Thursday, June 14, 2018

Multi-Drug Poisonings Over Time

I’ve written several previous posts about multi-drug interactions. My write-ups of the 2014, 2015, and 2016 drug overdose data all show cross-tables as well as tables displaying the distribution of numbers of substances found in accidental drug overdoses.

I didn’t have the ability to see what this looked like over time until recently. If you want to see, for example, how many people died from a combination of benzodiazepines and prescription opioids, you can’t just go to the CDC’s Wonder database and pull that information. You have to go to the large file that lists all (roughly) two-and-a-half million deaths that happen each year along with each cause listed on the death certificate. Then you have to define your own counting logic. To see trends over time, you have to pull together the detailed files for several years. That is what I have done.

I’m looking to see which drugs are “associated” with each other. (Below I use the term "correlation" loosely; I am not using Pearson correlation or any other well-known statistic. I'm literally saying "These two things are co-related.")  Roughly speaking, this means they appear together more often than I would expect by chance. Here’s an example of what I mean. In 2016, there were 55,071 accidental drug poisoning deaths (presuming those are correctly labeled, which is an open question). 14,483 of these were prescription opioid overdose deaths (which, by my own definition, includes the “Other Opioid” drug category or methadone, ICD10 codes T40.2 and T40.3); 9,177 of these were benzodiazepines-related (the death record listed the benzodiazepine code T42.4); 4,873 of these listed both. Taking some ratios, 26.3% of accidental drug poisonings involved prescription opioids (14483/55071), 16.7% involved benzodiazepines, and 8.8% involved both. What might it mean for two drugs to be associated with each other? To show up together more often than “by chance?” Crudely we can think of the drugs listed on death records as being randomly distributed across those records, in other words they are “independent” observations. (Maybe there is a better way of specifying what independence means, but I think it works for my purposes.) We “expect” to see 0.263 x 0.167 = 4.4% of accidental poisoning deaths to include both substances if the independence assumption holds. Instead we observed that 8.8% included both. So these substances show up together 2.02 times as often as we’d expect given independence. We can say, crudely speaking, that these substances are positively associated. (You can enlarge the tables below by clicking directly on them.)

That is no surprise, and it’s no coincidence that I chose these substances to illustrate my point. It’s well known that benzos can exacerbate the respiratory suppression of opioids and vice versa. (Benzos have their own effect on respiration independent of other substances, just not quite as extreme. Someone I know recently suffered altitude sickness in Colorado. He told me (presumably he heard this while getting medical attention) that you can’t have benzos if you’re having a panic attack at altitude. I was perplexed for a minute then realized, “Oh, right, they suppress your breathing.” Possibly dangerous at high altitude.)

What about other substances? Heroin and benzodiazepines seem to be roughly “independent.” When I do the same exercise explained above, I get something that’s close to 1.00. See the table below for details. (I'm counting deaths that involved either heroin or synthetic opioids as heroin deaths, unless I specify otherwise. The reason is that some of these people are intending to take heroin and end up taking fentanyl or something stronger instead due to black market uncertainties about drug quality. These deaths could be marked with either or both substances. There are technically codes to differentiate these: T40.1 for heroin and T40.4 for "other synthetic narcotics." But in practice I don't think medical examiners are scrupulously checking to see which substance was involved.)

Cocaine and heroin were slightly “dissociated” from each other in 1999, but the correlation seems to have grown over time. Joint cocaine/heroin overdoses appear 1.2 times as often as expected by chance in 2016. It's interesting that this jumped from ~1.0 to ~1.2 in 2011, the first year of heroin's very recent dramatic increase. (You can see the slope change dramatically in the charts in this post.) Notice that cocaine deaths peaked in 2006, then came back down. But their dramatic recent rise, as in the past couple of years, is largely being driven by heroin. If someone were just tracking cocaine deaths and reported some kind of cocaine epidemic, they would be mistaken. This phenomenon of multi-substance poisoning is important. It would be misleading to look at trends in any one substance in isolation.

