This paper has drawn on evidence available up to 12 November 2020. Further evidence on this topic is constantly published and DELVE will continue to develop this report as it becomes available. This independent overview of the science has been provided in good faith by subject experts. DELVE and the Royal Society accept no legal liability for decisions made based on this evidence.

Summary

Responding to the COVID-19 pandemic has required rapid decision-making in changing circumstances. Those decisions and their effects on the health and wealth of the nation can be better informed with data. Today, technologies that can acquire data are pervasive. Data is continually produced by devices like mobile phones, payment points and road traffic sensors. This creates opportunities for nowcasting of important metrics such as GDP, population movements and disease prevalence, which can be used to design policy interventions that are targeted to the needs of specific sectors or localities. The data collected as a by-product of daily activities is different to epidemiological or other population research data that might be used to drive the decisions of state. These new forms of data are happenstance, in that they are not originally collected with a particular research or policy question in mind but are created through the normal course of events in our digital lives, and our interactions with digital systems and services.

This happenstance data pertains to individual citizens and their daily activities. To be useful it needs to be anonymized, aggregated and statistically calibrated to provide meaningful metrics for robust decision making while managing concerns about individual privacy or business value. This process necessitates particular technical and domain expertise that is often found in academia, but it must be conducted in partnership with the industries, and public sector organisations, that collect or generate the data and government authorities that take action based on those insights. Such collaborations require governance mechanisms that can respond rapidly to emerging areas of need, a common language between partners about how data is used and how it is being protected, and careful stewardship to ensure appropriate balancing of data subjects’ rights and the benefit of using this data. This is the landscape of data readiness; the availability and quality of the UK nation’s data dictates our ability to respond in an agile manner to evolving events.

Citation

The DELVE Initiative (2020), Data Readiness: Lessons from an Emergency. DELVE Report No. 7. Published 24 November 2020. Available from https://rs-delve.github.io/reports/2020/11/24/data-readiness-lessons-from-an-emergency.html [pdf].

Table of Contents

• Summary
• Key points
• Challenges in accessing and analysing non-traditional data sources to inform COVID-19 policy
• Recommendations

1. Background
   • DELVE’s remit and data science for COVID-19 policy
   • Case Study 1: What is the COVID-19 reproduction rate in a local area?
   • Case Study 2: How much more are the elderly at risk in different countries?
     • Discoverability
     • Accessibility
     • Inconsistent data
     • Ephemeral reporting
     • Languages
     • Curating high-quality data sources
   • Case Study 3: How have COVID-19 policies affected consumer behaviour, and what are the implications for economic growth?
   • Understanding different data types
2. Challenges in Data Sharing
   • Streamline the data preparation pipeline
     • Establishing a common language
   • Implementing a framework for data readiness in DELVE projects
     • Accreditation mechanisms
     • Build capability for long-term data sharing
   • Box A. Coordinated collaboration to enable data use: using human mobility data to understand the impact of COVID-19 policies on the movement of people
   • Box B. Coordinated collaboration to enable data use: using transaction data to understand the impact of policy interventions on economic activity
   • The value of pathfinder projects
   • Accounting for data bias
   • Create incentives or duties to promote responsible data sharing
3. Conclusion

Appendices

• Appendix A: Python code for loading COVID-19 cases]
• Other Appendices
  • Organisational Data Maturity

Footnotes and References

Key points

Across the COVID-19 pandemic, data availability and quality have proven a bottleneck to understanding the current state of the UK’s health and economy and to providing policy advice for rapid decision making. This bottleneck has affected a wide range of data types, from early challenges with access to testing data that would enable estimates of COVID-19 incidence, Rt and mortality rates across the country, to the difficulty of obtaining data about monetary transactions that would help estimate the effects of the pandemic on GDP, to issues with the use of mobility data that would enable assessment of how movement of people across the UK population contributes to disease spread.¹

This report builds on the experience of the Royal Society’s DELVE group in providing rapid turnaround, data-driven answers to policy questions, and the challenges in access to data that arose during this work. It sets out the challenges we found that were common to these data sharing activities, focussing on the use of new data modalities, arising from, for example mobile phone usage or electronic payments data, rather than the classical challenges of collecting surveillance data. It then considers recommendations for addressing these challenges and creating an environment that supports the safe and rapid use of a range of different types of data in policymaking.

Challenges in accessing and analysing non-traditional data sources to inform COVID-19 policy

• The pipeline for data from discovery and acquisition to decision is fundamentally different for happenstance data created by everyday interactions, than for data collected specifically for research, but the two types of data are rarely differentiated in data sharing discussions. Access to happenstance data relies on successful interactions across the public sector, private sector and academia to establish trustworthy frameworks for data sharing that enable data access while managing concerns around security and privacy.

