipalities grew consistently faster than in richer counterparts, thus confirming, in general, that convergence occurred in a sizeable and significant magnitude, though the speed was faster after 2000 and heterogeneous between urban and rural municipalities and between those located in the north of the country and the rest. Secondly, the process of income convergence was relatively effective in reducing income disparities between municipalities. In general, regional disparities recorded their lowest levels by the late-2000s. Thirdly, growth in mean per capita income among poorer, converging mu- nicipalities was relatively efficient in reducing poverty headcount rates, thus suggesting that the process of income convergence effectively translated into an unambiguous process of poverty convergence. Fourthly, public spend- ing, in general, and the accelerated expansion of cash transfers in the first years of the 2000s and that of improved federal allocations to municipalities, in particular, had a positive impact on both convergence processes. Seem- ingly, increasing transfers had the double effect of bolstering sufficiently high growth rates in income among the poorest while exerting progressive changes in the distribution of income. The former effect, in a context of stagnant or disappointing overall economic growth, promoted sizeable reductions in food poverty rates, whereas declining inequality —and inequality convergence— eventually made growth rates more efficient in reducing subsequent poverty rates in the less advantaged municipalities. Finally, while all previous results are good news from an egalitarian per- spective, it is noticeable that the convergence processes partially took place because richer municipalities were losing ground or standing still at best. While this gives less cause for celebration, all subnational changes we anal- ysed in this paper reflect, in general, that the poorest regions in Mexico 44 have been able to achieve development gains, even in the face of non-trivial economic crises, which could have seriously undermined equity within the country. 45 References Aghion, P. and P. Bolton (1997). A Theory of Trickle-Down Growth and Development. The Review of Economic Studies 64 (2), 151–172. ¨ Alderman, H., M. Babita, G. Demombynes, N. Makhatha, and B. Ozler (2002). How Low Can You Go? Combining Census and Survey Data for Mapping Poverty in South Africa. Journal of African Economies 11 (2), 169–200. Alesina, A. and D. Rodrik (1994). Distributive Politics and Economic Growth. The Quarterly Journal of Economics 109 (2), 465–490. Banerjee, A. V. and E. Duflo (2003). Inequality and Growth: What Can the Data Say? Journal of Economic Growth 8 (3), 267–299. Barro, R. J. and X. Sala-i Martin (1991). Convergence across States and Regions. Brookings Papers on Economic Activity 22 (1), 107–182. Barro, R. J. and X. Sala-i Martin (1992). Convergence. Journal of Political Economy 100 (2), 223–251. Barro, R. J. and X. Sala-i Martin (1995). Economic Growth. Cambridge, MA: The MIT Press. Baumol, W. J. (1986). Productivity Growth, Convergence, and Welfare: What the Long-Run Data Show. The American Economic Review 76 (5), 1072–1085. Bedi, T., A. Coudouel, and K. Simler (2007). More than a Pretty Picture: Using Poverty Maps to Design Better Policies and Interventions. Wash- ington DC: World Bank Publications. enabou, R. (1996). Inequality and Growth. In B. S. Bernanke and J. J. B´ Rotemberg (Eds.), NBER Macroeconomics Annual 1996, Volume 11. Lon- don: The MIT Press. 