Taking a look at many possible combinations, I see the following table. (Other Opioids means specifically code T40.2; as stated above methodone has its own code T40.3. For anti-depressants, I'm looking at code T43.2, "other and unspecified antidepressants." Heroin in this table refers specifically to T40.1 and not T40.4. The last column shows how these two codes are correlated.)

I found the very last column interesting. Heroin and other synthetic narcotics were significantly dissociated from each other until 2014, when deaths listing T40.4 began to rise dramatically. That suggests these were very different phenomenon until 2013. There were some people dying of fentanyl and similar substances, but these were probably pain patients with legitimate prescriptions to the fentanyl patch or something like it. It looks like there is a sudden "regime switch." I would have expected the number to be bigger. A lot of the heroin deaths in the years 2014-2016 were actually taking heroin adulterated with fentanyl, but then again maybe there is a tendency for medical examiners to only write one substance on the death certificate. Like I had said, I don't think they are scrupulously checking to see if it's really heroin or if it's something else.

Take a look at the heroin and antidepressants column. It looks like these are dissociated, occurring together less frequently than we'd expect by chance. Other Opioids (code T40.2, not including methadone like the first table in this post) seems to overlap significantly with benzodiazepines and antidepressants. I'd summarize this table by saying that pharmaceuticals tend to go together, and tend not to go with heroin. I might be tempted to say something like, "There are street drug users, and then there are pharmaceutical drug users. These are distinct categories, even if they overlap." That's not quite right, because we're looking at just the population of users who died and not the full population. That's a pretty strong "filter" for our data, and presumably some kind of selection bias creeps in here. But I'd be surprised if you didn't see the same kind of pattern on SAMHSA drug survey data. Once again there is overlap on any given pair of substances, but we're trying to see if there is more or less than we'd expect by chance.

Here are some more substance correlations. Again "heroin" means just heroin (code T40.1) and Other Opioids means just code T40.2. "Cocaine Or Heroin" means just that, one or the other substance is present. So in the third column I'm roughly speaking looking for the correlation between prescription opioid deaths and street drugs. "Stimulants" refers to code T43.6, usually called "Psychostimulants with abuse potential", which includes things like methamphetamine and prescription ADHD medications. (Yes, the government counts them in the same category, because they are chemically similar and have very similar effects at similar doses. Insert stock comment about the hypocrisy of American drug policy here.) The second and third column sort of corroborate the story that "There are street drug users, and then there are prescription drug users." Methadone has interesting correlations with other substances. I wasn't sure what to expect, but I would have thought a lot of methadone poisonings were heroin users in rehab who maybe fell off the wagon. If that were true, I would have expected a number substantially higher than 1.00, but instead it's substantially below. Similar with Other Opioids and Methadone. (Does this mean methadone is often the sole drug found in an overdose death? Not particularly; see the tables in any of the three links at the top. Most methadone deaths involve multiple substances; the distribution of "number of substances involved" is similar to other things on the list.)

I wrote this post thinking the analysis is a good way of tracking this sort of question: Are deaths due to X being driven up by an increase in deaths from Y? If "Yes" I would expect some kind of change in this number. I don't exactly see any smoking gun here. It definitely looks like "other synthetic narcotics" went from being a pharmaceutical-type drug to a street-drug in the past three years, no surprise there. And it looks like heroin is driving up cocaine-related deaths; my "correlation" statistic jumps from ~1 to ~1.2 in 2011, which is when heroin-related deaths started really trending upward.

Something about this analysis bothers me, regarding how I've defined "independence." I have a certain number of bodies, say 55,071 in 2016. Then I'm effectively saying, "I have 14,483 'prescription opioid' tokens, 9,177 'benzodiazepine' tokens, 27,035 'heroin' tokens, 9,979 'cocaine' tokens. 'Independence' means I'm going to randomly distribute these tokens to the 55,071 bodies. 'Dependence' means having one kind of token makes it more likely that you'll get certain kinds of tokens and less likely that you'll get other kinds of tokens." I think this is right, but something about it feels weird. Like I should be fully specifying a rigorous statistical model. Someone more familiar with the tools of social science can pick up the torch from here if they like. If anyone knows, does this "correlation" statistic I've cooked up correspond to an existing concept? If so, please share and I'll try to tighten this analysis up with a more familiar set of tools and concepts.

