key: cord-0275750-172pzd6t authors: Rossi, Davide; Zacchiroli, Stefano title: Worldwide Gender Differences in Public Code Contributions date: 2022-02-15 journal: nan DOI: 10.1145/3510458.3513011 sha: 6864641517b4e59097712e639e44d2efed92e14e doc_id: 275750 cord_uid: 172pzd6t Gender imbalance is a well-known phenomenon observed throughout sciences which is particularly severe in software development and Free/Open Source Software communities. Little is know yet about the geography of this phenomenon in particular when considering large scales for both its time and space dimensions. We contribute to fill this gap with a longitudinal study of the population of contributors to publicly available software source code. We analyze the development history of 160 million software projects for a total of 2.2 billion commits contributed by 43 million distinct authors over a period of 50 years. We classify author names by gender using name frequencies and author geographical locations using heuristics based on email addresses and time zones. We study the evolution over time of contributions to public code by gender and by world region. For the world overall, we confirm previous findings about the low but steadily increasing ratio of contributions by female authors. When breaking down by world regions we find that the long-term growth of female participation is a worldwide phenomenon. We also observe a decrease in the ratio of female participation during the COVID-19 pandemic, suggesting that women's ability to contribute to public code has been more hindered than that of men. is, the difference in participation to public software development between men and women. Specifically, we study the development history of 160 million pieces of public software, developed over a period of 50 years by 43 million authors. We characterize the gender gap on this corpus over time and by world region. To determine author genders we rely on public data about name frequencies by gender around the world. To determine author locations we use email addresses, name frequencies around the world, and the timezone associated to each software change. We confirm that the gender gap in public code is huge. Female authors are only 8.1% of the total and have authored only 13.5% software versions. The gender gap is however shrinking, with women participation having increased steadily over the past 12 years. This improvement is a global phenomenon, observable in most world regions. We also observe a decrease in the ratio of female participation during the COVID-19 pandemic, suggesting that women have been more hindered than men in their ability to contribute to public code. Gender imbalance (or gender gap) is the situation in which, within a given group of people in society, a gender is significantly over-or under-represented with respect to the gender partition that exists in the world at large, which is close to a 50%/50% men/women split [24] . Gender imbalance tilted toward women under-representation and men over-representation in academia has been observed in several fields and is particularly severe across STEM disciplines (science, technology, engineering, and mathematics) [3, 5, 15, 31, 41] . This gender gap is even more acute in computing, where it has been so for many decades now [12, 21] , with only recent signs of being on the decline [42] and a long way to go before reaching gender parity. In the context of software development, Free and Open Source Software (FOSS) projects have been frequently analyzed from the gender imbalance angle, confirming multiple times [2, 8, 20, 25, 28, 32, 36, 39, 40, 42] the under-representation and very low participation (in relative terms) of women in FOSS. These results have been obtained via different techniques-from surveys to interviews and name-based analyses-and at very different scales-from individual projects and focus groups to very-large scale analyses of public code-leaving little to no doubts about the existence of a gender gap in FOSS and public software development. The geography of the gender gap in public code is a relatively under-explored angle of gender imbalance, which we explore in this paper at very large scale and along two orthogonal dimensions: time and space. Along the time axis, and following in the steps of recent related work [42] , we analyze more than 50 years of public code contributions from the Software Heritage archive [9] , consisting of more than 2.2 billion commits contributed by 43 million distinct authors, and we classify them by gender using a frequencybased approach applied to author names. This gives us a baseline worldwide historical trend of the evolution of the gender gap in public code contributions, which generally confirms previous results: the overall amount of contributions authored by female authors is very low, but is also steadily growing. Along the space axis we break down code contributions geographically to verify if there are significant differences in gender gap trends across different world regions. Specifically, we will answer the following research questions: RQ 1. What is the overall breakdown by gender and UTC offset in contributions (and contributors) to public source code? RQ 2. What is the overall breakdown by gender and by world regions in contributions (and contributors) to public source code? RQ 1 is our first approximation of the geographic position of contributors worldwide. The UTC offset is the difference in minutes between Coordinated Universal Time (UTC) and local time at a particular place on Earth. As UTC offset values are spread along the East-West axis of the planet, and following previous work [14] , we use them to group developers by (coarse-grained) longitude, before breaking down each group by author gender. RQ 2 is a more refined approximation of the location of commit authors, this time at a granularity of a division of the world in 12 regions, loosely based on the world sub-regions of the United Nations geoscheme [23] . To geolocate public code contributions and their authors at this granularity we use a heuristic based on commit timezones, name frequencies world-wide, and country code top-level domain (ccTLD) found in commit emails. In answering these two research questions we find