> If the institution is public, data should be public as long as individual PII is removed.

PII is much broader than most people understand because reidentification of what amateurs would see as deidentified data is easy (often trivial), and, as a consequence, to be useful for research data is often not fully deidentified.

EDIT: As an example, the HIPAA safe harbor deidentification standard requires removing 18 kinds of identifiers, including, as one of them:

All geographic subdivisions smaller than a state, including street address, city, county, precinct, ZIP code, and their equivalent geocodes, except for the initial three digits of the ZIP code if, according to the current publicly available data from the Bureau of the Census: (1) The geographic unit formed by combining all ZIP codes with the same three initial digits contains more than 20,000 people; and (2) The initial three digits of a ZIP code for all such geographic units containing 20,000 or fewer people is changed to 000

To add to this, PII isn't always even clear. Different jurisdictions identify PII differently, there isn't One Master Definition that you pass a unit of data through, upon which an authoritative "THIS IS PII" or "THIS ISNT PII" is returned.

+1

We had a multi-month project to get a subset of our data considered 'clean', and it required a consultant, a stats PhD and many dev hours. It was healthcare, so on the high end of paranoia (justifiably) but nowhere it is as simple as dropping the "name" column

HIPAA also allows for “expert determination” [0] for deidentification that differs from safe harbor and can allow for all sorts of things since there’s no definition of what an expert is.

And reidentification risk can be as high as even 1% and still be acceptable for hipaa. In a dataset of a million people that’s 10,000 people identified and still be “acceptable.”

But hipaa doesn’t apply to these CA data, it’s just the clearest example of deidentification regulations I know of.

But it’s totally possible to deidentify data suitable for release to these researchers. It’s just what CA considers deidentified and if it’s still useful enough to these researchers. For the topic they are researching it should be pretty straightforward to remove PII enough to protect individuals and only remove some really unique characteristics (ie, only a single 20 year old or a particular race and ethnicity).

But I’m guessing age groups by race and gender and socioeconomic are possible to preserve without tying back to an individual. Id go so far to say as it would be non-trivial, but pretty easy, for CA to produce this for the researchers, if not to the general public.

[0] https://www.hhs.gov/hipaa/for-professionals/privacy/special-...

The intent of the comment was not to say the process is trivial or that removing PII is sufficient. However, it is not as impossible as people are making it out to be. I’ve worked on datasets at social media companies where literally thousands of columns were considered PII but realistically removing/scrambling just a subset of columns would make it impossible to identify individuals.

> I’ve worked on datasets at social media companies where literally thousands of columns were considered PII but realistically removing/scrambling just a subset of columns would make it impossible to identify individuals.

Maybe, though I doubt it was that easy against any but the most trivial reidentification efforts, but since most privately held PII isn’t regulated (in the US at least), there's little consequence for a social media conpany getting it wrong other than PR.

> data should be public as long as individual PII is removed

This is one of those things when ideology doesn't match the real world. If the amount of data is large enough and with enough parameters, removing PII doesn't do anything to protect privacy.

What about medical records? What about protected classes? What about data about vulnerable people or victims?

Student data is protected by another layer of regulation and for good reason. Also, the judiciary is a 'public' institution in general. Should we not seal records for minors? 'No exceptions' - year right..

> If the amount of data is large enough and with enough parameters, removing PII doesn't do anything to protect privacy.

Honestly it doesn't have to be that large. We see this all the time with data websites or apps collect. Sure, you remove John Smith's name, but you still have his GPS coordinates. For the school, you remove Professor Smith's name, but you have a professor who teaches CS 123 and has 4 graduate students. You bet you can guess who that is.

I really do support open data, especially about public institutions, but at the same time we are in an era where this information is quite powerful. Seems to make a case for something like homeomorphic encryption or something, but will that even stop these collisions?

The appropriate notion here seems to be Differential Privacy, which is a mathematical definition informally saying "a scrambling of the dataset that is information theoretically indistinguishable from that where one arbitrary person is added or removed". It's a surprisingly deep topic, with entire (very good!) textbooks dedicated to it.

PDF (entirely legal): https://www.cis.upenn.edu/~aaroth/Papers/privacybook.pdf

We're getting a little off topic here, but the Federal government has published specific guidance on de-identification of medical records. You can construct some artificial scenarios where re-identification might be theoretically possible through record linkage with other data sources but in practice it's unlikely. In principle a similar approach could be used for student data, although I'm not familiar with the legal issues.

https://www.hhs.gov/hipaa/for-professionals/privacy/special-...

But all of this is orthogonal to the core issue of whether a state government should be allowed to prevent researchers from participating in lawsuits. There is no student privacy issue involved there. Witnesses in a civil suit still aren't allowed to violate student privacy laws regardless of the data they have access to, so it makes no sense to conflate those issues.

> We're getting a little off topic here, but the Federal government has published specific guidance on de-identification of medical records.

