Givewell wants to put lithium in your drinking water

Actually, they just want to look into the possibility.

Alexander Berger of Givewell writes:

In the past you’ve written a couple posts about GiveWell’s research, and we’ve recently posted something else that I thought might be of interest to your audience: an expression of interest in research on the impact of trace lithium on suicide rates.

The basic story is that there a number of non-experimental studies that find that higher levels of trace lithium in water is correlated with lower suicide rates (see footnote 1 for citations), and there’s strong evidence that much larger doses of lithium reduce suicide in some psychiatric populations, but the non-experimental studies don’t have strong causal identification and there don’t seem to have been any population-level experimental studies. Our rough back-of-the-envelope calculations suggest that if either of the models in Kapusta et al. (2011) or Blüml et al. (2013) are roughly correct, a small increase in the amount of trace lithium in drinking water in the U.S. could prevent thousands of suicides per year.

We’re interested in talking to researchers about what it would take to do a more sophisticated study—with the idea that we might eventually fund it—and I thought your audience might include many people who’d be qualified to weigh in. I also thought this might be of interest to you personally given some superficial similarity to the work you’ve done on radon.

I was curious so I did a quick google search and found this interesting opinion piece from 2014 by Anna Fels from which, among other things, I learned,

7-Up was originally called Bib-Label Lithiated Lemon-Lime Soda and contained lithium citrate right up until 1950.

82 thoughts on “Givewell wants to put lithium in your drinking water

  1. “higher levels of trace lithium” meaning how much higher? How much did it lower suicide rates?

    I wish people would get into the habit of replacing qualitative words like “higher”, :lower” etc. by numbers.

    Also, by my back-of-the-envelope calculation the usual psychiatric dose of Lithium is about 3000x larger than the typical dose one might get from drinking water. So extrapolating from the psychiatric evidence of suicide reduction to this sort of thing is a bit of a stretch.

    • Yes, it is a small dose, but it’s life-long, consistent, and the dose-response curve is not known here and may be quite flat since the mechanism might be preventing deficiency rather than acting as a cut-down psychiatric treatment. It’s unfortunate that there are no studies since this is such an easy thing to study – the existing natural variations, all way under any plausible safety limit, are enough to observe an effect if they were just randomized! Just another example of how it really is easy to run randomized experiments, and it’s not the difficulty which is the problem, but *people* who are the problem.

      (This reminds me of the Ebola vaccine: there was not enough to go around, to say the least, and there was considerable uncertainty over whether it was effective, so why can’t we just enter patients into a lottery, getting both unbiased fairness of scarce vaccines and also much better information on whether it worked? Instead, it was given out haphazardly, making interpretation more vexed and also triggering accusations of imperialism & racism when it was given to Africans (because you’re testing unproven treatments on poor black people HOW DARE YOU) and when it wasn’t given to Africans (you’re leaving poor black people to surely die and saving rich white people HOW DARE YOU) – how I wish I was making this up.)

      • What sort of study would you design? This sounds hard to study via an RCT.

        If it is indeed a life-long, low-dose effect, given the relatively low incidence of suicide (12 in 100,000) wouldn’t you need a fairly large cohort and very long duration tracking of subjects?

        • Yep, definitely hard. Would take tens of years at a minimum, possibly as much as say 50 years. And, you’d have all kinds of challenges on the reporting and classification of suicides.

          You’re probably better off spending some time to figure out some biologically plausible mechanism in animals before you go forward. like even just finding that mice given trace levels of lithium in the drinking water have different levels of lithium in their brains in areas known to be associated with emotional processing (fear, stress, etc). And to see how the dose-response works at varying levels of “trace”.

        • Besides, does a mass intervention even make sense here?

          Fluridation of drinking water was very different. Dental caries were super high incidence in the population (80% in children?). Also the target dose was in the same ballpark as the natural Fluoride levels in many locations. A modest boost in the level gave great efficacy.

          In this case we are talking of suicides, a relatively very low incidence event. Approx. 12 in 100,000. Doesn’t a targeted intervention make a hell lot more sense here than the broad non-discriminating brush of a drinking water doping program?

        • > What sort of study would you design? This sounds hard to study via an RCT.

          This is exactly the sort of defeatism which bugs me. Think about it a little more. This is a cluster-randomization set up, where you only need to add trace lithium to a central water supply to boost it to under established safe limits and then simply watch the already collected data about the populace’s mental health, suicides, and violent crime; since the effect should be fairly general, you get to look for a generalized decrease rather than being limited to a single end-point. Administratively, this is a centralized single intervention: you need to dump the dilute lithium into just a few places. No need to enroll subjects, track them, interact with each one, do testing or questionnaires, or in general spend the thousands of dollars per head that a clinical trial would. This is about as easy as an randomized experiment gets to run, especially when you compare with things like drug trials (try running a 10-year trial on thousands of healthy people for baby aspirin! now *that* is a hard randomized experiment to do). It boils down to walk into the water treatment plant, pull out the fluoride injector, put in some lithium, then a year later download a spreadsheet from the local authorities on suicides or murders or whatever and plug into R.

          Cheap, too: the cost to run is about the same as, say, doing water fluoridation, which costs something like under $0.5/person/year, so for a municipality or town of 50k, the intrinsic cost of running such an experiment is trivial, under $25k/yr (which wouldn’t even cover one postdoc’s university overhead). Since the amounts are so small and under natural levels of variation, and lithium is one of the cheapest elements around in the first place, I suspect that $25k would be a considerable overestimate.

