How to deal with sensitive individual data in open science?

Question

I work with governmental register data, which includes sensitive information about individuals, and therefore cannot make datasets available to the public lest I risk violating the privacy of my subjects. Even though the data are anonymised, it contains information such as addresses and dates of medical prescriptions that should not be made public. I myself need to work inside a secure computer lab and don't have free access to these data. What are the possibilities for open science in this case?

Edit: My terms of use prohibit any sharing of the dataset with others not authorised to use the data. I am not allowed, in addition, to transfer the data from the computer where they are supposed to be analysed.

Answer 1

Great question! Unfortunately, I think there are not going to be great answers. Three ideas have come up in my discussions with users of registry data:

Describe data access procedures completely: Be as transparent as you possibly can be about how you acquired the data, how others can acquire them, and what the policies and costs associated with acquisition would be.

Show everything that you can: While you cannot share the raw data, you can likely share a considerable amount of information from the data. For example, you can include descriptive statistics and graphics that convey the univariate and multivariate patterns in the data. You may also be able to share certain aggregated statistics (e.g., data aggregated at a block or city level).

Offer to collaborate: You may have privileged access to data (i.e., other simply will not be able to access it). In that case, to the extent allowed by data access rules, you should offer the ability to run analyses for others and collaborate with them using the data. This would mean that even your critics can use the data in a meaningful public exchange of ideas, even if the data themselves cannot be made public.

Answer 2

While you can't share the data or the sensitive bits, you can share your results for whatever analyses you're doing.

For each analysis you make of the data, check your results. Remove any identifying or sensitive information, and then publish what you've got.

If others contact you about your results, and ask what your data was, you just have to make very clear that you have the data as part of your job, which gives you access to some sensitive material. If put nicely, no reasonable person will complain about the fact that you can't share it with them.

Answer 3

I will agree that sensitive HIPPA data can't be published, even if anonymized. However, many institutions have groups to provide assistance in this area. One possibility is to produce a reporting database that aggregates the entries according to a set of criteria. Look at the U.S. Census data. They produce a large number of datasets that are aggregated according to various criteria. The criteria for the "group by" (This is a keyword in SQL used to explain data.) There have to be guidelines out there for aggregating data for release.

As I said, look at how the U.S. Census publishes reports. I have seen a few and they contained all of the rules used for aggregation.

Answer 4

One other solution for this problem is for the data owner to produce and release a synthetic data set, based on the original, which adequately replicates the properties of that original (e.g. joint distributions of variables) but does not disclose information about real individuals. A literature search on "synthetic data" should find plenty of discussion of methods.

It's a fair bit of work, and as you're not the data owner all you could do here is ask the relevant agency to consider producing a synthetic data set, but if there's enough demand they might consider it. I know quite a few government agencies are looking at synthetic data for this purpose, because sometimes the hassle of producing synthetic data is less than the hassle of running things like secure data labs.