• The lack of a common language for understanding data quality or the action needed before a dataset can be used in analysis holds back research collaborations. Guidelines for data quality that do exist focus on intentions - or set out the desired end-point - rather than providing mechanisms for improving data accessibility. This contributes to the prevalence of perceived barriers to data sharing.

• The pipeline for data from discovery and acquisition to sharing includes ethical, legal, technical, commercial and quality issues. But a lack of a coherent vocabulary for identifying and resourcing issues means projects are often inappropriately resourced and staffed.

• Where successful multiparty data sharing collaborations between academia, the private sector and government have been established at pace, they have often relied on relationships that existed before the pandemic, which could be rapidly adapted to pandemic needs.In the long-term, new initiatives are needed to cultivate such partnerships, so that data sharing efforts can be repurposed at times of crisis, in ways that encourage public dialogue about data use. Creating such partnerships will require further effort to build capability at all levels of decision-making within government, if it is to promote the trustworthy use of data in policymaking.

• Many of the barriers to data sharing that have been experienced during the pandemic are long-standing issues in data policy, including the lack of common technical standards, concerns about privacy or security breaches, and the lack of organisational capabilities. Together, these contribute to a cluster of coordination problems between organisations that are felt acutely in efforts to share happenstance data, due to the need for multiparty coordination.

Recommendations

• Government should update the statutory objective of the Office for National Statistics (ONS) to accommodate trustworthy access to happenstance data to generate national and local statistics. Such statistics are required on very short time frames to facilitate fast decision-making for the nation in the rapidly evolving circumstances of a national emergency.

• The ONS should collaborate closely with the Information Commissioner’s Office (ICO) to formulate a standardized qualification for data access, equivalent to a ‘data driving license’ that would demonstrate trustworthiness and ensure that qualified experts can get rapid access to different data types with the appropriate standardized ethical and legal training in place.

• Government should fund interdisciplinary pathfinder data projects. These projects should require collaborations between industries, run across government departments and integrate different academic expertise. Each project should target a specific policy question. Beyond the pathfinder role, the projects will leave a legacy in the form of expertise and guidance in understanding the stages of the data-sharing pipeline. Priority areas for pathfinder projects include:

  • Nowcasting of economic metrics: At least one of these pathfinder projects should create a close collaboration between Cabinet Office and Treasury around nowcasting of classical economic metrics (such as GDP) from happenstance data (e.g. payments data). Efficient resourcing and strategic implementation of data sharing projects will only be possible if Treasury and Cabinet Office are aligned on plausible benefits and costs of data sharing projects.

  • Mobility data: Another project should drive a step-change in the use of mobility data for public policy. To achieve this, the ONS should act as the trusted body to convert happenstance data into high-frequency population mobility statistics. One pathfinder project should produce daily views of population mobility between geographic regions, aggregated from origin to destination counts from mobile phone operators.

1. Background

DELVE’s remit and data science for COVID-19 policy

The Royal Society’s DELVE group was formed in response to the COVID-19 crisis with a remit to explore specific policy questions in a data-driven manner.² The DELVE Initiative was established with the ambition that data science could play a role in helping develop policy responses to the COVID-19 pandemic, by identifying lessons from the responses of other countries or by combining datasets to generate novel insights. Such analysis requires access to data, which could come from both official statistics, or from so-called happenstance data, generated as a by-product of daily activities. Drawing from a multidisciplinary team of domain experts in policy, public health, economics, education, immunology, epidemiology, and social science, alongside statisticians, mathematicians, computer scientists and machine learning scientists, DELVE set out to provide advice and analysis that could feed into live policy decisions.

By the middle of August 2020 DELVE had completed five reports: on the use of face masks by the general public; the test and trace system; the control of hospital and health-care acquisition of the disease; the risk associated with a return to schools; and the economic aspects of the COVID-19 crisis. These reports have grappled with a wide range of questions, including:

• Should face masks be worn to help contain COVID-19, and who should wear them where and when?

• Which kinds of households and firms would be most impacted by non-pharmaceutical interventions, and by how much?

• Are there gaps in emergency fiscal measures introduced to cushion the impact of lockdown?

• What is the impact of future case burdens on ICU and ventilator capacity?

• Can early signals of the effect of policy interventions such as lockdown measures and the easing of such measures be obtained?

The breadth of these questions highlights the intersections of expertise that were required as well as the diversity of data sources that were potentially in scope of these efforts.