          > Would take tens of years at a minimum, possibly as much as say 50 years.

          I don’t know what you are basing this extreme claim on. The studies do not seem to require 50 years of data to turn in positive correlations… Since no one has done a meta-analysis of the correlations yet, it’s hard to give a solid estimate of how many years/people you need, but the Texan studies suggest that the effect may be fairly quick (apparently lithium levels in any locality fluctuate considerably over the year due to things like rain leaching lithium out of the ground) so that would help considerably in efficient designs like alternating high/low levels in each locality.

          The problem, as so often, is the politics. You can point out the possible benefits and no evidence of harm, but as soon as you suggest any kind of supplement or fortification which isn’t the traditional fluoride/iodine/iron/vitamin D/folic set of additions which Americans have come to accept as natural and just, then kneejerk responses like Menzies’s kick in. Any attempt to study this experimentally will spend all their time and resources just doing outreach and getting permission and cooperation and complying with regulations and requirements.

          > And, you’d have all kinds of challenges on the reporting and classification of suicides.

          I also don’t know where you are getting this idea that suicide reporting is so extraordinarily bad that no reliance can be put on it and increases in official numbers on suicides are meaningless.

          > You’re probably better off spending some time to figure out some biologically plausible mechanism in animals before you go forward.

          The animal studies are positive. Look at the reviews linked. (IIRC, Schrauzer 2002 is the one that spends the most time on the animal literature.) Not, of course, that finding animal results should bolster one’s confidence all that much…

          > In this case we are talking of suicides, a relatively very low incidence event. Approx. 12 in 100,000. Doesn’t a targeted intervention make a hell lot more sense here than the broad non-discriminating brush of a drinking water doping program?

          No, because you can’t target these people and the programs we have do not work. Consider the base rates alone: how do you plan to reach exactly those 12 in 100k people? (And even if we could: you really think that we have fantastic programs which eliminate suicide, violent crime, schizophrenia, etc, and it’s just a matter of targeting them right…? Psychiatrists would love to know about these interventions you have in mind, because the ones they use on patients who have actually tried suicide sure don’t seem to work all that well.) I would also point out that if the existing solutions did work to the point where there was no possible benefit from lithization, then the causal relationship here should not be perceptible as a correlation.

        • It would depend on the location’s natural levels. You want to change levels as much as possible from year to year to get greater power (for example, alternating years from ~0 to as high as you dare), but if you go above its past peak level, then you open yourself to ethical issues if lithium turns out to be harmful rather than neutral or beneficial.

        • @gwern

          From the Kapusta article, the peak for Austrian districts was Mistelbach at 0.08 ppm.

          So let’s say we take the typical district with a 0.01 ppm natural level & boost that 8x to 0.08 ppm.

          You really think by observing the next year’s suicide rate you can tell something meaningful?

        • I don’t know, Rahul, but I think before I so confidently declared that everything was futile and utterly hopeless like most of the commenters here seem to without reading the papers or thinking much about the topic, I would like to see a power analysis first indicating how much data is necessary – because I have not found peoples’ intuitions to be a good guide to statistics, to say the least.

          (At some point I would like to do a set of meta-analysis/power-analysis/decision-analyses to fully assess lithium, but there are still studies in the pipeline and I’m more interested at the moment in some other topics like vitamin D & baby aspirin’s effects on all-cause mortality which are much more practical and easier to do since I don’t have to do the meta-analysis myself.)

        • @gwern

          I’d invoke the Precautionary Principle. I think it’s up to you (or whoever thinks mass doping of Li is a good idea for a public trial) to demonstrate that their study is safe AND potentially has the power to reveal anything useful.

          e.g. I’d wager 0.08 ppm would be safe but useless (at least for gleaning anything out from a 1 year study). Maybe you’d see some effect at a higher dose or much longer durations but then we come to uncharted territory for safety.

        • But it brings me to a complaint, that people don’t properly cast their models in dimensionless form (you knew it was coming). Using median concentration among the observed locations as the scale factor would be very appropriate. Median concentration would be around 0.01 mg/L, so the dosages would range from 0 to 8 times the scale factor, the equation of the line would be something like 18 – 2x on that scale, and you could easily determine that if you took the median person, and increased their dose by an additional median amount, you’d decrease risk from 16/100k to 14/100k or so…

          instead of whatever totally wacky number they have (-143.39 x + 17.974 ??? uh ok)

        • > What sort of study would you design? This sounds hard to study via an RCT.

          This is exactly the sort of defeatism which bugs me. Think about it a little more. This is a cluster-randomization set up, where you only need to add trace lithium to a central water supply to boost it to under established safe limits and then simply watch the already collected data about the populace’s mental health, suicides, and violent crime; since the effect should be fairly general, you get to look for a generalized decrease rather than being limited to a single end-point. Administratively, this is a centralized single intervention: you need to dump the dilute lithium into just a few places. No need to enroll subjects, track them, interact with each one, do testing or questionnaires, or in general spend the thousands of dollars per head that a clinical trial would. This is about as easy as an randomized experiment gets to run, especially when you compare with things like drug trials (try running a 10-year trial on thousands of healthy people for baby aspirin! now *that* is a hard randomized experiment to do). It boils down to walk into the water treatment plant, pull out the fluoride injector, put in some lithium, then a year later download a spreadsheet from the local authorities on suicides or murders or whatever and plug into R.

