23andme and the FDA and me

23andMe provides a service wherein you send them a sample of your spit, they run it through a machine that detects 550,000 genetic markers, and then they express a likelihood that you are susceptible to certain genetic disorders.

To do this, they combed through the genetics literature for GWASes: genome-wide association studies, searching for the link between certain genes and certain disorders. I was on the team that did one of them: a study searching for genetic causes of bipolar disorder. Our study involved interviewing and drawing blood from over 2,500 subjects (a thousand at NIH and 1,500 at a German sister lab). As of this writing, Google Scholar says it’s been cited 418 times. If you have a 23andMe account, you can find our study as one of a few under the section on bipolar disorder.

The FDA considers this service—running spit through the Illumina machine, then comparing the data to correlations like those we reported—to be a “medical device”, and has ordered 23andMe to cease and desist from distributing this medical device.

After the genotyping itself, this is an informational product, so it gives us a chance to see how an institution built to handle food and drugs that people ingest will handle a product that is almost entirety a list of statistical correlations.

We have limited information about the storyline to date, and I’ll try to avoid hypothesizing about what’s not in the letter, which already gives us enough for discussion. The 23andMe people may have been uncooperative, or it might even be a simple case of clashing personalities, or who knows what else went on. The labeling regulations in the CFR go on for pages, and I won’t pretend to know whether 23andMe complies or not. [More useful trivia from 21 CFR 801(164): “artificial flavorings, artificial sweeteners, chemical preservatives, and color additives are not suitable ingredients in peanut butter.” That’s the law.]

Snake oil

The FDA should exist. It should be policing health claims. We joke about snake-oil salesmen, but that’s a phrase because there was once a time when people really did sell snake oil and people really were dumb enough to use it instead of seeing a real doctor (not that real doctors at the time were much better…). Pseudoscience lives to this day, and if the FDA didn’t exist, we’d see a lot of late-nite ads for guaranteed cures for cancer for the low low price of $19.95, which would cause people to delay seeking real treatment, which would kill people.

But there are differences. As above, the genotyping service is almost purely informational. You can go to the drug store and buy a thermometer, a bathroom scale, a mirror, or any of a number of other self-inspection tools to find out about yourself. While you’re at the drug store, you can check your blood pressure on that automated machine in the back. Then you can go home and compare your data to Wikipedia pages about hundreds of different maladies. As the first pass of several, Illumina’s genotyping machine is exactly like these other tools for measuring the body, apart from the fact that it is unabashedly a flippin’ miracle of modern science. Perhaps there was a time when people said the same thing about the thermometer.

Is the machine somehow unreliable? When working with the Illumina machines(*) what blew my mind is that the machine is very accurate, as in over 99.9% correct over 550,000 data points. So it’s like a thermometer, except the results are more accurate and reliable. A big chunk of academia’s publications are about surmounting problems in doing research, and along that vein it is easy to find abundant articles on the process and reliability of using the machines themselves. Its quirks are well understood.

[(*) They’re BeadStations, but we always called them the Illumina. It’s good to be reminded that the provision of cheap, universal genotyping is a plot by the Illuminati.]

Diagnostics

But 23andMe provides more than just a list of SNPs (single-nucelotide polymorphisms, which I colloquially refer to as genetic markers here). They also provide the odds of being susceptible to certain diseases, using studies such as the one I worked on.

The FDA’s letter explains how this can cause problems:

For instance, if the BRCA-related risk assessment for breast or ovarian cancer reports a false positive, it could lead a patient to undergo prophylactic surgery, chemoprevention, intensive screening, or other morbidity-inducing actions, while a false negative could result in a failure to recognize an actual risk that may exist.

Here, it seems that the FDA is protecting against doctor incompetence, because providing a mastectomy or chemotherapy because somebody walked in to the office with a printout from 23andMe is not the action of a competent doctor. So there is an implication here that the FDA feels that doctors can’t be trusted. What research there is shows that the ability of MDs to do basic statistical analysis is not stellar. Perhaps the FDA should require that patients may only receive medical test results under the supervision of a trained statistician.

How it works

The false negative part is a little closer to the story of the snake-oil doses that prevent somebody with a real problem from seeking real treatment. However, here we run into a problem as well: the results from 23andMe do not report positives or negatives. They only report probabilities, binned into average, above-average, &c.

After all, our paper doesn’t report anything definite either. We report a statistical association between certain SNPs having certain values and the odds ratio for bipolar disorder.

Given a series of such odds ratios, we can apply Bayes’s rule. You may not remember it from Stats 101, but the gist is that we begin with a prior likelihood of a state (such as ‘the person who mailed in this spit has blue eyes’), then update that likelihood using the data available. If we have several studies, we can apply Bayes’s rule repeatedly. In reality, we would do this with probability distributions—typically bell curves indicating the mean likelihood of blue eyes and confidence bands around that mean. Repeated addition of new data, using multiple SNPs described in multiple studies, typically narrows the variance of the distribution, increasing our confidence in the result.