46 Bourguignon, F. (2003). The Growth Elasticity of Poverty Reduction: Ex- plaining Heterogeneity Across Countries and Time Periods. In T. S. Eicher and S. J. Turnovsky (Eds.), Inequality and Growth, Theory and Policy Im- plications. Cambridge, MA: The MIT Press. Chiquiar, D. (2005). Why Mexico’s Regional Income Convergence Broke Down. Journal of Development Economics 77 (1), 257–275. Clarke, G. R. G. (1995). More Evidence on Income Distribution and Growth. Journal of Development Economics 47 (2), 403–427. Cuaresma, J. C., S. Klasen, and K. M. Wacker (2017). Is There Poverty Convergence? Discussion paper 1711, Gutenberg School of Management and Economics, Mainz. Deininger, K. and L. Squire (1998). New Ways of Looking at Old Issues: Inequality and Growth. Journal of Development Economics 57 (2), 259– 287. Dollar, D., T. Kleineberg, and A. Kraay (2016). Growth Still is Good for the Poor. European Economic Review 81, 68–85. Dollar, D. and A. Kraay (2002). Growth is Good for the Poor. Journal of Economic Growth 7 (3), 195–225. Durlauf, S. N. (1996). A Theory of Persistent Income Inequality. Journal of Economic Growth 1 (1), 75–93. Durlauf, S. N. and P. A. Johnson (1995). Multiple Regimes and Cross-country Growth Behaviour. Journal of Applied Econometrics 10 (4), 365–384. Elbers, C., J. O. Lanjouw, and P. Lanjouw (2003). Micro-Level Estimation of Poverty and Inequality. Econometrica 71 (1), 355–364. opez-Calva, C. Rodr´ Enamorado, T., L. F. L´ an, and H. Winkler ıguez-Castel´ (2016). Income Inequality and Violent Crime: Evidence from Mexico’s Drug War. Journal of Development Economics 120, 128–143. 47 Esquivel, G. (1999). exico, 1940-1995. Convergencia Regional en M´ El omico 66 (264(4)), 725–761. Trimestre Econ´ Esquivel, G., N. Lustig, and J. Scott (2010). Mexico: A Decade of Falling Inequality: Market Forces or State Action? opez-Calva and In L. F. L´ N. Lustig (Eds.), Declining Inequality in Latin America: A Decade of Progress?, A Decade of Progress?, pp. 175–217. Brookings Institution Press. Ferreira, F. H. G. and M. Ravallion (2011). Poverty and Inequality: The Global Context. In B. Nolan, W. Salverda, and T. M. Smeeding (Eds.), The Oxford Handbook of Economic Inequality. Oxford: Oxford University Press. Foster, J., J. Greer, and E. Thorbecke (1984). A Class of Decomposable Poverty Measures. Econometrica 52 (3), 761–766. ekely (2008). Is Economic Growth Good for the Poor? Foster, J. E. and M. Sz´ Tracking Low Incomes Using General Means. International Economic Re- view 49 (4), 1143–1172. Fosu, A. K. (2017). Growth, Inequality, and Poverty Reduction in Developing Countries: Recent Global Evidence. Research in Economics 71 (2), 306– 336. Galor, O. (1996). Convergence? Inferences from Theoretical Models. The Economic Journal 106 (437), 1056–1069. Galor, O. and J. Zeira (1993). Income Distribution and Macroeconomics. The Review of Economic Studies 60 (1), 35–52. ıa-Verd´ Garc´ u, R. (2005). Income, Mortality, and Literacy Distribu- tion Dynamics Across States in Mexico: 1940-2000. Cuadernos de ıa 42 (125), 165–192. Econom´ Grimm, M. (2007). Removing the Anonymity Axiom in Assessing Pro-poor Growth. The Journal of Economic Inequality 5 (2), 179–197. 48 Hoff, K. (1996). Market Failures and the Distribution of Wealth: A Per- spective from the Economics of Information. Politics & Society 24 (4), 411–432. Knowles, S. (2005). Inequality and Economic Growth: The Empirical Rela- tionship Reconsidered in the Light of Comparable Data. The Journal of Development Studies 41 (1), 135–159. Kraay, A. (2006). When is Growth Pro-poor? Evidence from a Panel of Countries. Journal of Development Economics 80 (1), 198–227. Lin, P.