          Cheap, too: the cost to run is about the same as, say, doing water fluoridation, which costs something like <$0.5/person/year, so for a municipality or town of 50k, the intrinsic cost of running such an experiment is trivial, You’re probably better off spending some time to figure out some biologically plausible mechanism in animals before you go forward.

          The animal studies are positive. Look at the reviews linked. (IIRC, Schrauzer 2002 is the one that spends the most time on the animal literature.) Not, of course, that finding animal results should bolster one’s confidence all that much…

        • Wait for real you are encouraging researchers to dump chemicals in public water supplies to experiment on people without their knowledge?

        • “this is a centralized single intervention…….No need to enroll subjects, track them, interact with each one, do testing or questionnaires,”

          Hey, it’s a feature not a bug! :)

        • Let me remind you that since lithium levels already vary naturally and most water systems make no attempt to control it below the very high levels usually established, people are already being experimented on every day without their knowledge. I am not proposing anything that Mother Nature is not doing to everyone every day; I am merely proposing that we could try to learn something from it rather than endlessly suffer potential substantial harms.

        • Yeah but then you use the natural variation in Lithium already present, you don’t manipulate it.

          That is the difference between observational research and experimental research.

        • Well, but if it is a life-long gradual effect, why do you expect your t0 + 1 year spreadsheet to reveal any effect save random fluctuations?

        • > Yeah but then you use the natural variation in Lithium already present, you don’t manipulate it.

          I have no idea what you mean here. A researcher randomly manipulating lithium levels within natural ranges is causally the same thing, but statistically it’s totally different: we can infer causality from one but not the other, even though the actual fact of causality is the same in both.

          > Well, but if it is a life-long gradual effect, why do you expect your t0 + 1 year spreadsheet to reveal any effect save random fluctuations?

          We don’t know it’s only a gradual effect but even if it is, a shorter experiment can still show it. If it accumulates, there’s still a short-run amount!

        • “A researcher randomly manipulating lithium levels within natural ranges is causally the same thing, but statistically it’s totally different: we can infer causality from one but not the other, even though the actual fact of causality is the same in both.”

          I doubt you could justifiably infer causality because I suspect the effect would be small enough that various other things would be different as well between different times and locations. Controlling one variable doesn’t mean everything else was the same. Really you need a precise prediction about what the effects would be to build confidence that is the primary explanation for any difference.

        • > I doubt you could justifiably infer causality because I suspect the effect would be small enough that various other things would be different as well between different times and locations.

          Randomized experiments are run all the time on different locations; it just means you have a multilevel structure (what is this – multilevel models coming up on Andrew Gelman’s blog? impossible!) which must be taken into account so you get your units right. For example, the African economic and philanthropy experiments discussed so often on both this blog and the Givewell blog often involve whole villages as the treatment units. (This has some implications for lower power but again, one should do an actual analysis here before a priori declaring defeat and giving up.)

          > Controlling one variable doesn’t mean everything else was the same.

          Randomization balances the other variables on average.

        • @Anoneuid

          +1.

          Even going by the Kapusta results, the Li concentration variation is only explaining 4% of the observed suicide rate variation.

        • > That is a myth afaict, instead it makes extremely unbalanced and exactly balanced samples unlikely

          I don’t think you understand what you’re linking. The point it is making is *not* that randomization does not balance on average; it does. But that it is inefficient compared to if you could balance. (Read the Hurlbert paper linked there. You could also check out Ziliak’s papers on Gosset and blocking – Ziliak is partisan but he still explains what the point is.) Everyone is well aware that randomization does not produce exact balance in the same way that explicit blocking or pairing can; but of course, you can’t, because *you don’t know all variables*. This is why for optimal designs, you balance/block on all the variables you know about (eg twin experiments which get you much better power because all the genetic confounds are balanced; or for a hypothetical lithium experiment, repeated-measures design on each locality, where pairs of years are randomized to the intervention of lowering/raising) and then randomize to take care of the rest. The randomization helps because it balances all the unknown and unobserved variables on average.

        • “The randomization helps because it balances all the unknown and unobserved variables on average.”

          This comment explained it best:
          “The primary goal of randomization is to simulate the effect of independence. It does this by eliminating biases that arise through systematic assignment of treatments to subjects. These biases produce inaccurate estimates—most importantly, biased variance estimates—and loss of control over Types I and II error. Even confounding variables (which really amount to a lack of independence) are simply a case of omitted variable bias.”

          It does not balance out differences from the null of no difference on average in practice, only in a hypothetical scenario.

        • So, I don’t disagree that there’s an ethical issue to be considered, but it’s not clear without actually spending some time on the issue, that there is a real “ethical problem”.

          For example, would there be an ethical problem if you took people who had a disease, and there were two hospitals that had a certain treatment, and you randomized them to move to location 1 for treatment or location 2, while not telling them that at location 1 the naturally occurring lithium level in the drinking water was the same as their starting point 0.01 mg/L, while the naturally occurring level in town 2 was 8 times higher at 0.08 mg/L?

          I doubt very much that people doing that kind of research look at the trace mineral content of the waters. Is this a case where ignorance = no ethical issue, but if you knew there was a difference then you have an ethical issue? I find that ethically problematic

          Trace elements and other low-level exposures are really interesting actually. I vaguely remember reading about how exposing mice to one time low level doses of radiation followed later by a higher dose could have measurable benefits vs just the second dose in controls… I guess the thing being that there are all these low level effects out there that can go up or down. Some things can be protective at low doses and toxic at high doses (like most drugs for example). It’s just a really interesting issue as to how “normal background variation” in a wide variety of dimensions might affect people.