that the longterm growth of female participation is a world-wide phenomenon that is observable across all UTC offsets and all zones in the analyzed corpus albeit with trends showing appreciable regional differences. While analyzing the data used to answer RQs 1 and 2, we noticed a worldwide phenomenon: the decrease of the ratio of women participation to the production of public code for the year 2020-the year when the COVID-19 pandemic struck. Hence, although that was not part of the initial study design, we also separately studied this phenomenon and address it in the paper as the following research question: RQ 3. Has the impact of the 2020 COVID-19 pandemic on contributions to public code been quantitatively different by gender? The historical trends, both world-wide and region-by-region, that emerge from our analyses suggest that women have been more negatively hindered in their ability to contribute to public code during the pandemic than men, reverting for the first time in 2020 the positive trend of the ratio of contributions by female authors, which had been growing steadily since the 90s. Paper structure. We review relevant related work in section 2. We detail our analysis methodology in section 3. We present the obtained raw results and discuss them in light of our research questions in section 4. Before concluding, we discuss limitations and threats to validity in section 5. Data availability. A replication package for this paper is available from Zenodo [34] . In early work on this topic, Hill et al. [15] summarized the underrepresentation of women in STEM, documenting the quantitative extent of the phenomenon and discussing when female students drop out from an initially well-balanced funnel of students. More recent work [3, 5, 31, 41] review the status quo for STEM, including quantitative and qualitative analyses of the phenomenon, as well as analyses of which practices in society (including in education) contribute to it. Margolis and Fisher [21] looked into the case of computer science students, at college level and below, characterizing the gender gap in the field and also exploring one of its main causes (i.e., the widespread and self-reinforcing assumption that computing is a "boys' clubhouse") based on interviews with a pool of more than 200 college students. From a theoretical framework point of view, the gender gap in FOSS has been explained by Nafus [22] as a consequence of interaction practices that hinder women's inclusion. Empirically, the rapid increase of free/open source software has attracted since the early 2000s' attention to the gender gap in FOSS, which has been verified to be more severe than in computing at large. Researchers have resorted to different techniques to characterize the FOSS gender gap. Multiple survey-based studies of FOSS contributors have reported low ratios of women respondents. Surveys [8] up to 2003 reported that 95-99% FOSS participants selfidentified as men. A more recent survey of 2000 FOSS contributors in 2013 [32] reported a ratio of 10% women respondents. In addition to surveys who invited FOSS contributors at large without preselecting projects, the FOSS gender gap has been verified also within specific FOSS communities, such as Debian [25] , KDE [28] , and OpenStack [19] . Specific artifacts resulting from software development activities, in both FOSS and collaborative software development in general, have been analyzed to quantify the gender gap. Mailing lists of early FOSS projects have been analyzed for gender differences by Kuechler et al. [20] , finding evidence of declining female participation over time. Stack Overflow and GitHub teams have been studied for analogous reasons by Vasilescu et al. [39, 40] , finding compatible evidence. Terrell et al. [36] studied pull requests on GitHub, finding evidence of gender bias against code contributions coming from women outsiders. Bosu and Sultana [2] mined code reviews of ten popular FOSS projects, determining that only 10% of active contributors were women. Trinkenreich has designed [37] a holistic research agenda to investigate how FOSS communities can actively increase the participation of women in their projects. In preliminary work along that path Trinkenreich et al. [38] identified different career paths in FOSS, as each of them might require different engagement strategies to attract and retain diverse contributors. Canedo et al. [4] conducted a mixed-methods study to investigate the gender gap for women core developers in FOSS, confirming its existence but noting that interviewed women core developers did not report having experienced gender discrimination in FOSS. Rastogi [30] has studied a different form of bias in public code development, that related to geographical location, showing that it has an impact on maintainers' decisions on whether to accept a pull request or not. Lacking pull request data in our dataset it is not something we can explore, but we provide baseline data about developer origins that can support future similar studies. A very large scale gender study of public code contributions has been conducted by Zacchiroli [42] , studying all public code commits available from Software Heritage to observe the evolution over time of the gender gap. In this study we use the same data source and the same tool (gender-guesser) for gender identification, with the following differences: (1) we characterize the geography of the FOSS gender gap, allowing comparisons across world regions; (2) we use a more recent data set, covering 1 more year (specifically: 2020, allowing to answer RQ 3). The impact of the COVID-19 pandemic on the gender gap due to an uneven distribution of caregiver responsibilities among parents has been investigated in several studies [1, 7] . Ralph et al. [29] have investigated the effects of the COVID-19 pandemic on software development specifically, via a survey of more that 2000 developers worldwide. They tentatively suggest that "women, parents and people with disabilities may be disproportionately affected". In answering RQ 3 we provide the first body of empirical evidence that the pandemic has indeed disproportionately affected women's ability to contribute to public code. Across several works [13, 14, 33] Barahona