An assessment of the Bolsa Família program’s main social effects

(1)

1

An assessment of the Bolsa Família

program’s main social effects

Fabio Mangia

10418148

BSc Economics

(2)

2 1 - Introduction

In his book Em Busca de um Novo Modelo (2002) Brazilian economist Celso Furtado discussed two fundamental features of Latin American peripheral economies: tendencies toward indebtedness and income concentration. While the former implicated in higher vulnerability to foreign capital, the latter reflected the failure of growth models at fostering social development. Although not negative in itself, income inequality has been linked to crime incidence (Fajnzylber, Lederman and Loayza, 2002, p.25) and sustaining historical economic oppressions. Conditional cash transfer systems have become an increasingly popular partial solution to these problems related to income inequality and extreme poverty.

In Brazil the very first CCT program was implemented in 1995 when the government commenced to financially provide for families that earned half the minimum wage and had offspring between 7 and 14 years of age (Friend, 2002, p.1043). In the following year the program was expanded to a national level and was named Programa de Erradicação do Trabalho Infantil (Program to Eradicate child labor). By 1997 these initiatives had drawn considerate attention and the government was granted authorization by the national legislature to cover up half the costs of CCT programs on a municipal level. In 2001 Fernando Henrique Cardoso, Brazil’s then president issued a law that contributed to the establishment of Bolsa Escola (School Grant), the main precursor to Bolsa Família (Family Grant). This new program differed from the others to the extent that it actively promoted school attendance as it introduced a specific requirement - recipients had to provide evidence that their children attended school regularly.

In 2003 president Lula inherited several CCT programs from Cardoso’s administrations, each of which employed different criteria for selection. This thus meant an inefficiency and logistic problem which was solved by the unification of all the programs into a single one with the same criteria (Friend, 2002, p.1043). Besides unifying these different social programs into one, president Lula also raised the stipend value that each family would receive, thus officially creating the Bolsa Família. This endeavor required substantial effort should one analyze the figures: 5 million families received up to R$45 from Bolsa Escola in 2003 and those numbers respectively increased to 11.1 million and R$107 by the end of 2004.

Organization wise, the Ministério de Desenvolvimento Social (Ministry of Social Development or MDS) administers the Bolsa Família by registering all recipients in the Cadastro Único (Single File) (Friend, 2002, pp. 1043-1044). Families are divided into two categories: those in poverty, that is, with a per capital monthly income lower than R$140 and those in extreme poverty, who have a monthly income lower than R$70. In order to be eligible to receive the stipend families need to satisfy specific conditions that aim to enhance the prospects of future generations. Building on the conditions established by the Bolsa Escola, children need to meet minimum school attendance requirements and receive essential healthcare, such as vaccinations and regular checkups (2002, p.1044). On top of the educational and health requirements, women are the designated recipients of the cash transfers (Brauw, Gilligan, Hoddinott and Roy, 2013, p. 487). These measures in particular aims to not only foster the social development of marginalized families but also to empower women and thus mitigate the gender gap that still affects Brazil. Based on the completion of all requirements, the family income and the number of children in the household, recipients’ stipend varies from R$22 and R$200.

(3)

3

It is clear that conditional cash transfer systems are policies designed to empower marginalized groups and thereby reduce severe income inequality. However, despite its goals, the Bolsa Família has faced much criticism and led to a political polarization towards the issue (see Folha de São Paulo, 1/05/2014). While some media outlets deem it a populist demagogic policy, others praise its attempt to eradicate poverty in a country as unequal as Brazil. No policy is above scrutiny and I aim to examine to what extent the CCT program has met its goals one decade after its implementation. More specifically, I will analyze whether the Bolsa Família has had a statistically significant effect on inequality measured by the GINI index, secondary education enrollment, poverty , unemployment and mortality rate through multivariate regression analysis.

The structure of this research paper is as follows. Section 2 consists in a literary overview of previous relevant research regarding CCT effects. Section 2a offers a more in depth conceptualization of the program, whereas 2b-d respectively illustrate what has been verified with respect to the CCT program’s effects on poverty and inequality, school enrollment and unemployment. Section 3 defines the methodology implemented, underlines some research limitations and specifies the format of the data used. It also provides an intuition behind the regression designs and describes the initial expectations. Then, section 4 consists in the main regression tables and the results analysis. Finally, section 5 concludes the paper and summarizes the Bolsa Família’s main social effects.

2- Theoretical framework

2a) A conceptualization of Conditional Cash Transfer Programs

According to Das, Do and Ozler (2005, p. 58) Conditional Cash Transfer programs are employed when individual and society’s preferences differ. An example of this would be when a family decides to send their son to work as opposed to ensure he attends school. While the wage earned by the offspring is relevant to this short-term orientated family, society will benefit from human capital accumulation and thus prefers that he continues his education. As such, conditional cash transfer systems stem from a paternalistic framework and thus aim to align conflicting interests in order to maximize welfare. The authors defended that this asymmetry between society’s wishes and individual behavior is often entailed by irrationality. From this perspective then, a CCT program is necessary to protect people from their own irrationalities. Its efficiency, however, depends on how it manages market failures ensued as a result of mismatched preferences and how it allocates resources to a target group.

Several market failures may require the use of CCT programs. Firstly, Das, Do and Ozler (2005, p. 64-65) underscored direct externalities, specifically physical and learning ones. Two different experiments are mentioned to illustrate their effects. In one study conducted in Kenya a program provided free deworming treatment to rural primary school students. In this case the households did not consider the benefits for the whole community of their taking deworming pills. In other words, individuals were not aware of the positive externality of their choices and thus deworming pills were under consumed. Hence the CCT program provided them the adequate incentive to contribute more of this public good. Another study found a similar issue with learning externalities. Because the adoption of new technologies involves a learning curve and thereby costly

(4)

4

experimentation and opportunity cost, fewer individuals or households are willing to make the first move. This clearly is related with the free rider problem and leads much like the case with deworming pills to the under-provision of a public good and to inefficiency. CCT programs in this case too can be used to overcome the free rider problem by providing subsidies for the adoption of new technology and thus making use of the positive learning externalities (Foster and Rosenweig in Das et al, 2005, p.65).

Household bargaining is another market failure that can be rectified by CCT programs. The authors (2005, pp. 65-66) mentioned literature on child labor and how it arises from a mismatch between parent’s and children’s preferences. Essentially, were children from low income families able to pay their families for their education, education enrollment and completion levels would be higher. Notwithstanding, this is not tangible and in most cases education decisions are made by parents who may be more short run orientated. In fact, Kochar (in Das et al, 2005, p. 65) defended that parents’ and children’s rate of return to their education are significantly different, which logically leads to potentially higher rates of child labor. The authors then argued that conditional cash transfer program can remedy this conflict of interests by reducing the parents’ incentives to make their offspring work.

Assuming that a policymaker has thoroughly analyzed the externalities and the target group and decided to implement a CCT program, potential problems still remain. The researchers (2005, pp. 66-67) raised two relevant issues on this stage, namely participation and fungibility of the conditioned-on commodity. The former is related to the amount of the transfer and the cost of the condition. Essentially, the target group will only choose to take part of the program if the benefit of doing so is higher than the cost. If we consider the previously mentioned problem of child labor, participation will only be low if the amount of the cash transfer is relatively inferior to the foregone wage the child could earn for the household. Fungibility, however, arises when the individual has access to a close substitute to the conditioned-on commodity and can therefore offset the distortion imposed by the CCT program through a lower consumption of the substitute. By doing so the recipient ensures that overall amounts are not changed by the program even if the established condition is fulfilled.