But releases (even without patient consent, with an IRB waiver) of non-deidentified PHI data for research is allowed, and this is specifically because deidentification necessarily destroys elements that would often be necessary in research.

> You can construct some artificial scenarios where re-identification might be theoretically possible through record linkage with other data sources but in practice it's unlikely.

It is explicitly part of the HIPAA safe harbor standard that, in addition to removing the required identifiers, you cannot come up with such a scenario, and if you can, the data is not deidentified. (The last criterion of the standard is “The covered entity does not have actual knowledge that the information could be used alone or in combination with other information to identify an individual who is a subject of the information”.)

What does any of that have to do with the legal issue of whether a state should be able to prohibit participation in certain lawsuits as a condition of gaining access to research data? Neither party has raised re-identification as a concern, nor have there been any allegations of privacy law violations.

> What does any of that have to do with the legal issue of whether a state should be able to prohibit participation in certain lawsuits as a condition of gaining access to research data?

As you yourself noted upthread, you had already taken this subthread afield from that topic.

I also completely lost the plot here...

AFAICT, the thread went something like this:

The top-level concern is something like this: professors use their trusted relationship to schools in order to make bank on expert witness fees, which feels a bit corrupt and calls into question the researcher's motives.

A rebuttal to this concern is that we can side-step that issue entirely because these data sets should be public anyways (anonymized, of course!). This obviates the above concern, since the researchers won't need to compromise themselves in order to get exclusive access to data that allows them to be expert witnesses and rake in $$$$.

But the problem with that proposal is re-identification: if we can't make the data anonymous, then we all agree that it shouldn't be released (implicit in the "anonymized, of course!" caveat to "just release all the data" proposal).

Then you pointed out that even for more important data like healthcare data, FDA apparently has ways of allowing release of data that takes into account the risk of re-identification risk (I didn't know this; thanks for sharing!)

Then dragonwriter and you got deep into the weeds on HIPPA stuff.

TBH I have no idea which of you is most correct here. But anyways, there are two ways for this conversation to go:

1. You are correct, good enough anonymization is possible: Stanford researchers should not be silenced; it is problematic that they have access to data other people cannot access, but the correct solution is to negate the originally problematic distinction between those researchers and the general public by making data public. Then there is no reason for the researchers to agree to these contract clauses, because they will have access to the data.

2. dragonwriter is correct, good enough anonymization is not possible: We can go back up to the top-level concern and observe that "just release all the data with anonymization" isn't a feasible solution to this problem. Or maybe there isn't actually a problem here at all. IDK. But in any case, "obviate the problem in the top-level post by releasing anonymized data" isn't a workable solution.

Again, not following closely enough to have an opinion, but that's where we are now.

I think a good compromise position is that we should have a law stating that K12 data should be available to certain education researchers -- subject to IRB approval and so on -- without any other strings attached. Including "don't sue me" clauses in releases of public data sets does feel like an inappropriate abuse of student privacy concerns.

The researchers don't have a trusted relationship with schools. They have a contractual relationship with the state government. The fundamental issues underlying the lawsuit are First Amendment freedom of expression and contract law; expert witness fees and researchers' motives are irrelevant.

Whether student data de-identification is good enough or not is a total red herring. No one has accused the researchers in this case of violating privacy rules. The comments here about such privacy issues are largely hypothetical and tangential.

If you think that California needs a new law expanding research access to educational data then feel free to suggest that to your state legislators, or sponsor a ballot initiative.

LOL what? I'm so confused.

It's not a red herring. It's a side conversation about a different but related topic.

Someone proposed just releasing all the data.

Someone else replied with why that wouldn't work.

Ie, a conversation happened and the topic of discussion shifted.

FWIW I agree with you on the object level question. No idea why you're being so abrasive, especially when you're the one who initiated/continued the conversational thread about deanonymization and even prefaced with "We're getting a little off topic here".

Presumably at that point you understood that the topic of conversation had shifted, and people's agreement/disagreement didn't necessarily have anything to do with the original topic... since you literally said so and no one disagreed... so your reaction here is pretty odd and off-putting.

Neutral bystander here, I didn’t get any of that from op.

This is true but it’s quite possible to correct for unique or infrequently occurring combinations so privacy is still preserved and data are made available.

It’s not that hard to design data release to compensate for privacy protections and statistically test for a specific level of risk. There’s a whole body of work on statistical disclosure control and there’s plenty of open source or cheap enough privacy enhancing technology available.

I’m not familiar with CA, but I expect they have someone on staff who can produce a “safe” dataset that preserve privacy and still allows for this question to be researched by low level geography, demographic, and socioeconomic factors.

> as long as individual PII is removed

The devil is in the details. These records were likely not designed to be shared and I'd assume the entire system contains vulnerabilities that could create leakage. Leakage that could be used to harm individuals in a variety of ways - from discriminating future prospects to harassment and much much more.

I agree in principle to what you're saying but we need to be truthful about what these current systems are capable of.

