[Summary: Thinking aloud about a pragmatic / humanist approach to data infrastructure building]
Stephen Abbott Pugh of Open Knowledge International has just blogged about the Open Data for Tax Justice ‘design sprint’ that took place in London on Monday and Tuesday. I took part in the first day and a half of the workshop, and found myself fairly at-odds with the approach being taken that focussed narrowly on the data-pipelines based creation of a centralised dataset, and that appeared to create barriers rather than bridges between data and domain experts. Rather than the rethink the approach, as I would argue is needed, the Open Knowledge write up appears to show the Open Data for Tax Justice project heading further down this flawed path.
In this post, I’m offering an (I hope) constructive critique of the approach, trying to draw out some more general principles that might inform projects to create more participatory data infrastructures.
As the OKI post relates:
“Country-by-country reporting (CBCR) is a transparency mechanism which requires multinational corporations to publish information about their economic activities in all of the countries where they operate. This includes information on the taxes they pay, the number of people they employ and the profits they report.”
Country by Country reporting has been a major ask of tax justice campaigners since the early 2000s, in order to address tax avoidance by multi-national companies who shift their profits around the world through complex corporate structures and internal transfers. CBCR got a major boost in 2013 with the launch of reporting requirements for EU Banks to publicly disclose Country by Country reports under the CRD IV regulations. In the extractives sector, campaigners have also secured regulations requiring disclosure of tax and licensing payments to government on a project-by-project basis.
Although in the case of UK extractives firms, reporting is taking place to companies house as structured data, with an API available to access reports, for EU Banks, reporting is predominantly in the form of tables at the back of PDF format company reports.
If campaigners are successful, public reporting will be extended to all EU multinationals, holding out the prospect of up to 6000 more annual reports that can provide a breakdown of turnover, profit, tax and employees country-by-country. If the templates for disclosure are based on existing OECD models for private exchange between tax authorities, the data may also include information on the different legal entities that make a corporate group, important for public understanding the structure of the corporate world.
Earlier this year, a report from Alex Cobham, Jonathan Gray and Richard Murphey set out a number of use-cases for such data, making the case that “a global public database on the tax contributions and economic activities of multinational companies” would be an asset for a wide range of users, from journalists, civil society and investors.
Sprinting with a data-pipelines hammer
This week’s design sprint focussed particularly on ‘data extraction’, developing a set of data pipeline scripts and processes that involve downloading a report PDF, marking up the tables where Country by Country data is stored, describing what each column contains using YAML, and then committing this to GitHub where the process can then be replicably run using datapipeline commands. Then, with the data extracted, it can be loaded into an SQL database, and explored by writing queries or building simple charts. It’s a technically advanced approach, and great for ensuring replicability of data extraction.
But, its also an approach that ultimately entirely misses the point, ignoring the social process of data production, creating technical barriers instead of empowering contributors and users, and offering nothing for campaigners who want to ensure that better data is produced ‘at source’ by companies.
Whilst the OKI blog post reports that “The Open Data for Tax Justice network team are now exploring opportunities for collaborations to collect and process all available CRD IV data via the pipeline and tools developed during our sprint.” I want to argue for a refocussed approach, based around a much closer look at the social dynamics of data creation and use.
An alternative approach: crafting collaborations
I’ve tried below to unpack a number of principles that might guide that alternative approach:
Principle 1: Letting people use their own tools
Any approach that involves downloading, installing, signing-up to, configuring or learning new software in order to create or use data is likely to exclude a large community of potential users. If the data you are dealing with is tabular: focus on spreadsheets.
More technical users can transform data into database formats when the questions they want to answer require the additional power that brings, but it is better if the starting workflow is configured to be accessible to the largest number of likely users.
Back in October I put together a rough prototype of a Google spreadsheets based transcription tool for Country by Country reports, that needed just copy-and-paste of data, and a few selections from validated drop-down lists to go from PDFs to normalised data – allowing a large user community to engage directly with the data, with almost zero learning curve.
The only tool this approach needs to introduce is something like tabula or PDFTables to convert from PDF to Excel or CSV: but in this workflow the data comes right back to the user to be able to work with it after it has been converted, rather than being taken away from them into a longer processing pipeline. Plus, it brings the benefit of raising awareness of data extraction from PDF that the user can adopt for other projects in future, and allowing the user to work-around failed conversions using a manual transcription approach if they need to.