        • @Daniel
          “… would there be an ethical problem if you took people who had a disease, and there were two hospitals that had a certain treatment, and you randomized them to move to location 1 for treatment or location 2, while not telling them that …”

          There would be an ethical problem if you randomized them without their informed consent.

        • I disagree, if the hypothesized effect is due to a low dose over a long time. You’re going to have to run this lithium supplementation for 20 years, and then you’re going to have to contend with people moving into and out of the area over that time period, and you’re going to have to un-confound it all with whatever societal declining levels of suicide and crime etc there are overall, with spatial variability, and the effect of local economic conditions over a 20 year period.

          It’s the duration that’s difficult. if giving people a single glass of trace-lithium water was an “effective drug” this would be easy.

        • > You’re going to have to run this lithium supplementation for 20 years, and then you’re going to have to contend with people moving into and out of the area over that time period, and you’re going to have to un-confound it all with whatever societal declining levels of suicide and crime etc there are overall, with spatial variability, and the effect of local economic conditions over a 20 year period.

          These are all good reasons to object to the correlational studies, but randomization balances confounds, that’s the point. It’s also not clear that you have to study for 20 years (again, where are you pulling all these numbers and claims from?), but even if you did, so what?

          > if giving people a single glass of trace-lithium water was an “effective drug” this would be easy.

          If you could give people a single glass of trace-lithium and see huge effects observable in a small sample, wouldn’t that immediately refute all the correlational studies for having much smaller effects…?

        • “again, where are you pulling all these numbers and claims from?”

          Mostly from your own point: “Yes, it is a small dose, but it’s life-long, consistent”

          under the assumption that only a “life-long, consistent” dose or something approximating that would have an effect, you’d need to run this experiment for tens of years. The observational studies show small but nontrivial effects, and they ARE based on something approximating “lifelong”.

          “These are all good reasons to object to the correlational studies, but randomization balances confounds”

          People move into and out of areas so “lifelong integrated dosage” could vary widely and not under your randomized influence, entire areas undergo swings in economic conditions. Randomization balances confounds at a given timepoint, but a single randomization at point t=0 doesn’t necessarily balance time-integrated confounds over 20 years! Also, how many municipalities are you going to test, to avoid spatially correlated confounds you’re going to want to test across hundreds of municipalities across the country I’d think, otherwise “texas had a big economic boom starting a few years after we initiated, and 3 out of the total of 5 treated municipalities were in texas” becomes a problem.

          “If you could give people a single glass of trace-lithium and see huge effects observable in a small sample, wouldn’t that immediately refute all the correlational studies for having much smaller effects…?”

          Right, that was my point. Short-term studies (1 glass, or even maybe 1 year) are not likely to be informative just based on the fact that the observational studies DO involve a near lifetime of dosing and show moderate effects. For example the first study cited:

          http://bjp.rcpsych.org/content/bjprcpsych/198/5/346.full.pdf

          and especially the supplement:

          http://bjp.rcpsych.org/content/bjprcpsych/suppl/2011/04/11/198.5.346.DC1/bjp_198_ds346.pdf

          Figure DS4 shows you might expect a reduction of suicide rate from say 15 to 10 per 100k in this population by setting the lithium level from basically zero to the max observed.

          In a test population of 10M people, with half having the treatment and half not, you’d be looking to detect 250 prevented suicides in a given year. But, what is the lag time until this effect takes effect? If you can detect it in a year, great, that helps, but if you need to wait 5 to 25 years before this effect really builds up… you’d better do it in hundreds of municipalities to avoid “something special and unrelated happened a few years after we started in the municipalities we tested”

          Long-term studies are just hard because “long term” anything is hard.

        • All that is to say, it sounds like a really interesting study to design, because it has lots of great statistical issues (time series for example) but it sounds a lot harder than “swap out the flouride treatment device with a fluoride+lithium treatment device and download a spreadsheet in a year”

        • @Daniel Lakeland
          “Randomization balances confounds at a given timepoint, but a single randomization at point t=0 doesn’t necessarily balance time-integrated confounds over 20 years!”

          The types of confounding you describe would indeed make it difficult to estimate the causal effect of exposing the entire population to lithium for an extended period, and that is the causal effect of most interest. We agree there. However, with random assignment and a large enough sample of towns, any difference in suicide rates between the treated and untreated towns can be said to be caused by lithium. The direction of the causal effect of interest would most likely be the same as the direction of the observed difference between towns randomly assigned to lithium vs not. So I still think it’s possible to learn something from randomization in this context. It’s similar to how the intention to treat effect in a clinical trial does not tell you the effect had there been full compliance, but it’s still useful even if the latter is of more interest.

        • > Mostly from your own point: “Yes, it is a small dose, but it’s life-long, consistent”

          I was pointing out to your skepticism that there are a wide range of possibilities here for how a small daily dose could cause effects.

          > The observational studies show small but nontrivial effects, and they ARE based on something approximating “lifelong”.

          They also suggest lots of fluctuation and it’s often noted that lithium levels are variable, which is a suggestion that these effects could be relatively quick and we cannot jump straight to ‘this study will take forever’.

          > People move into and out of areas so “lifelong integrated dosage” could vary widely

          As I already pointed out to AJG, this sort of measurement error of dosage will bias downwards the observed correlation from the true causal effect and be common to both the correlational and experimental studies.

