Famous schizophrenia researcher E. Fuller Torrey recently wrote a paper trying to cast doubt on whether schizophrenia is really genetic. His exact argument is complicated, but I feel like it sort of equivocates between “the studies showing that schizophrenia are genetic are wrong” and “the studies are right, but in a philosophical sense we shouldn’t describe it as ‘mostly genetic’”.

Awais Aftab makes a clearer version of the philosophical argument. He’s not especially interested in debating the studies. But he says that even if the studies are right and schizophrenia is 80% heritable, we shouldn’t call it a genetic disease. He says:

Heritability is “biologically vacuous” (Matthews & Turkheimer, 2022), and I think we would be better off if more of us hesitated to assert that schizophrenia is a “genetic disorder” based predominantly on heritability estimates.

I think about questions like these through the lens of avoiding isolated demands for rigor. There are always complicated ways that any statement is false. So the question is never whether a statement is perfectly true in every sense. It’s what happens when we treat it fairly, using the same normal criteria we use for everything else.

For example, we commonly use language like “smoking causes lung cancer”. So when I ask “do genes cause schizophrenia?”, I’m not asking whether this is so perfectly and platonically true that no philosopher could ever nitpick any of its implications. I mean - do genes cause schizophrenia in the normal sense of causation that smoking causes lung cancer?

Aftab makes several arguments against causal language in his post, some others in this discussion in the comments, and there are a few more in the original Torrey paper. Let’s go through them, apply the smoking standard, and see how they fare.

Argument 1: Genes are a risk factor, not a cause.

I challenge this distinction. We say that smoking is a risk factor for lung cancer, but also a cause of lung cancer.

Why might a risk factor / cause distinction be tempting? In infectious disease, we distinguish between the cause of flu (the influenza virus) and various risk factors for flu (eg age).

Even though this distinction seems intuitively obvious, I can’t find a good definition anywhere! Some sources say risk factors are merely correlational whereas causes are causal. But age is certainly more than correlational with flu (if we magically de-aged someone, they would get the flu less).

Other sources use a necessary-and-sufficient standard: flu virus is necessary to get the flu, and it’s sufficient even if you don’t have any other risk factors. In comparison, you can get the flu without necessarily being old, and age isn’t sufficient to give you the flu unless there’s a flu virus going around.

But this example is rigged. Flu-like symptoms are a common feature of many conditions (eg SSRI withdrawal), but we define flu as “flu-like symptoms that you get from the influenza virus”. So the seemingly-necessary virus → flu causation only works by definitional fiat. Nor is flu virus sufficient. Lots of people get flu virus but never develop symptoms, maybe because their immune system is good enough to fight it off. This, too, is definitional: we define someone as “having the flu” if they have the virus, even when they don’t have symptoms.

Here it helps to remember Symptom, Condition, Cause. The flu virus (cause) can cause respiratory inflammation (condition). We could imagine someone who had lots of respiratory issues just because they were very old and decrepit, in which case we might call age (cause) the cause of their respiratory inflammation (condition).

So the flu (cause) vs. age (risk factor) distinction only makes sense in the context of definitional fiat where we’re talking about respiratory-inflammation-caused-by-the-flu-virus as a natural category.

Lung cancer is a condition with many possible causes - smoking, radiation, genetics, bad luck. Of these, smoking is the most common. So we colloquially say “smoking causes lung cancer”.

Schizophrenia is also a condition with many possible causes, and I think it makes sense to colloquially say “genetics causes schizophrenia”.

Argument 2: Statistics about variance explained by heritability in general don’t necessarily apply to individuals

You also can’t point to any individual lung cancer patient and say “smoking caused this person’s lung cancer”. Even if they’re a smoker, they might have gotten the cancer through bad luck, such that they would have still gotten it even if they didn’t smoke.

Still, on a society-wide level we know that smoking causes most lung cancer, so it’s still okay to say “lung cancer is mostly caused by smoking”.

In the same way, we can never say for sure that genetics caused any specific case of schizophrenia1, but on a society-wide level we can still say “schizophrenia is mostly genetic”.

Argument 3: The genes that cause schizophrenia aren’t specific to schizophrenia.

Smoking isn’t specific to lung cancer. It can also cause throat cancer, stroke, heart attacks, COPD, etc. Still, since it’s the main cause of lung cancer, we still say “smoking is the main cause of lung cancer”.