You'd have to remove so much PII as to make any examination worthless "A student of age <redacted> and gender <redacted> at school <redacted> has a GPA of <redacted>". As little information as "a 16 year old black male at Main Street High School" can be narrowed down to a handful of possible candidates at a lot of CA schools.

I disagree as it depends on how many 16 year old black males there are in that high school. It’s pretty simple to apply k-anonymity to control for an acceptable risk level. And add in generalization of age into groups and many questions can be answered.

I think you could definite answer race x gender x grade but it will be harder when you factor in more unique characteristics like household income or vaccination status, etc.

Removing PII would probably also involve removing individual grades and other information that's necessary for any research to be effective. Thanks to predatory data collection practices on the internet, we know how little information you actually need to deanonimize someone. The problem only becomes worse when we're talking about kids.

That said, research that cannot be reproduced is useless. There's a balance to be struck here, and it's somewhere between "make all data public" and "lock the data in a vault".

Not when there is an existing law limiting it. FERPA specifically protects student records.

If FERPA makes it illegal to share the data with researchers, then certainly it shouldn't have been shared.

If FERPA allows sharing the data with researchers, is it right/proper/legal to share only on the condition it can't be used to harm the schools in court? Presumably that part is not in FERPA.

(And to be clear, California here didn't just say they couldn't use the data in court, they said the researchers could not testify in any court case at all against the state. But we're talking hypothetically)

> as long as individual PII is removed

https://www2.ed.gov/policy/gen/guid/fpco/ferpa/index.html, "The Family Educational Rights and Privacy Act (FERPA) (20 U.S.C. § 1232g; 34 CFR Part 99) is a Federal law that protects the privacy of student education records"

Now the question is if aggregated data with no PII is protected.

Data aggregation does not guarantee privacy.

Of course not, but there are tests that can be applied to determine if privacy is protected.

It’s not possible to just aggregate and be done. But it is possible to set some privacy threshold and then insure that all records conform to that acceptable risk level.

Even anonymous records? That would seem to preclude studying the effectiveness of the education system. And if they weren't anonymous, what possible conclusions could you draw compared to anonymous records that would warrant that access?

I did some consulting on this WAAAAY back in the day.

In the case of education, good enough anonymization often isn't really possible. A lot of information about school students is public -- yearbooks are gold mines, as are results from extra-curricular activities such as sports, class notes about whether/where they went to college or first jobs after college, etc. Still more can be purchased. As yet more can be inferred from public data (eg home address, rough estimates at parental income, etc.). This was back in the day. I'm sure now it's much worse.

Most of the questions you want to ask about education are about treatments and outcomes. If these treatments (eg extra-curriculars) and outcomes (eg attended college, graduated HS, etc) are public then you can often figure out which student corresponds to each supposedly anonymous data-point.

We didn't have a name for it when I did this work, but the basic point was "if you publish the dataset everyone will know little Johnny Table's test scores and GPA". Today this is called a reidentification attack.

(I don't really know enough about the topic to have an opinion on the article per se, but "just publish everything" is definitely not a workable solution :))

Provide proof by de-anonymizing one data set from research within the past five years.

LOL. I'm not going to deanonymize datasets for an internet stranger. Certainly not for free, but also not at any hourly rate unless I know that the organization employing me either has the original non-anonymized dataset or at least have very strict internal controls about how deanonymized data will be handled.

(I also haven't done this work in a LONG time, and there's now a whole lot of academic work on the topic that didn't exist back then, so there are probably much better consultants for legitimate organizations looking to hire for this work.)

If you want proof, you can use google to find LOTS of papers along these lines analyzing real datasets. I think https://arxiv.org/pdf/cs/0610105.pdf is a fairly typical example.

I work with large private K-12 datasets regularly. Even as a passer-by, we had one for a big government-issued dataset in Australia. It might’ve been census data? It might’ve been just over five years ago? I actually don’t care. The principles of effective data anonymisation, esp. in education, are understood by people that actually work in this area. It seems odd to be so demanding.

I’ve noticed a trend on commenters making these sorts of demands. I can’t tell if it’s a new (particularly dumb) rhetorical ploy or if they’re just looking for free labor.

Another thing to consider about deanonymization:

HIPAA says that 4 or more digits of a zipcode is PII. The people who protect your healthcare think most of a zipcode is too cardinal to reveal.

How many schools serve more than one zipcode? More than a few?

I very very strongly disagree. It's important that researchers get access to diverse data and industry collaboration is often crucial for this. If companies are required to make all their data public they will be far less willing to collaborate with research. It's already hard enough as is to convince corporations that it's worth their time.

In theory this makes sense, and I agree with the spirit. In practice, you can often re-identify individuals even in an anonymized dataset. For example, if you're dealing with a very rare disease, or a small minority group, you can usually figure out an anonymized row of data is referring to if you really try. So, it's not so simple, and the responsible thing to do is not have a blanket policy that takes human judgment and accountability out of the loop.