          > Randomization balances confounds at a given timepoint, but a single randomization at point t=0 doesn’t necessarily balance time-integrated confounds over 20 years! Also, how many municipalities are you going to test, to avoid spatially correlated confounds you’re going to want to test across hundreds of municipalities across the country I’d think, otherwise “texas had a big economic boom starting a few years after we initiated, and 3 out of the total of 5 treated municipalities were in texas” becomes a problem.

          For cluster randomization, I think somewhere upwards of 4 works well in the African studies, but you can also randomize each by year as well. This is where information on how fast doses kick in would be very useful for study design: if it’s relatively acute and you can observe most of the effect with units of years, then you don’t need to include a lot of a different locations since you can get good information by alternating years and a total time period of 10 years could be thoroughly adequate (5 pairs) and it’s much easier to get permission etc from a single location. If effects are slow, then you would need more locations – you might not need to follow them too long, but I figure it’s probably much easier to continue the experiment for a year in a location that has granted permission than it is to get permission in the first place, so needing a lot of locations is painful.

          > Figure DS4 shows you might expect a reduction of suicide rate from say 15 to 10 per 100k in this population by setting the lithium level from basically zero to the max observed. In a test population of 10M people, with half having the treatment and half not, you’d be looking to detect 250 prevented suicides in a given year.

          Not just suicide, remember. Any real study would be tracking the other stuff like violent crime for two reasons: to get greater power in looking at them all for reductions, and to deal with concerns about side-effects. (If you can show a reduction in not just suicide, but rape, murder, depression, etc, that helps assuage concerns about negative side-effects; it seems unlikely that any micronutrient which could help with all that would have some horrible catch and is not simply just plain good for you. For example, iodine has not been shown in any rigorous way to have no side-effects, and in fact we’re fairly sure that the introduction of iodization has killed a number of people with thyroid/goiter problems, but it’s hard to read all the experiments showing greater IQs, achieved education, employment, etc and come away with anything but the belief that the iodine really was an important micronutrient that people were deficient in and there is no horrible catch where iodization turns people into robots or shortens their lifespans or something.) And that’s a fairly reasonably detectable effect: a few hundred thousand people for 50% power in the worst-case scenario using the simplest design, quickly checking in `power.prop.test`. That’s not a terribly large city or county, and a preliminary experiment can help justify more extensive/expensive experiments as the posterior probability increases.

          > but it sounds a lot harder than “swap out the flouride treatment device with a fluoride+lithium treatment device and download a spreadsheet in a year”

          I don’t see why. The actual experiment is still trivial. You haven’t pointed out any mechanical difficulty, just statistical. It’s not like everyone drinks exclusively from their own well and you have to distribute lithium tablets and hope people use them while maybe half of them will comply like in a clinical trial.

        • Z, yes so that’s why I mentioned how you’d have to do many towns, to cover a large enough range of geographical and socio-political environments, that you had some chance to balance the confounds as well as the time-trends in the confounds. You’d probably like even to do matched pairs of towns. One where you treat, and one somewhat nearby in similar political districts (states/counties etc) where you don’t, under the assumption that the physically nearby towns would have similar trajectories on the confounding variables. Then you’d have to consider issues of mobility between the nearby towns…

          It’s an interesting experiment to think about designing, but it’s not as easy as “dump some lithium and download a spreadsheet a year later”

        • gwern: I agree that it’s mostly if the effect takes a long time that the statistical issues make the whole thing difficult. If you really can test this in 1 year… we’ve got a whole different ballgame (ethical issues still are an issue).

          So, shouldn’t we probably run a “distribute low level lithium to at-risk people and monitor for a year” before putting it in the drinking water? I’d say yes. Offer to give out free bottled water to veterans at a VA and randomize the bottles or something like that. With appropriate disclosure, recruitment, and monitoring of course.

        • Yeah, the power calculation is one of the things we’d be curious about.

          For what it’s worth, one of the ideas we had was that, given reasonable concerns about adding anything to public water supplies, even well within natural variation, perhaps an experiment could measure the effect of filtering out lithium (again, within the range of natural variation), rather than adding it: http://www.givewell.org/health-effects-of-trace-lithium-in-drinking-water. I’m not sure whether a trial to filter lithium out would be considered to violate clinical equipose though; it seems like it should if the goal would be to look for an increase in suicide rates.

          As we say in the link above, it seems like it should be possible to do more persuasive observational research than has been done to date (e.g. by finding some plausibly exogenous natural variation in water lithium levels, perhaps at the border of service by two different water sources), and that seems like an obvious step to take before any intervention would be appropriate.

          And I agree with the issue you raised below about the need to collect evidence for negative side effects, and the idea that it would be quite hard to do so persuasively, given the risk of small, diffuse negative effects.

        • Naive question:

          How much is the year to year variation in Lithium readings at the same location?

          e.g. The Kapusta work seems to correlate suicide rates versus single readings at one point in time? Are the readings in a single town / district fairly consistent?

          How about the sensitivity to sampling location within the water supply system? Trace metal quantitation can be tricky.

        • Good questions – I’ve wondered the same things, and it’s not totally clear to me. Kapusta et al. took an average of 65 samples per district (their unit of analysis) over 5 years, and then just did the analysis on the average lithium reading at the district level; they don’t report on the time or intra-district variation. I haven’t seen anything on variation within the same water supply system – it’s totally possible that there’s a high level of variation there (and accordingly that borders between systems wouldn’t be meaningful).

