A study on the geographical distribution of Brazil’s prestigious software developers

2.1 Network prestige

In this paper, we are interested in measuring the prestige of GitHub users located in Brazil. In social network analysis, prestige can be measured based on directional relations among actors. We measured prestige using a graph of follow-relationships, in which there is an arc linking user A to B if A follows B. On GitHub, this implies that user A receives notifications from B’s development activities, which means that there is interest from A in assessing B’s contributions.

There are different network measures that can be computed to quantify the prestige of an actor in a social network. The simplest actor-level measure of prestige is the in-degree of a vertex i [43] in a graph, which is often referred to as i’s popularity. However, popularity is a very restricted measure of prestige because it takes only direct choices into account. With popularity it does not matter whether choices are received from popular people. The overall structure of the network is disregarded [6].

Another prestige measure is proximity. It defines an influence domain of actor i as the set of actors from whom i is reachable and considers the distance these actors are from i. It ignores actors who cannot reach i, thus it is defined even if the network is not connected (when some actors are not reachable from other actors) [43].

We used Pajek [17] to calculate the proximity prestige for each vertex in the graph of follow-relationships described above. In Pajek, the proximity prestige of a vertex is the proportion of all vertices (except itself) in its input domain divided by the mean distance from all vertices in its input domain.

Maximum proximity prestige is achieved if a vertex is directly chosen by all other vertices. This is the case, for example, in a star-network in which all choices are directed to the central vertex. Then, the proportion of vertices in the input domain is 1 and the mean distance from these vertices is 1, so proximity prestige is 1 divided by 1. Vertices without input domain get minimum proximity prestige by definition, which is zero [6].

2.2 GitHub

GitHub is a web-based hosting service that allows developers to host their software project repositories using the Git revision control system. Since its launch in April 2008, GitHub has become one of the most popular source code hosting services with over twenty million projects maintained by over eight million registered developers [18]. It is now the largest code host in the world [9].

In addition to revision control, GitHub acts as a social network site that enables developers to connect and collaborate with each other. Developers can search for software projects that they are interested in, easily fork those projects to make their own contributions, and follow the work of others. We are particularly interested in follow-relationships, as they represent a deliberate interest from one developer in another’s work and denote the prestige of a developer in GitHub’s social network.

The site organizes software repositories by software developer or organization, rather than by project, showing a list of each developer’s repositories and their activity on GitHub in a news feed. For a developer, this effectively turns their GitHub profile into an easily accessible public portfolio of their open source development activities [36].

A GitHub user profile includes information on their repositories (i.e., projects) and their recent public activities, such as committing code to a repository or opening an issue report, which are usually not visible in other development environments. The profile page also shows several statistics that are often used on social networking sites, such as the number of other developers following a user or the number of projects they are watching. Such transparency is an interesting feature of GitHub and other social coding sites [42].

GitHub is particularly attractive for researchers because it provides access to its internal data stores through an extensive REST API [19], which researchers can use to access a rich collection of unified and versioned process and product data [9].

Before we detail how we accessed data on Brazilian developers using the GitHub API, we introduce basic demographics about Brazil in the next section to frame our research.

2.3 Brazil’s demographics

Brazil is the fifth-biggest country in the world in terms of area and population. With more than 200 million inhabitants, it is also the biggest country in South America and covers almost half (47.3 %) of the entire continent. Except for Chile and Ecuador, Brazil shares a border with every other country in South America.

Roughly 90 % of Brazil’s inhabitants live in states on the eastern and southern coasts of Brazil, where the population density varies from 20 to 300 residents per square kilometre. The rest of Brazil, i.e., the Amazon and the mountain regions, offers a lot more space with a population density of less than 5 residents per square kilometre in some cases. In contrast, the Federal District of the capital Brasília and the state of Rio de Janeiro have population densities of more than 300 inhabitants per square kilometre.