In early April 2020, DELVE formed an initial list of datasets we felt would be informative in addressing policy questions. This data included mobility data (e.g. the aggregated movements of mobile telephones between different cell towers) and payments transaction data (e.g. aggregated credit card payments being made in different geographic regions). Such data originates from individual citizens but is collated by commercial companies in the course of their normal business. The data is invaluable in determining, for example, an estimate of the effect of lock down measures on GDP in real time³ or for aiding localized estimation of the disease reproduction number (R) through understanding the movement of people between localities.

Despite the potential benefits of using such data in the COVID-19 response, access to many forms of data has been problematic across the pandemic. Recognising the shortcomings of the UK’s COVID-19 data readiness, in an evidence session with the Commons Science and Technology Committee in July, Government Chief Scientist Patrick Vallance explained that “an important lesson across not just pandemics but every emergency, is data; you must have the data flows and you must understand data ownership and how data is going to get to people for the information you require”.⁴

DELVE’s work across the pandemic illustrates this lesson, and the action that is needed to ensure the UK is better able to use its data resources in response to emergencies in future. It also illustrates the scale of the successes that have been achieved elsewhere, which may offer lessons for the UK’s data strategy. For example, the French interbank network, Groupement des Cartes Bancaires (CB), made nearly five billion payment card transactions from approximately 70 million cards issued by all banks in France available for analysis.⁵ Their partnership with analysts complied with the EU GDPR (Article 89), and produced an intricately detailed view of the impact of COVID-19 on consumer spending and business sales on a daily and weekly scale. We examine such examples of coordinated collaboration to enable data use in Boxes A and B in this report. While there are examples of the successful deployment of similar data types to support COVID-19 policy development in the UK, the scale of collaborations established elsewhere demonstrates the wider opportunities at stake.⁶

We revisit three representative questions the DELVE group considered below. Each illustrates different types of challenges that data scientists and policymakers might encounter when using data in policymaking. These scenarios cover:

• COVID-19 prevalence in different localities across the UK: Important statistics, such as local COVID-19 reproduction rate, are currently derived from case numbers collected through official national studies. Happenstance data, such as high-quality origin to destination mobility data, can be used to refine these estimates, enabling more targeted policy responses at a local level. Designing such tailored responses requires ready access to multiple data sources, which in turn relies on strong pre-existing relationships and connections across the public sector, industry and academia.

• International comparisons on the impact of COVID-19 on different communities: International comparisons can help policymakers understand what types of policy interventions might be helpful in managing the COVID-19 pandemic, or benchmark performance of approaches taken in the UK. Performing such analyses requires access to data of sufficient quality to enable cross-national comparisons. Data published by governments across the world may have different levels of accessibility, different formats, or cover different population samples. Before analysis, it must first be wrangled into comparable forms. A common understanding of data readiness could help speed up this process.

• The pandemic’s economic impact: Official statistics play an important role in assessing the health, wealth and wellbeing of the nation. At times of crisis, these can change at pace. While vital, these official measures may not be responsive enough to inform rapid policy development. Alternative ways of measuring economic activity exist - for example, consumer behaviour data can indicate patterns of economic growth or constriction - but they rely on access to data held by private companies.

Understanding different data types

The COVID-19 response has demanded rapid creation of new data systems. The early stages of the pandemic were characterized by uncertainty about the incidence and prevalence of COVID-19 at both local and national levels, raising question marks over the apparent and actual hospitalization and mortality rates. These questions have been addressed through increasing the scale and scope of the UK’s testing program.⁷ Similarly, probability surveys that we speculated may be necessary, for instance to assess the prevalence of COVID-19, were either enacted as proposed or modified appropriately as prioritizations changed.⁸ Such core studies are foundational to implementing any public health response. These areas are not the focus of this report. Our aim is to characterize the new data modalities, which we come by through happenstance data, that could be used to improve policy responses.

So-called happenstance data is not originally collected with a particular research or policy question in mind but is created through the normal course of events in our digital lives, and our interactions with digital systems and services. This happenstance data pertains to individual citizens and their daily activities. To be useful it needs to be anonymized, aggregated and statistically calibrated to provide meaningful metrics for robust decision making. Compounding these technical issues associated with the use of such data are differences in understanding of what data might usefully be deployed for which purposes: unlike traditional forms of research data, collected with a clear purpose in mind, happenstance data is repurposed from another function. It may not be obvious to researchers or policymakers what forms of data might be useful, or to those holding data that they have an asset that could be helpfully deployed in response to a crisis.