          My instinct is that someone should know the answer to these questions off the top of their head, but I’m not sure who that person is. A lot of the literature on this seems to be by psychiatrists rather than, e.g., water quality engineers, though I may just not know the right place to be looking.

        • Lots of water supply companies have to monitor their water for a variety of things, you can get reports on these measurements as public documents sometimes. And if not, you could certainly start calling those agencies and ask them to get ahold of their water quality records.

          Here’s my local “mutual” water company’s report: http://www.rclwa.org/uploads/2013WQ.pdf

          they mention that they take “thousands of water quality tests for more than 100 different contaminants.. weekly, monthly, annually, and every 3 years depending on the substance”

          I bet that’s data that they could be either required or at least persuaded to give out.

        • “it seems like it should be possible to do more persuasive observational research than has been done to date (e.g. by finding some plausibly exogenous natural variation in water lithium levels, perhaps at the border of service by two different water sources), and that seems like an obvious step to take before any intervention would be appropriate.”

          Agreed

        • I think that given that our best available observational evidence is that lithium prevents suicide. Any experiment in which it’s removed from the water is strictly unethical (ie. as of now, we EXPECT that to kill people).

          Given that we have limited idea of what the side-effects will be on the majority of the population, any experiment in which it’s added to the water within the “normal” range needs clear analysis and ethical justification. One way to get that would be to investigate side effects and main effects in an animal model. Another way would be to do more careful observational studies, not so much to look at the main effect (reduced suicide), but to look for evidence of any side effects!

        • Here’s a slightly different suggestion:

          Aren’t there well documented risk factors for suicides. Even genetic ones. What about a study that targets such cohorts & then dopes their drinking water via bottled water or tap attachments or something?

          Wouldn’t that be a better approach to start with than mass doping?

      • It’s really quite depressing that human ethical intuition is set up this way, that as soon as you touch the existing variation, even just to randomise, the guilt is on your shoulders. But hard to get around.

        What about bipolar patients on microdoses? It’s plausible that after taking high doses for a while you just need to maintain.
        I am guessing such research is not carried out, because you would not risk new episodes, but I’ve frequently heard that patients adjust their own dosages downward to cope better. Not experimentally of course…

        • Ruben – it isn’t just that you are experimentally manipulating people that is the problem. It is that the experiment is designed specifically to affect how people’s brains are working, how they are thinking, how they are feeling. These are fundamental aspects of being human, and intentionally manipulating that without consent or supervision is, to me, unethical. If the dose is enough to affect suicidal decisions, it is almost certainly enough to alter other aspects of people’s mental state.

      • gwern:

        I do think you are pointing to the real problem and I think it would to worth reviewing experiences in the vaccine field. What is different is the option to opt out of vaccination (though sometime with penalties).

        (I used to work in that field and yes the delays in Ebola vaccine testing are something I wish you were making up – it was very dismaying).

  2. Separate from the scientific questions… the proposal to put something in the drinking water with the express intent of altering mental functioning… is it April already?

    • Yeah. This.

      I mean, suppose fortification would prevent 1 out of 100 suicides, what else are you changing about the other 99+% of the population?

      Now if you had a group of moderately chronically depressed individuals identified as well-above-average risk of suicide, and you offered them a slot in the study along with background on the scientific evidence and then randomly allocated RealLithiumWater(tm) and PlaceboLithiumWater(tm) (aka “Water”), maybe that would be OK.

      But it wouldn’t be a test of fortifying drinking water for a whole population – it would just be a drug trial. Because you want to give people drugs to keep them from killing themselves. So that is what you should do, a drug trial, and you should do it clearly and transparently and openly.

  3. A cost-benefit analysis might be an interesting hurdle here. Even a tiny, innocuous side effect would have to contend with the 100,000-to-12 amplification factor.

    e.g. Say the Li doping caused a trivial itch on the pinky finger. In a targeted drug that’d be a negligible cost to pay to save your own life. But when multiplied by 100,000x how does the collateral damage look?

    • > A cost-benefit analysis might be an interesting hurdle here. Even a tiny, innocuous side effect would have to contend with the 100,000-to-12 amplification factor.

      There are also large benefits to consider. One murder prevented may not sound like much value spread over a large number of people, but murders themselves are extremely costly to society and economic estimates start in the low millions just for the victims and soar into the dozens or scores if you try to take into account the rippling costs of the entire prison and justice system apparatus, extra private spending on security, corrosive effects to social trust and cooperation, etc. And it’s not just murder which is being reported as inversely correlated but all kinds of violent crime, suicide, and mental issues.

    • Or say Li causes a short-term decrease in suicides but has a long-term consequence that is otherwise bad. Like maybe the suicide decrease shows up in a rigorous study that everyone believes in and the study takes 5 years, but then we find out later that by causing a few less people per year to commit suicide we did something really bad to kids’ development or pregnant mothers or the elderly or something, but that effect builds up over like 10 years. I certainly can’t think of a good RCT design that won’t take forever and run into ethical issues even if you don’t make any methodological mistakes. Like you forgot to account for people moving away and moving into your town of choice, or the effect just gets swallowed up in noise.

      • > Like maybe the suicide decrease shows up in a rigorous study that everyone believes in and the study takes 5 years, but then we find out later that by causing a few less people per year to commit suicide we did something really bad to kids’ development or pregnant mothers or the elderly or something, but that effect builds up over like 10 years.