Brazil is divided into 26 states and a Federal District, which can be divided into five major regions: North. The North accounts for almost half of the area of Brazil (45 %), but it is the region with the fewest inhabitants. In particular the Northwest is not industrially developed. Instead, the region is home to the Amazon basin, the largest ecosystem on earth. The following states are in the North: Acre (AC), Amapá (AP), Amazonas (AM), Pará (PA), Rondônia (RO), Roraima (RR), and Tocantins (TO). Northeast. Almost a third of Brazilians live in the Northeast, a region that is culturally very diverse. It is characterized by Portuguese, African, and indigenous influences. The following states are in the Northeast: Alagoas (AL), Bahia (BA), Ceará (CE), Maranhão (MA), Paraíba (PB), Pernambuco (PE), Piauí (PI), Rio Grande do Norte (RN), and Sergipe (SE). Center-West. The Center-West of Brazil owes its importance to its wealth in raw materials. Nevertheless, the region is not particularly well developed. However, intensive efforts, such as the move of the capital to Brasília, are being made to strengthen the region. The following states are in the Center-West: Distrito Federal (DF), Goiás (GO), Mato Grosso (MT), and Mato Grosso do Sul (MS). The capital, Brasília, is located in the DF. Southeast. The Southeast of Brazil is home to more people than any other South American country. With the metropolitan areas of São Paulo and Rio de Janeiro, this region is the economic engine of the country. The following states are in the Southeast: Espírito Santo (ES), Minas Gerais (MG), Rio de Janeiro (RJ), and São Paulo (SP). South. The South is the smallest region of Brazil with climatic conditions similar to those of southern Europe. The region shows significant cultural influences from German, Polish, and Italian immigrants. The following states are in the South: Paraná (PR), Santa Catarina (SC), and Rio Grande do Sul (RS).

Brazil’s most populous metropolitan areas are São Paulo with about 20 million inhabitants, Rio de Janeiro with about 12.5 million inhabitants, and Belo Horizonte with about 5 million inhabitants, making São Paulo the largest city in the southern hemisphere.

Nowadays, Brazil’s economy is the seventh largest in the world in terms of nominal gross domestic product (GDP), and the seventh largest in terms of purchasing power parity. A member of the BRIC countries, Brazil had one of the world’s fastest growing major economies until about 2010 with economic reforms that gave the country new international reputation and influence. However, the economy has slowed down to modest growth over the last four years.

2.4 Brazil’s IT industry

According to a recent study [1], Brazil ranked 7th in IT investments worldwide and 1st in Latin America, with an investment of 61.6 billion US dollars in 2013. Of this, 10.7 billion came from the software market and 14.4 billion from the services market.

The domestic market is operated by approximately 11,230 companies, dedicated to the development, production and distribution of software and services. From those companies, about 93 % can be categorized as micro and small enterprises. Finance, Services and Telecom accounted for almost 51 % of the user market, followed by Industry, Government and Commerce.

The study also pointed out the regional concentration of investments in the IT market. The Southeast region of Brazil met the largest volume of funds allocated to the sector in 2013, with 64.6 %. The North of the country was the least invested in the sector, with a percentage of 2.2 %; the Northeast recorded 8.6 %; South and Center-West accounted for 13.4 % and 11.0 % respectively.

2.5 Hypotheses

There are many challenges associated with the Brazilian software industry and its growth. The vast majority of software companies are located in the Southeast and South regions of Brazil. In 2008, these two regions accounted for 84.3 % of all software companies in the country with more than 20 employees [37]. This result emphasizes the well-known inequality across regions in Brazil.

Assuming that the uneven distribution of software n and their employees across Brazilian states is related to the socio-economic situation in each of these states, our first hypothesis tests whether developers’ prestige is associated with the development level of the state they are located in:

H1. Developers’ prestige is associated with the development level of the state they are located in.

Our following hypotheses explore the follow-relationships between Brazilian developers on GitHub in more detail. Previous work on follow-relationships found that developers tended to connect with people with similar levels of performance and experience [35]. In fact, the presence of homophily, i.e. the tendency of individuals to associate and bond with similar others, has been discovered in many other network studies in sociology (see McPherson et al. [31] for a review).

For our paper, we examined the homophily of follow-relationships by focusing on two different attributes: geographic location and programming language choice. In particular, our second hypothesis investigates whether developers tend to follow other developers located in the same state:

H2. Developers tend to follow developers located in the same state.

As an alternative explanation of why developers might follow each other, we investigate whether they might use common programming languages as a decision factor:

H3. Developers tend to follow developers who use the same programming languages.

Prestige itself might be a factor for a developer when deciding who to follow. In a study on Twitter’s follow-relationships, Hopcroft et al. [13] found that the likelihood of two prestigious users creating a reciprocal relationship is nearly 8 times higher than the likelihood of two ordinary users. Our forth hypothesis tests whether, in Brazil, a prestigious developer tends to follow other prestigious developers.

H4. In Brazil, the prestige level of a developer who is following is associated with the prestige level of a developer who is being followed.

In a study with open-source software communities, Shen and Monge found that project leaders tend to follow more people, showing that project leaders are more well-connected than developers in other roles [35]. Our fifth hypotheses does a similar test by focusing on the association between network prestige and the number of people developers follow:

H5. In Brazil, developers’ prestige is associated with the number of developers they follow.

