Race and Competitiveness in Brazilian Elections: Evaluating the Chances of Black and Brown Candidates through Quantile Regression Analysis of Brazil's 2014 Congressional Elections*

Although the proportion of black, brown and indigenous electoral candidates in Brazil is close to the proportion of blacks, browns and indigenous in the general population, the proportion elected to the country’s Federal Congress is significantly lower. Statistical techniques such as linear or logistic regression are typically used to estimate the effect of a particular variable such as color/race or gender on a candidate’s electoral performance. However, in Brazilian elections, characterized by substantive, asymmetrical differences such as extreme variations in campaign finance distribution, the efficacy of these types of regression models is limited. Such being the case in Brazil's open list proportional representation system, we propose quantile regression as the most suitable means for estimating the relationship between voting and other variables such as race/color, because it enables us to estimate relationships between the variables of interest across several distribution quantiles. Quantile regression models show that black and brown candidates get as many as 40% fewer votes than white candidates in higher vote distribution quantiles. Furthermore, analysis of access to campaign financing finds that black and brown candidates on average garner only 75% of the funds available to white candidates at quantile 80 of campaign finance distribution. This drops to 65% at quantile 90.

phenomenon 1 , there is some consensus that women are often excluded from politics even before elections occur, when party lists are being drawn up (LAWLESS, 2012).
Something quite different seems to happen with Afro-Brazilian candidates.
Research on the topic shows that the proportion of non-whites on party lists is relatively close to that of the general population (BUENO and DUNNING, 2013;MACHADO, 2015a, 2015b). Thus, the filters that impede the representation of this group seem to be different from those imposed on women.
Non-white women are subjected to a double filter to political access; analyses of candidacy nominations find that the largest discrepancy between the candidate percentages and population size occurs in this category (CAMPOS and MACHADO, 2015b).
Researchers have traditionally used statistical techniques such as linear or logistic regression (BUENO and DUNNING, 2013;CAMPOS and MACHADO, 2015a) to estimate the effect of traits such as color/race or gender on a candidate's election chances. Linear regression models estimate the correspondence between the averages of independent variables and those of a dependent variable (MOSTELLER and TUKEY, 1977, p. 266). However, analyses focused on distribution averages are limited when there are substantial differences in significance for extreme points of the distribution (KOENKER and HALLOCK, 2001). In these cases, averages are not good descriptive parameters for the complete series. One example of this is the typical vote distribution in Brazilian elections. In a logistic model with a binary dependent variable, it is necessary to define an arbitrary prior cut in a continuous variable of interest. This is the case with the The Brazilian open list proportional representation system distributes seats in municipal, state and federal legislatures to political parties and coalitions based on the total votes received by said parties and coalitions. Therefore, it is in the parties' and coalitions' best interests to register as many candidates as possible and ______________________________________________________________________________________________ 1 On this subject, see Pinto (1994), Miguel (2000), Miguel and Queiroz (2006), and Araújo e Alves (2007). where the object of interest is an explanation of such competition as exists between candidates that have real chances of being elected. One way to address this issue is through logistic regression that distinguishes between competitive (exceeding a certain number of votes) and non-competitive (CAMPOS and MACHADO, 2015a) candidacies. However, this approach necessitates an arbitrary cut to define the profiles of the candidates to be compared in the regression model.  Elected candidates exclusively proceed from high performance candidacies. This means that elected MoCs almost entirely come from a group of just over one thousand candidates and that the other four thousand had no real chance of ever being elected, that is, they did not participate competitively in the race.

Race and Competitiveness in Brazilian
Based on the average of the variables included in a given model, linear regression is unable to accurately estimate the weight of a set of independent variables on the chances of obtaining campaign votes or resources in such an unequal context. To help clarify this point, let us look at the result of a linear regression that estimates the impact of candidate color/race on their vote. We would like to point out that this first model has a more didactic than analytical character. From a simplified model, it is easier to understand the potential differences between the standard linear regression model and the quantile one 8 .   (2019) 13 (3) e0005 -10/31 education levels and occupational status than non-whites. The low representation of non-whites in elected office could, therefore, be explained by social lags that generate inequality in conditions of electoral competition. Controlling for the variables of class and education enables us to test if there is inequality due to candidates' color/race beyond these structural disadvantages.
In addition to the aforementioned socio-economic aspects, elements specific to electoral dynamics can help explain voting differences, such as incumbency and Electoral funding's explanatory capacity for voting patterns is illustrated in various studies, as summarized by Mancuso (2015) 12 . It is worth noting that the information about campaign financing in Brazil released by the TSE presents data on both fundraising and campaign spending. There is no consensus on which of these two variables has more explanatory potential, but the small variation in effect observed by their distinct use implies a low impact on the final results of the analysis ______________________________________________________________________________________________ 11 Jacobson (1978) says that the expenses of candidates for reelection are less impactful than those of challengers, but later studies question these findings (LEVITT, 1994), including information from Brazil (SAMUELS, 2001). For a more detailed discussion see Silva (2010 pp. 27-30). 12 These findings were made by Samuels (2002) The following analysis will come from the data evaluated in this article that supports this view.
14 This classification was based on the work of Robert Erikson, John Goldthorpe and Luciene Portocarero (1979), and is applied to Brazilian professional categories associated with politicians in Campos and Machado (2015a). Although the EGP class model is divided into seven divisions, when dealing with the social distinction that allows for reputational gains in politics, the professions that are in the two highest strata of the model were understood as high-income political classes. As shown in Table 01, the number of votes for non-whites is, on average, lower than that for whites. The estimate of this relationship in standard linear regression models ranges from 48% to 17%, according to the specifications of the model.   The largest adjusted R2 is that of Model 07, which is the model that shows the greatest causal relationship. However, it is hard to isolate campaign spending as an independent variable from voting potential and other candidate characteristics such as race/color. More promising electoral candidates tend to be more likely to obtain campaign financing, which may lead to the endogeneity problems previously pointed out by Mancuso (2015). Thus, we have chosen to address this with two models: one that does not consider financial resources,

Race and Competitiveness in
Model 05, and one that does, Model 07, since there is a possibility of bias in the estimates depending on whether financing-related information is included.