The authors (2005, p. 67) offered an example to better elucidate this problem. They delineated a program which provides parents with funds to purchase school material in order to foster children’s academic performance. Such a program may have a high participation rate but minimum effects on learning outcomes if the households reduce their own expenditure on school supplies. For instance, were the family already purchasing school materials, then the cash transfer would not necessarily make them buy more. This substitution may take form of a lower consumption of a close substitute, changing patterns of consumption altogether or reallocating investments within the household. The researchers offered one illustration for each respective type of substitution: eating less at home when given food in school and sending fewer boys to school when girls are afforded a stipend.

2b) Bolsa Família, poverty and inequality

Poverty is a function of inequality and the main issue a conditional cash transfer system aims to solve in both the short and long-term (Modesto & Cardoso, 2010, p.26). Given its importance, many a study has been published in this matter reaching similar conclusions while employing

(5)

5

different methodologies. Hoffman (2007), Modesto and Cardoso (2010) and Soares (2006) used a parametric approach which decomposed the GINI index in order to estimate the effects of the Bolsa Família on inequality. More specifically, their approach was to calculate the concentration curve. Similarly to the Lorenz curve, this analytical tool indicates how progressive each type of income is. With the aid of the concentration curve it becomes possible to verify what extent of a given cash transfer system was part of the income of a given percentage of society (Modesto and Cardoso, 2010, p.34).

Unlike the previously mentioned authors, Barros, Carvalho and Franco (2007) employed a different approach. By performing non-parametric simulations they decomposed the income distribution in all its components. These researchers firstly defined (2007, p.42) that a fall in inequality is decomposed in two main factors: changes in the income earned in work and income earned elsewhere. Then, they divided this second category into seven others which ranged from rent and dividends to conditional cash transfer system and investigated how each different income source affected inequality. This approach is arguably more sophisticated and complex, which is why it has not been used by many other studies. However, regardless of complexity, Barros et al. (2007), Modesto et al. (2010) and Soares (2006) all found that the Bolsa Família was responsible for a reduction in both poverty and inequality since its implementation in October 2003. From 2002-2008 Hoffman (2006, p.55) found a reduction in the GINI index of 0.0185 and that 31.4% of this reduction is due to cash transfer systems. Similarly, Soares verified that 20% of the GINI reduction from 2001-2004 was due to the Bolsa Família.

2c) Bolsa Família and school enrollment

Higher education enrollment and progression can yield higher human capital and in doing so contribute to breaking the inter-generational transmission of poverty (Hall, 2006, p.4). Different studies have addressed to what extent conditional cash transfer systems induce higher school enrollment and progression. In the case of the Bolsa Família, Glewwe and Kassouf (2012, p.516) found that the program increased both enrollment and grade promotion while also reducing dropping out. The authors highlighted that in grades 1-4 and 1-8 respectively enrollment increased by 5.5% and 6.5% and dropout rates decreased by 0.5%. Their research method consisted in running a regression on the education outcome of interest, be it the enrollment or dropout rate, and explanatory variables that ranged from family characteristics to whether or not the given family receives a stipend (2012, p.509).

A more complex research approach was carried out by Neto (in: Castro and Modesto, 2010, p.54). Based on variables concerning family and child characteristics, he estimated the probability that a given child is eligible to the Bolsa Família and then used a propensity score (logit) model and matching process that aimed to synthesize the information within the variables that affected participation in the program. Neto found a positive effect of the conditional cash transfer program on school attendance (2010, p.68). However, this positive effect was larger in the Northeast then in the Southeast of Brazil – Neto speculates that this is due to historical differences. The Northeast has usually faced more poverty and inequality then the South, and therefore the CCT effects would logically be lower in the overall wealthier South.

It is worth mentioning another study regarding this matter but that was conducted in Mexico. Brauw and Hoddinot analysed the effects of the Mexican conditional cash transfer program

(6)

6

named Progresa – which is similar to the Bolsa Família with regards to its requirements and targeting. The authors (2011, p.368) verified considerable benefits associated with employing CCT programs for entry into secondary lower school. They also underscored that a growing body of evidence suggests that programs such as Progresa or Bolsa Família have a wide range of positive effects on welfare.

2d) Bolsa Família and unemployment

Part of the criticism towards the Bolsa Família program concerns the idea that it motivates the beneficiaries not to work and therefore that it leads to unemployment supported by tax-payers. Teixeira (2010, p.92) found evidence that conflicts this view. By using a model proposed by Becker (1976), she (2010, p.103) verified that the PBF has no effect on the probability of men or women working. There was, however, a significant effect on the program’s impact on weekly working hours but Teixeira argued it was small in magnitude. The model applied in her analysis consisted in a regression on labor hours and compared the predictions between the beneficiary (treatment) and non-beneficiary (control) households (2010, p.93).

Brauw, Gilligan, Hoddinott and Roy found similar results. The authors (2015, p.424) studied the impacts of the CCT program on the labor supply amount of recipient. More specifically, they gathered data from both rural and urban areas and analyzed the Bolsa Família effects across them. They (2015, p.445) concluded that the program has no aggregate effect on labor force participation. In addition, they detected a considerable shift of household working hours from the formal sector to the informal one. There was also a reallocation of labor supply within the household in the case of rural areas – female recipients refrained from working whereas male ones worked more hours.

3-Data, methodology and expectations 3a) Data

The aim of this paper is to test and gather the main results verified by previous researches on the effects of the Bolsa Família with respect to inequality, poverty, education and unemployment. In order to achieve this goal I ran a total of 5 regressions using different data sets, which were collected from the World Bank website, Instituto de Pesquisa Econonômica Aplicada (IPEA) and the Portal da Transparência (Transparency Portal). The data analyzed here refers to the period of 1982 up to 2013. Most of the data was gathered at the World Bank Website, except for the Bolsa Família yearly expenditure and the real minimum wage. The former was collected at the website that displays the Brazilian government budget and annual expenditures, that is, Portal da Transparência. The yearly real minimum wage was calculated on monthly figures found at the IPEA website.

By looking at the data set of the GINI index and the Bolsa Família annual expenditure, one can notice negative correlation between them – as the expenditure with the CCT program increased, the GINI index decreased. Figures 1A and 1B bellow evidence precisely that. Although inequality itself is affected by other factors including unemployment, minimum wage and economic shocks, this already indicates that the program might be partially responsible for its decrease since 2004.

(7)

7

Brazilian Gini Index from 1982 up to 2013

Figure 1A

Yearly expenditure on the Bolsa Família program in Reais from 1982 up to 2013

50 55 60 65 g in i 1980 1990 2000 2010 2020 year 0 5 .0 e + 0 9 1 .0 e + 1 0 1 .5 e + 1 0 2 .0 e + 1 0 2 .5 e + 1 0 b ff u ll 1980 1990 2000 2010 2020 year Figure 1B

(8)

8

In spite of being motivated by previous authors who applied complex methodologies, such as non-parametric simulations (Barros et. Al, 2007) or GINI decompositions (Hoffman, 2007), the research hereby presented made use of a more simple approach due to specific logistical limitations. The main limitations were lack of technical expertise and the availability of the data itself. Data on social indicators Government websites including the Institute of Geography and Statistics (IBGE) and Institute of Applied Economic Research (IPEA) provided little information on the Brazilian social context over the second half of the twentieth century. Most data series began only at the 1990’s and specific years were often missing. This meant that the World Bank became the primal data source of this paper. While more reliable and with more abundant figures, the time series were only consistent from 1982 onwards, which in turn meant that a limited sample of 33 observations was used.