Deploying these happenstance data resources at times of crisis requires careful cooperation across the private sector, which collects much of the relevant data, the public sector, which can provide an institutional base for data aggregation and governance, and academia, which can supply resources to analyse it. These multiparty data access agreements need to be developed in accordance with data protection legislation and wider concerns about data security, privacy, and value to different parties. The scenarios discussed above illustrate the potential usefulness of two types of happenstance data - origin to destination mobility data and card transaction data - in developing different types of policy response.

Despite the usefulness of this data and the good intent of some commercial partners, these forms of data remain difficult to access in the UK. In contrast, the relative success of some other countries in creating structures to use this data for public good suggests there are policy integrations, practices or relationships that could increase the accessibility of this data. The sections that follow draw from DELVE’s experiences in aggregating and analysing COVID-19 policy-relevant data. Through collaborators in those countries we were also able to compare the challenges experienced in their data sharing journey with ours. Those experiences inform the key points outlined at the top of this report and the recommendations that follow.

2. Challenges in Data Sharing

Happenstance data can be very difficult to share.

Firstly, happenstance data is often collected at large scale and potentially across many distributed devices. There can be technical challenges to assimilation and sharing of the data for analysis. Even if the data can be shared, the quality and comprehensiveness of the underlying data may not be sufficient to address a given policy question.

Secondly, the data originates from the activities of individual citizens whose privacy and personal data rights must be respected. Careful stewardship of this data is required, based on collaborations across organisations to establish research questions, agree ways of working and set in place formal contracts that together enable its safe and rapid use. Such collaborations take time to establish and rely on pre-existing relationships.

Thirdly, such data is often commercially sensitive, or is perceived to be commercially valuable. This creates business concerns that discourage data sharing across organisations.

Finally, once the quality of the data is understood and inadequacies in the data collection have been accounted for, there remain statistical challenges in the interpretation of the data. In particular population-level studies collecting bespoke research data will be analysed to take account of study design, including potential confounders, bias and study size to minimise the threat of invalid inference and present findings that acknowledge the limitations.

Addressing these challenges requires effort to:

• Streamline the data preparation pipeline

• Build capability for long-term data sharing

• Create incentives or duties to promote responsible data sharing

Streamline the data preparation pipeline

DELVE’s experience during the Covid-19 pandemic is that the process by which data scientists, statisticians and other researchers obtain data is frequently the most time-consuming part of attempting to arrive at data-driven answers to any question. This is either because of the difficulty in accessing data, or the process of collecting, cleaning and combining various diverse sources of data.

Establishing a common language

Happenstance data is subject to various levels of “data accounting”. Financial data corresponds directly to monetary units and is subject to regular audit, both from corporate entities and their accountants. Other forms of data, with much less direct value, are rarely subject to careful audit. There, data may have missing values or outlier values that stem from incorrect calibration or inappropriate thresholding. In the most extreme case, the data may have been improperly recorded, or vital values may have been deleted from log files to reduce storage requirements.

Data quality can improve when data sharing guidelines are used. Some attempts to formulate such guidelines for the use of data in research can already be found emerging from the Bioinformatics community. A range of organisations have also been active in developing and promoting the ‘FAIR’ principles for data use, seeking to ensure that open data is made “Findable”, “Accessible”, “Interoperable” and “Reusable”.⁹ Each of these terms indicates a desired state for data by which an individual data set can be judged, in particular around the meta data and the licensing of the data set. These principles describe a useful destination where we would like our datasets to be, but they are less helpful when it comes to characterizing the nature of the journey to get there, what capabilities and resources will be needed, what milestones should be monitored. In particular, they are of little help when it comes to resourcing a project, either from the perspective of the generators of the data set, or from the point of view of the putative users of the data.

The pipeline for data from discovery and acquisition to decision is fundamentally different for happenstance data created by everyday interactions, than for data collected specifically for research, but the two types of data are rarely differentiated in data sharing discussions. Instead, researchers, civil servants and organisations often believe they are discussing similar types of data or data features, when in fact each is referring to data being held in a different state, which requires different types of processing or management to make it useful. A common language is needed - one that enables common goals to be built around the processing of data - in order to make faster progress in negotiations around data access.

To fill this gap, a system is needed to create a common reference point for both the generators of the dataset and the putative users to discuss its maturity and suitability for analysis. DELVE’s approach to filling this gap was the development of a data readiness framework – a data maturity marker that helped coordinate and prioritise the efforts of our research software engineers in bringing datasets to a state of availability.

