In a comment based conversation with Anne-Marie Cunningham/@amcunningham last night, it seems I’d made a few errors in the post Demographically Classed, mistakenly attributing the use of HES data by actuaries in the Extending the Critical Path report to the SIAS when it should have been a working group of (I think?!) the Institute and Faculty of Actuaries (IFoA). I’d also messed up in assuming that the HES data was annotated with ACORN and MOSAIC data by the researchers, a mistaken assumption that begged the question as to how that linkage was actually done. Anne-Marie did the journalistic thing and called the press office (seems not many hacks did…) and discovered that “The researchers did not do any data linkage. This was all done by NHSIC. They did receive it fully coded. They only received 1st half of postcode and age group. There was no information on which hospitals people had attended.” Thanks, Anne-Marie:-)
Note – that last point could be interesting: it would suggest that in the analysis the health effects were decoupled from the facility where folk were treated?
Here are a few further quick notes adding to the previous post:
– the data that will be shared by GPs will be in coded form. An example of the coding scheme is provided in this post on the A Better NHS blog – Care dot data. The actual coding scheme can be found in this spreadsheet from the HSCIC: Code set – specification for the data to be extracted from GP electronic records and described in Care Episode Statistics: Technical Specification of the GP Extract. The tech spec uses the following diagram to explain the process (p19):
I’m intrigued as to what they man by the ‘non-relational database’…?
As far as the IFoA report goes, an annotated version of this diagram to show how the geodemographic data from Experian and CACI was added, and then how personally identifiable data was stripped before the dataset was handed over to the IFoA ,would have been a useful contribution to the methodology section. I think over the next year or two, folk are going to have to spend some time being clear about the methodology in terms of “transparency” around ethics, anonymisation, privacy etc, whilst the governance issues get clarified and baked into workaday processes and practice.
Getting a more detailed idea of what data will flow and how filters may actually work under various opt-out regimes around various data sharing pathways requires a little more detail. The Falkland Surgery in Newbury provides a useful summary of what data in general GP practices share, including care.data sharing. The site also usefully provides a map of the control-codes that preclude various sharing principles (As simple as I [original site publisher] can make it!):
Returning the to care episode statistics reporting structure, the architecture to support reuse is depicted on p21 of the tech spec as follows:
There also appear to be two main summary pages of resources relating to care data that may be worth exploring further as a starting point: Care.data and Technology, systems and data – Data and information. Further resources are available more generally on Information governance (NHS England).
As I mentioned in my previous post on this topic, I’m not so concerned about personal privacy/personal data leakage as I am about trying to think trough the possible consequences of making bulk data releases available that can be used as the basis for N=All/large scale data modelling (which can suffer from dangerous (non)sampling errors/bias when folk start to opt-out), the results of which are then used to influence the development of and then algorithmic implementation of, policy. This issue is touched on in by blogger and IT, Intellectual Property and Media Law lecturer at the University of East Anglia Law School, Paul Bernal, in his post Care.data and the community…:
The second factor here, and one that seems to be missed (either deliberately or through naïveté) is the number of other, less obvious and potentially far less desirable uses that this kind of data can be put to. Things like raising insurance premiums or health-care costs for those with particular conditions, as demonstrated by the most recent story, are potentially deeply damaging – but they are only the start of the possibilities. Health data can also be used to establish credit ratings, by potential employers, and other related areas – and without any transparency or hope of appeal, as such things may well be calculated by algorithm, with the algorithms protected as trade secrets, and the decisions made automatically. For some particularly vulnerable groups this could be absolutely critical – people with HIV, for example, who might face all kinds of discrimination. Or, to pick a seemingly less extreme and far more numerous group, people with mental health issues. Algorithms could be set up to find anyone with any kind of history of mental health issues – prescriptions for anti-depressants, for example – and filter them out of job applicants, seeing them as potential ‘trouble’. Discriminatory? Absolutely. Illegal? Absolutely. Impossible? Absolutely not – and the experience over recent years of the use of black-lists for people connected with union activity (see for example here) shows that unscrupulous employers might well not just use but encourage the kind of filtering that would ensure that anyone seen as ‘risky’ was avoided. In a climate where there are many more applicants than places for any job, discovering that you have been discriminated against is very, very hard.
