Tools that love to be together

[updates in brackets below]
[see also follow-up post: Stringing beads: from tool combinations to workflows]

Our survey data analyses so far have focused on tool usage for specific research activities (e.g. GitHub and others: data sharing, Who is using altmetrics tools, The number games). As a next step, we want to explore which tool combinations occur together in research workflows more often than would be expected by chance. This will also facilitate identification of full research workflows, and subsequent empirical testing of our hypothetical workflows against reality.

Checking which tools occur together more often than expected by chance is not as simple  as looking which tools are most often mentioned together. For example, even if two tools are not used by many people, they might still occur together in people’s workflows more often than expected based on their relatively low overall usage. Conversely, take two tools that each are used by many people: stochastically, a sizable proportion of those people will be shown to use both of them, but this might still be due to chance alone.

Thus, to determine whether the number of people that use two tools together is significantly higher than can be expected by chance, we have to look at the expected co-use of these tools given the number of people that use either of them. This can be compared to the classic example in statistics of taking colored balls out of an urn without replacement: if an urn contains 100 balls (= the population) of which 60 are red (= people in that population who use tool A), and from these 100 balls a sample of 10 balls is taken (= people in the population who use tool B), how many of these 10 balls would be red (=people who use both tool A and B)? This will vary with each try, of course, but when you repeat the experiment many times, the most frequently occurring number of red balls in the sample will be 6. The stochastic distribution in this situation is the hypergeometric distribution.

Figure 1. Source: Memrise

For any possible number x of red balls in the sample (i.e. 1-10), the probability of result x occurring at any given try can be calculated with the hypergeometric probability function. The cumulative hypergeometric probability function gives the probability that the number of red balls in the sample is x or higher. This probability is the p-value of the hypergeometric test (identical to the one-tailed Fisher test), and can be used to assess whether an observed result (e.g. 9 red balls in the sample) is significantly higher than expected by chance. In a single experiment as described above, a p-value of less than 0.05 is commonly considered significant.

In our example, the probability of getting at least 9 red balls in the sample is 0.039 (Figure 2).  Going back to our survey data, this translates to the probability that in a population of 100 people, of which 60 people use tool A and 10 people use tool B, 9 or more people use both tools.

Figure 2 Example of hypergeometric probability calculated using GeneProf.

In applying the hypergeometric test to our survey data, some additional considerations come into play.

Population size
First, for each combination of two tools, what should be taken as total population size (i.e. the 100 balls/100 people in the example above)? It might seem intuitive that that population is the total number of respondents (20,663 for the survey as a whole). However, it is actually better to use only the number of respondents who answered the survey questions where tools A and B occurred as answers.

People who didn’t answer both question cannot possibly have indicated using both tools A and B. In addition, the probability that at least x people are found to use tools A and B together is lower in a large total population than in a small population. This means that the larger the population, the smaller the number of respondents using both tools needs to be for that number to be considered significant. Thus, excluding people that did not answer both questions (and thereby looking at a smaller population) sets the bar higher for two tools to be considered preferentially used together.

Choosing the p-value threshold
The other consideration in applying the hypergeometric test to our survey data is what p-value to use as a cut-off point for significance. As said above, in a single experiment, a result with a p-value lower than 0.05 is commonly considered significant. However, with multiple comparisons (in this case: when a large number of tool combinations is tested in the same dataset), keeping the same p-value will result in an increased number of false-positive results (in this case: tools incorrectly identified as preferentially used together).

The reason is that a p-value of 0.05 indicates there is a 5% chance the observed result is due to chance.  With many observations, there will be inevitably be more results that may seem positive, but are in reality due to chance.

One possible solution to this problem is to divide the p-value threshold by the number of tests  carried out simultaneously. This is called the Bonferroni correction. In our case, where we looked at 119 tools (7 preset answer options for 17 survey questions) and thus at 7,021 unique tool combinations, this results in a p-value threshold of 0.0000071.

Finally, when we not only want to look at tools used more often together than expected by chance, but also at tools used less often together than expected, we are performing a 2-tailed, rather than a 1-tailed test. This means we need to halve the p-value used to determine significance, resulting in a p-value threshold of 0.0000036.

Ready, set, …
Having made the decisions above, we are now ready to apply the hypergeometric test to our survey data. For this, we need to know for each tool combination (e.g. tool A and B, mentioned as answer options in survey questions X and Y, respectively):

a) the number of people that indicate using tool A
b) the number of people that indicate using tool B
c) the number of people that indicate using both tool A and B
d) the number of people that answered both survey questions X and Y (i.e. indicated using at least one tool (including ‘others’) for activity X and one for activity Y).