Another issue that deserves mentioning is the use of interpolation. Because few samples were entirely complete, it was necessary to resort to interpolation so that regressions could be run. Specifically, it was necessary to interpolate the following variables: gini, unemployment rate, poverty, progression to secondary education. The list of all variables and their specification is offered in the following table:

Variable Specification

gini gini index – ranging from 0-100

poverty2 percentage of the Brazilian population who lives with under 2 dollars a day

unemployment rate percentage of unemployment individuals in the Brazilian population

mortality rate number of under-five year olds deaths per 1000 births

progression2nd percentage of students who concluded their primary education and continued to secondary school

minwage Brazilian yearly real minimum wage

gdpgrowth Brazilian gross national product yearly growth urban yearly percentage growth of Brazilian urban

population

bfgdp Bolsa Família expenditure as a percentage of gross national product

inflation Percentage increase in consumer prices

3b) Methodology and expectations

The model design of each regression in this paper was shaped by intuition and data limitation. As such, the multivariate regressions analyzed here utilized specific independent variables that were not necessarily used by previous researchers including Barros (2007) or Hoffman (2007). In this section the intuition behind each of the five models will be defended. The main tables for the regressions can be found in the next section while the full tables of all regressions – including those

(9)

9

that treated the Bolsa Família as a binomial variable – and the regression diagnostics can be found in the appendix.

While ten regressions were run, five are found in this session and the other half is in the appendix. This division stems from the fact that they were nearly the same, except for a slight alteration on the explanatory variable of the Bolsa Família. The CCT program was measured firstly as a percentage of GDP (bfgdp) and then as binomial variable (bf) in order to test whether the small yearly changes on the Bolsa Familia jeopardized its ability to explain the inequality, poverty, unemployment, life expectation and enrollment to secondary education. Although different in absolute values, the coefficients for the CCT program shared the same sign, which means they both indicated the same trend of the effect on inequality. Next the five regressions will be individually addressed.

3b.1) Regression 1 – measuring BF’s effect on inequality

𝑔𝑖𝑛𝑖 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑚𝑖𝑛𝑤𝑎𝑔𝑒 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡 + 𝛽5𝑔𝑑𝑝𝑔𝑟𝑜𝑤𝑡ℎ + 𝛼

Other variables expected to affect this index were the minimum wage, inflation and unemployment rate and GDP growth. Although the exact effects minimum wage can have on inequality are still debatable, there is little question that equality is at least partially determined by minimum wage matters. Hamidi and Terell (2001, p.10) for example found that raises in the minimum wage in the case of Costa Rica decreased overall inequality while Golan, Perloff and Wu (2001, p.29) found the inverse in the United States labor market. With little information other than that the Brazilian informal sector represents about fifteen to twenty percent of GDP and that Brazil has a relatively low income per capita, I expected minimum wage to have a negative effect on inequality. The inflation rate, on the other hand, was expected to have a positive effect on inequality as a rise in prices, particularly of non-superfluous goods, hits the poor the hardest. Given Brazil’s overall lack of high skilled workers, I too assumed that unemployment would have a positive effect on inequality because the supply of low-skilled to high-skilled workers is high. In addition, I expected GDP growth to have a positive effect on inequality because Brazil’s growth model has led to much income concentration and the Bolsa Família is one of the few policies that aim to mitigate this issue on a larger scale. Therefore, I expected the CCT program to decrease inequality.

3b.2) Regression 2 – measuring BF’s effect on the morality rate

𝑚𝑜𝑟𝑡𝑎𝑙𝑖𝑡𝑦 𝑟𝑎𝑡𝑒 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑚𝑖𝑛𝑤𝑎𝑔𝑒 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑔𝑑𝑝𝑔𝑟𝑜𝑤𝑡ℎ + 𝛽5𝑝𝑜𝑣𝑒𝑟𝑡𝑦2 + 𝛼

The government income distribution policy was expected to have a negative effect on the mortality rate because ideally the stipends’ would take their children to health centers more often and thus decrease the death rate of under-five year olds. For similar reasons to those mentioned on the previous paragraph minimum wage was assumed to have a negative effect. It is fair to assume that the majority of stipends have a low level of education and thus most of their income comes from the informal sector and minimum wage jobs. Should this be true, a raise in the minimum wage would entail higher living standards for the target population and thereby decrease infant deaths. Given this economic fragility of the target population, inflation can be expected to have a positive effect on the mortality rate. Finally, GDP growth can be expected to have a negative effect if one assumes that it will be followed by higher spending on health care. However, given Brazil’s history of

(10)

10

low government spending in social services, it could be reasonable to assume that GDP-growth would not necessarily lower the mortality rate. The final variable, poverty, logically is expected to have a positive effect.

3b.3) Regression 3 – measuring BF’s effect on enrollment to secondary education

𝑝𝑟𝑜𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛2𝑛𝑑 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑚𝑖𝑛𝑤𝑎𝑔𝑒 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑔𝑑𝑝𝑔𝑟𝑜𝑤𝑡ℎ + 𝛼

Because of its requirements the CCT program was expected to increase the rate of enrollment to secondary education. The same, however, cannot be said about the other variables. A higher minimum wage increases the opportunity cost for low family incomes who may be more tempted to send their children to work. A similar argument can be made for inflation – higher prices may lead to a time inconsistency problem for low income families which may lead to the offspring working instead of resuming to study. Once again the expectations for GDP-growth are mixed; on the one hand, a booming economy may stimulate government spending on social services and increase the rates of return on education but on the other hand Brazil has not historically spent much on public education. It is worth mentioning that when this model included the variable unemployment the Bolsa Família effect was statistically non-significant. However, one could consider that higher unemployment is not necessarily likely to prevent children from low income families from studying because unemployment implies a lower opportunity cost for education. As such, unemployment was removed from this model.

3b.4) Regression 4 – measuring BF’s effect on unemployment

𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡 𝑟𝑎𝑡𝑒 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑚𝑖𝑛𝑤𝑎𝑔𝑒 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑔𝑑𝑝𝑔𝑟𝑜𝑤𝑡ℎ + 𝛽5𝑢𝑟𝑏𝑎𝑛 + 𝛼

Regression 4a analyzed the legitimacy of one of the main criticisms towards the Bolsa Família – that it somehow induces stipends to not work and live off government benefits. Considering that the Brazilian minimum wage in 2013 was 678 reais per month and that the average amount received by stipends in the same year was 152.67 reais, it is hardly conceivable that recipients would be inclined not to work. It is also possible to argue that assuming unemployment will increase with the Bolsa Família is a short-sighted view that dehumanizes the target groups for it implies that they have no ambition to develop themselves through education or work. Figures 2a and 2b bellow suggest graphically that there is no positive correlation between the amount invested on the Bolsa Família and the unemployment rate in Brazil. As a result, I expect the CCT program to have little to no effect on unemployment; it could even be possible that it has a negative effect on unemployment as the stipends might be more motivated to work with this governmental aid.

The other variables assumed to have an effect on unemployment include minimum wage, inflation rate, GDP growth and the percentage growth of Brazilian urban population. Minimum wage can be associated with unemployment when it is not effectively coordinated between unions and employers (Nickel, p. 72, 1997). Considering the Phillips curve, one could assume a negative relationship between the inflation and unemployment in the short-term but given that the time frame of the sample is over three decades this prediction can be put into question.