I don’t have the report in front of me, but they found my brother to be at over 50% risk for obesity. We thought it was funny: he lifts weights and can talk your ear off on all matters of diet. But that’s just how the system works. His behavior has had a greater influence on his odds of becoming obese than the genetic markers, but 23andMe just has a spit sample from a U.S. resident. In the United States, over 30% of males are obese, so if I only knew that a person is male and living in the U.S., I’d guess that he has 30% odds of being obese. That’s my prior, which could then be updated using the genetic information to produce a more personalized image. It looks like his genes raised the odds of obesity from baseline. Perhaps this percentage over 50% is the sort of “false positive” that the FDA letter referred to.

[To balance this “false positive”, the tests also did a good job of picking up some rare risks that we know from our family’s medical history to really be risks. It’s probably TMI to go into a lot of detail here.]

FDA v NIH

To summarize the story so far, by my conception there are three steps to the genotyping service:

  1. Run spit through the Illumina machine.
  2. Gather data about marker/disorder correlations from the literature.
  3. Use statistical methods like Bayes’s rule to aggregate the data and studies to report the best guess at the odds of each disorder.

From the FDA letter: “we still do not have any assurance that the firm has analytically or clinically validated” the genotyping service.

Starting from step three, it is reasonable for the FDA to take a “guilty until proven innocent” attitude toward the data analysis and to require that 23andMe show its work to the FDA. But although the updating problem is far from trivial in practice, a review in good faith by both sides could verify or disprove the statisticians’ correctness in well under the five years that the FDA and 23andMe seem to have been bickering. Some of the results may not even need to go as far as applying Bayes’s rule, and may be simple application of a result from a paper.

The application of the FDA’s statement to steps one and two, where the bulk of the science happened, is the especially interesting part from a bureauphilic perspective. As above, the academic literature has more than enough on the analytic and clinical validity of step one (get good SNPs from the machine) and abundant studies such as the one I worked on that used 2,500 subjects to verify the analytic and clinical validity of step two (calculate disorder odds from SNP results). E.g., here is a survey of 1,350 GWASes.

Yet the authors of the FDA letter do not have “any assurance” of validity.

A paper’s admissability as evidence to the FDA depends on the “regulatory pathway” under which the study was done. If a study is done to approve a drug on a European market, that study is not admissible as evidence at the FDA. Evidently, our study done at the NIH is not admissible as evidence at the FDA. Perhaps I need to parse the sentence from the letter more carefully: “we still do not have any assurance that the firm has analytically or clinically validated” the genotyping service, where the use of “the firm” indicates that research from the NIH doesn’t count, but if the product vendor replicates the study under the explicit purview of the FDA (which, given the scale of some of these studies and their sheer number, is entirely impossible) then that does count.

It’s good that the FDA has standards on testing, because drug companies have strong incentives to put their thumb on the scale in every trial. It could indeed be the case that the FDA has determined the correct way to test medical devices and derive statistical results and, at the same time, everybody else is doing it wrong. Or it could be that the FDA as an institution has an acute case of NIH syndrome.

Advertisements

500 million lines of code

Here at Bureauphile, we care about the measurement of performance. This post is about one terrible way to do it: lines of computer code. You may have seen the claim that healthcare.gov is backed by “about 500 million lines of software code”. That figure is from the last line of an NYT article, from an unnamed source. This was picked up by many people, probably because it makes for a nice headline about the bloat of government contracts, with a punch that a simple statement like this site doesn’t work for a lot of people doesn’t have. I ran into it again at Putative, which prompted me to give this some debunking.

First, let’s do the simplest of calculations. If it takes you one second to write a line of code, and you are a contractor working a solid eight hours/day shift, it will take you 17,361 days to write half a billion lines. With a 250-day person year, that’s 69 person-years. Of course, it takes a lot more than a second to write a line of code: more typical would be—I am not making this up— ten lines of code per day. Let’s give them 100 lines of code per day, then we’re still at 5 million days, or 20,000 person years to write all this up. The contractors started this project at the beginning of the year, and did not have 20,000 people working on it.

So the number is prima facie fishy.

But lines-of-code counts, even on the best of days, should not be taken seriously, because it is so difficult to define what is a line of code.

First, different languages have different whitespace customs. In C, you’ll find people who write

if (x==0)
 {
 return INFINITY;
 }
 else
 {
 return y/x;
 }

whereas in other languages the custom is much closer to

if (x==0) return INFINITY; else return y/x;

So this simple sentiment could be one or eight lines of code.