These numbers were extracted from the cleaned survey data either by filtering in Excel (a,b (12 MB), d (7 MB)) or through an R-script (c, written by Roel Hogervorst during the Mozilla Science Sprint.

The cumulative probability function was calculated in Excel (values and calculations) using the following formulas:

=1-HYPGEOM.DIST((c-1),a,b,d,TRUE)
(to check for tool combination used together more often than expected by chance)

and
=HYPGEOM.DIST(c,a,b,d,TRUE)
(to check for tool combination used together less often than expected by chance)

Figure 3 – Twitter

Bonferroni correction was applied to the resulting p-values as described above and conditional formatting was used to color the cells. All cells with a p-value less than 0.0000036 were colored green or red, for tools used more or less often together than expected by chance, respectively.

The results were combined into a heatmap with green-, red- and non-colored cells (Fig 4), which can also be found as first tab in the Excel-files (values & calculations).

[Update 20170820: we now also have made the extended heatmap for all preset answer options and the 7 most often mentioned ‘others’ per survey question (Excel files: values & calculations)]

Figure 4 Heatmap of tool combinations used together more (green) or less (red) often than expected by chance (click on the image for a larger, zoomable version).

Pretty colors! Now what?
While this post focused on methodological aspects of identifying relevant tool combinations, in future posts we will show how the results can be used to identify real-life research workflows. Which tools really love each other, and what does that mean for the way researchers (can) work?

Many thanks to Bastian Greshake for his helpful advice and reading of a draft version of this blogpost. All errors in assumptions and execution of the statistics remain ours, of course 😉 

GitHub and more: sharing data & code

A recent Nature News article ‘Democratic databases: Science on GitHub‘ discussed GitHub and other programs used for sharing code and data. As a measure for GitHub’s popularity, NatureNews looked at citations of GitHub repositories in research papers from various disciplines (source: Scopus). The article also mentioned BitBucket, Figshare and Zenodo as alternative tools for data and code sharing, but did not analyze their ‘market share’ in the same way.

Our survey on scholarly communication tools asked a question about tools used for archiving and sharing data & code, and included GitHub, FigShare, Zenodo and Bitbucket among the preselected answer options (Figure 1). Thus, our results can provide another measurement of use of these online platforms for sharing data and code.

Figure 1 – Survey question on archiving and sharing data & code

Open Science  – in word or deed

Perhaps the most striking result is that of the 14,896 researchers among our 20,663 respondents (counting PhD students, postdocs and faculty), only 4,358 (29,3%) reported using any tools for archiving/sharing data. Saliently, of the 13,872 researchers who answered the question ‘Do you support the goals of Open Science’ (defined in the survey as  ‘openly creating, sharing and assessing research, wherever viable’), 80,0% said ‘yes’. Clearly, for open science, support in theory and adoption in practice are still quite far apart, at least as far as sharing data is concerned.

Figure 2 Support for Open Science among researchers  in our survey

Among those researchers that do archive and share data, GitHub is indeed the most often used, but just as many people indicate using ‘others’ (i.e. tools not mentioned as one of the preselected options). Figshare comes in third, followed by Bitbucket, Dryad, Dataverse, Zenodo and Pangaea (Figure 3).

Figure 3 – Survey results: tools used for archiving and sharing data & code

Among ‘others’, the most often mentioned tool was Dropbox (mentioned by 496 researchers), with other tools trailing far behind.  Unfortunately, the survey setup invalidates direct comparison of the number of responses for preset tools and tools mentioned as ‘others’ (see: Data are out. Start analyzing. But beware). Thus, we cannot say whether Dropbox is used more or less than GitHub, for example, only that it is the most often mentioned ‘other’ tool.

Disciplinary differences

As mentioned above, 29,3% of researchers in our survey reported to engage in the activity of archiving and sharing code/data. Are there disciplinary differences in this percentage? We explored this earlier in our post ‘The number games‘. We found that researchers in engineering & technology are the most inclined to archive/share data or code, followed by those in physical and life sciences. Medicine, social sciences and humanities are lagging behind at more or less comparable levels (figure 4). But is is also clear that in all disciplines archiving/sharing data or code is an activity that still only a minority of researchers engage in.