(11)

11

Unemployment in Brazil between 1982 and 2013

Figure 2a

Yearly expenditure with the Bolsa Família in Reais from 1982 up to 2013

Figure 2b 2 4 6 8 10 u n e mp lo ym e n t 1980 1990 2000 2010 2020 year 0 5 .0 e + 0 9 1 .0 e + 1 0 1 .5 e + 1 0 2 .0 e + 1 0 2 .5 e + 1 0 b ff u ll 1980 1990 2000 2010 2020 year

(12)

12

GDP growth, however, is expected to have a negative effect on unemployment due to the following reasoning: as the economy grows, so does the internal demand for goods and services, which in turn leads to a higher demand for labor. The opposite can be argued for the percentage growth of urban population: as labor supply in urban areas increases for a given fixed demand, unemployment can follow.

3b.5) Regression 5 – measuring BF’s effect on poverty

𝑝𝑜𝑣𝑒𝑟𝑡𝑦2 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑢𝑟𝑏𝑎𝑛 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡𝑟𝑎𝑡𝑒 + 𝛼

The final model 5a analyses whether poverty measured as the percentage of the Brazilian population earning less than two dollars a day is affected by the CCT program, the percentage growth in urban population, inflation and unemployment rate. The Bolsa Família is expected to have a negative effect on poverty given that it provides extra financial assistance to target groups and aims to eradicate intergenerational poverty by requiring that children attend school frequently. The opposite can be said with regards to the expectation of all the other variables’ effect. Inflation and unemployment rate can all impoverish fragile populations further. Similarly, the growth of urban Brazilian population is often associated with the migration of low skilled workers, which will lead to unemployment in case of low economic growth (Martignoni, Carvano & Januzzi, p. 295, 2006).

As previously stated, each regression was run twice – once with the Bolsa Família as a binomial variable and once as a percentage of GDP – in order to verify whether the estimated coefficients and p-values varied. Once it was determined that the coefficient signs were the same, it was more relevant to employ the specification of the CCT program as a percentage of GDP. In order to verify whether the OLS regression would be the best estimation method, some of its main assumptions were tested. According to Stock and Watson (2012, p.240) the Least Squares assumptions in the multiple regression model are as follows:

(i) The error term must have a conditional mean zero given the explanatory variables (ii) The independent variables are independently and identically distributed from their

joint distribution

(iii) Large outliers are unlikely (iv) No perfect multicollinearity

The first two assumptions are assumed to hold. Whether or not many outliers are present in the data is investigated by graphical analysis, which is mentioned in the next section while the graphs themselves can be found in the appendix. The fourth assumption is checked by the Variance Inflation Factor (VIF) test. In addition to these OLS assumptions, I will check 4 other potential issues, namely heteroscedasticity, autocorrelation between residuals, non-normal residuals and model misspecification. The Breusch-Pagan/Cook-Weisberg test will be used to determine whether the error terms are homoscedastic, that is, if their variance is constant. The Durbin-Watson test in turn will be used to detect the presence of autocorrelation in the residuals. Similarly to what is done for the case of outliers, I will again use graphical analysis in order to check if residuals are normally distributed. Finally, model specification and omitted variable bias will be verified by the Ramsey Regression Specification error test. All test statistics hitherto mentioned are found in the appendix next to their respective regression models. They will be referred to in the next section.

(13)

13 4 – Results and analysis

4a) Gini index from 1982 up to 2013 regressed on the Bolsa Família as a percentage of GDP, minimum wage, inflation and unemployment rate and GDP growth

𝑔𝑖𝑛𝑖 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑚𝑖𝑛𝑤𝑎𝑔𝑒 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡 + 𝛽5𝑔𝑑𝑝𝑔𝑟𝑜𝑤𝑡ℎ + 𝛼 Regression 1

Gini Coefficient P-value R-squared

bfgdp -8.636385 0.000 0.7098 Minwage -.0074995 0.064 Inflation .0002421 0.430 Unemployment -.1652281 0.089 Gdpgrowth -.012531 0.908 _cons 63.27138 .00

The output above indicates that some of the prior expectations were confirmed. Most importantly, the Bolsa Família was found to have a statistically significant effect on inequality measured by the Gini index. More specifically, the model predicted that a one unit change in the CCT program will theoretically lead to a 8.64 reduction of the Gini index. This might seem excessive at first but one should consider the specification of the variables in this model. The Bolsa Familia is measured a percentage of GDP and therefore a one unit increase implies a considerable investment in the program. In this case all the other variables did not have a statistically significant effect on inequality. This is surprising given that the minimum wage is relevant specifically to the low skilled and vulnerable workers and that inflation theoretically hinders the poor the hardest.

A possible explanation for the non-significance of most variables is that they indeed do not considerably affect income inequality in spite of the original expectations. Minimum wage for instance, is relevant to cater for the basic needs of the low-income groups which are a considerable component of a low income per capita nation such as Brazil. It does not, however, necessarily imply that wealth is better distributed. A higher minimum wage then does not necessarily mean that the income gap is considerably reduced. Similarly, unemployment perhaps affects a large part of the social-economic spectrum, which in turn may not affect income equality per se. Inflation had a very had p-value and a low coefficient, which suggests that it has no effect on income equality. It is possible that this result stems from the fact that Brazil has a low per capita income which in turn means that higher price levels hit the Brazilian population reasonably equally.

Although the R-squared does not ultimately determine the quality of an estimation model, the verified value indicates that almost 71 per cent of the variance is explained by the model. Considering that the main finding is consistent with those of Hoffman (2007), Modesto and Cardoso (2010) and Soares (2006), one could assume that the model is a good enough estimation.

With regards to the assessment of potential statistical issues, the VIF test did not indicate a strong multicollinearity since no variable got a vif test statistic close to or higher than 10. However, heteroscedasticity was verified since the p-value of the Breusch-Pagan/Cook-Weisberg test was lower than the 5 percent significance level, so a robust regression was run. The Durbin-Watson test

(14)

14

statistic of 1.896561 lied just below the critical value of 1.9, so autocorrelation turned out not be an issue. Similarly, Ramsey RESET´s p-value of .7 indicated that there is no omitted variable bias. However, graphical analysis of the residuals suggested that they are not normally distributed.

4b) Mortality rate between 1982 and 2013 regressed on the Bolsa Família, minimum wage, inflation rate, GDP growth and poverty

Regression 2

Mortality Rate Coefficient P-value R-squared

bfgdp -62.61555 .005 .9240 Minwage .1220775 .000 Inflation -.0005039 .833 Gdpgrowth .26056 .544 Poverty2 2.546682 .000 _cons -59.05814 .000

Model 2A offers a higher number of significant results. It indicates that a one unit increase in the Bolsa Familia expenditure as a percentage of GDP leads to a 62.61 reduction of the mortality rate. This likely stems from the multicolinnearity problem verified by the VIF statistic. The bfgdp and the poverty2 variables respectively presented a VIF test static of 8.21 and 6.75, which might imply they are correlated to a certain extent. The minimum wage had a negative effect on the mortality rate just as expected. The expectation for the effect of poverty was also verified. Nevertheless, inflation did not have a significant effect on mortality rate – perhaps because the original assumption was too extreme. It is fair to argue that inflation hinders the life quality of the fragile population the worst, but it does not necessarily lead to a higher number of deaths. GDP growth too had no significant effect, which suggests that economic growth does not increase under-five year olds mortality rate.