This last part is a larger privacy issue – health data is just a part of the equation, and can be added to an already potent mix of data, from the self-profiling of social networks like Facebook to the behavioural targeting of the advertising industry to search-history analytics from Google. Why, then, does care.data matter, if all the rest of it is ‘out there’? Partly because it can confirm and enrich the data gathered in other ways – as the Telegraph story seems to confirm – and partly because it makes it easy for the profilers, and that’s something we really should avoid. They already have too much power over people – we should be reducing that power, not adding to it. [my emphasis]
There are many trivial reasons why large datasets can become biased (for example, see The Hidden Biases in Big Data), but there are also deeper reasons why wee need to start paying more attention to “big” data models and the algorithms that are derived from and applied to them (for example, It’s Not Privacy, and It’s Not Fair [Cynthia Dwork & Deirdre K. Mulligan] and Big Data, Predictive Algorithms and the Virtues of Transparency (Part One) [John Danaher]).
The combined HES’n’insurance report, and the care.data debacle provides an opportunity to start to discuss some of these issues around the use of data, the ways in which data can be combined, the undoubted rise in data brokerage services. So for example, a quick pop over to CCR Dataand they’ll do some data enhancement for you (“We have access to the most accurate and validated sources of information, ensuring the best results for you. There are a host of variables available which provide effective business intelligence [including] [t]elephone number appending, [d]ate of [b]irth, MOSAIC”), [e]nhance your database with email addresses using our email append data enrichment service or wealth profiling. Lovely…
So it seems that in a cost-recovered data release that was probably lawful then but possibly wouldn’t be now* – Hospital records of all NHS patients sold to insurers – the
Staple Inn Actuarial Society Critical Illness Definitions and Geographical Variations Working Party (of what, I’m not sure? The Institute and Faculty of Actuaries, perhaps?) got some Hospital Episode Statistics data from the precursor to the HSCIC, blended it with some geodemographic data**, and then came to the conclusion that “that the use of geodemographic profiling could refine Critical illness pricing bases” (source: Extending the Critical Path), presenting the report to the Staple Inn Actuarial Society who also headline branded the PDF version of the report? Maybe?
* House of Commons Health Committee, 25/2/14: 15.59:32 for a few minutes or so; that data release would not be approved now: 16.01:15 reiterated at 16.03:05 and 16.07:05
** or maybe they didn’t? Maybe the data came pre-blended, as @amcunningham suggests in the comments? I’ve added a couple of further questions into my comment reply… – UPDATE: “HES was linked to CACI and Experian data by the Information Centre using full postcode. The working party did not receive any identifiable data.”
CLARIFICATION ADDED (source )—-
“In a story published by the Daily Telegraph today research by the IFoA was represented as “NHS data sold to insurers”. This is not the case. The research referenced in this story considered critical illness in the UK and was presented to members of the Staple Inn Actuarial Society (SIAS) in December 2013 and was made publically available on our website.
“The IFoA is a not for profit professional body. The research paper – Extending the Critical Path – offered actuaries, working in critical illness pricing, information that would help them to ask the right questions of their own data. The aim of providing context in this way is to help improve the accuracy of pricing. Accurate pricing is considered fairer by many consumers and leads to better reserving by insurance companies.
There was also an event on 17 February 2014.
Via a tweet from @SIAScommittee, since deleted for some reason(?), this is clarified further: “SIAS did not produce the research/report.”
The branding that mislead me – I must not be so careless in future…
Many of the current agreements about possible invasions of privacy arising from the planned care.data release relate to the possible reidentification of individuals from their supposedly anonymised or pseudonymised health data (on my to read list: NHS England – Privacy Impact Assessment: care.data) but to my mind the
SIAS report presented to the SIAS suggests that we also need to think about consequences of the ways in which aggregated data is analysed and used (for example, in the construction of predictive models). Where aggregate and summarised data is used as the basis of algorithmic decision making, we need to be mindful that sampling errors, as well as other modelling assumptions, may lead to biases in the algorithms that result. Where algorithmic decisions are applied to people placed into statistical sampling “bins” or categories, errors in the assignment of individuals into a particular bin may result in decisions being made against them on an incorrect basis.