3.1 Data collection

To obtain data on software developers in Brazil, we accessed the GitHub API using the PyGithub module [20] to search for users who publicly stated their location in their GitHub profile [21]. We ran two different queries, both on November 5, 2014. The first query searched for users who had created their accounts between January 1, 2009 and November 2, 2014, and whose location contained the word “Brasil”. The second query was similar to the first one, but searched for a different spelling: “Brazil”. Whereas the first spelling is the Portuguese way of spelling Brazil, the latter is the one used in English.

The first query returned 8,815 unique users, and the second query returned 12,064 unique users. Merging these lists resulted in a total of 20,875 users that had either indicated “Brazil” or “Brasil” in their profile. Of these, 8,634 did not specify their location any further, i.e., their location only consisted of one word indicating the country. Since we are interested in state-specific information of developers in Brazil, we eliminated those from the dataset, leaving 12,241 users.

To test our third hypothesis, we used the Github API to get the number of lines of code by programming language for each repository owned by a given user. We summed up the number of lines of code by programming language and then assigned the language associated with the highest number of lines of code to that user.

3.2 Data preparation

To ensure that only active users were included in our dataset, we further filtered out GitHub users that had not made any contribution to a public repository within the last three months, i.e., after August 2, 2014. While we may ignore GitHub users that only contribute to private repositories this way, the decision was made in order to avoid noise in our dataset from individuals with a GitHub account that do not contribute to projects. We used GitHub’s definition of a contribution [28] in this step: GitHub considers it a contribution when a user pushes to a repository (PushEvent), when a user makes a pull request (PullRequestEvent), or when a user creates an issue (CreateIssueEvent). Of the 12,241 users left after the previous step, 7,977 had not made any contribution to a public repository on GitHub within the last three months, leaving us with 4,264 active users.

As an example, our pre-processing steps would transform “Franca/SP - Brasil” into “FRANCA SP BRAZIL” and “João Pessoa, PB, Brasil” into “JOAO PESSOA PB BRAZIL”.

Finally, we generated a table containing all follow-relationships among these 4,016 active developers in Brazil. The generated table contained tuples with unique user identifiers in the form <user1,user2> if User 1 followed User 2.

4.1 Developers in Brazil

Figure 1 illustrates the geographical distribution of active developers in Brazil per state. The state of São Paulo has the highest numbers of developers (1,424), followed by Rio de Janeiro (526) and Rio Grande do Sul (337). We were able to assign at least one developer to each of the 26 states and the Federal District, but the states at the lower end of the spectrum have very few developers: Roraima (4), Acre (2), and Amapá (1). It is worth noting that the latter states are all in the north region of the country.

Figure 2 shows how the active developers are distributed over the five regions of Brazil. The Southeast region has most of the Brazilian developers, whereas the Center-West and North regions have the least.

Network prestige was not equally distributed across the Brazilian states. São Paulo is the state with the most significant levels of network prestige (0.06672), followed by Rio de Janeiro (0.06213) and Rio Grande do Sul (0.06573). JavaScript is the main programming language among Brazilian active developers (21.0 %), followed by Java (14.3 %) and Ruby (11.9 %).

4.2 Hypotheses tests

To verify the correlations proposed in hypotheses H1, H4, and H5, we first evaluated the normality of our data by applying the Kolmogorov-Smirnov test [44]. The tests were significant for all variables in analysis for these three hypotheses, confirming the non-normality of data. Therefore, following the Bishara and Hittner recommendation [4], we applied Spearman’s rho (r _s ) to evaluate the association between variables.

H1. Developers’ prestige is associated with the development level of the state they are located in.

As presented in Section 3, the development level of each state is reflected in its characteristics of GDP, HDI, percentage of urban population, population density, number of higher education institutions, and speed of the Internet connection.

Table 3 and Fig. 3 show the correlation between prestige level (proximity prestige index) and all other variables related to hypothesis H1. Note that all variables were positive and significantly associated with proximity prestige index by Spearman’s rho. However, because the magnitudes of all correlations were not strong, indicating a possible important set of other variables with influence on the prestige in the social network, we conclude that H1 is supported with a very weak correlation.

State characteristic	Spearman’s rho
State GDP	.094**
State HDI	.062*
State Urban %	.091**
State Population Density	.094**
State Higher Education Institutions	.082**
State Average Internet Speed	.091**

*indicates significant correlation at the 0.05 level (2-tailed)

**indicates significant correlation at the 0.01 level (2-tailed)

H2. Developers tend to follow developers located in the same state.