Tuesday, June 12, 2018

The Nature of the Firm

Why don’t auto workers just manufacture pieces of automobiles and sell their work product on the open market? Why don’t I provide “actuarial analysis” from my home, sitting in my pajamas at a laptop, and sell it to the highest bidder? Why don’t software engineers write code in their basement and sell it piecemeal to interested buyers? No doubt some of this kind of freelance labor exists, but for the most part people work for firms. That is, most workers have long-term arrangements to work exclusively for a single company. They are paid an hourly wage or a salary. Some might get a sort of “performance pay” or rate for piece work, but by and large they are not selling their work product in anything like an open market.

A long time ago, this looked like an interesting puzzle to Ronald Coase. The economy as a whole is characterized by open markets and the free buying and selling of goods and services. But within the economy there are these tiny islands of socialism called “firms.” If open-market buying and selling is relatively efficient for the economy as a whole, why should individual firms resort to centralized planning? In his paper The Nature of the Firm, Coase tries to solve this puzzle.

I won’t give you a full review of the paper, mostly because I haven’t read it. You can find good discussions of it elsewhere, see here or here for example. (Seriously, listen to the Econtalk with Mike Munger. "Bunny slippers. Tell me about bunny slippers." It's priceless.) And besides it’s only 20 pages, so shame on me for not having read it yet. But I’ll give the cliff-notes version and talk about why corporate bureaucracy in practice is disappointing.

Coase’s answer was that firms reduce transaction costs between individuals within the organization. The programmer writes and maintains code for his employer for a specific purpose. He gets paid whether it’s a busy time or a relatively easy time. If the department that uses his code finds an error, he’s right there to fix it. Versus if he were selling code piecemeal, there would be no guarantee he’d have the free time to fix it. Mr. Freelance Programmer could perhaps shove a contract in a disappointed clients face and say, “Honor served. I did what you asked of me.” Of course the client could sue the programmer if they disagree about the contract's fulfillment, but this would take time and resources. Transaction costs loom large in this world of freelance labor. It would be time-consuming to pre-specify every likely contingency or disappointment in a contract, and it’s even more time-consuming to sue for breach of contract. Long-term relationships and dominance hierarchies can solve these problems. If the code breaks, your boss simply tells you to fix it. He doesn’t need to sue you, because the threat of losing your job for insubordination keeps you in line. Your annual pay-raises and bonuses are, at least vaguely and on average, related to the quality of your work, so you have an incentive to create a good work product. This is the ultimate repeat-business relationship.

The same goes for many other kinds of relationships. The auto-worker needs to reliably show up to work on the assembly line, which requires inputs from a hundred other people to keep moving. Someone could try to write a contract binding all these people together on a day-to-day basis and distribute the proceeds of selling the piece-work equitably. But it’s a lot easier for a single employer to simply buy all the machines and other capital and employ the workers on some kind of ongoing basis. Potential conflicts and legal disputes between n(n-1)/2 pairs of individuals are thus minimized. Those workers are more productive, thus there is a greater surplus for all involved parties to share, thus their pay is higher than it would be in a freelance world. Or consider a free-lance capital owner, someone who buys or builds the assembly line with the intention of renting it out at "the market price" to freelance laborers. In this messy and confusing world, I think everyone would quickly see the advantage of forming long-term relationships and dominance hierarchies. 

So why does this go so horribly wrong in the real world? Why is corporate bureaucracy so stifling? Why do some bureaucracies actively fight the lessons of Coase and erect barriers to communication? If corporate bureaucracy is supposed to minimize transaction costs, why does it seem to increase them? 