        Can you name 3 examples of trace nutrients which had beneficial effects on broad measures of mental well-being like suicide or violent crime or depression but which turned out to have larger back-fire effects on things no one noticed? Or is this purely a hypothetical?

        > Like you forgot to account for people moving away and moving into your town of choice, or the effect just gets swallowed up in noise.

        Measurement error like this (I would also mention that different people will be exposed to different amounts of tap water based on their eating and drinking habits) dilutes the effect to zero, but this is already weakening the observed correlations. So if a experiment has been properly designed to be well-powered based on the observed correlations, that is already accounted for.

        • I can name several minerals that have broad negative effects on people which we nevertheless in the past spewed into the environment at “lowish” levels (ie. not at levels that were immediately toxic):

          lead (from auto fuel, paints),
          mercury (fluorescent lamps, batteries, coal power plants),
          cadmium (protective coatings on metal parts) for example.

          So, the assumption that Lithium is a “nutrient at low levels” and that it has “beneficial effects on broad measures of well-being” should maybe be investigated in a lab setting in an animal model before we set everyone’s exposure to the max naturally observed level. I don’t necessarily trust an observational regression sufficiently to justify the experiment. To justify spending money on the animal model? absolutely. To justify a well-informed volunteer human group? Maybe. To justify forcible consumption at the high end of naturally occurring concentrations? yeah, not so much. To justify altering the level to control it between 0.5 and 2x the median concentration (0.005 to 0.02 mg/L)… maybe but you’re limiting your effect size a lot there

        • > I can name several minerals that have broad negative effects on people which we nevertheless in the past spewed into the environment at “lowish” levels (ie. not at levels that were immediately toxic):

          None of those had any support for benefits from animal studies, correlation studies, or hypothetical experiments (all of them, as I recall, indicated harm and it was heavy industry lobbying which tried to override scientific concern and any need for experimental verification), and so they are irrelevant to either AJG’s hypothetical or my question.

        • They had tons of support for benefits: lead prevented knocking in engines, reduced the cost of building efficient engines and made better transportation available to a large number of people. Better transportation meant better health too, as people got wealthier by taking advantage of new technologies they could afford better food, more access to health care, and yadda yadda.

          mercury allowed us to create fluorescent lights which reduced the need for electricity production, reduced pollution from power plants, improved the performance of people doing indoor tasks, improved education performance, education and economic prosperity produces benefits in health and welfare throughout life… etc

          Cadmium reduced the cost of shipping things by reducing the rate of corrosion in seawater environments, thereby reducing transportation costs, increasing access to imported foods, vegetables, medicines… whatever.

          I’d argue that given how healthy we are now relative to 200 years ago, spewing lead and mercury and soforth into the environment probably had huge huge huge net benefits. Still, if we could have kept it out of the environment and still achieved most of those benefits, it would have been great.

          So, just as “creating efficient and cheap engines without spewing lead into the air” would have been a good idea, we have to evaluate the possibility that “using lithium to reduce suicides without spewing it in the drinking water” would be a good idea too.

        • How about something like chronic Vitamin D hypervitaminosis? Does that count?

          PS. Is Lithium even classified as a nutrient currently?

    • “Even a tiny, innocuous side effect would have to contend with the 100,000-to-12 amplification factor.”

      That’s the annual suicide rate in the US, averaged across sex and age (there are large age and sex differences in this rate; there are also large cross-national differences). The fraction of the global population that dies by suicide is, if I recall correctly, on the order of 1%. Also, the annual rate of hospitalizations from suicide attempts can be as much as 10 to 100 times higher than the completion rate, depending on sex and age.

      • The variation across cohorts is a great reason why not to spike water supplies generically: you get zero targeted intervention of at-risk groups.

        A global suicide rate of 1% sounds terribly high to me. Have a cite? Maybe I am misreading the numbers.

        S. Korea has one of the highest incidences of suicide but even there it looks like two orders of magnitude lower than your “fraction of the global population that dies by suicide is…. on the order of 1%” statistic.

        • The 1% figure is not the annual global rate of suicide. It is the (very approximate) lifetime probability of dying by suicide. You can see that it is of the right order of magnitude because the annual rate is about .01%, and a lifetime is (roughly) 100 years (all back of the envelope).

          By the way, I am not arguing for or against the proposal.

  4. I admire your zest, @gwern, but I have doubts stemming, partly, from my own experience with psychotropic medication. When a drug reduces a person’s suicidality and propensity for violent behavior, this is a psychotropic effect. The person will be mentally changed, and will notice that change from the unmedicated state. To get such an effect in drug therapy, you need doses of lithium thousands of times higher than the naturally occurring levels in the drinking water. And you need to take this dose every day. I simply don’t see how the radically lower exposure you get from drinking water could have any psychotropic effect, let alone a psychotropic effect strong enough to influence someone’s tendency to commit suicide or act violently towards others. In other words, my prior on this would be highly informative. (If I were polemically inclined, I might say you’re advocating homeopathic doses of lithium.)

  5. jrc: Yeah but then you use the natural variation in Lithium already present, you don’t manipulate it.