We first summarized for each state the number of follow-relationships where the followed developer was from the same state as the follower. Then, we divided this number by the number of possible relationships between developers from the same state. We also summarized for each state the amount of follow-relationships where the follower developer was from the state and the followed developer was from a different state. Accordingly, we divided this number by the number of possible relationships between follower developers of the state and followed developers from different states. In other words, we created a coefficient for follow-relationships in the same state and follow-relationships across states.

To evaluate H2, we applied a pairwise Wilcoxon signed-rank test. Result showed a significant difference between the two generated coefficients (t=−4.211; p<0.001). All coefficients of followers in the same state were greater than coefficients of followers of developers from different states, except for three states (Roraima, Paraíba, and Mato Grosso). Therefore, we conclude that H2 is supported.

H3. Developers tend to follow developers who use the same programming languages.

To verify H3, we followed the same approach applied in H2. Similarly, we generated a coefficient of relationships for each programming language for follow-relationships with developers of the same programming language and follow-relationships from one programming language to others. Result pointed to a non-significant difference between the two coefficients (t=−1.461; d f=39; p=0.152). This result did not support H3.

H4. In Brazil, the prestige level of a developer who is following is associated with the prestige level of a developer who is being followed.

For this hypothesis we considered the above mentioned table of follow-relationships. We estimated the association between the proximity prestige indices of each developer in existent following relationships. As shown in Fig. 4 (left), although Spearman’s rho was positive and significant (r _s =0.231; p<0.001), its magnitude was considered weak [7]. Therefore, we conclude that H4 is supported with a weak correlation.

H5. In Brazil, developers’ prestige is associated with the number of developers they follow.

To test H5, we related the prestige level of developers with their out degree centrality level. As shown in Fig. 4 (right), result for the correlation test were positive and significant (r _s =0.394; p<0.001), but its magnitude was at most moderate [7]. Hence, H5 is supported with a moderate correlation.

6.1 External validity

External validity reflects the extent to which the results of a study can be generalized to other settings. We cannot claim that the results of our study are generalizable beyond Brazil’s borders, however, many emerging economies (in particular the BRIC-countries) face challenges similar to those of Brazil, and it is plausible that many of our conclusions apply to those countries as well.

In addition, our results are limited to the population of software developers found on GitHub, and cannot readily be generalized to every software project, open source or not. We chose to study GitHub’s population of users because their geographical location information was available on their profiles which enabled us to automatically mine it. However, GitHub is now the largest code host in the world [9], and we are not aware of any other publicly available data source for our study.

6.2 Construct validity

Construct validity reflects the extent to which our study actually measures what we claim to measure. For this study, the issues related to construct validity are mostly connected to the way in which users on GitHub specify their location. A GitHub user can provide any text as location, and the texts we collected may not represent a valid location in the world. In fact, we were not able to find state information for 248 of the 4,264 active GitHub users located in Brazil that had included additional information about their location.

In addition, we investigated the participation of developers based on the information they provided on their GitHub profile page. This information may not be frequently updated by the user. This means that the user may be working in other locations while contributing to projects on GitHub. Therefore, in our study, locations are merely indicators of where developers lived when they signed up for GitHub, and may not correspond to the their actual locations. However, a manual approach would have been infeasible for the amount of data needed to generate statistically significant conclusions, and we believe that the location information entered by GitHub users is at least a good approximation of where Brazil’s developers are actually located.

Another issue is the construct validity of the network prestige measure. For our study, this measure considered only the follow-relationships among developers. This measure may not correspond to the actual prestige of a developer in his social world or community. Further work is needed to assess whether follow-relationships are good predictors of one’s prestige among peers in development communities such as GitHub.

Finally, for the investigation of H3, we assigned the most used programming language for each user in our dataset. A developer on GitHub can use a variety of programming languages, and we ignored those languages except for the most used one. However, we believe our procedure offered a good approximation of developers’ preferences regarding programming languages.

6.3 Internal validity

Internal validity reflects the extent to which a causal conclusion based on the study and its methods is possible. Based on our statistical tests, we cannot infer causal relationships between socio-economic data and developers’ prestige, for example. However, because of our use of well-recognized statistical techniques, we are confident that the correlations we found hold for the data we collected and analyzed. To investigate whether the socio-economic situation caused developers to become more prestigious or vice versa, will be a goal of future work.

Also, we must point out that the correlation level estimated regarding H1, H4 and H5 should be considered as very weak, weak and moderate respectively. Hence, although the correlations were both positive and significant, those results must be considered with some care.