Case Studies 1 to 3 tell different tales of data readiness. To provide the language with which we could describe the state of data, we broadly classify data into a spectrum of Bands of readiness.¹⁰ The starting point for anyone wishing to provide an answer could be in any of these Bands:

Band D: a blank slate, where the data does not exist at all;

Band C: from ‘we’ve heard via the grapevine that data exists’, to verifying its existence and obtaining access to it;

Band B: data is accessible, but it is not clear whether it is appropriate for the question or faithfully described;

Band A: data is readily accessible, documented and useful for a specific task.

The data Band depends on the context in which a question is asked, and the role of those analysing it. Data could be in Band A for the group who owns it, but the same data might be in Band C for someone who doesn’t have clearance to access it, or appear in Band B when published online in a report.

The categorization is useful because it frames the expectations on whether any data-driven conclusions could be drawn, and the time and effort involved in the process. The effort is usually spent at the transitions between the Bands: between C and B, finding means to access the data digitally, and between B and A, defining questions that could be answered from the data. As an example of a transition, one might hear that some entity has potentially useful data (Band C), then follow a long path of negotiating digital access, only to discover that the data would not ultimately be useful for the original question.

Within DELVE projects, the data readiness bands provided a common language in interdisciplinary collaborations; from the mutual understanding that hearsay data is only hearsay data, to the properties that make it fit for purpose. As most effort is usually spent at the transitions between the Bands, it frames the challenges to readying data. Foremost, Case Study 1 highlighted the downstream impact of the practices of data producers and publishers: substantial “data wrangling” is required to bring data from online dashboards and PDFs into research-friendly formats. A coherent vocabulary about the readiness of data helps the resource planning, time estimation and staffing of projects.

Accreditation mechanisms

There are access barriers between qualified experts and publicly held data, but the steps to access data held by the ONS have been simplified during the early months of the COVID-19 pandemic. Prior to the pandemic, a researcher had to travel to the ONS headquarters in Wales to do an induction followed by a test to access data already held in the Secure Research Service (SRS). After the induction, data could be accessed in approved “safe rooms”, which, in the worst case, meant traveling to ONS headquarters again. Now, induction is done remotely, and access to SRS data is possible from a researcher’s home office. The one caveat is that if data is shared with the ONS but is not held in the SRS, it can’t be accessed through this protocol.

However, the processes to access national data are not streamlined. The access and induction process has to be repeated for data sources that could possibly be in Band A for a question, but are held by different public bodies. Iterative onboarding, without the guarantee that data, once accessed, might be useful, is a bottleneck when research agility is of the essence.

These challenges are further compounded by the multiparty nature of the data access arrangements required to use many forms of happenstance data. Such agreements require that data be aggregated from private sector organisations, standardised and stewarded by public sector bodies, and analysed by authorised third parties, while taking into account the requirements of data protection legislation and concerns about protecting individual rights, data security, and intellectual property.

In this respect, there may be lessons from recent developments in the governance of health data and the role of accreditation mechanisms in enabling rapid data access while continuing careful data stewardship.¹¹ One way to reach this goal is for the ONS to collaborate with the Information Commissioner’s Office (ICO) to formulate their accreditation process into a standardized qualification for data access. The qualification could be equivalent to a “data driving license” that would ensure that qualified experts can get rapid access to different data types with the appropriate standardized ethical and legal training in place. The authorisation and authentication of researchers is emerging as an international best practice, with the ONS, HDR UK and other organisations already subscribing to the “five safes” framework.¹² In practice the ONS would form part of the convening group that generalizes their framework to a national qualification that would include other data sets, with scope for other institutions to join later.

Build capability for long-term data sharing

Access to happenstance data will typically require the coordinated cooperation of many parties. In the case of mobility data and transaction data, for example, effective collaborations during COVID-19 have relied on the skills and expertise of individuals from across the public sector, private sector and academia. Boxes A and B outline the coordinated collaboration required to enable data use, and explore the requirements to set up such collaborations across different countries.

With notable exceptions, barriers to data sharing are perceived rather than real. Several factors likely contribute to this dynamic:

• Data scientists and data protection officers in organisations may perceive different risks associated with data use. This may be because the roles of data scientists and data protection officers are separated. Following our running example of card transaction data, a typical legal concern is that research would require explicit consent from each individual card holder. However, as analysis can often be conducted with fully anonymized subsamples, or aggregates in which no individual could be traced, the risk of disclosure can be directly managed. The challenge in making progress has two sides. On one side, GDPR can too easily be called upon where there is a lack of will to find a solution. On the other side, there is the uncertainty of being in breach of how data protection legislation is interpreted by regulators and Government.¹³

• While many organisations aspire to use data effectively, there may be a mismatch between these strategic aspirations and the business-as-usual practices that contribute to organisational data maturity. The actions described earlier in this report in relation to data readiness can contribute to an organisation’s data maturity. In addition to these, there are wider considerations in relation to organisational competence for data use that can contribute to the ease with which data resources are deployed at times of crisis. Some of these are considered further in Addendum 1.