et al. have established a methodology to pinpoint (coarse-grained) developer locations using various signals that include commit timezones, email domains, and mailing list participation. We adopt similar heuristic, using as signal only information coming from public code commits (namely: names, timezones and email ccTLDs). A very recent study by Prana et al. [27] ("to appear" at the time of writing) pursues similar objectives to ours, and RQ 2 specifically, albeit with a different approach. The authors also study the gender gap in FOSS worldwide, but do so using a mixed-methods approach consisting of repository mining and targeted developer surveys. The scale of the present study is much larger both in terms of analyzed repositories (160 M v. 22 K) and time period (50 years v. 7); the drawback of this scale is that we could not further corroborate our findings with data coming from developers via surveys. In terms of findings we observe the same general trends of worldwidelow gender diversity and fast(er) increase in specific world regions, although details differ slightly across regions. The data flow and main components of the methodology adopted to answer RQs 1 to 3 are depicted in fig. 1 . To answer our research questions we need to assign a gender and a world region to the authors of commits in the dataset. Some terminology considerations are in order about both axes. Regarding gender, in the following we will refer to automated classification decisions described as "gender detection" or "gender assignment". With that we do not intend to arbitrarily define people within a binary gender confinement regardless of their preferences and sensitivity. None of the gender-related decisions made by the automated techniques used in this paper make sense when applied to individuals included in the analyzed corpus. The meaning of the exercise is statistical in nature and aims only to address the stated research questions. The used approach makes sense only in aggregate form and carries with it the unavoidable limitations that name-based gender detection entail; we elaborate on those limitations in section 5. Regarding the geolocation of commits and their authors, we only consider the macro geographical areas in which contributions are likely to have been made, at the granularity of large world regions. As geolocation targets we use the 12 regions shown in fig. 2 , namely (in alphabetical order): Africa, Australia and New Zealand, Central and South America, Central and South Asia, China, East Asia, Europe, North America, Pacific, Russia, South-eastern Asia, West Asia. To obtain them we started from the United Nations geoscheme [23] as devised by the United Nations Statistics Division, to which we applied some merges and split based on geographical distance and/or on the sharing of preeminent cultural identification features, such as spoken language, only when needed to avoid under-or over-represented data samples (e.g., to avoid that China dominates the East Asia subsample or Russia the East Europe one). Other than that, considerations similar to those made about gender also apply to national or cultural identities: no assessment about the identity of individuals in the corpus is implied, obtained figures only make sense in aggregate form, and limitations (discussed in section 5) apply. As starting point we retrieved from Software Heritage [26] a snapshot of all the commits archived until 2021-07-07. It consists of 2 198 808 389 commits, unique by SHA1 identifier, harvested from about 160 million public projects coming from major development forges (GitHub, GitLab, etc.) and source code distributions (Debian, PyPI, NPM, NixOS, etc.). Commits in the dataset have been contributed by 43 381 366 authors, unique by ⟨name, email⟩ pairs. Obtained commits came as two relational tables, one for commits and one for authors, with the former referencing the latter via a foreign key. Each row in the commit table contains the following fields: commit SHA1 identifier, author and committer timestamps, author and committer identifiers (referencing the author table). The distinction between commit authors and committers come from Git, which allows to commit a change authored by someone else. For this study we focused on authors and ignored committers, as the difference between the two is not relevant for our research questions and the amount of commits with a committer other than its author is negligible. For each entry in the author table we have author full name and email as two separate strings of raw bytes. Looking into the raw author full names we realized that some of them are not real author names, but rather emails or gibberish strings, likely coming from misconfigured VCS tools. Hence as a preliminary analysis steps we filtered out implausible or unusable author names in the dataset, such as: names that cannot be decoded as UTF-8 strings, email addresses used as names, names consisting of only blank characters, names containing more than 10% nonletters, and names longer than 100 characters. We did not perform any other data filtering or selection. After filtering, 33 351 300 Figure 1 : Analysis data flow. Starting from the Software Heritage dataset we detect the gender of commit authors using gender-guesser on name tokens and a majority criterion. We geolocate commits with two methods: (1) using country-level top-level email domains (e.g., .uk) extracted from commits; and (2) comparing commit author names to the most popular names among the countries that have a timezone compatible with the commit timestamp offset. "plausible" authors remained (76.9% of the initial authors), having authored 1 735 130 408 commits (78.9% of the initial commits). Figure 3 shows the evolution over time of the commits in the dataset, for the period 1970-2020 1 . The dataset appears to grow exponentially over time, both in terms of commits and authors in it, and has done so for almost 50 years now. Exceptions are the first and last (complete) years in the dataset, respectively 1970 and 2020, for different reasons. 1970 contains the UNIX epoch (1 January 1970 at midnight UTC), which is often used as "default" timestamp for older or missing points in time, ending up being over-represented. 