-C. and H.-C. R. Huang (2011). Inequality Convergence in a Panel of States. The Journal of Economic Inequality 9 (2), 195–206. Ljungqvist, L. (1993). Economic Underdevelopment: The Case of a Missing Market for Human Capital. Journal of Development Economics 40 (2), 219–239. en (2006). A Normal Relationship? Poverty, Growth, Lopez, J. H. and L. Serv´ and Inequality. Policy Research Working Paper 3814, World Bank, Wash- ington, D.C. L´ ´ Mel´ opez-Calva, L. F., A. on, L. Rodr´ endez, E. G. Rasc´ ıguez-Chamussy, and ekely (2008). El ingreso de los hogares en el mapa de M´ M. Sz´ exico. El omico LXXV (4) (300), 843–896. Trimestre Econ´ opez-Calva, and E. Ortiz-Juarez (2016). Deconstructing Lustig, N., L. F. L´ the Decline in Inequality in Latin America. In K. Basu and J. Stiglitz (Eds.), Inequality and Growth: Patterns and Policy. Volume II: Regions and Regularities. Nueva York: Palgrave Macmillan. Persson, T. and G. Tabellini (1994). Is Inequality Harmful for Growth? The American Economic Review 84 (3), 600–621. Piketty, T. (1997). The Dynamics of the Wealth Distribution and the Interest Rate with Credit Rationing. The Review of Economic Studies 64 (2), 173– 189. 49 Quah, D. T. (1993). Galton’s Fallacy and Tests of the Convergence Hypoth- esis. The Scandinavian Journal of Economics 95 (4), 427–443. Quah, D. T. (1996). Convergence Empirics across Economies with (Some) Capital Mobility. Journal of Economic Growth 1 (1), 95–124. Quah, D. T. (1997). Empirics for Growth and Distribution: Stratification, Polarization, and Convergence Clubs. Journal of Economic Growth 2 (1), 27–59. Ravallion, M. (1995). Growth and Poverty: Evidence for Developing Coun- tries in the 1980s. Economics Letters 48 (3-4), 411–417. Ravallion, M. (1997). Can High-inequality Developing Countries Escape Ab- solute Poverty? Economics Letters 56 (1), 51–57. Ravallion, M. (1998). Does Aggregation Hide the Harmful Effects of Inequal- ity on Growth? Economics Letters 61 (1), 73–77. Ravallion, M. (2001). Growth, Inequality and Poverty: Looking Beyond Averages. World Development 29 (11), 1803–1815. Ravallion, M. (2003). Inequality Convergence. Economics Letters 80 (3), 351–356. Ravallion, M. (2004). Pro-Poor Growth: A Primer. Policy Research Working Paper, World Bank. Washington, DC. Ravallion, M. (2007). Inequality IS Bad for the Poor. In J. Micklewright and S. P. Jenkins (Eds.), Inequality and Poverty Re-Examined. Oxford: Oxford University Press. Ravallion, M. (2012). Why Don’t We See Poverty Convergence? American Economic Review 102 (1), 504–523. ıguez-Castel´ Rodr´ an, C., K. Cadena, and L. Moreno (2017). Evaluacion de largo plazo de los impactos distributivos del Fondo de Infraestructura So- cial Municipal (FISM). Technical Report, Banco Mundial, Washington, DC. 50 ıguez-Oreggia, E. (2007). Winners and Losers of Regional Growth in Rodr´ on Econ´ Mexico and their Dynamics. Investigaci´ omica LXVI (259), 43–62. ıguez-Pose, A. and J. S´ Rodr´ anchez-Reaza (2005). Economic Polarization Through Trade: Trade Liberalization and Regional Growth in Mexico. In R. Kanbur and A. J. Venables (Eds.), Spatial Inequality and Development. New York: Oxford University Press. Sala-i Martin, X. (1996). Regional Cohesion: Evidence and Theories of Regional Growth and Convergence. European Economic Review 40 (6), 1325–1352. Sala-i Martin, X. (2006). The World Distribution of Income: Falling Poverty and... Convergence, Period. The Quarterly Journal of Economics 121 (2), 351–397. Su, J.