The regression’s R-squared was high, implying that over 92 percent of the variation is explained by the model. However, given the multicollinearity problem between variables bfgdp and poverty2 and the large difference between the coefficients of the bf and bfgdp variables, it is not possible to infer that this model offers good estimations. However, despite these problems it does suggest the same main finding supported by previous research – that the CCT program decreases the children mortality rate. However, regarding model specification, Ramsey Reset test´s p-value of .0085 indicates that there is some omitted variable bias. Similarly, the Durbin-Watson test lies within the rejection region, which means that residuals are correlated and thus the OLS assumption is violated. Thus in this case it is no longer the best linear unbiased estimator.

(15)

15

4c) Progression to secondary education between 1982 and 2013 regressed on the Bolsa Família, minimum wage, inflation rate and GDP growth.

Regression 3

Progression2nd Coefficient P-value R-squared

bfgdp 32.9251 .000 0.5305

Minwage -.0282291 .003

Inflation .0003518 .751

Gdpgrowth -.2052849 .391

_cons 96.61765 .000

Regression 3A offers a mix of significant and insignificant results. A one unit increase in the Bolsa Familía as a percentage of GDP leads to nearly a 33 percentage increase in the percentage of students who concluded their primary education and continued to secondary school. A similar reasoning can be used here to that of model 1A: the specification of the bfgdp variable implies that a one unit increase in it leads to a considerable increase in the Bolsa Família expenditure. Another factor that might have been partially responsible for this high coefficient is the fact that the Bolsa Família was only implemented for a decade in 2013 while the variable progression2nd is measured since 1981. In addition, it is possible that other variables not available in the data set explain the change in progression to secondary education better than bfgdp but are still related to the CCT program.

Minimum wage had a negative effect on progression to secondary education as expected. It would appear that the premise that a raise in minimum wage leads to a higher opportunity cost for low income families is verified. Inflation and GDP growth, however, did not have a significant effect on progression to secondary education. The model thus suggests that higher economic growth in Brazil is not reflected into a higher percentage of students furthering their education nor that inflation is an important factor in an individual’s decision to continue studying or not.

The R-squared in this regression is lower than on the others, which means that the models only explains 53 per cent of the variation verified. This could be considered evidence that the given dependent variable is more accurately explained by variables that were not available in the used data set. This view is supported by Ramsey Reset test´s p-value of .0016. Correlation between the residuals is again verified by the Durbin-Watson test of .8529, which violates the OLS assumption and entails that the regression is not the best unbiased estimator. Specific estimations can then be put into question, but it is worth reminding that despite this issue the main predictions are supported by previous research. Heteroskedasticity was not verified and neither was perfect multicollinearity.

(16)

16

4d) Unemployment rate from 1982 up to 2013 regressed on the Bolsa Família, minimum wage, inflation rate, GDP growth and urban population growth

Regression 4

Unemployment Coefficient P-value R-squared

bfgdp -7.118031 .127 0.7363 Minwage -.0050572 .251 Inflation -.0012587 .002 Gdpgrowth -.0960111 .225 Urban -3.063351 .000 _cons 17.4122 .000

Model 4A indicates only two significant results. Its main finding is that the Bolsa Família does not lead to higher unemployment, much like Teixeira (2010) and Brauw, Gilligan, Hoddinott and Roy (2015) found. Similarly, the minimum wage does not have a significant effect on unemployment either. Inflation, however, has and it thus seems that Philips curve short term tradeoff between inflation and unemployment is verified even in the long term in this case. Variable urban had an expected coefficient. Instead of increasing unemployment, it was actually verified that it decreases it. This supports Davidson’s view (2015) that rather than raising unemployment by increasing labor supply, immigration increases the overall size of the economy. In other words, immigration simultaneously rises and increases demand for labor by spending their wages on other goods and services and thus raising demand for labor. This reasoning may also explain why GDP growth did not have a significant effect on unemployment – the positive effect of an increase in economy is already indirectly accounted for growth in urban population.

Having said that, it is worth highlighting that variable bfgdp had a multicollinearity problem according to the VIF test statistic of 11.85. It is possible that the CCT program variable is correlated with variables urban and minwage for they too had relatively high values for their VIF test statistic. This then casts doubt over the accuracy of its coefficient’s estimation. Nevertheless, the errors in this model were homoscedastic and the Ramsey Reset test did not indicate omitted variable bias. Furthermore, the Durbin-Watson test did not suggest any autocorrelation. It is possible that the Bolsa Família does not raise unemployment itself, but is associated with a higher participation of the stipends in the informal sector and thus increases the Brazilian underground economy. Whether all participants in the informal sector are accounted as employed or unemployed is debatable.

(17)

17

4e) Percentage of Brazilian population living under two dollars a day regressed on the Bolsa Família, urban population growth, inflation rate and unemployment rate

𝑝𝑜𝑣𝑒𝑟𝑡𝑦2 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑢𝑟𝑏𝑎𝑛 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡𝑟𝑎𝑡𝑒 + 𝛼

Regression 5A

Poverty2 Coefficient P-value R-squared

bfgdp -13.91701 .022 .9118

Urban 5.898273 .001

Inflation .0030029 .000

Unemployment .3759354 .282

_cons 4.818059 .441

The final regression 5a indicates that many variables have a significant effect on poverty among the Brazilian population. Most importantly, a one unit increase in the CCT program is predicted to have almost a 14 percentage decrease in the population who lives with less than 2 dollars a day in Brazil. The growth of urban population has the inverse effect – it is predicted to increase poverty. This is reasonable considering the historical reasons for immigration in Brazil – usually low skilled workers moved to more industrialized areas in search of employment but with few if any contacts. Given the lack of infrastructure and overall gentrification policies undertaken by the major urban areas including São Paulo and Rio de Janeiro, many immigrants found themselves unemployed and marginalized in peripheral areas. The VIF test indicated that the variable urban has a multicollinearity problem, possibly with unemployment, which may explain why only one of them was found to have a significant effect. However, it is relevant to underscore that no variable had a VIF test statistic higher than 10, so overall multicollinearity was not an issue.

Inflation, just as expected, has a positive effect on poverty. This makes sense as a rise in prices of non-superfluous goods will curtail the budget of the low-income families the most. As such, the expectation was verified, albeit the predicted effect of inflation itself is limited - .3 percent. Furthermore, the R-squared indicates that 91.18 percent of the variation is explained by the model. This partially suggests that the regression predicts reasonably well, but this view can be questioned if the Durbin-Watson test statistic is taken into account. Similar to previous models, errors in this case were found to be correlated. In other words, this prevents the OLS estimators from being unbiased in this case. Despite this issue, the Ramsey test did not indicate omitted variable bias as it had a p-value of .4465. Heteroskedasticity was also not verified.

5-Conclusion

It is clear that conditional cash transfer programs aim to align asymmetrical individual and society’s preferences and in so doing foster development and eradicate intergenerational poverty. This study has verified that overall the use of such programs in the Brazilian context has proved fruitful. Although with a more limited data set and more simple research methodology, the models here provided results supported by previous research. It indicates that the Bolsa Família has decreased income inequality, poverty, mortality rate of under-five year olds and increased

(18)

18

enrollment to secondary education. Furthermore, the research illustrated that the Bolsa Família has not had an effect on unemployment, despite this being one of its main criticism in Brazilian media. The fact that women in the households are the preferred recipients is also worth mentioning as it attempts to empower marginalized women and thus partially offset gender inequality.