Most companies have some kind of acquisitions process, perhaps even an entire department. This is to some degree necessary. You can’t just say “yes” every time an employee asks for a new tool or shiny new toy. But often the process is ill-defined and clumsy. There aren’t enough acquisitions staff to handle the requests, or they don’t know how to prioritize the flood of requests. Almost everyone has seen something like this: “If the Acquisitions Department would just buy the damned software, the company would save thousands of dollars worth of employee hours every year. What’s the damn hold-up!?” But someone has to sift through the requests to see if they have merit or don’t. And sometimes the requester is mistaken about the value-added. Still, the process often takes way too long, or it stops because you didn’t fill out the right request form and nobody told you your request got halted. There are times when I’ve said, “If I could pay out-of-pocket and install the software or hardware myself, I would do so, because the added productivity is worth that much to me. The surplus to the company as a whole is even greater than that, but whatever. That’s their loss.” And in fact I paid for a year’s subscription to Datacamp rather than waiting for my employer to do so, because I felt I was wasting time by waiting. It was a good investment. Why did my employer erect these huge barriers to free communication? Why did it fail to reduce transaction costs? It's pretty bad when your employees are longing for an open market in order to actually get shit done. 

Some departments come up with schemes for keeping track of special requests. These can be stifling. “Fill out this form, and submit it via this web portal.” Fuck you, how about this. How about I have a conversation with one of my colleagues, one professional to another, and that colleague performs a task for me because we’re both trying to add value for our employer? “Oh, you want me to submit a ticket through the obscure internal portal that I use once a year? Okay, do you remember the instructions. No? Okay, whose department is that? Oh, I don’t have permission to use that portal? I’ll have to get access, then set up a user name and password. Oops, it won’t add me as a user. Which apparently means I already have an account? Okay, how do I recover my username or reset my password?” 

I have tremendous respect for one of my former colleagues. He basically "hired" one of our programmers to build something useful for him by buying him a few rounds of golf. This guy knew how to get some shit done, and he went outside the bureaucracy to do it. On the other hand, this is a sign of a dysfunctional bureaucracy. My friend should have been able to get the programming resources necessary to complete his project. Bad management made this impossible. The company was failing, in this instance, to perform its Coasian role of reducing transaction costs between employees. Some particularly dedicated employees essentially said, "Screw it, I'll just go out and buy one."

I don't think my past employers were particularly dysfunctional. I think it's probably like this everywhere. I don't have a solution, either. It is genuinely difficult to "manage" a very large group consisting of thousands of individuals, who sometimes have conflicting goals. Maybe organizations should "institutionalize" the process of going around the bureaucracy. Each department has a flexible spending budget for instances of "Screw it, I'll go out and buy one." The procurement department tracks these to identify where its processes are broken. Revealed preference helps identify requests that are truly needed, as opposed to requests that are frivolous. As in, "If they're willing to spend out of their own budget for this, even out of their own pockets sometimes, maybe that means it's really useful. We should probably just shell out for it."

Wednesday, June 6, 2018

More On Piling On

I wrote a piece about piling on yesterday. I posted it thinking it was one of my weaker posts. Then I saw this excellent piece at Coyote Blog and felt a bit vindicated:
I have a guy, who I won't name for privacy reasons, who works for me in Arizona.  Over 10 years ago, barely over 18, he was convicted of some non-violent drug crimes and locked away.  Had I done the same things in my youth, my rich dad likely would have kept me out of jail but as a poor Hispanic in the world of Sheriff Joe's Phoenix, he went to jail.  Over ten plus years later, he had a stable marriage and had his civil rights restored, but was still mostly doing minimum wage labor.   He has been a good, reliable maintenance person at one of our campgrounds, in a job where he could work with his wife.  One day a customer got in some sort of dispute with this man's wife, looked him up online, and found he had a prison record.   This customer then started sending me messages that I must fire this person immediately or else this customer would file suit against us for creating a dangerous environment for her.  When I refused, she then started posting yard signs around town that we hired felons and telling people on social media that they needed to shun our maintenance guy in any number of ways and accusing him of running a narcotics ring out of the campground.
 This is the kind of crap that non-violent drug offenders face their entire lives.  And you wonder why some of them, after finding no work and being shunned by civil society, might turn to crime?

This nosy customer is an awful person. She is effectively demanding that an ex-con never be allowed to re-engage with society. What’s the solution? Keep everyone locked up forever for even the most trivial offense? This is exactly the kind of piling on that I’m complaining about in yesterday’s post. Kudos to the author for ignoring her complaints and standing behind his worker. The temptation to cave to this kind of pressure must be strong. (Am I "piling on" on her? I don't think so, because I don't know who she is. The post is shaming a behavior, not a person. I would attack the behavior and not the person even if I did know who she was.)