-J. (2003). Convergence Clubs among 15 OECD Countries. Applied Economics Letters 10 (2), 113–118. Sz´ ekely, M., L. F. L´ ´ Mel´ opez-Calva, A. ınez, E. G. Rasc´ endez Mart´ on, and ıguez-Chamussy (2007). L. Rodr´ Poniendo a la pobreza de ingresos y exico. a la desigualdad en el mapa de M´ ıa Mexicana. Nueva Econom´ ´ Epoca XVI (2), 239–303. ıa, C., S. Klasen, and S. Vollmer (2016). The Distribution Villalobos Barr´ Dynamics of Human Development in Mexico 1990–2010. Review of Income and Wealth 62 (S1), S47–S67. World-Bank (2018). Mexico. Systematic Country Diagnostic. Report 132634- MX, World Bank, Washington, D.C. 51 Annex: Summary statistics of the income, poverty and inequality dataset 1992 2000 2005 2010 2014 Mean S.D. Mean S.D. Mean S.D. Mean S.D. Mean S.D. a. Descriptive statistics of the income, poverty, and inequality dataset Simple averages across municipalities Real per capita income (MXN$ August 2014) 1,738.0 842.7 1,578.3 932.2 1,699.3 902.4 1,663.5 881.0 1,966.9 967.4 Gini coefficient 0.426 0.055 0.385 0.062 0.379 0.053 0.341 0.045 0.384 0.039 Poverty headcount (% of population) Food poverty 41.6 21.4 44.5 25.4 37.7 22.0 38.7 23.7 35.6 19.5 Capabilities poverty 49.6 21.4 52.3 24.9 46.0 22.4 47.5 24.0 44.0 20.1 Assets poverty 68.6 17.9 70.7 20.2 67.0 19.3 69.7 20.1 65.4 18.1 Population-weighted averages of municipality figures Real per capita income (MXN$ August 2014) 2,765.2 1,424.6 2,870.8 1,492.9 2,966.1 1,295.8 2,755.6 1,215.5 3,245.6 1,452.1 Poverty headcount (% of population) Food poverty 25.9 19.3 24.1 21.7 19.7 17.4 20.8 17.7 20.9 15.6 Capabilities poverty 33.6 20.4 31.5 23.0 26.7 18.8 28.9 18.9 28.3 17.0 Assets poverty 55.7 19.4 53.3 21.7 49.1 18.6 54.1 18.3 50.6 17.7 Average population by municipality 34,016 100,736 40,525 120,304 42,839 127,807 45,817 131,249 49,742 141,243 b. Summary statistics on municipalities’ public spending and revenues 52 Simple averages across municipalities Public spending (per capita, MXN$ August 2014) 80.3 103.2 158.3 124.3 254.1 160.0 343.6 199.7 418.8 308.6 Public sector payroll 25.0 50.3 41.6 41.4 76.4 69.0 91.4 80.4 105.1 92.0 Transfers and subsidies 7.2 14.1 21.9 23.8 25.2 22.5 34.5 45.1 25.5 32.9 Public investment 21.4 30.6 40.2 48.2 76.0 52.4 123.8 93.3 162.8 185.0 Public debt 5.8 10.6 7.0 18.3 11.5 15.3 15.5 18.6 14.0 19.0 Public revenues (per capita, MXN$ August 2014) 80.3 103.2 157.9 124.5 254.0 160.4 344.1 200.3 419.3 309.7 Taxes 8.1 16.9 5.4 11.5 9.4 17.8 10.8 19.4 12.8 24.1 Unconditional federal transfers (participaciones ) 53.8 76.4 95.8 86.5 126.1 122.1 145.5 131.2 163.5 164.1 Conditional federal transfers (Ramo 33 ) − 17.5 59.2 50.0 88.4 44.5 140.7 85.2 202.1 175.2 Average CCT beneficiary families by municipality − − 1,156 1,607 2,077 2,815 2,413 3,439 2,527 3,915 Number of municipalities covered in the dataset 2,361 2,361 2,361 2,361 2,361 Total population covered in the analysis 80,310,818 95,678,853 101,144,021 108,174,343 117,439,680 Total population in the country 81,249,645 97,483,412 103,263,388 112,336,538 119,530,753 Total CCT beneficiary families in the country − 2,437,297 4,892,284 5,682,617 5,965,275 Source : Author’s calculations. Panel a : based on ENIGH and census datasets; Panel b : based on the public finance dataset of the National Institute of Statistics and Geography (INEGI) and on administrative records from the flagship CCT program —introduced as Progresa in 1997 and rebranded as Oportunidades in 2002 and more recently as Prospera.