2020 being close to when we obtained our dataset, we estimate its dip corresponds to a Software Heritage archival delay: other 2020 public code commits exist, but they had not been archived yet by Software Heritage at the time. The exponential growth of the dataset will be relevant when discussing historical trends, like gender ratios, as more recent data points will correspond to exponentially larger amounts of commits. With these considerations in mind, we proceeded to partition our corpus along the orthogonal gender and world region axes. To detect the gender of a contribution we use the name of its author, as recorded in the corresponding commit, and apply gender-guesser [10] to it. gender-guesser is an open source Python tool and library for gender detection, which is based on first name frequencies around the world and that is frequently used in related work. The availability of the package as open source was an important point in favor of its adoption: alternatives based on commercial APIs might give better accuracy (for some world regions), but would both hinder replicability and be very expensive on such a large corpus. A detailed description and a comparative benchmark of gender-guesser and its main competitors has been conducted by Santamaria and Mihaljevic [35] . It shows that gender-guesser works comparatively well with geographically diverse datasets, as ours is by construction. gender-guesser takes as input a Unicode string, which is expected to be a first (given) name, and returns a detected gender among 6 possible values, depending on the tool's confidence about the result: male, mostly male, unknown, mostly female, female, andy (the last one for unisex names). Author names in our corpus are full names, not split into first versus family names. Aside from the "API" mismatch problem, the first/family name distinction is not meaningful across all world cultures represented in our corpus [18] . Hence, to determine the gender of an author we use a majority criterion. Specifically, we tokenize full name strings into name tokens, splitting at each blank, hyphen, or case change (as in CamelCase notation, which we have verified to be used by several authors in the dataset), and then use gender-guesser to determine the gender of each token. If and only if a strict majority of name tokens for a given author full name is detected as belonging to one gender (no matter how strongly) we associate the majority gender to the author; otherwise their gender will remain unknown. After this step all commit authors get associated to either a gender or unknown. As each commit is associated to exactly one author, we can also partition commits by detected gender, by making them inherit the gender of their authors. As shown in fig. 1 region detection is performed using two different techniques. The first geolocation technique, ccTLD, uses the country code top-level domain (ccTLD) found in the domain of the email address of commit authors, as recorded by version control system (VCS) commits in our dataset. We relied on the IANA list of Latin character ccTLDs [17] and manually mapped each corresponding country, sovereign state, or dependent territory to one of the world regions in fig. 2 . The second geolocation technique, tz&name, uses the UTC offset and author name of each commit to detect the most likely world region in which the commit was authored. The UTC offset is the difference in minutes between Coordinated Universal Time (UTC) and local time at a particular place on Earth. In the initial dataset each commit is associated to a UTC offset. From it we determine a list of compatible places that, at the time of the commit timestamp, had the same UTC offset as that of the commit. For making this determination we use the IANA time zone database (or zoneinfo) [16] , which includes as places countries and territories worldwide. For example the following places (with their zoneinfo name in parentheses) had a UTC offset of +240 minutes when their local time was 2012-01-01 12:00:00: Russia MSK (Europe/Moscow), Azerbaijan (Asia/Baku), United Arab Emirates (Asia/Dubai), Russia SAMT (Europe/Samara), Oman (Asia/Muscat), Georgia (Asia/Tbilisi), Armenia (Asia/Yerevan), Mauritius (Indian/Mauritius). When considering local time 2012-08-01 12:00:00 Azerbaijan is removed from the list because the country moves to a different offset due to daylight saving time (DST). Then, on local time 2016-08-01 12:00:00, Russia MSK is not included anymore, as whole Russia stopped using DST in 2014 and Samara changed offset in 2016. Then we assign to each compatible place a score that captures the likelihood that a given author name is assigned to a person in that place. To this end we use a dataset of the frequencies of the most common first and family names by Forebears which, quoting from [11] : provides the approximate incidence of forenames and surnames produced from a database of 4 044 546 938 people (55.5% of living people in 2014). As of September 2019 it covers 27 662 801 forenames and 27 206 821 surnames in 236 jurisdictions. As for gender detection, lacking a first/family name split, we first tokenize names as before and then lookup individual name tokens in both first and family names frequency lists. For each element found in name lists we multiply the place population 2 with the name frequency to obtain a measure that is proportional to the number of persons bearing that name (token) in the specific place. We sum this figure for all elements to obtain a place score ending up with a list of ⟨place, score⟩ pairs. We then partition this list by the world region that a place belongs to and sum the score for all the places in each region to obtain an overall score, corresponding to the likelihood that the commit belongs to a given world region. We assign the starting commit as coming from the world region with the highest score. The two geolocation techniques-ccTLD and tz&name-suffer from different weaknesses. The main problem with email is recall: for only about 13% of the commits it was possible to associate an author email with a ccTLD. While at the scale of the full dataset an eviction of about 87% leaves a very large dataset to work with (≈ 300 M commits), when projecting to specific years and world