Overcoming these barriers to collaboration takes time, energy, and human resources, all of which are in high demand at times of crisis.

The common denominator between the successful examples of data sharing cited in this report is the pre-existence of collaborative relationships or simple mechanisms by which relationships could be initiated. There have been successful examples of data sharing frameworks that have been established at pace to enable access to data held in the private sector. The experiences of other countries, set out in this report, illustrate the opportunity to pursue this at scale across multiple stakeholders.¹⁴

At a time of national crisis, with events proceeding at pace, there is very little time to build the foundational infrastructure for collaboration. It takes time to agree on data sharing agreements, and it takes time to build data infrastructure. This process is made easier if there has previously been ‘in principle’ agreements around the forms of data that can be accessed and for what purpose.

The value of pathfinder projects

Pathfinder projects between key segments of industry, academia and government establish ways of working and common agreement around the types of data that could be made available between all parties. Once they exist, they could be adapted to changing policy needs in an agile way. To be successful, such projects require resourcing and strategic leadership, with the experience of other countries suggesting political leadership can play an important role in aligning incentives between stakeholders.¹⁵ Beyond the pathfinder role, these projects can leave a legacy in the form of expertise and guidance in understanding the stages of the data-sharing pipeline.

Through implementation, pathfinder projects provide examples of addressing governance issues by practical necessity. Some are:

• Multi-party data sharing. In the examples in Boxes A and B that go beyond prior academic-corporate relationships, data sharing has worked best for organisations that had already invested in the infrastructure and governance mechanisms to enable data sharing between multiple parties in a secure and privacy-aware way.

• Protocols for contributors. We noted in Box A that data given in kind could stand in competition with business products that use the same data, for example aggregated human mobility data. A pathfinder project will contextualize the criteria by which data is contributed, including timelines and compensation.

• Data privacy. In Box A the INE, as a trusted third-party, ensured that anonymized and aggregated data wasn’t re-identifiable, even when combined with other datasets.

We recommend that pathfinder interdisciplinary data projects are funded. Each project should target a specific policy question,be linked closely to senior decision-makers in relevant government departments, and seek to build on existing work to develop trustworthy data access and linkage frameworks.¹⁶

At least one of these projects should involve a close collaboration between Cabinet Office and Treasury on nowcasting of classical economic metrics (such as GDP) from happenstance data like payments data. The project may not require any law change, but would require a dedicated effort from Government to institutionalise a process that will enable the same kind of collaboration between academia, government and the key private sector players as the examples in Box B.

A second pathfinder project should produce daily views of population mobility between geographic regions, aggregated from origin to destination counts from mobile phone operators. Many measures to contain COVID-19 targeted the movement of people, and high frequency data about general population mobility is required to inform everything from statistical models to the assessment of the impact of policy decisions. In light of such data’s broad utility, the ONS should act as the trusted body to convert happenstance data into high-frequency population mobility statistics.

Accounting for data bias

A dataset is only as valuable as the design by which it was collected. Happenstance data is collected for a different purpose than the questions for which it provides a basis for answers to. This necessitates calibration through surveys or complimentary data, much like the REACT-1¹⁷ studies calibrate and correct for the systematic underreporting of COVID-19 cases in the UK. The examples of mobility and transaction data in Boxes A and B preclude the movement and economic activity of anyone who is digitally disenfranchised. Without a careful correction, data will be biased by the uneven use of smartphones or credit cards.

Create incentives or duties to promote responsible data sharing

Data that might be useful for a question that’s asked in an emergency is usually held by organisations, where it is considered an asset, costing up to millions of pounds to curate and store, and is an integral part of the organisations’ operations or business. This kind of data is often time-person-place data and is of a sensitive nature, regulated by user agreements.

The barriers to data sharing between organisations have been well-explored prior to the COVID-19 crisis. A 2017 report by HM Treasury described how data is “an under-exploited asset”,¹⁸ and a range of forces can act as disincentives to data use. For example:

• A lack of common technical standards or the efforts needed to make data interoperable may contribute to a ‘market failure’ in data sharing¹⁹ or result in data being of insufficient quality to share.