(Sidenote: from discussions, I understand that one of the reasons the OKI team made their technical choice was from envisaging the primary users as ‘non-experts’ who would engage in crowdsourcing transcriptions of PDF reports. I think this is both highly optimistic, and relies on a flawed analysis of the relatively small scale of the crowdsourcing task in terms of a few 1000 reports a year, and the potential benefits of involving a more engaged group of contributors in creating a civil society database)
Principle 2: Aim for instant empowerment
One of the striking things about Country by Country reporting data is how simple it ultimately is. The CRD IV disclosures contain just a handful of measures (turnover, pre-tax profits, tax paid, number of employees), a few dimensions (company name, country, year), and a range of annotations in footnotes or explanations. The analysis that can be done with this is data is similarly simple – yet also very powerful. Being able to go from a PDF table of data, to a quick view of the ratios between turnover and tax, or profit and employees for a country can quickly highlight areas to investigate for profit-shifting and tax-avoidance behaviour.
Calculating these ratios is possible almost as soon as you have data in a spreadsheet form. In fact, a well set up template could calculate them directly, or the user with basic ability to write formula could fill in the columns they need.
Many of the use-cases for Country by Country reports are based not on aggregation across hundreds of firms, but on simply understanding the behaviour of one or two firms. Investigators and researchers often have firms they are particularly interested in, and where the combination of simple data, and their contextual knowledge, can go a long way.
Principle 3: Don’t drop context
On the topic of context: all those footnotes and explanations in company reports are an important part of the data. They might not be computable, or easy to query against, but in the data explorations that took place on Monday and Tuesday I was struck by how much the tax justice experts were relying not only on the numerical figures to find stories, but also on the explanations and other annotations from reports.
The data pipelines approach dropped these annotations (and indeed dropped anything that didn’t fit into it’s schema). An alternative approach would work from the principle that, as far as possible, nothing of the source should be thrown away – and structure should be layered on top of the messy reality of accounting judgements and decisions.
Principle 4: Data making is meaning-making
A lot of the analysis of Country by Country reporting data is about look for outliers. But data outliers and data errors can look pretty similar. Instead of trying to separate the process of data preparation and analysis, these two need to be brought closer together.
Creating a shared database of tax disclosures will involve not only processes of data extraction, but also processes of validation and quality control. It will require incentives for contributors, and will require attention to building a community of users.
Some of the current structured data available from Country by Country reports has been transcribed by University students as part of their classes – where data was created as a starting point for a close feedback loop of data analysis. The idea of ‘frictionless data’ makes sense when it comes to getting a list of currency codes, but when it comes to understanding accounts, some ‘friction’ of social process can go a long way to getting reliable data, and building a community of practice who understand the data in more depth.
Principle 5: Standards support distributed collaboration
One of the difficulties in using the data mentioned above, prepared by a group of students, was that it had been transcribed and structured to solve the particular analytical problem of the class, and not against any shared standard for identifying countries, companies or the measures being transcribed.
The absence of agreement on key issues such as codelists for tax jurisdictions, company identifiers, codes and definitions of measures, and how to handle annotations and missing data means that the data that is generated by different researchers, or even different regulatory regimes, is not comparable, and can’t be easily combined.
The data pipelines approach is based on rendering data comparable through a centralised infrastructure. In my experience, such approaches are brittle, particularly in the context of voluntary collaboration, and they tend to create bottlenecks for data sharing and innovation. By contrast, an approach based on building light-weight standards can support a much more distributed collaboration approach – in which different groups can focus first on the data that is of most interest to them (for example, national journalists focussing on the tax record of the top-10 companies in their jurisdiction), easily contributing data to a common pool later when their incentives are aligned.
Campaigners also need to be armed with use-case backed proposals for how disclosures should be structured in order to push for the best quality disclosure regimes
What’s the difference?
Depending on your viewpoint, the approach I’ve started to set out above might look more technically ‘messy’ – but I would argue it is more in-tune with the social realities of building a collaborative dataset of company tax disclosures.
Fundamentally (with the exception perhaps of standard maintenance, although that should be managed as a multi-stakeholder project long-term) – it is much more decentralised. This is in line with the approach in the Open Contracting Data Standard, where the Open Contracting Partnership have stuck well to their field-building aspirations, and where many of the most interesting data projects emerge organically at the edge of the network, only later feeding into cross-collaboration.
Even then, this sketch of an alternative technical approach above is only part of the story in building a better data-foundation for action to address corporate tax avoidance. There will still be a lot of labour to create incentives, encourage co-operation, manage data quality, and build capacity to work with data. But better we engage with that labour, than spending our efforts chasing after frictionless dreams of easily created perfect datasets.