regions we have to deal with troublesomely small sample sizes. As for tz&name a relevant issue is that of the UTC offset 0, or time zone zero (TZZ), which is overrepresented in the dataset. This is due to a number of reasons, including: wrong time zone settings in development environments, incorrect timestamp migrations when a repository was converted across VCS systems (e.g., from Subversion to Git), archival errors, and more. The over-representation of TZZ commits in the dataset decreases over time: 2000 had 96% TZZ commits, 2010 had 64%, and 2020 only 22% (with a dataset grown about 3 orders of magnitude since 2000). Due to their respective strength and weaknesses, we analyzed commits using both geolocation techniques, as well as a mixture of them (such as using ccTLD for TZZ commits and tz&name for 2 To obtain population totals, as the notion of "place" at hand is heterogeneous, from a full country to a slice of a larger country spanning multiple timezones, we used a mixture of primary sources (e.g., government websites), and non-primary ones (e.g., Wikipedia articles). Figure 4 : Breakdown of the entire corpus by detected gender for authors (left) and authored commits (right), excluding those for which gender could not be determined. the rest) and compared results. We obtained results that, although with noticeable differences in absolute values, exhibit consistent long-term trends, with the sole exception of cases in which the sample size is very small (10 commits or less). Using the technique (gender-guesser + majority criterion) discussed in section 3 we were able to detect the gender for 64.1% of the authors in the dataset, corresponding to 21.4 M author names. Those authors have contributed 50.3% of the commits in the dataset, for a total of 872.3 M, which we can further partition by gender. Figure 4 shows the extent of the gender gap in our corpus as a whole, after the exclusion of unknown-gender authors and commits, but before any other subsampling. The vast majority (86.5%) of authors is detected as being male for 10.4 M authors v. only 1.6 M female authors. The imbalance is even larger when looking at commits, where 91.9% of commits (for a total of 10.4 M commits) have been authored by men v. 8.1% (1.6 M commits) by women. But how does this massive imbalance evolve over time, and how does it change around the world? We start by answering RQ 1, which is a first approximation of contributor locations on the east-west world axis. We remind that, by looking only at commit UTC offsets, as needed to answer RQ 1, we lack any information about timezones, because they change over time (e.g., due to daylight saving time (DST)) and depend on country-specific regulations. For what is worth, since DST is mainly adopted in the northern hemisphere from March to November, we repeated the analysis discussed below filtering out all commits performed over that period (which should not introduce relevant gender bias), obtaining results that are analogous to those presented below for the entire dataset. Figure 5 (a) shows the evolution of the ratio of commits by female authors over the 20 most active time offsets (in terms of available commits in the dataset), displayed as a stacked bar chart. The solid line in each chart is a loess regression curve [6] that materializes the data trend. The small pie chart to the right of each diagram shows, in black, the ratio of the number of commits that is contributed by that time offset to the entire dataset. In the following we show and discuss only trends for the period 2000-2020. Data and charts for the 1970-2000 are available in the replication package, but fall into the pitfall of (1) dealing with partitions derived from an exponentially smaller starting dataset (see fig. 3 ), which (2) is further partitioned by UTC offsets, resulting in slices that are so small to be of limited significance. Figure 5 (b) shows the same evolution, but this time in terms of the number of distinct authors, as opposed to the number of authored commits. To avoid outliers due to sporadic contributors we only count authors that have contributed at least 5 commits in a given year. This threshold is a compromise between the ability to filter out anomalies introduced by drive-by contributors and the significance of the data after filtering. The value we adopted is the result of a qualitative process started with a large threshold, iteratively lowered until no appreciable outliers were present and yet all data slices were still reasonably represented. The replication package also includes charts obtained with different threshold values. We propose these two views-commits in fig. 5 (a) and authors in fig. 5 (b)-to overcome inherent potential biases when only discussing one of the two metrics. Indeed, commit-based gender imbalance can be influenced by prolific contributors (i.e., authors contributing a large number of commits in a given year), whereas author-based gender balance can be influenced by a large number of less prolific contributors. Note that the stacked graphs for different UTC offsets have different maximum values on their Y axes and hence difference scales. This is because name-based gender detection is subject to appreciable biases when operating in different world cultures, due to phenomena that include: the adoption of typical female names also for males, the presence of gender-neutral names, a gender-imbalanced reference database, and others. This makes the comparison of the absolute value of ratios between different world zones of little significance, without undermining the comparison of the evolution trends of these ratios, because the gender detection approach does not vary over time. This is what the stacked charts of fig. 5 allow to do, without having to vertically squeeze some charts too much due to a common scale. With one exception (discussed below) both views highlight a common trend. The participation of women to the production of public code has grown steadily over the past 12 years, across all UTC offsets, in terms of both contributed commits and active yearly authors. The sole exception is offset UTC+240, where the trend is less clear in the authors chart, due to the dip in the 2011-2014 period; after that, the growth is