Rather than focussing always on the ‘can I personally be identified from the supposedly anonymised or pseudonymised data’, we also need to be mindful of the extent to, and ways in, which:
1) aggregate and summary data is used to produce models about the behaviour of particular groups;
2) individuals are assigned to groups;
3) attributes identified as a result of statistical modelling of groups are assigned to individuals who are (incorrectly) assigned to particular groups, for example on the basis of estimated geodemographic binning.
What worries me is not so much ‘can I be identified from the data’, but ‘are there data attributes about me that bin me in a particular way that statistical models developed around those bins are used to make decisions about me’. (Related to this are notions of algorithmic transparency – though in many cases I think this must surely go hand in hand with ‘binning transparency’!)
That said, for the personal-reidentification-privacy-lobbiests, they may want to pick up on the claim in the
SIASIFoA report (page 19) that:
In theory, there should be a one to one correspondence between individual patients and HESID. The HESID is derived using a matching algorithm mainly mapped to NHS number, but not all records contain an NHS number, especially in the early years, so full matching is not possible. In those cases HES use other patient identifiable fields (Date of Birth, Sex, Postcode, etc.) so imperfect matching may mean patients have more than one HESID. According to the NHS IC 83% of records had an NHS number in 2000/01 and this had grown to 97% by 2007/08, so the issue is clearly reducing. Indeed, our data contains 47.5m unique HESIDs which when compared to the English population of around 49m in 1997, and allowing for approximately 1m new lives a year due to births and inwards migration would suggest around 75% of people in England were admitted at least once during the 13 year period for which we have data. Our view is that this proportion seems a little high but we have been unable to verify that this proportion is reasonable against an independent source.
Given two or three data points, if this near 1-1 correspondence exists, you could possibly start guessing at matching HESIDs to individuals, or family units, quite quickly…
– ACORN (A Classification of Residential Neighbourhoods) is CACI’s geodemographic segmentation system of the UK population. We have used the 2010 version of ACORN which segments postcodes into 5 Categories, 17 Groups and 57 Types.
– Mosaic UK is Experian’s geodemographic segmentation system of the UK population. We have used the 2009 version of Mosaic UK which segments postcodes into 15 Groups and 67 Household Types.
The ACORN and MOSAIC data sets seem to provide data at the postcode level. I’m not sure how this was then combined with the HES data, but it seems the
SIASIFoA folk found a way (p 29) [or as Anne-Marie Cunningham suggests in the comments, maybe it wasn’t combined by SIASIFoA – maybe it came that way?]:
The HES data records have been encoded with both an ACORN Type and a Mosaic UK Household Type. This enables hospital admissions to be split by ACORN and Mosaic Type. This covers the “claims” side of an incidence rate calculation. In order to determine the exposure, both CACI and Experian were able to provide us with the population of England, as at 2009 and 2010 respectively, split by gender, age band and profiler.
This then represents another area of concern – the extent to which even pseudonymised data can be combined with other data sets, for example based on geo-demographic data. So for example, how are the datasets actually combined, and what are the possible consequences of such combinations? Does the combination enrich the dataset in such a way that makes it easier for use to deanonymise either of the original datasets (if that is your primary concern); or does the combination occur in such a way that it may introduce systematic biases into models that are then produced by running summary statistics over groupings that are applied over the data, biases that may be unacknowedged (to possibly detrimental effect) when the models are used for predictive modelling, pricing models or as part of policy-making, for example?
Just by the by, I also wonder:
– what data was released lawfully under the old system that wouldn’t be allowed to be released now, and to whom, and for what purpose?
– are the people to whom that data was released allowed to continue using and processing that data?
– if they are allowed to continue using that data, under what conditions and for what purpose?
– if they are not, have they destroyed the data (16.05:44), for example by taking a sledgehammer to the computers the data was held on in the presences of NHS officers, or by whatever other means the state approves of?