Figure 4 – Share of researchers archiving/sharing data & code

Do researchers from different disciplines use different tools for archiving and sharing code & data? Our data suggest that they do (Table 1, data here). Percentages given are the share of researchers (from a given discipline) that indicate using a certain tool. For this analysis, we looked at the population of researchers (n=4,358) that indicated using at least one tool for archiving/sharing data (see also figure 4). As multiple answers were allowed for disciplines as well as tools used, percentages do not add up to 100%.

While it may be no surprise that researchers from Physical Sciences and Engineering & Technology are the most dominant GitHub users (and also the main users of BitBucket), GitHub use is strong across most disciplines. Figshare and Dryad are predominantly used in Life Sciences, which may partly be explained by the coupling of these repositories to journals in this domain (i.e. PLOS to Figshare and Gigascience, along with many others, to Dryad).

Table 1: specific tool usage for sharing data & code across disciplines

As a more surprising finding, Dataverse seems to be adopted by some disciplines more than others. This might be due to the fact that there is often institutional  support from librarians and administrative staff for Dataverse (which was developed by Harvard and is in use at many universities). This might increase use by people who have somewhat less affinity with ‘do-it-yourself’ solutions like GitHub or Figshare. An additional reason, especially for Medicine, could be the possibility of private archiving of data in Dataverse, with control over whom to give access. This is often an important consideration when dealing with potentially sensitive and confidential patient data.

Another surprising finding is the overall low use of Zenodo – a CERN-hosted repository that is the recommended archiving and sharing solution for data from EU-projects and -institutions. The fact that Zenodo is a data-sharing platform that is available to anyone (thus not just for EU project data) might not be widely known yet.

A final interesting observation, which might go against the common idea, is that among researchers in Arts&Humanities who archive and share code, use of these specific tools is not lower than in Social Sciences and Medicine. In some cases, it is even higher.

A more detailed breakdown, e.g. across research role (PhD student, postdoc or faculty), year of first publication or country is possible using the publicly available survey data.

The number games

In our global survey on innovations in scholarly communication, we asked researchers (and people supporting researchers, such as librarians and publishers) what tools they use (or, in the case of people supporting researchers, what tools they advise) for a large number of activities across the research cycle. The results of over 20,000 respondents, publicly available for anyone analyze, can give detailed information on tool usage for specific activities, and on what tools are preferentially used together in the research workflow. It’s also possible to look at results for different disciplines, research roles, career stages and countries specifically.

But we don’t even have to dive into the data at the level of individual tools to see interesting patterns. Focusing on the number of people that answered specific questions, and on the number of tools people indicate they use (regardless of which tools that are) already reveals a lot about research practices in different subsets of our (largely self-selected) sample population.

Number of respondents

In total, we received 20,663 responses. Not all respondents answered all questions, though. The number of responses per activity could be seen to reflect whether that activity plays a role in the research workflow of respondents, or at least, to what extent they use (or advise) concrete tools to carry out that activity (although we also included all answers like ‘manually’, ‘in person’ etc).

On methodology

For each question on tool usage, we offered seven preselected choices that could be clicked (multiple answers allowed), and an ‘and also others’ answer option that, when clicked, invited people to manually enter any other tools they might use for that specific research activity (see Figure 1).

Figure 1 – Example question with answer options

We did not include a ‘none’ option, but at the beginning of the survey stated that people were free to skip any question they felt did not apply to them or did not want to answer. Nonetheless, many people still answered ‘none’ (or any variation thereof) as their ‘other’ options.

Since methodologically, we cannot make a distinction between people who skipped a question or who answered ‘none’, we removed all ‘none’ answers from the survey result. We also adjusted the number of respondents that clicked the ‘and also other’ option to only reflect those that indicated they used at least one tool for the specific research activity (excluding all ‘nones’ and empty answers).

Figure 2 shows the percentage of respondents that answered each specific research activity question, both researchers (PhD students, postdocs and faculty) and librarians. The activities are listed in the order they were asked about, illustrating that the variation in response rate across questions is not not simply due to ‘survey fatigue’ (i.e. people dropping out halfway through the survey).

Figure 2 – Response rate per survey question (researchers and librarians)

What simple question response levels already can tell us

The differences in response levels to the various questions are quite marked, ranging from barely 15% to almost 100%. It is likely that two effects are at play here. First, some activities are relevant for all respondents, e.g. writing and searching information, while others like sharing (lab) notebooks are specific to certain fields, explaining lower response levels. Second, there are some activities that are not yet carried out by many or for which respondents choose not to use any tool. This may be the case with sharing posters and presentations and with peer review outside that organized by journals.