    Gwern: I have no idea what you mean here. A researcher randomly manipulating lithium levels within natural ranges is causally the same thing, but statistically it’s totally different: we can infer causality from one but not the other, even though the actual fact of causality is the same in both.

    jrc (new): Let me google “Natural Experiment” for you…. ahhh, here we go, from Wiki:

    A natural experiment is an empirical study in which individuals (or clusters of individuals) exposed to the experimental and control conditions are determined by nature or by other factors outside the control of the investigators, yet the process governing the exposures arguably resembles random assignment. Thus, natural experiments are observational studies and are not controlled in the traditional sense of a randomized experiment. Natural experiments are most useful when there has been a clearly defined exposure involving a well defined subpopulation (and the absence of exposure in a similar subpopulation) such that changes in outcomes may be plausibly attributed to the exposure. In this sense the difference between a natural experiment and a non-experimental observational study is that the former includes a comparison of conditions that pave the way for causal inference, while the latter does not. Natural experiments are employed as study designs when controlled experimentation is extremely difficult to implement or unethical, such as in several research areas addressed by epidemiology…

    https://en.wikipedia.org/wiki/Natural_experiment

    • I think this is a reason to consider the existing observational studies on lithium to be “pretty good” studies. That is, few people are making life choices based on the trace content of lithium. So, unlike trying to determine whether chocolate reduces heart disease risk, and having to contend with the fact that people self-select into “consuming chocolate” very clearly, so you have to contend with the possibility “people who have some unknown trait X both prefer to eat a lot of chocolate, and have lower heart disease risk” the situation with trace lithium in the water isn’t so clear.

      However, there is significant spatial variability in lithium content. See the study map:

      http://bjp.rcpsych.org/content/bjprcpsych/suppl/2011/04/11/198.5.346.DC1/bjp_198_ds346.pdf

      So it might be a good idea to look at differences in nearby pairs of districts, where the trace content is dramatically different, because the farther two districts are apart, the more likely some other factors vary which are more important. For example:

      Allergies? (which might affect general health negatively and lead to higher suicide rates)
      Sun exposure? (which has known effects such as seasonal affective disorder, and vit D deficiency)
      Economic conditions? (which is a well known suicide risk factor)
      General societal differences? (ie. different primary religions, different legal systems..)

      It’s even very likely that within Austria you’ll see this kind of variation. More vs less forested areas (sunlight, allergies), more vs less mountainous areas (altitude, allergies, sunlight), more vs less employment … etc

      And, key here, is that it’s not plausible that the natural variation in trace lithium balances all those other confounds through “random like” assignment. It is in fact totally plausible that different elevations, different forested areas, and different economic conditions affect the soil lithium content, the water lithium content, and the type of equipment installed in water distribution systems…

      So, it’s pretty good, but it’d be better to try to pick a few districts where the differences in lithium were large but they are right next to each other, and see if you could come up with a better estimate of the effect than the regression that was done there.

      • Daniel,

        Yes. I totally agree. I was not suggesting that you have to use ALL of the natural variation, just that there is sufficient natural variation that it can be harnessed in useful ways, over long periods of time, and that any massive-scale experimenting on the mental functioning of the entire populace in the name of Science is totally unnecessary (and this is one of the reasons it is unethical, though there are others).

        In fact, I would argue that the first (as far as I know) natural experiment in the history of science is a pretty good guide here for how to think about these things, and it kills me that people would advocate for massive experimentation on people’s mental states at the population level before they have even begun to scratch the surface of what can be done in the absence of experimentation. So let me leave you with John Snow (circa 1854, quoted by David Freedman:

        “The pipes of each Company go down all the streets…A few houses are supplied by one Company and a few by the other…Each Company supplies both rich and poor, both large houses and small; there is no difference either in the condition or occupation of the persons…
        The Experiment too was on the grandest scale. No fewer than three hundred thousand people ..of every age…every rank and station…were divided into two groups without their choice….To turn this grand experiment to account, all that was required was to learn the supply of water to each individual house…”

        http://www.math.rochester.edu/people/faculty/cmlr/Advice-Files/Freedman-Shoe-Leather.pdf

  6. @Daniel Lakeland & others: why do you think clinicians give lithium in the doses they do? Why is there a known therapeutic dose range? Do you think no one would have noticed if the drug were effective at even just half the standard dose? Or as little as a tenth of it? How likely is it that a thousandth or less of the standard dose would have any measurable effect on sth like suicidality and violent behavior?

  7. Here’s a 45 year old article from “Time” about lithium in El Paso’s water supply:

    “The Texas Tranquilizer”
    October 4, 1971:

    By legend Texans are a grandiose breed with more than the natural share of megalomaniacs. But University of Texas Biochemist Earl B. Dawson thinks that he detects an uncommon pocket of psychological adjustment around El Paso. The reason, says Dawson, lies in the deep wells from which the city draws its water supply.

    According to Dawson’s studies of urine samples from 3,000 Texans, El Paso’s water is heavily laced with lithium, a tranquilizing chemical widely used in the treatment of manic depression and other psychiatric disorders. He notes that Dallas, which has low lithium levels because it draws its water from surface supplies, has “about seven times more admissions to state mental hospitals than El Paso.” But state mental health officials point out that the mental hospital closest to Dallas is 35 miles from the city, while the one nearest El Paso is 350 miles away—and the long distance could affect admission figures.

    But FBI statistics show that while Dallas had 5,970 known crimes per 100,000 population last year, El Paso had 2,889 per 100,000. Dallas (pop. 844,000) had 242 murders, El Paso (pop. 323,000) only 13. Dr. Frederick Goodwin, an expert on lithium studies for the National Institute of Mental Health, doubts that “lithium has these magical properties in the population.” Others are not so sure. If lithium does have anything to do with the relative peace in El Paso, what would it do for other cities like New York and Chicago?

Leave a Reply to gwern Cancel reply

Your email address will not be published. Required fields are marked *