Cryptographically Certified Hypothesis Testing [pdf] | Hacker News Reader

Cryptographically Certified Hypothesis Testing [pdf]

sachaservanschreiber.com

58 points3s6 comments

This is basically "Do preregistration and put it in a blockchain-like structure".

It's not a bad idea. It adds cryptographic integrity to good scientific practice.

But hardly anyone outside medicine does hypothesis preregistration yet. In computer science it's basically unknown to do it. All the A/B testing you heard about in user studies is almost certainly p-hacked. So the first step would be to convince people that they actually have to do something about p-hacking.

Most fields are not accustomed to confirmatory statistical analysis as medicine, psychology or finance are. The poor understanding of p-hacking is a mere consequence of this for most fields.

A good custom and first step for any field would be to elaborate more on the validity of an analysis per se, or to give studies a "reality check" at all.

I mean, dealing with the multitude of irreducible results in the natural sciences is a major source of what I would consider to be very bad science, owing to institutionalized biases formed from 'conclusions' made from previously unwarranted results. While it as well is a major issue in medicine, no one dies when some ones claims of a soil carbon characteristic or a stomatal conductance trait end up being irreducible; in-fact it can cost you a career to even open up the can of worms which is validating the previous results on which your work claims to be based.

There are two issues at odds here as I see it. One, science is a career, and if you don't show results, well, it will be a short career indeed. Two, science is a philosophy which governs the determination of the best `truth` possible given the evidence. The two are highly at odds, and perhaps this may give a way to at least resolve some issues related to irreconcilability in all science.

>But hardly anyone outside medicine does hypothesis preregistration yet.

And that's where this idea fails from a practical standpoint. This system gives scientists the means to do something that hurts their chances at publishing/funding.

Now, if it were somehow enforced, then you'd have a different situation, and it could give some serious benefits.

Complexity systems researcher, Didier Sornette, did something similar in 2009 for [The Financial Bubble Experiment: advanced diagnostics and forecasts of bubble terminations](0):

• ...We do not make this document public. Instead, we make its digital fingerprint public. We generate three digital fingerprints for each document, with the publicly available (1) MD5 hash algorithm [1] and (2) 256 and 512 bit versions of the SHA-2 hash algorithm [2] [3]. This creates three strings of letters and numbers that are unique to this file. Any change at all in the contents of this file will result in different MD5 and SHA-2 signatures.

• We create the first version of our main document, containing the first two sections of this document, a brief description of our theory and methods, the MD5 and SHA-2 hashes of our first forecast and the date (1 May 2010) on which we will make the first original .pdf document public.

• We upload this main ‘meta’ document to http://arxiv.org. This makes public our experiment and the MD5 and SHA-2 hashes of our first forecast. In addition, it generates an independent timestamp documenting the date on which we made (or at least uploaded) our forecast. arxiv.org automatically places the date of when the document was first placed on its server as ‘v1’ (version 1). It is important for the integrity of the experiment that this date is documented by a trusted third party.

• We continue our research until we find our next confident forecast. We again put the forecast results in a .pdf document and generate the MD5 and SHA-2 hashes. We now update our master document with the date and digital fingerprint of this new forecast and upload this latest version of the master document to arxiv.org. The server will call this ‘v2’ (version 2) of the same document while keeping ‘v1’ publicly available as a way to ensure integrity of the experiment (i.e., to ensure that we do not modify the MD5 and SHA-2 hashes in the original document). Again, ‘v2’ has a timestamp created by arxiv.org.

• Notice that each new version contains the previous MD5 and SHA-2 signatures, so that in the end there will be a list of dates of publication and associated MD5 and SHA-2 signatures.

• We continue this protocol until the future date (1 May 2010) at which time we upload our final version of the master document. For this final version, we include the URL of a web site where the .pdf documents of all of our past forecasts can be downloaded and independently checked for consistent MD

0: https://arxiv.org/abs/0911.0454

Wonderful approach!

I would just change it to work with notebooks instead of PDF (to include the data and the algorithm and make replication easier) - through the notebooks could certainly generate PDFs.

Then it would bring into publishing the problem of replicable builds: the PDF would not be very helpful if the MD5 didn't match the published signature!

The field seems ripe for disruption, with a new model for scientific research and publication.

Reply on news.ycombinator.com