Then there are also notable differences between researchers and librarians. Researchers expectedly more often indicate tool usage to publish and librarians are a bit more active in using or advocating tools for reference management. Perhaps more interestingly, it is clear that librarians are “pushing” tools that support sharing and openness of research.

The differences in response levels to the various questions are quite marked, ranging from barely 15% to almost 100%. Some activities seem relevant for all respondents, e.g. writing and searching information. Others, like sharing (lab) notebooks, sharing posters and presentations and peer review outside that organized by journals, may be carried out by . fewer people, or may be done without using specific tools.
Then there are also notable differences between researchers and librarians. As expected, researchers more often indicate tool usage to publish and librarians are a bit more active in using or advocating tools for reference management and selection of journals to publish in. Perhaps more interestingly, it is clear that librarians are “pushing” tools that support sharing and openness of research.

Disciplinary differences
When we look at not just the overall number of respondents per activity, but break that number down for the various disciplines covered (Figure 3), more patterns emerge. Some of these are expected, some more surprising.

Figure 3 – Response rate per survey question across disciplines (researchers only)

As expected, almost all respondents, irrespective of discipline, indicate using tools for searching, getting access, writing and reading. Researchers in Arts & Humanities and Law report lower usage of analysis tools than those in other disciplines, and sharing data & code is predominantly done in Engineering & Technology (including computer sciences). The fact that Arts & Humanities and Law also score lower on tool usage for journal selection, publishing and measuring impact than other disciplines might be due to a combination of publication culture and the (related) fact that available tools for these activities are predominantly aimed at journal articles, not books.

Among  the more surprising results are perhaps the lower scores for reference management for Arts & Humanities and Law (again, this could be partly due to publication culture, but most reference management systems enable citation of books as well as journals). Scores for sharing notebooks and protocols were low overall, where we would have expected this activity to occur somewhat more often in the sciences (perhaps especially life sciences). Researchers in Social Sciences & Economics and in Arts & Humanities relatively often use tools to archive & share posters and presentations and to do outreach (phrased in the survey as: “tell about your research outside academia”), and interestingly enough, so do researchers in Engineering & Technology (including computer science). Finally, peer review outside that done by journals is most often done in Medicine, which is perhaps not that surprising given that many developments in open peer review are pioneered in biomedical journals such as The BMJ and BioMedCentral journals.

You’re using HOW many tools?

How many apps do you have on your smartphone*? On your tablet? Do  you expect the number of online  tools you use in your daily life as a researcher to be more or less than that?

Looking at the total number of tools that respondents to our survey indicate they use in their research workflow (including any ‘other’ tools mentioned, but excluding all ‘nones’, as described above), it turns out the average number of tools reported per person is 22 (Figure 4). The frequency distribution curve is somewhat skewed as there is a longer tail of people using higher numbers of tools (mean = 22.3; median = 21.0).

Figure 4 – Frequency distribution of total number of tools used per person (20,663 respondents)

We also wondered whether the number of tools a researcher uses varies with career stage (e.g. do early career researchers use more tools than senior professors?).

Figure 5 shows the mean values of the number of tools mentioned by researchers, broken down by career stage. We used year of first publication as a proxy for career stage, as it is a more or less objective measure across research cultures and countries, and less likely to invoke ‘refuse to answer’ then asking for age might have been.

Figure 5 – Number of tools used across career stage (using year of first publication as proxy)

There is an increase in the number of tools used going from early to mid career stages; peaking for researchers who published their first paper 10-15 years ago. Conversely, more senior researchers seem to use less tools, with the number of tools decreasing most for researchers who first published over 25 years ago.  The differences are fairly small, however, and it remains to be seen whether they can be proven to be significant. There also might be differences across disciplines in these observed trends,  depending on publication culture within disciplines. We have not further explored this yet.

It will be interesting to correlate career stage not only with the number of tools used, but also with type of tools: do more senior researchers use more traditional tools that they have been accustomed to using throughout their career, while younger researchers gravitate more to innovative or experimental tools that have only recently become available? By combining results from our survey with information collected in our database of tools, these are the type of questions that can be explored.

Mozilla Science Lab Global Sprint 2016 – getting started with analysis

On June 2-3 between 9:00-13:00, we want to bring a group of smart people together in Utrecht to kickstart the analysis of our survey data. This event will be part of the Mozilla Science Lab Global Sprint 2016.