Another thing happened yesterday. My four-year-old had had a bad day at his daycare. He’s sometimes aggressive and takes toys from the other kids, and this was just one of those days. He was crying and sitting with the teacher. As soon as I walked in, the other children were regaling me with his list of offenses. The teacher told them to quiet down, that he was the teacher, that he’d already punished my son. (It was very sweet. He picked up my son, who was sobbing, and they gave each other a big hug. “Tomorrow is another day. We can do better.”) My boy knew he’d done bad. He didn’t need his entire peer group piling on.

I see the nosy customer in the same light as I see my kid’s tattling class-mates. The impulse to indulge moral superiority is strong, particularly when it's against someone you don't like or have an outstanding dispute with. But it’s childish to cave to this impulse.

The title of this post is kind of mean if you read it as a pun. It's intentional.

Tuesday, June 5, 2018

Scold the Scolding Scolds

I wanted to write a post about how much I hate "piling on." Some people are adequately punished or shamed, but sometimes the internet (or some other mob) takes it way too far.

Scolding is a means of unofficially sanctioning unacceptable behavior. You either explicitly state or imply that someone has violated some social norm, then roll your eyes or scoff or somehow imply that the behavior is unacceptable. Scolding is natural enough, and it's probably necessary to some extent. The problem is that there can be way too much piling on. Maybe your misdeed goes viral, and you feel the wrath of the entire internet. You did something bad, sure, but only bad enough to earn the scorn of, say, two dozen people in your immediate social circle. Instead you get ten million people deciding that your single misdeed completely sums up your entire existence, and they collectively impose the "career death sentence" on you. Come on, guys. Maybe the impulsive statement or the ill-considered Tweet is being adequately punished. Maybe it's not the only thing in the world this person has ever done. Maybe it doesn't actually summarize the person's entire existence (though it does to you because it's all you happen to know about them). I think the proper approach to this stuff is to look the other way. If someone presents a toxic opinion, try to attack the idea and not the person expressing it. Actively avoid calling attention to the specific person, because you know that the internet is full of outrage-mongers who will pile on.

Another problem is that some things are not misdeeds at all, and changing norms have made previously unacceptable behavior normal. It's been a while, but in my adult life I've heard people make snide, disapproving comments (or perhaps innuendo) about another person's sexuality. I've also heard a crack about someone's oldest child being born before his marriage to the mother. Sexual norms have changed. Nowadays you get scolded for commenting on these things (rightly so); a generation ago you got scolded for actually doing them.

I think attitudes about drug use are also changing. If someone were to insinuate that someone else was an active pot-smoker, I think the accuser in that scenario would lose status. This is a good trend, and I'd love to see it extended to other things. If someone insinuates that an ex-con can't be trusted because of their ex-con status (in and of itself), I think the scold should be scolded. An independent reason for distrusting the person is fine, but piling on on someone who has been adequately punished and served his time is not. The piling-on for convicted "sex-offenders" is also pretty atrocious. Is this person not being humiliated enough? Don't you think other people are doing this job adequately? Didn't the criminal justice system destroy the person's life enough? Glancing sideways and gossiping about someone because they are on a sex offender registry strikes me as gratuitous. Jacob Sullum has written at length about the excessive punishment of sex offenders. The rationale often given for their humiliating post-conviction monitoring is their supposedly high recidivism rate, which is a myth.

I am reminded of a scenario that plays out at my house a few times a week. I will punish my four-year-old son for something bad he did. I'll lecture/scold him, and my six-year-old will jump in and repeat my admonishment to the four-year-old. I have to remind my six-year-old that he is not the parent and that the punishment has already been handled. The feeling of moral superiority can be irresistible, especially to someone with low levels of maturity and poor impulse control. I have to teach my children to actively resist this feeling and recognize when enough punishment has been doled out.