However, the regression models presented a few issues with regards to multicollinearity and autocorrelation but given that the main predictions of the CCT program are in line with Barros (2007), Sorares (2010), Hoffman (2006) and others, it could be argued that they are good enough predictors. The limitation of the data set constricted the possible model designs and thus acted as a main cause for misspecification and multicollinearity problems. For example, in the case of regression 4, bfgdp presented a high VIF tests statistic. It was likely correlated with variables urban and minwage. It would have been produced better results to use different variables, but very few were available in the World Bank or in the IBGE website. The variables used on this paper were among the very few that had a reasonably consistent sample since 1982. The majority of the samples available at both sources started in the mid-nineties and was often limited to fewer than 15 observations.

Regardless of the data limitations, the correlation between residuals in regressions 2 and 3 specifically cast doubt over their validity. Were it not for previous research, they alone would offer weak evidence of the main social effects of the Bolsa Família. Similarly, the non-normal residuals of regressions 1, 2, 4 and 5 also hinder the unbiasedness of the estimators. The OLS assumptions were then partially violated by a number of the models, apart from the one regarding unlikely outliers. A different estimator could have offered more precise and valid results. Nevertheless, it is important to remember once again that despite all these issues, the main findings are supported by previous research. In other words, it seems that the Bolsa Família indeed has has some positive social effects in the context of Brazil.

There are few arguments against the implementation of conditional cash transfer programs in low per capita income such as Brazil, but it is still necessary to address all its repercussions thoroughly in order to optimize the welfare gain. For example, while little evidence indicates that the Bolsa Família raises unemployment levels, it is relevant to analyze to what extent the recipients are working in the informal sector. This can be considered a problem because informal sector workers underpay taxes and thus tax revenues that can be used for further social investments are lost. A relevant question for further research would then be how the Bolsa Família can be altered in order to maximize the likelihood of recipients working in the formal sector. It would be also interesting to study how much the returns on education can lead to less income inequality. Brazil has a historically high level of income inequality and better understanding the extent which education and CTT programs can mitigate it is necessary.

(19)

19

Bibliography

Barros, R. P., Carvalho, M. d., & Franco, S. (2007). O Papel das Transferências Públicas na Queda Recente da Desigualdade de Renda Brasileira. In R. P. Barros, M. N. Foguel, & G. Ulyssea, Desigualdade de Renda no Brasil: uma análise da queda recente (volume 2) (pp. 41-86). IPEA. Barros, R. P., Carvalho, M. d., Franco, S., & Mendonça, R. (2006, August). Uma das principais causas

da queda recente na desigualdade de renda brasileira. IPEA, pp. 1-29.

Brauw, A. d., & Hoddinott, J. (2011). Must conditional cash transfer programs be conditioned to be effective? The impact. Journal of Development Economics, 96, pp. 359–370.

BRAUW, A. D., GILLIGAN, D. O., HODDINOTT, J., & ROY, S. (2014). The Impact of Bolsa Famı´lia on Women’s Decision-Making Power. World Development, 59, pp. 487–504.

Brauw, A. d., Gilligan, D. O., Hoddinott, J., & Roy, S. (2015, April). Bolsa Família and Household Labor Supply. Economic Development and Cultural Change, 63(3), pp. 423-457.

Das, J., Do, Q.-T., & Özler, B. (2005). Reassessing Conditional Cash Transfer Programs. The World Bank Research Observer, 20(1), pp. 57-80.

Davidson, A. (2015, 3 15). New York Times. Retrieved 6 2015, 2015, from

http://www.nytimes.com/2015/03/29/magazine/debunking-the-myth-of-the-job-stealing-immigrant.html?_r=0

Ferro, A. R., & Nicollela, A. C. (2007, March). The impact of conditional cash transfer programs on household work decisions in Brazil. Research in Labor Economics, pp. 1-54.

Filmer, D., & Schady, N. (2011). Does more cash in conditional cash transfer programs always lead to larger impacts. Journal of Development Economics, 96, pp. 150-157.

Filmer, D., & Schady, N. (2011, May 13). Does more cash in conditional cash transfer programs always lead to larger impacts? Journal of Development Economics, 96, pp. 150–157. folha.uol.com.br. (2014, May 1). Retrieved June 2015, from folha de são paulo:

http://www1.folha.uol.com.br/paywall/login.shtml?http://www1.folha.uol.com.br/poder/2 014/05/1448304-ft-diz-que-reajuste-do-bolsa-familia-e-populista-e-contra-ataque-a-oposicao.shtml

FRIED, B. J. (2012). Distributive Politics and Conditional Cash Transfers: The Case of Brazil’s Bolsa Famı´lia. World Development, 40(5), pp. 1042–1053.

FRIED, B. J. (2012). Distributive Politics and Conditional Cash Transfers:The Case of Brazil’s Bolsa Famı´lia. World Development, 40(5), pp. 1042–1053.

Glewwe, P., & Kassouf, A. L. (2012). The impact of the Bolsa Escola/Familia conditional cash transfer program on enrollment, dropout rates and grade promotion in Brazil. Journal of

(20)

20 Guia Trabalhista. (n.d.). Retrieved 6 30, 2015, from

http://www.guiatrabalhista.com.br/guia/salario_minimo.htm

Hall, A. (2006, November). From Fome Zero to Bolsa Família: Social Policies and Poverty Alleviation under Lula. Journal of Latin American Studies, 38(4), pp. 689-709.

Hoffmann, R. (2006, June). Transferências de renda e a redução da desigualdade no Brasil e cinco regiões entre 1997 e 2004. Econômica, 8(1), pp. 55-81.

Kakwani, N., Neri, M., & Son, H. H. (2009). Linkages between pro-poor growth, social programmes and labour market: The recent Brazilian experience. The World Institute for Development Economics Research, pp. 1-33.

Martignoni, E. M., Carvano, L. M., & Jannuzzi, P. d. (2006). Força de trabalho e desemprego na Região Metropolitana do Rio de Janeiro nos anos 1990: o efeito dos fatores demográficos. 287-299.

Neto, R. d. (2010). Impacto do Bolsa Família sobre a frequência à escola: estimativas a partir de informações da pesquisa nacional por amostra de domicílio (PNAD). In Bolsa Família 2003-2010: avanços e desafios (pp. 53-71). IPEA.

Nickell, S. (1997). Unemployment and Labor Market Rigidities: Europe versus North America . Journal of Economic Perspectiv, 55-74 .

Portal Brasil. (2013, 10 22). Retrieved 6 30, 2015, from Brasil.gov.br:

http://www.brasil.gov.br/cidadania-e-justica/2013/10/bolsa-familia-paga-r-2-1-bilhoes-em-outubro

Reynolds, S. A. (2015). Brazil’s Bolsa Familia: Does it work for a dolescents and do they work less for it? Economics of Education Review, 46, pp. 23–38.

Soares, S., & Sátyro, N. (2009, October). O programa Bolsa Família: desenhol institucional, impactos e possibilidades futuras. IPEA, pp. 1-40.

Soares, S., Souza, P. H., Osório, R. G., & Silveira, F. G. (2010). Os impactos do benefício do programa Bolsa Família sobre a desigualdade e pobreza. In Bolsa Família 2003-2010: avanços e desafios (pp. 27-52).