What is Mozilla Science Lab?
Mozilla Science Lab is a community of researchers, developers, and librarians making research open and accessible and empowering open science leaders through fellowships, mentorship, and project-based learning.

What is the Global Sprint?
This two-day sprint event brings together researchers, coders, librarians and the public from around the globe to hack on open science and open data projects in their communities. This year, it has four tracks anyone can contribute to: tools, citizen science, open educational resources and open data. There are about 30 locations participating in this year’s sprint – our event in Utrecht is one of them.

What will we do during the sprint?
Quick exploration of the survey results can be done in an interactive dashboard on Silk (http://dashboard101innovations.silk.co), but many more in-depth analyses are possible. Some examples can already been found on Kaggle. During this Mozilla Science Lab Sprint, we intend to make a head-start with these analyses by bringing together people with expertise in numerical and textual analysis.

The survey results can provide insights into current practices across various fields, research roles, countries and career stages, and can be useful for researchers interested in changing research workflows. The data also makes it possible to correlate research tool usage to stance on Open Access and Open Science, and contains over 10,000 free-text answers on what respondents consider the most important developments in scholarly communication.

So: two half-days (mornings only) of coding/hacking t0 discover patterns, make connections and get information from our data. If you have ideas of your own on what to do with our data, or want to help us realize our ideas, join us! You bring your laptop and your mind, we supply a comfortable space, 220V, coffee, WiFi and something healthy (or less so) to eat – and of course, the data to work with!

You can join us on either of these days or both days if you like. On both days there will be a short introduction to get people started. Ideally the first day will result in some analyses/scripts that participants can build on during the second day.

Who can join?
We invite all smart people, but especially those with experience in e.g. R, Python, Google Refine, NVIVO or AtlasTI. Any and all ideas for analysis are welcome!

You can register here, or just contact us to let us know you’re coming.

Where? When?
The sprint will take place on June 2-3 on the Uithof in Utrecht, either in the Utrecht University Library building, or another space close by.

Please note that the sprint will take place between 9:00-13:00 on both days. Participants are of course free to continue working after these hours, and online support will be given wherever possible.

Data are out. Start analyzing. But beware.

Now that we have our data set on research tool usage out and shared the graphical dashboard, let the analysis start! We hope people around the world will find the data interesting and useful.

If you are going to do in depth analyses, make sure to read our article on the survey backgrounds and methods. It helps you understand the type of sampling we used and the resulting response distributions. It also explains the differences between the raw and cleaned data sets.

For more user friendly insights, you can use the graphical dashboard made in Silk. It is easy to use, but still allows for quite sophisticated filtering and even supports filtering answers to one question by answers given to another question. Please be kind on Silk: it crunches a lot of data and may sometimes need a few seconds to render the charts.

Example chart that also shows filter options in the dashboard

When looking at the charts and when carrying out your analyses, please note two things.

First, whatever you are going to do, make sure to reckon with the fundamental difference between results from preset answers (entered by simply clicking an image) and those from specifications of other tools used (entered by typing the  tool names manually). The latter are quite probably an underestimation and thus cannot be readily compared with the former. [Update 20160501: This is inherent to the differences between open and closed questions, of which ease of answering the question is one aspect. Specifications of ‘others’ can be seen as an open question]. This is why we present them separately in the dashboard  Integrated lists of these two types of results, if made at all, should be accompanied with the necessary caveats.

Frequency distribution of 7 preset answers (dark blue) and the first 7 ‘other’ tools (light blue) per survey question

Second, basic statistics tells that when you apply filters, the absolute numbers in some cases can become so low as to render the results unfit for any generalization. And the other way around: when not filtering, please note that usage patterns will vary according to research role, field, country etc. Also, our sample was self-selected and thus not necessarily representative.

Now that we are aware of these two limitations, nothing stops you (and us) to dive in.

Our own priorities, time permitting, are to look at which tools are used together across research activities and why that is, concentration ratios of tools used for the various research activities, and combining these usage data with data on the tools themselves like age, origin, business model etc. More in general, we want to investigate what tool usage says about the way researchers shape their workflow: do they choose tools to make their work more efficient, open and/or reproducible?  We also plan to do a more qualitative analysis of the thousands of answers people gave to the question what they see as the most important development in scholarly communication.

By the way, we’d love to get your feedback and learn what you are using these data for, whether it is research, evaluation and planning of services or something else still. Just mail us, or leave your reply here or on any open commenting/peer review platform!