(I don't know how to apply this lesson to "the internet" without it becoming an infinite regress. "Hey, yeah! The new etiquette is that we don't scold logical fallacies. You just calmly point them out and move on. So stop doing it, you fucking jerk!" "Hey, do you even realize that you just..."

Maybe silence is best.)

Sometimes we have to adjust our scolding norms to scold the scolders. Prudes and internet vigilantes can take things way too far. Sometimes we need to tell them, "No, no. That's already been handled. Your piling on doesn't make you virtuous or brave." The object of your scorn might be a legitimately horrible person, but chances are they've already been told that by enough people, some of whom had the power to actually punish them.

Wednesday, May 30, 2018

The Doodle

[This post will contain some very crude humor. Please feel free to pass it over if you don’t care for that. You’ve been warned, and I’m placing the content beneath the fold so you don’t accidentally see it. If you read this post, you are deliberately “opting in.”]

Tuesday, May 29, 2018

Gwern is Excellent

I’ve been reading Gwern recently and it’s fascinating.

When reading the site’s About page I realized that he’s doing what I actually want to be doing with my own blog.
Blog posts might be the answer. But I have read blogs for many years and most blog posts are the triumph of the hare over the tortoise. They are meant to be read by a few people on a weekday in 2004 and never again, and are quickly abandoned - and perhaps as Assange says, not a moment too soon. (But isn’t that sad? Isn’t it a terrible ROI for one’s time?) On the other hand, the best blogs always seem to be building something: they are rough drafts - works in progress.
His site has many long essays, and he is perpetually updating them. On my blog I rarely comment on the daily news ("daily noise" as I call it) or the latest outrage, unless it’s to point out or reiterate something I’ve already said about that particular topic. I’ve written a few long-ish posts that should hold up well regardless of the latest headlines. But actually I’ve been tempted to write things in response to viral news stories, and I often end up re-writing the same post several times, only realizing at the end that I'd written an identical post before. My two longest opioid pieces are my response to a Vox piece and my response to an essay by verBruggen. There is considerable overlap, and in those cases it was worth specifically targeting claims in both pieces. But the option of “perma-linking” to a few long essays and perpetually updating them never really occurred to me. Then when an outrage piece about the “opioid epidemic” or about a tiny cut in the welfare state comes out in the next news cycle, I can dust off an old essay and say “Oh, that old chestnut again. Here’s how I dealt with it last time.” It looks like Gwern is actually using version control software to keep track of updates to his essays. It's probably useful to keep track of how things have changed. On a long enough timeline, this might capture important changes of viewpoint that might otherwise be scattered across several older posts.

The content itself is excellent. You’d do well to read any few random essays under Most Popular. His piece on the Zeo sleepself-experiments was brilliant. It actually encouraged me to keep better track of my sleep and try to get better sleep at night. Meditating before bed seems to help, but I have nowhere near his discipline for controlled self-experimentation. He describes the quantified self (QS) using the example of Stephen Wolfram:
One failure mode which is particularly dangerous for QSers is to overdo the data collection and collect masses of data they never use. Famous computer entrepreneur & mathematician Stephen Wolfram exemplified this for me in March 2012 with his lengthy blog post The Personal Analytics of My Life in which he did some impressive graphing and exploration of data from 1989 to 2012: a third of a million (!) emails, full keyboard logging, calendar, phone call logs (with missed calls include), a pedometer, revision history of his tome A New Kind of Science, file types accessed per date, parsing scanned documents for dates, a treadmill, and perhaps more he didn’t mention.
 Wolfram’s dataset is well-depicted in informative graphs, breathtaking in its thoroughness, and even more impressive for its duration. So why do I read his post with sorrow? I am sad for him because I have read the post several times, and as far as I can see, he has not benefited in any way from his data collection, with one minor exception:
 Very early on, back in the 1990s, when I first analyzed my e-mail archive, I learned that a lot of e-mail threads at my company would, by a certain time of day, just resolve themselves. That was a useful thing to know, because if I jumped in too early I was just wasting my time.
 Nothing else in his life was better 1989-2012 because he did all this, and he shows no indication that he will benefit in the future (besides having a very nifty blog post).
He seems to share my hesitation about fully “scientific medicine”, in that he seems to recognize that there will be heterogeneous responses to any treatment:
The point is making your life better, for which scientific certainty is not necessary: imagine you are choosing between equally priced sleep pills and equal safety; the first sleep pill will make you go to sleep faster by 1 minute and has been validated in countless scientific trials, and while the second sleep pill has in the past week has ended the sweaty nightmares that have plagued you every few days since childhood but alas has only a few small trials in its favor - which would you choose? I would choose the second pill!
 To put it in more economic/statistical terms, what we want from a self-experiment is for it to give us a confidence just good enough to tell whether the expected value of our idea is more than the idea will cost. But we don’t need more confidence unless we want to persuade other people!
 Amen. I love his piece on LSD microdose self-experimentation. He even calls out a famous advocate of the treatment, basically saying: How dare you not have done this yet?