confirmed for 2015-2020. We note that Russian regions are by far the most populated among those found in this UTC offset and that the years are compatible with when Russia changed their DST adoption, resulting in regions changing UTC offsets. We therefore speculate that this irregular behavior might be caused by a drastic change of areas (and population) included in this dataset slice. Offset +360 also shows a recent consistent growth, but the overall commits trend is dominated by the spike in 2001, making subsequent years appear low in comparison. Looking at the data for 2001 in this offset shows the presence of a single strong female author and only other 6 less prolific authors. Moreover, the name of the strong author in question, although detected as female with our approach, is a name used for both genders in most of the countries at those latitudes. A similar effect is visible in the commit chart at offset -480, which is dominated by a 2003 spike. A detailed analysis of the data shows the large influence of a single strong author (who contributed 495 785 commits), detected as female but likely either a bot or somebody committing changes authored by others. Overall, the charts also show a growth trend for the 2000-2007 period, although a less steep one and less ubiquitous across UTC offsets. We observe that, due to exponential dataset growth, the amount of 2008-2020 commits is about 25 times larger than that of 2000-2007, which makes it far easier for the data of less populated regions to be subject to erratic trends due to specific outliers. For example, the previously discussed commit spike at +240 in 2005 is due to two authors who together contributed 92% of the commits performed by all women that year. This is the kind of effect that can affect the commits chart; we expect this not to affect the authors chart and, indeed, fig. 5(b) shows no signs of it. Another outlier is at offset +420 in 2005 (along with two minor ones in 2000 and 2001) that reverses the trends. A detailed look at the data shows a very small data sample (23 commits in total) and a case of doubtful gender detection. Outliers aside some offsets (including the well-represented +60) show a 2000-2008 trend of stable and even decreasing ratio of women contributions, with no obvious anomalies in the data, which lead to wonder if there are underlying regional phenomena apply. We explore this aspect by addressing RQ 2. For the reasons discussed in section 3 we use a mixed strategy to assign authors to the world regions of fig. 2 : for commits with UTC offset zero we rely on the ccTLD strategy, whereas for other commits we use the tz&name strategy. Figure 6 shows the ratio of both yearly female authors and that of commits authored by women as a set of stacked charts, this time broken down by detected world regions rather than by UTC offset (as it was the case in fig. 5 ). Previously discussed caveatsabout different Y-axis scales and the filtering on active authors by at least 5 yearly commits-still apply to these charts. The breakdown by world regions confirm the presence of a stable growth trend, with few exceptions (discussed below). Over the past 12 years both the ratio of female authors and that of their commits have grown steadily across most world regions. Relevant differences across regions, in both the volume and tendencies of the trends, are noticeable this time. The recent trend for Central and South Asia (which includes India) shows a very slow growth which even seems to have plateaued in the last few years. A slow growth (at least in one of the two studied metrics) also characterizes West Asia, recent years in Southeast Asia, and partially China. Taken together these seem to suggest that Asian zones are subject to a different recent dynamic characterized by a slower growth of female contribution. When taking into account the whole 2000-2020 time span, it appears that the growth that characterizes 2008-2020 was not present in 2000-2007. In some cases growth is less evident, in some cases there is even a reduction in the ratio of women participation, resulting in an overall trend with an upward concave curve hitting a minimum at about 2008. Further investigation shows that for some of the regions this behavior can originate from local anomalies. For instance, the West Asia subsample is dominated by Israel, where gender-guesser detects as female "Eli", which is an Hebraic male name that is also incidentally the name for several prolific committers in those years. China is subject to effects related to being a very limited subsample: the only region with less commits than China is Pacific. This translates, for the first years of the studied period, to a very limited number of commits and authors (less than a dozen is some cases) so the overall trend can be severely affected by even small anomalies. In all other cases there seem to be enough good-quality data to support the conclusion of a legitimate shrinking/expanding trend, which can be seem for commits (and partially authors) in the Americas, South East Asia, and Africa. Overall, we find evidence that female participation to public code production has been growing stably in most world regions over the past 12 years, with a less pronounced trend in some Asian countries. Looking further back to the past 20 years, women participation to public code is subject to different regional developments. Previous results consistently show a marked growth trend in female participation to the production of public code over the last 12 years. However, when looking closely at previous charts a recurrent worldwide anomaly also jumps to the eye: women participation in the production of public code has decreased everywhere in 2020 with respect to 2019. Whereas this decrease is not enough to change the loess trend, it is nevertheless present. More precisely: across all world regions the ratio of yearly active female authors has decreased in 2020 w.r.t. 2019; and across all regions but one, the ratio of commits contributed by women has decreased in 2020 w.r.t. 2019. To better highlight this phenomenon we visualize it in fig. 7 , for authors only, zooming into the 2016-2020 period, and using stacked line charts. Note that this decrease is in stark contrast with