Stock, J. H., & Watson, M. W. (2012). Introduction to econometrics. Essex: Pearson Education Limited.

Teixeira, C. G. (2010, March). A heterogeneity analysis of the Bolsa Famíia programme effect on men and women´s work supply. International Policy Centre for Inclusive Growth, 61, pp. 1-22.

(21)

21 _cons 65.1467 3.382073 19.26 0.000 58.20726 72.08615 gdpgrowth -.016048 .1075625 -0.15 0.883 -.2367481 .2046521 unemployment -.2151029 .2049252 -1.05 0.303 -.6355747 .2053689 inflation .00011 .0005878 0.19 0.853 -.0010959 .001316 minwage -.011103 .0049546 -2.24 0.033 -.021269 -.0009369 bf -2.456465 1.202473 -2.04 0.051 -4.923735 .0108057 gini Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 267.118071 32 8.34743973 Root MSE = 1.7808 Adj R-squared = 0.6201 Residual 85.6194976 27 3.1710925 R-squared = 0.6795 Model 181.498574 5 36.2997148 Prob > F = 0.0000 F( 5, 27) = 11.45 Source SS df MS Number of obs = 33

Appendix

Below one can find all the 5 regressions in two different specifications “a” and “b” – the former with the Bolsa Família as a percentage of gdp (bfgdp) and the latter as binomial variable (bf). In addition, the four test statistics for each model are displayed after model specification “b”. It is important to underscore that these test statistics are related to models “a”, that is, to the regressions with variable bfgdp. Finally, there are two graphs for each regression model that offer some insight on whether or not the residuals are normally distributed and whether there were many outliers and thus a violation of the OLS assumption. The avplots graph indicated that the CCT program variable did not have many outliers in general, albeit this cannot be said about all the other variables in each model. Overall, however, there were not many outliers in the case of the statistically significant variables. The other graph verified whether or not residuals were normally distributed – the closer are the points to the fitted line, the more normal are the residuals. It is clear that some models had this issue – particularly regressions 1, 2, 4 and 5.

Regression 1a Regression 1b Root MSE = 1.6945 R-squared = 0.7098 Prob > F = 0.0000 F( 5, 27) = 46.58 Linear regression Number of obs = 33

_cons 63.27138 2.388703 26.49 0.000 58.37017 68.1726 gdpgrowth -.012531 .1071949 -0.12 0.908 -.2324769 .2074148 unemployment -.1652281 .0936987 -1.76 0.089 -.3574819 .0270257 inflation .0002421 .0003025 0.80 0.430 -.0003785 .0008628 minwage -.0074995 .0038814 -1.93 0.064 -.0154635 .0004645 bfgdp -8.636385 2.084527 -4.14 0.000 -12.91348 -4.35929 gini Coef. Std. Err. t P>|t| [95% Conf. Interval] Robust

(22)

22

𝑔𝑖𝑛𝑖 = 𝛽1𝑏𝑓𝑔𝑑𝑝 + 𝛽2𝑚𝑖𝑛𝑤𝑎𝑔𝑒 + 𝛽3𝑖𝑛𝑓𝑙𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝑢𝑛𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡 + 𝛽5𝑔𝑑𝑝𝑔𝑟𝑜𝑤𝑡ℎ + 𝛼

Prob > F = 0.7001 F(3, 24) = 0.48 Ho: model has no omitted variables

Ramsey RESET test using powers of the fitted values of gini . ovtest

Prob > chi2 = 0.0047 chi2(1) = 8.00

Variables: fitted values of gini Ho: Constant variance

Breusch-Pagan / Cook-Weisberg test for heteroskedasticity

Durbin-Watson d-statistic( 6, 33) = 1.896561 . estat dwatson Mean VIF 2.63 gdpgrowth 1.13 0.886053 inflation 1.80 0.556034 unemployment 2.26 0.441788 bfgdp 3.58 0.279209 minwage 4.38 0.228441 Variable VIF 1/VIF . vif -1 0 -5 0 5 e( g in i | X ) -.2 -.1 0 .1 .2 e( bfgdp | X ) coef = -8.6363849, se = 3.1686653, t = -2.73 -6 -4 -2 0 2 4 e( g in i | X ) -150 -100 -50 0 50 100 e( minwage | X ) coef = -.00749951, se = .00514637, t = -1.46 -6 -4 -2 0 2 4 e( g in i | X ) -1000 -500 0 500 1000 1500 e( inflation | X ) coef = .00024214, se = .00056289, t = .43 -6 -4 -2 0 2 4 e( g in i | X ) -4 -2 0 2 4 e( unemployment | X ) coef = -.16522812, se = .18984141, t = -.87 -6 -4 -2 0 2 4 e( g in i | X ) -5 0 5 e( gdpgrowth | X ) coef = -.01253105, se = .10040656, t = -.12 0 .0 0 0 .2 5 0 .5 0 0 .7 5 1 .0 0 N o rm a l F [(r -m )/ s ] 0.00 0.25 0.50 0.75 1.00

(23)

23 Regression 2a Regression 2b _cons -59.05814 10.46795 -5.64 0.000 -80.5366 -37.57969 poverty2 2.546682 .4272104 5.96 0.000 1.670119 3.423246 gdpgrowth .26056 .4242535 0.61 0.544 -.6099363 1.131056 inflation -.0005039 .0023605 -0.21 0.833 -.0053472 .0043394 minwage .1220775 .0169925 7.18 0.000 .0872118 .1569432 bfgdp -62.61555 20.52368 -3.05 0.005 -104.7267 -20.50444 mortalityr~e Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 18656.4753 32 583.014852 Root MSE = 7.2468 Adj R-squared = 0.9099 Residual 1417.92971 27 52.5159152 R-squared = 0.9240 Model 17238.5456 5 3447.70911 Prob > F = 0.0000 F( 5, 27) = 65.65 Source SS df MS Number of obs = 33

_cons -61.48922 10.76512 -5.71 0.000 -83.57744 -39.40101 poverty2 2.964608 .3568014 8.31 0.000 2.232512 3.696704 gdpgrowth .3579858 .4446681 0.81 0.428 -.5543978 1.270369 inflation -.0021073 .0022819 -0.92 0.364 -.0067893 .0025747 minwage .1051166 .0149642 7.02 0.000 .0744126 .1358205 bf -15.41857 6.026074 -2.56 0.016 -27.78305 -3.054087 mortalityr~e Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 18656.4753 32 583.014852 Root MSE = 7.5391 Adj R-squared = 0.9025 Residual 1534.64258 27 56.8386142 R-squared = 0.9177 Model 17121.8327 5 3424.36654 Prob > F = 0.0000 F( 5, 27) = 60.25 Source SS df MS Number of obs = 33

0 .0 0 0 .2 5 0 .5 0 0 .7 5 1 .0 0 N o rm a l F [(r -m )/ s ] 0.00 0.25 0.50 0.75 1.00

(24)

24

Ramsey RESET test using powers of the fitted values of mortalityrate . ovtest Durbin-Watson d-statistic( 6, 33) = .921046 Mean VIF 4.08 gdpgrowth 1.10 0.907724 inflation 1.73 0.578314 minwage 2.61 0.383251 poverty2 6.75 0.148214 bfgdp 8.21 0.121729 Variable VIF 1/VIF . vif