One day, when avoiding work, I wrote a one-line script that counts lines of C-family code. The gist is that I omit comments, then look for lines that have something more than a }, a), or whitespace on them. Here it is, in one line:

sed -e 's|/\*.*\*/||' -e 's|//.*||' $* | awk -e '/\/\*/ {incomment=1}' -e '{if (!incomment) print $0}' -e '/\*\// {incomment=0}' - | grep '[^}) \t]' | wc -l

Following the ingrained UNIX tradition, this is four programs piped together, one of which (awk) has a three-line program specified on the command line. Readers good with awk will notice cases where this is inaccurate; it’s not worth caring.

I’m coming to a relative stopping point with my work on the the Apophenia library for statistical and scientific computing. How many lines of code does it take to implement—if I may be immodest for a moment—a darn solid statistical library?

14,860: using the nontrivial line counter above.

27,934: Simply counting lines, regardless of their content, including documentation and blanks [wc -l for the POSIX geeks].

39,947: There is a testing suite to verify that Apophenia calculates the right numbers, which includes several data sets, totalling 10,013 lines, which we can add in. More on this below.

41,397: It is a not uncommon technique to write code that generates other code. I use the m4 macro language to autogenerate 1,450 lines of code that gets distributed in the package (by a rough count), which we can add to the above.

56,330: I use GNU Autotools for installation, which takes the generation of code using m4 to the extreme: it produces a 16,933 line script based on a 119 line pre-script that I wrote. Can’t use the library without installing it, so add that on.

325,687: When I sit down to a fresh computer (say, one that just solidified in Amazon’s cloud), I have to install other libraries before Apophenia will run. Apophenia relies upon the GNU Scientific Library, which is 197,452 lines by my nontrivial line counter (they use the super-sparse format—it’s 299,716 lines by raw line count), and SQLite 3, which is 71,905 nontrivial lines of code. One could argue that those are a necessary part of Apophenia.

So Apohenia is somewhere between about 15,000 and 325,000 lines of code. If I’m bragging to friends about how efficient my codebase is, it’s the former; If I’m on a Dickensian government contract, it be the latter.

Getting back to healthcare.gov, one question we might really want to ask is: what lines of code might a maintainer one day have the responsibility of manually revising? The press has been reporting that 5 million lines of code have to be changed, which might be getting toward this question, though I am comfortable assuming that somebody just made up that number too.

We know that the site relies on code lifted from other projects, because the press has reported that they screwed up a copyright attribution for one of them. Apophenia is at arm’s length from the GNU Scientific and SQLite libraries, but web projects are more likely to function by cutting/pasting code from javascript libraries into what gets served up. However, if a bug is found in one of these libraries, the .gov maintainers would first file a bug report with the library maintainers, not try to fix it themselves (depending on the situation).

Is data code? The reader will be unsurprised to hear that there are about 141,000 procedure and diagnosis codes in the ICD-10 system. If there’s a database with 50 lines of description for each procedure (also plausible, especially in a sparse-on-the-page format like some XML), then you’ve got 7 million lines of “code” that needs maintaining right there.

The people running healthcare.gov surely bought whatever underlying data sets from a provider, and if an insurance company submits bad pricing files to healthcare.gov, the legal contracts probably say that it is the responsibility of the insurer to fix it. But in managing the real-world project, responsibility isn’t quite so clear-cut: end users will just see a wrong number and declare that healthcare.gov is broken.

Nobody expects the site to run in fifty lines of javascript that the developers tweeted to each other. Healthcare.gov is no doubt tens of thousands of lines of code by any measure, because health care insurance in the United States is in the running for the most complex system on Earth, and this web site has to simplify it and deliver it securely to millions of people in diverse contexts. But defining where this project ends and the projects, databases, and other underlying structures begin is futile, as is measuring the complexity of the task by lines of code.

Peer reviewer incentives and anonymity

Last time, I sketched a model of the peer review process as an extensive-form game. The model described the review process as a noisy measurement: the paper has some quality, and the review measures that quality plus some bias and some variance. With greater effort, the review’s variance can be lowered. The game I described was one-shot, about a single paper going through the process.

I didn’t describe the reviewer’s incentive to exert effort to carefully evaluate a paper, because within the one-shot game, there is none. To get the referee to exert nonzero effort, there has to be another step inserted into the game:

  • Based on referee’s observed effort level, the editor, author, or reader reward or punish the referee.

This post will discuss some of the possible ways to implement this step.

My big conclusion is that anonymity in peer review is more of a barrier than a help. Having reviewers sign their names opens up the possibility of publishing the reviews, which turns a peer reviewer into a public discussant of the paper, and turns the review itself into a petit publication. Journals in the 1900s couldn’t do this because of space limitations, but in the world where online appendices are plentiful, this can be a good way to reward reviewers for putting real effort into helping readers and editors understand and evaluate the paper.

Continue reading