the worldwide trend of increased women participation over the past 12 years. Remember also that the observed variations are in ratios, which could not be explained by an incomplete dataset or other cross-cutting phenomena: after getting better for almost 12 years the gender gap has worsened again in 2020. While not originally part of our study design, this recurrent anomaly led us to state RQ 3 about the impact of the COVID-19 pandemic on women participation in the production of public code. The decrease in women participation in 2020 stands out as an anomaly in the observed trends up to 2019 and is correlated with the insurgence of the pandemic in 2020. While with the data we have at hand we cannot verify a causal relation between the two, it is our educated guess that this decrease has been caused by the COVID-19 pandemics. The closures of schools and daycare facilities have increased responsibilities for caregivers which, as many studies have shown (among them [1, 7] ), have not been equally split between genders, impacting more on women's ability to continue working than men's. A plausible interpretation of our data is hence that women's ability to contribute to public code-either as part of their day job, or as an activity conducted during their spare time-has also been negatively impacted by the COVID-19 pandemic, increasing the gender gap in 2020. Interestingly enough, the impact of this imbalance seems to be more limited in Asiatic regions than elsewhere in the world. Whether this is due to a different impact of the pandemic on gender imbalance in those societies, or to different local contribution patterns to public code between genders, is something we cannot assess and that remains to be explored as future work. In this section we discuss threats to the validity of the obtained results, starting with construct validity concerns. Accuracy/scale trade-off. This is the first study, to the best of our knowledge, that has analyzed at this scale the geographic dimension of the gender gap in public code contributions. Dealing with such a large dataset (43 million authors and 2.2 billion commits, before any filtering) calls for the crude, but fully automated methods, that we have adopted for detecting both author genders and their coarse-grained location. Alternative approaches found in the literature rely on crawling individual author information (e.g., from social media or development platform profiles) and even one-to-one interviews with them. There appears to be a clear trade-off between the accuracy of the latter methods and the ability to scale to much larger datasets of the methods we have chosen for in this study. Considering that: (1) we have relied for the most parts on methods and techniques that are also found in the literature (even though they have not been applied at this scale before), (2) we are only drawing conclusions about long-term aggregate trends, and (3) the trends we observe appear to be statistically stable; we do not consider our general choice of analysis methods a significant threat to the validity of our answers to the stated research questions. It is still worth reviewing specific methodological choices. Gender detection. Regarding gender detection, the choice of gender-guesser as building block is based on a preexisting benchmark [35] of automated gender detection tools and on the fact that, being open source, it enables replicating our findings without depending on third-party APIs. On top of the tool itself we added a majority criterion, due to the need of working with non-parsed author names. It is trivial to come up with handcrafted cases that break this heuristic, for example with family names composed by name tokens that are also common first names detected as belonging to the "wrong" gender w.r.t. actual author gender (which is a phenomenon mostly affecting Chinese names). Other than that, though, most family names are detected by gender-guesser as being of an "unknown" gender, which does not skew the gender majority of an author in any direction. Hence in practice, especially at this scale, we do not expect this aspect to significantly impact experimental results. Another limit of gender-guesser only marginally impacts our study: it only operates on latin alphabet names making it unapplicable to names adopting other alphabets. However we observed that most developers from countries such as Japan, China, South Korea, Thailand and others, usually adopt Western names when contributing to public code, limiting the impact of the problem. The proposed heuristic is unable to determine a gender for a large part of the starting dataset (35.9% authors and 49.7% commits), which might impact results. The remaining part of the dataset still corresponds to the largest scale study of this kind, which we believe is important to report about. We have also mitigated this risk by just looking at trend ratios within the subset of authors for which we could determine a gender. It will be trivial in the future to integrate upcoming improvements in name-based gender detection in replications of our experiments. Region detection. The two techniques used for region detection has complementary strengths and weaknesses. On the one hand, ccTLD-based region detection, in addition to having limited applicability, is affected by the Internet practices that surround the use of national top-level domains (TLDs). Some of them are rarely used, such as .us for the USA, where generic TLDs (.com, .net, .org, etc.) originated from and remain to this date much more popular than elsewhere. As a consequence ccTLD-based region detection leads to the underrepresentation of countries like the USA. Conversely some ccTLDs are used worldwide as part of popular "domain hacks", e.g., .io from British Indian Ocean Territory used for computer-related websites, or .tv from Tuvalu used for television ones. On the other hand, tz&name-based region detection relies on population totals retrieved from different and potentially heterogeneous public sources. In particular official population figures are usually aligned with national census that, in some cases, can be up to 15 years old. As a consequence we might have ended up comparing, say, 2020 population figures with 2005 figures. For