His method for self-blinding is simple but brilliant:
But how to blind myself? I used my pill maker to make 9 OO pills of piracetam mix, and then 9 OO pills of piracetam mix+the Adderall, then I put them in a baggy. The idea is that I can blind myself as to what pill I am taking that day since at the end of the day, I can just look in the baggy and see whether a placebo or Adderall pill is missing: the big capsules are transparent so I can see whether there is a crushed-up blue Adderall in the end or not. If there are fewer Adderall than placebo, I took an Adderall, and vice-versa. Now, since I am checking at the end of each day, I also need to remove or add the opposite pill to maintain the ratio and make it easy to check the next day; more importantly I need to replace or remove a pill, because otherwise the odds will be skewed and I will know how they are skewed.
Even granting that we don't all have the time or inherent motivation to do self-experimentation (much less the appreciative following who will admire our written-up work), I feel like we should all be doing some version of this all the time. Even I can't bring myself to do it. Since reading a few of Gwern's essays a month ago, about the only thing I've figured out is that meditation before bed seems to make me sleep better. (How to "self-blind" on this one? How would I give myself a placebo "20 minutes of quiet meditation"?)

There's still quite a lot I haven't read yet but want to. (Well, "want to" in the economist joke sense: "I want that car." "No you don't!" It's a real knee-slapper.) I've made two false starts at his Why Correlation Does Not Equal Causation essay. I was distracted both times and failed to absorb it, so I think I'll need to find a quiet weekend when the kids are napping and try to read it. The topic is really important. I build statistical models for a living, but even the most sophisticated of these does not tease out causation. A neural net or a gradient boosting model will find deeper and more intricate patterns in the data than, say a generalized linear model (or something cruder like raw averages by demographic). But they don't ultimately tell you whether some relationship is causal. For my purposes (setting insurance rates) I typically don't care. Raw correlations allow you to price insurance for a large population independent of knowing anything about what causes high accident rates. But for other purposes, it's important. Should our society spend more or less on prisons? Should my business double advertising, or perhaps halve it? These are important questions about how the world will change if we take or decline to take some action.

Gwern, if you're reading this I salute you! Keep up the excellent work.

Monday, May 28, 2018

Self-Driving Cars, Trolley Problems and the Veil of Ignorance

Several recent pieces raise what seems like a very silly question: If faced with the choice, should your self-driving car kill you or instead kill multiple other people? If, say, your car is careening toward a family of five, but it can swerve at the last minute and instead crash into a brick wall and kill the "driver", should it do so?

This seems categorically different from asking what a random person should do if he finds himself in a situation where he can choose to save himself or save five other people. (In the traditional trolley problem, the decision-maker is not contemplating self-sacrifice but is trading one stranger's life for several others.) I think this "problem" also goes away if we are capable of total automation. The reason is that we can specify our desired outcome ahead of time and pre-commit to it.

The trolley problem is an interesting mental exercise because the scenario is novel and unlikely to actually happen. A human being with limited capacity for rationality is plopped into this novel scenario with no context, no legal advice, and no information about similar cases and how they were adjudicated. And this person is asked to derive some moral philosophy on the spot and make a decision. But with the question of self-driving cars, we're discussing ahead of time what the default should be. It's not some mysterious moral dilemma presented in a vacuum.