• Organisations may fear privacy or security breaches, or believe that the risks of sharing data outweigh the benefits.²⁰

• Organisations may lack the skills or management systems to create data infrastructures that enable its use.²¹

Together, these issues create a cluster of coordination problems between organisations, resulting in misaligned incentives and incomplete contracts.²²

Corporations and public institutions have indicated willingness to contribute data to projects seeking to help the UK Government in responding to COVID-19, either by making such data public or through establishing collaborations.²³

For example, the Google COVID-19 Community Mobility Reports²⁴ can be used to analyse the relative change in the number of visits to certain places – such as transport hubs in a region – over a specified time period. While not providing the granularity of origin to destination mobility data, Google Mobility Reports data can be used to monitor daily changes in mobility across a region, for example over a period of lockdown, to understand how local populations are responding to policy interventions. This data is similar to aggregated, anonymized insights that are used in products such as Google Maps. Without the proactive sharing of such data, it is unlikely that the research community would be sufficiently aware of the existence of this data to seek to build analyses around it.

In many cases, access to a full dataset is not required to be able to draw insights from it. It is instead often sufficient to share useful aggregates under a common interface, alongside information about relevant pre-processing functions and detailed data descriptors.²⁵ In the Spanish mobility example, we’ve seen that separate daily mobility matrices were computed inside each telecommunications provider. The daily totals were shared with the INE who, as a trusted third party, combined and truncated them (so that very low origin to destination counts could not be used to infer individual movements by any means) to give national matrices.

The common thread between these successful collaborations is the pre-existence of relationships between researchers, policymakers and companies, considered in section c, below. In the absence of such ‘bottom-up’ collaborations, alternative approaches are needed to shift institutional behaviours towards data sharing.

There are a range of ways in which policymakers could seek to increase access to data, for example, by imposing a duty to share certain types of data at times of national crisis.

The Office for National Statistics (ONS) is responsible for collection and publication of statistics related to the UK’s economy, population and society. Under 2017’s Digital Economy Act, the ONS has “permissive and mandatory gateways to receive data from all public authorities and Crown bodies and new powers to mandate data from some UK businesses”. It also has a cluster or responsibilities that require it to engage with businesses, for example to understand patterns of trade, and to obtain statistical information about the population of Great Britain.²⁶ In pursuit of these functions, the ONS has already established mechanisms that enable access to potentially sensitive data for research purposes.²⁷

Recognising the importance of timely access to data, in April 2020 the Secretary of State for Health issued notifications to healthcare organisations, GPs, local authorities and arm’s length bodies “that they should share information to support efforts against coronavirus (COVID-19)”.²⁸ Where there is data that could be similarly vital in developing public health interventions, similar direction may be required in ensuring that organisations prioritise information-sharing activities.

Given the range of organisations that may hold ‘happenstance’ data, the ability to issue such a notice - to enable access to data that could support the COVID-19 response - would require expansion of the powers available to the ONS to access data held by organisations, beyond its current powers in relation to sample and survey data. If the ONS is to take this activist role in promoting data sharing at times of crisis, its statutory remit will need to be updated, to cover the compilation of happenstance data, or its aggregates, for use in policymaking on very short time frames. When circumstances are rapidly evolving in a national emergency, such data will contribute to fast decision-making for the nation.

3. Conclusion

In the normal course of business, there can be misalignments between the generators and users of a given data set in terms of their objectives and incentives for data sharing. The COVID-19 pandemic has heavily impacted industries, individuals, their households and their family relationships. In some ways, this has removed a barrier to the process of data sharing; in our interactions with companies, we found them keen to help in addressing the important policy questions on which we were working. Agreements in principle on what data might be shared for which purposes were easy to make, but agreement in practice proved very difficult to implement.

While there are multiple stumbling blocks in addressing the challenges above, we were able to find examples of successful data sharing endeavours from other European countries that are currently subject to the same regulatory environment as far as data privacy and personal data rights are concerned. These countries also have the same commercial providers for their mobile phones and there is a considerable overlap between their banking systems and ours.

Core to the success of these efforts are teams of highly-skilled people from across academia, the public sector and companies, with skills covering data science, data governance and the development of digital tools for use in policymaking, who are able to rapidly create and scale research collaborations. Demand for people with such skills across sectors of the economy is high²⁹ and building capability in these areas is a significant challenge for governments.³⁰ In recent years, the UK Government has introduced a variety of initiatives aiming to build its digital capabilities, including professional development programmes and changes to recruitment.³¹ However, it has also recognised that further work is required to equip civil servants with the skills they need to make effective use of data in policymaking.³² Similar challenges exist across other sectors.³³ In this context, pathfinder projects can play a crucial role in building expertise at all levels in how best to manage and deploy data resources.