Prob > chi2 = 0.6768 chi2(1) = 0.17

Variables: fitted values of mortalityrate Ho: Constant variance

-1 0 0 10 20 e ( m o rt a lit y ra te | X ) -.2 -.1 0 .1 .2 e( bfgdp | X ) coef = -62.615549, se = 20.523677, t = -3.05 -2 0 -1 0 0 10 20 30 e ( m o rt a lit y ra te | X ) -200 -100 0 100 200 e( minwage | X ) coef = .12207752, se = .01699249, t = 7.18 -2 0 -1 0 0 10 20 e ( m o rt a lit y ra te | X ) -1000 0 1000 2000 e( inflation | X ) coef = -.00050392, se = .00236047, t = -.21 -2 0 -1 0 0 10 20 e ( m o rt a lit y ra te | X ) -5 0 5 10 e( gdpgrowth | X ) coef = .26055998, se = .42425348, t = .61 -3 0 -2 0 -1 0 0 10 20 e ( m o rt a lit y ra te | X ) -4 -2 0 2 4 6 e( poverty2 | X ) coef = 2.5466824, se = .42721036, t = 5.96

(25)

25 Regression 3a Regression 3b _cons 96.61765 4.135757 23.36 0.000 88.14594 105.0894 gdpgrowth -.2052849 .235409 -0.87 0.391 -.6874985 .2769286 inflation .0003518 .0010973 0.32 0.751 -.0018959 .0025995 minwage -.0282291 .0088452 -3.19 0.003 -.0463477 -.0101104 bfgdp 32.9251 5.959772 5.52 0.000 20.71706 45.13313 progressio~d Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 973.186001 32 30.4120625 Root MSE = 4.0397 Adj R-squared = 0.4634 Residual 456.943154 28 16.3193984 R-squared = 0.5305 Model 516.242847 4 129.060712 Prob > F = 0.0002 F( 4, 28) = 7.91 Source SS df MS Number of obs = 33

_cons 91.41351 4.536873 20.15 0.000 82.12015 100.7069 gdpgrowth -.1878141 .2814313 -0.67 0.510 -.7643001 .3886718 inflation .0005126 .0013014 0.39 0.697 -.0021532 .0031784 minwage -.0156733 .0094791 -1.65 0.109 -.0350902 .0037437 bf 9.208008 2.479263 3.71 0.001 4.129467 14.28655 progressio~d Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 973.186001 32 30.4120625 Root MSE = 4.7803 Adj R-squared = 0.2486 Residual 639.822238 28 22.8507942 R-squared = 0.3425 Model 333.363763 4 83.3409408 Prob > F = 0.0163 F( 4, 28) = 3.65 Source SS df MS Number of obs = 33

0 .0 0 0 .2 5 0 .5 0 0 .7 5 1 .0 0 N o rm a l F [(r -m )/ s ] 0.00 0.25 0.50 0.75 1.00

(26)

26

Ramsey RESET test using powers of the fitted values of progression2nd . ovtest

Prob > chi2 = 0.1092 chi2(1) = 2.57

Variables: fitted values of progression2nd Ho: Constant variance

Durbin-Watson d-statistic( 5, 33) = .8529521 Mean VIF 1.70 gdpgrowth 1.09 0.916160 inflation 1.20 0.831623 bfgdp 2.23 0.448598 minwage 2.28 0.439533 Variable VIF 1/VIF . vif -1 0 -5 0 5 10 e ( p ro g re s s io n 2 n d | X ) -.2 -.1 0 .1 .2 e( bfgdp | X ) coef = 32.925096, se = 5.9597717, t = 5.52 -5 0 5 10 15 e ( p ro g re s s io n 2 n d | X ) -100 0 100 200 e( minwage | X ) coef = -.02822907, se = .00884523, t = -3.19 -1 0 -5 0 5 10 e ( p ro g re s s io n 2 n d | X ) -1000 0 1000 2000 e( inflation | X ) coef = .00035178, se = .0010973, t = .32 -1 0 -5 0 5 10 e ( p ro g re s s io n 2 n d | X ) -5 0 5 e( gdpgrowth | X ) coef = -.20528494, se = .23540902, t = -.87

(27)

27 Regression 4a Regression 4b _cons 17.41222 1.460068 11.93 0.000 14.4164 20.40803 urban -3.063351 .7172715 -4.27 0.000 -4.535071 -1.591632 gdpgrowth -.0960111 .077331 -1.24 0.225 -.2546812 .062659 inflation -.0012587 .0003754 -3.35 0.002 -.002029 -.0004884 minwage -.0050752 .0043301 -1.17 0.251 -.0139598 .0038095 bfgdp -7.118031 4.514741 -1.58 0.127 -16.38151 2.145453 unemployment Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 180.332791 32 5.63539971 Root MSE = 1.327 Adj R-squared = 0.6875 Residual 47.5476214 27 1.76102301 R-squared = 0.7363 Model 132.785169 5 26.5570338 Prob > F = 0.0000 F( 5, 27) = 15.08 Source SS df MS Number of obs = 33

_cons 17.22105 1.558648 11.05 0.000 14.02297 20.41913 urban -2.080014 .593746 -3.50 0.002 -3.29828 -.8617474 gdpgrowth -.1069742 .0816907 -1.31 0.201 -.2745897 .0606412 inflation -.0013968 .0003829 -3.65 0.001 -.0021825 -.000611 minwage -.0108532 .0032557 -3.33 0.002 -.0175333 -.0041731 bf -.0943151 1.313846 -0.07 0.943 -2.790105 2.601474 unemployment Coef. Std. Err. t P>|t| [95% Conf. Interval] Total 180.332791 32 5.63539971 Root MSE = 1.3866 Adj R-squared = 0.6588 Residual 51.9151407 27 1.92278299 R-squared = 0.7121 Model 128.41765 5 25.68353 Prob > F = 0.0000 F( 5, 27) = 13.36 Source SS df MS Number of obs = 33

0 .0 0 0 .2 5 0 .5 0 0 .7 5 1 .0 0 N o rm a l F [(r -m )/ s ] 0.00 0.25 0.50 0.75 1.00

(28)

28

Ramsey RESET test using powers of the fitted values of unemployment . ovtest

Prob > chi2 = 0.8794 chi2(1) = 0.02

Variables: fitted values of unemployment Ho: Constant variance

Durbin-Watson d-statistic( 6, 33) = 1.067176 Mean VIF 5.11 gdpgrowth 1.09 0.916160 inflation 1.30 0.766646 minwage 5.05 0.197911 urban 6.25 0.160008 bfgdp 11.85 0.084355 Variable VIF 1/VIF . vif -4 -2 0 2 4 e ( u n e m p lo y m e n t | X ) -.15 -.1 -.05 0 .05 .1 e( bfgdp | X ) coef = -7.1180307, se = 4.5147414, t = -1.58 -4 -2 0 2 e ( u n e m p lo y m e n t | X ) -150 -100 -50 0 50 100 e( minwage | X ) coef = -.00507515, se = .00433012, t = -1.17 -4 -2 0 2 4 e ( u n e m p lo y m e n t | X ) -1000 0 1000 2000 e( inflation | X ) coef = -.00125868, se = .00037542, t = -3.35 -3 -2 -1 0 1 2 e ( u n e m p lo y m e n t | X ) -5 0 5 e( gdpgrowth | X ) coef = -.09601106, se = .07733099, t = -1.24 -4 -2 0 2 4 e ( u n e m p lo y m e n t | X ) -1 -.5 0 .5 e( urban | X ) coef = -3.0633512, se = .71727147, t = -4.27