Suppose you ask someone: "Occasionally there will be scenarios where one person can die or five people can die, depending on a split-second decision. You are no more likely to be any single one of those people. Should the one die, or the five?" It's a silly question. Even if you are completely selfish and  your goal is to favor yourself at any cost, you opt to save the five rather than the one. Obviously we want to set the default rules so that the maximum number of people survive these kinds of scenarios. (That's even the "selfish" answer once we remove any mechanism for specifically favoring ourselves.) Generally we want to set society's default rules to prevent as much harm as possible. Of course, you can get people to flip their answers by stipulating, "By the way, you're not one of the five. You're the one." People are selfish. And people will impulsively make the selfish choice when presented with a split second decision. I think this is what's happening when people raise the "Should self-driving cars sometimes kill their passengers?" question. The proper way to raise this question is to not tell the person whether they are a pedestrian or a passenger in the self-driving car. (Or perhaps even more clearly, make the hypothetical choice be between a self-driving car containing one passenger and another containing five.)  We should try to answer this question from behind a Rawlsian Veil of Ignorance, which basically means not telling people who they are in the hypothetical. The Veil of Ignorance avoids self-serving answers and the associated self-serving rationalizations. Doing otherwise is like asking, "Would you sometimes cheat to benefit yourself, even though the socially optimal rule is to be self-sacrificing?" Duh. Most would cheat a little, and many would cheat a lot.

This video invites the viewer to "imagine shopping for a car and having the choice between a car that sacrifices the driver on occasion or a 'preserve the driver at all costs' car." If presented with this survey question, our courage may fail us and we may tick the 'save my own ass' box. But this is the wrong hypothetical. Let's try to fix it. Imagine you're buying a car from an autonomous car company that has done its due diligence on the "liability" question. It knows it will be liable for the occasional injury to passengers and pedestrians, so this company has engineered its software to minimize overall injury (within other constraints, of course). Then another company offers to sell you a car that will favor the driver at all costs. It faces a greater liability, because the car occasionally plows into a crowd rather than killing the passenger. So it has to charge you more for the privilege. Will you buy the more expensive car? It probably depends on the price difference. My best guess is that this would be an obscure feature and almost nobody would think to ask for it. I think even a modest price difference (say, a few hundred dollars) would deter most would-be purchasers. They'd give a different answer when actually shelling out for the privilege of saving themselves in the rare event of a trolley-like scenario, different from the one they gave on a bullshit survey. I also think a company making that sort of car would be a lightning rod for lawsuits. It would incur far more than the expected costs of the additional liability it's intentionally assuming. ("Our vehicles kill an additional ten people per year, multiply ten by the average value of a wrongful death claim, divide by the number of vehicles...Our actuaries tell us this is the expected surcharge per vehicle.") No, I think every single accident involving this manufacturer's cars would be under a black cloud. Every marginal case would be adjudicated to their detriment, and then some. Plus there would be social censure for people buying such cars. "Oh, rich boy thinks he can shell out to save his own skin rather than plow into a group of nuns, does he?" Everyone would know that that brand is the "coward who plows into nuns and orphans to save his own skin" brand. You'd likely see an escalated version of painting "EARTHKILLER" on Humvees.

Change the scenario this way. Cars are sold with the "correct" software, the kind that saves the greatest number of lives. But perhaps it's easy to tweak the code to favor the driver. (Maybe all you need to do is wire someone $10 worth of Bitcoin and someone on the dark web sends you a software patch. Your car connects to your laptop by Bluetooth and installs the patch. Insane hypothetical, I know.)  Is anyone going to do this, even if it's super easy? In the rare even that you do plow into a crowd of nuns, everyone is going to suspect your car's software was tampered with. An investigation confirms there was tampering, and suddenly you become liable for the actual harm caused. Anticipating this sequence of events, almost nobody tries it.

I think this particular trolley problem is extremely rare, making it all the more silly. Why focus attention on something so inconsequential when there are bigger fish to fry? But I think the "correct" answer is obvious enough. The people who are raising this as some kind of serious moral dilemma are being kind of absurd.