places with relatively stable population this is not an issue, but it could become one in more unstable regions. We expect the impact of these weaknesses to be marginal in our dataset overall. Nonetheless we have mitigated their potential effects by combining the two techniques as detailed in section 3. We have also compared results obtained with each technique separately, obtaining similar results outside of the peculiar timezone zero. The dataset we have analyzed does not correspond to the fully body of neither public code nor free/open source software. Our findings hence inherit the limitations of Software Heritage as a research archive, in particular related to the (lack of) coverage of specific development platforms or software distribution technologies, and to archival lag (which we have observed for the last "complete" year in our dataset: 2020). The study dataset is, however, the largest approximation of contributions to public code that is readily and publicly available for analysis, which we believe validates our choice of starting point. Larger samples of public code can be analyzed (and probably will in the future), but we do not expect significant differences to emerge from similar incremental improvements in coverage. It would be more relevant to complement our analysis adding significant bodies of non-public code, such as those developed via large in-house private forges, which can potentially constitute a very different population in term of gender gap evolution. We make no claim about the generalizability of our findings to those contexts. We have studied the gender gap in public code contributions along the orthogonal axes of time and geographic location of contributors. To that end, we have used heuristics based on name frequencies, email domains, and timestamp offsets, that enabled us to analyze 2.2 billion commits contributed by 43 million authors over a period of 50 years. We confirm previous results about the gender gap in public code: women have contributed less than 10% of public code overall, but the ratio of their involvement is growing steadily. We provide novel evidence that this growth-both in terms of active female authors and of their commits-is a global trend, shared by most world regions over the past 12 years. However, 2020 has been a setback year, with the ratio of women participation decreasing everywhere in the world, most likely due to the COVID-19 pandemic disproportionately affecting women. Future work. As future work, region-specific analyses to understand local trends would be useful in particular, but not only, to understand why the COVID-19 contraction in women participation appears to have impacted Asiatic regions less than others. The ability to study large bodies of non-public, but still collaboratively developed, code is also to be pursued, in order to compare collaboration dynamics before versus away from the public eye. We also intend to attempt large-scale validation, which remains challenging on datasets of this scale, as well as quantitative comparisons with findings by other studies in selected world region. Provided that gender diversity results for identifiable subsets of the population analyzed in this study are available from other studies, one can cross-check results to either reinforce the respective findings or pinpoint the causes of discrepancies, informing future studies. The impact of COVID-19 on gender equality Diversity and Inclusion in Open Source Software (OSS) Projects: Where Do We Stand Gender diversity in STEM disciplines: A multiple factor problem Work Practices and Perceptions from Women Core Developers in OSS Communities Cracking the code: Girls' and women's education in science, technology, engineering and mathematics (STEM) Robust Locally Weighted Regression and Smoothing Scatterplots Scarborough. 2021. COVID-19 and the gender gap in work hours Community-based production of open-source software: What do we know about the developers who participate? Software Heritage: Why and How to Preserve Software Source Code Online at https:// forebears.io/about/name-distribution-and-demographics, accessed during Women and computing Geographic origin of libre software developers Determining the Geographical distribution of a Community by means of a Time-zone Analysis Why so few? Women in science, technology, engineering, and mathematics Country code top-level domains Personal names around the world Gender Differences in Early Free and Open Source Software Joining Process Unlocking the clubhouse: Women in computing Patches don't have gender': What is not open in open source software Standard country or area codes for statistical use Department of Economic and Population Division Social Affairs Preliminary Report on the Influence of Capital in an Ethical-Modular Project: Quantitative data from the 2016 Debian Survey The Software Heritage graph dataset: public software development under one roof Rahul Purandare, and Nachiappan Nagappan. 2021. Including Everyone, Everywhere: Understanding Opportunities and Challenges of Geographic Gender-Inclusion in OSS Joining and Socialization in Open Source Women's Groups: An Exploratory Study of KDE-Women Amin Milani Fard, and Rana Alkadhi. 2020. Pandemic programming Do biases related to geographical location influence workrelated decisions in GitHub The gender gap in STEM fields: Theories, movements, and ideas to engage girls in STEM Women in Free/Libre/Open Source Software: The Situation in the 2010s Geographic location of developers at SourceForge Worldwide Gender Differences in Public Code Contributions -Replication Package Comparison and benchmark of name-to-gender inference services Gender differences and bias in open source: Pull request acceptance of women versus men Please Don't Go -A Comprehensive Approach to Increase Women's Participation in Open Source Software Hidden Figures: Roles and Pathways of Successful OSS Contributors Gender, representation and online participation: A quantitative study Gender and tenure diversity in GitHub teams Gender gap in science, technology, engineering, and mathematics (STEM): Current knowledge, implications for practice, policy, and future directions Gender Differences in Public Code Contributions: A 50-Year Perspective