Quantile regression
We will now examine how these coefficients behave in the different voting quantiles. For this, we present the linear and quantile regression in    The variable with the greatest effect on votes is incumbency, the rate of which tends to fall as we move down through the distribution quantiles. This relationship raises a question that needs further investigation. Although the statistical model calculates the effect of incumbency in each of the quantiles, it is important to consider that these cases are concentrated in the highest cuts. Only one case is observed in the 10% decile with the rest located above 75%. That is, the most robust variable of incumbency cannot be used in the initial distribution quantiles because it does not exist there. This implies that the linear regression coefficient should be viewed with caution, because when we look at the distribution effect in quantiles in which incumbent candidates actually exist, there is a relevant reduction, such as in the 95% quantile, where incumbency guarantees over 70% more votes than other candidates.
As expected, controlling the relationship between candidate votes and racial profile reduces the impact of being non-white, taking into consideration that part of the electoral inequality linked to this characteristic is due to specific conditions that are more prevalent among non-white candidates, including the fact that fewer non-  In the models that include campaign spending, we observe that there is almost no difference between the estimates of race/color coefficients from quantile models in relation to the linear regression model. The F tests show that there is no statistical significance between the different quantile models themselves and between them and the linear model. Figure 03 shows this visually.
If the negative effects of the variables of race/color, and gender diminish in the higher quantiles, this is due to greater access to campaign financing for these more competitive nonwhites and women. As discussed earlier, the relationship between incumbency and votes gradually reduces. The linear model estimates an over 100% vote gain for incumbent candidates, while in the quantile regression model the impact at the 90 quantile is approximately 95% more votes compared to challenger candidates. Figure 03. Estimates of the relationship between race/color and votes, using Model 07 with campaign expenses -linear regression and quantiles from 10 to 95 with 0.5point increments Source: Elaborated by the authors using 2018 electoral data from the TSE. Note: Key: y axis = color/race; x axis = quantile.
However, due to possible endogeneity problems caused by adding campaign spending to the model, we have to explain the campaign funding distribution. On the one hand, the effect of campaign spending on voting shows the need to explain the constraints related to access to campaign funding. On the other, the issue of endogeneity provides an important justification for looking at campaign financing as a dependent variable. This is depicted from Table 05. We will now look at campaign fundraising instead of campaign spending as a dependent variable, because the effect of financing that we aim to explain is the difference in the profiles of candidates who receive unequal funding for their campaigns. Variation in campaign spending is related to this, but it depends on other factors, such as campaign strategies 17 , the perception of real chance of victory 18 and specific electoral conditions in each district. Although revenues and expenses have similar results, confirmed by Mancuso (2015) in the comparison between Table 05 and Table A2 in the Annex, the above argument provides a theoretical basis for designating campaign fundraising as a dependent variable.
As in the case of campaign spending, the campaign fundraising indicator was calculated as a proportion of the total revenue declared in the electoral district, to enable comparison between different contexts of the elections. The variable was also transformed into a natural log, in view of its distribution ( Figures A6 and A7 in the Annex). As in the case of vote distribution, elected candidates concentrate at the top in terms of campaign fundraising distribution, as shown in Figure 04. candidates depending on size of city, while Nelson Rojas de Carvalho (2003) identifies different patterns of territorial distribution of voting among federal congressional candidates. 18 Wescrey Portes Pereira (2018) notes in his case studies that unelected candidacies tend not to spend the total amount of funds raised while elected candidates tend to spend the full amount on their campaigns.
The comparison of linear models specifications in Table 05 shows that Model 05 has the highest explanatory capacity. It is important to note here that the estimate of the effect of race is reduced to the point of losing statistical significance when it is controlled by these socioeconomic and political variables.  Returning to the analysis of the effect of race on campaign fundraising, Figure   05 shows how there is either no statistically relevant difference between non-whites and whites in the lower funding strata or there is a slight tendency of advantage for non-white candidates. However, this relationship becomes a clear disadvantage for non-white candidates from quantile 70 of the financing distribution forwards, leading to significantly greater inequality between whites and non-whites in the funding range above 90% of the distribution.
In the higher levels of voting and campaign funding, where competition practically dissipates and shows signs of confirmation of favorites in their respective constituencies, a more qualitative analysis is needed to understand better the stories behind these campaigns 19 . Other crucial variables could be looked ______________________________________________________________________________________________ at, such as the amount of free commercial airtime allocated to each candidate, political experience, professional experience, or belonging to a politically connected family. Nevertheless, the above data can be used to clarify an important question related to elections and race. The low electoral representation of the non-white population cannot be simply attributed to the material inequality in Brazilian society. Neither education nor profession are enough to explain the lower levels of voting for non-whites, because even when these variables are used to control the racial effect on electoral dynamics, non-white status persists as a vote reducing factor. It is important to emphasize that this effect is even more pronounced when analyzing black women candidates in the electoral context.

Conclusions
The structure of Brazilian electoral competition, which combines proportional representation, open lists and the distribution of seats by electoral quotient, poses specific challenges to the analysis of factors that affect the success of certain types of candidates. Most analytical models for elections measure electoral success either as obtaining legislative seats or as obtaining large amounts of votes (SPECK and CERVI, 2016, p. 60). In majority or proportional closed-list systems, a