The COVID-19 pandemic has highlighted the best and worst practices in data access: our experience of obtaining similar data from different countries is of a wide range of practices, from screen-scraping individual PDFs to parsing well-documented CSV files directly with minimal effort.

Recent years have seen a variety of efforts to establish principles or guidelines for well-governed data sharing. However, many of these remain statements of good intention, rather than providing mechanisms for improving data quality and accessibility. DELVE used the data readiness bands, outlined in this report, as a common language for understanding progress in a data sharing project. Other efforts, such as the Oxford’s Blavatnik School of Government’s COVID-19 Government Response Tracker,³⁴ show alternative means of compiling agile data resources in an emergency.

Ensuring these efforts are successful and sustainable will require trustworthy governance of data assets, both happenstance and traditional, creating frameworks that can give members of the public confidence that their data is being used safely and rapidly. Evidence from public dialogues on data and its use suggest that key concerns include the purposes for which data is used and how the benefits of its use are shared across society.³⁵ Such dialogues suggest, for example, that individuals are generally content for data to be used to improve healthcare services that result in public benefit.³⁶ Embedding such dialogue in the development of data governance frameworks can help develop and steward data assets in ways that maintain public confidence.³⁷

In addition to the lessons from the COVID-19 response presented in this report, Government could take further action to bolster its planning for civil emergencies by reviewing the National Risk Register³⁸ through the lens of data readiness. Such a review would assess how data could be used in policy interventions for crisis response, the data resources that would need to be readily-available to enable such response, and the action needed to improve data readiness or organisational data maturity in those areas.

The availability and quality of the UK nation’s data dictates our ability to respond in an agile manner to evolving events. If the UK is to effectively deploy its data resources and its expertise in data science to tackle future emergencies – whether from the continuing spread of COVID-19 or other rapidly-emerging policy challenges – action is needed to create a culture of careful stewardship of these resources. Such stewardship would seek to enable safe and rapid use of data to achieve public benefit, while managing concerns about individual rights and respecting the commercial value of these resources.

Appendices

Appendix A

Python code snippet for loading COVID-19 cases by age and sex for Belgium:

pd.read_csv('https://epistat.sciensano.be/Data/COVID19BE_CASES_AGESEX.csv',
            parse_dates=['DATE'])

Python code snippet for loading COVID-19 cases by age and sex for Spain for a single day. Note that the coordinates listed in table_areas can change from report to report and need to be established manually.

url = 'https://www.mscbs.gob.es/en/profesionales/saludPublica/ccayes/alertasActual/nCov-China/documentos/Actualizacion_54_COVID-19.pdf'

table_areas = [
 # (top left bottom right)
 [1.85*72, 0.8*72, 4.75*72, 7.65*72],
 [4.75*72, 0.8*72, 7.5*72, 7.65*72],
 [7.5*72, 0.8*72, 10.1*72, 7.65*72],
]
 
 url = base_url.format(report=report_number)
 with urlopen(url) as f:
   b = BytesIO(f.read())
 reader = PdfFileReader(b)
 num_pages = reader.getNumPages()
 found_page = -1
 date = None
 for i in range(num_pages):
   page = reader.getPage(i)
   text = page.extractText()
 
   if i == 0:
     match = re.search(r'.*(\d{2}\.\d{2}\.(?:\d{2}){1,2}).*',text)
     if not match:
       print("COULDN'T FIND DATE MATCH!")
       print(text)
       break
     date_text = match.groups()[0]
     if len(date_text) == 8:
       date_text = date_text + '20'
     date = pd.to_datetime(date_text, format='%d.%m.%Y')
 
   if re.search('.*edad y sexo.*', text):
     print(f'Found age data on page {i} in report {report_number}')
     found_page = i
 
 if found_page >= 0:
   print(f'Attempting to read table from page {found_page}')
   dfs = []
 
   for area in table_areas:
     b.seek(0)
     dfs.append(tabula.read_pdf(b, output_format='dataframe',
                pages=found_page+1, area=area, 
				pandas_options={'dtype': 'str','header': None}, 
				multiple_tables=False, stream=True)[0])

Other Appendices

The following are materials prepared by individual members of DELVE as inputs into this report.

• DATA-TD1: Organisational Data Maturity

  Prepared for the DELVE Initiative by Neil D. Lawrence, Jessica Montgomery, Ulrich Paquet

Footnotes and References