In the same way, even though schizophrenia genes also cause other things, we can still say that schizophrenia is mostly genetic2.

Argument 4: No particular gene contributes very much to schizophrenia

No particular cigarette you smoke contributes very much to your risk of lung cancer, but smoking 10,000 cigarettes does. This is just the sorites paradox.

Argument 5: Variance in heredity is only meaningful relative to particular societies.

The famous example: if you plant a lot of seeds in poor soil, all the variance in how high they grow will be genetic. If you plant a lot of seeds in good soil, all the variance will still be genetic. But if you plant them scattered across soils of various qualities, much of the variance will be environmental.

But this is true of everything. In a society where everyone had magical cancer-preventing nanobots, smoking wouldn’t cause lung cancer. In a society where everyone inhaled radioactive dust all day, smoking would be an utterly insignificant cause of lung cancer compared to all the radioactive dust inhalation. In a society where no one smoked, 0% of lung cancer would be explained by smoking status.

And also, schizophrenia is a terrible condition to make this argument for. It is famously nearly the same prevalence in every society studied. And studies from 1970s China find the same heritability as Western studies.

Although one can imagine some super-exotic society where the impact of environmental factors on schizophrenia outweighs genetic factors, I’m not sure this is any more interesting than hypothetical cancer-preventing nanobots that could eliminate the effect of smoking. I think it’s reasonable to discuss the causes of diseases relative to real societies that exist on Earth.

Argument 6A: Heritability doesn’t tell us about causal processes

“Smoking causes lung cancer” doesn’t tell us about causal processes. You have to do difficult research work and expand it out into some specific chain, like “the tar in tobacco causes DNA to mutate in such-and-such a way”. Still, we were happy to say “smoking causes lung cancer” even before we knew the causal process.

Even “A gene causes cystic fibrosis” doesn’t immediately reveal causal processes after the gene. You have to do the work and find that the gene affects the consistency of lung mucus. You could call it a lung mucus disease if you wanted to. But you could also call it a genetic disease. Everyone calls it both of these things, and they’re not in conflict.

In the same way, “genes cause schizophrenia” doesn’t reveal causal processes after the genes. But if it’s true, we’ll one day learn something like “this particular gene makes your glutamate receptors 0.0001% less efficient, and you need efficient glutamate receptors for various brain processes, so if you combine that with 9,999 other genes, your brain processes stop working well.

Argument 6B: Genes could just be a proxy for some more satisfying cause of schizophrenia

The argument here is: imagine schizophrenia is caused by some extremely common virus, that everybody has all the time. In most people, the immune system controls the virus, and it has no effect. But in some people, with genes for bad immune systems, the virus gets out of control and causes schizophrenia. Here we would say the virus causes schizophrenia3. But it would show up in twin studies as 100% genetic, because all the differences in who gets schizophrenia and who doesn’t are caused by genes.

This is possible, but I think it’s false. We know many apparent risk factors for schizophrenia: cannabis use, birth canal asphyxia, social defeat, toxoplasma infection, poor prenatal nutrition. If you try to combine these into one big picture, it doesn’t work.

The only thing that makes sense to me (and I’m definitely not an expert, and E. Fuller Torrey is, so take this with a grain of salt) is thinking of schizophrenia as cumulative damage to some abstract computational organ. Consider an analogy to kidney disease. The cause of kidney disease is “anything that hurts your kidney”. It can be caused by kidney infections, kidney cancer, trauma to the kidney, autoimmune diseases affecting the kidney, etc. Kidney disease is just a catch-all term for “anything causing your kidney to work less well than it should”. If you examined its heritability, it would probably be substantially genetic. Some of these genes would be for bad kidneys. Others would be weirder: probably a gene for risk-seeking impulsive behavior would show up, since that increases your chance of trauma to the kidney.

Schizophrenia is less comprehensible than kidney disease; all kidney disease causes meet in one clear nexus (the kidney, an easily visible physical organ), whereas it’s unclear what all schizophrenia causes have in common. I think when we know more, we’ll find that there’s some sense in which they all damage the same computational system in the brain. There won’t be a more satisfying cause like a virus. And when we figure out what’s going on with the genes, we’ll find that they’re genes involved in that same computational system, or for things that disturb it.

(see also Ontology Of Psychiatric Conditions: Dynamical Systems for more on how I think about these kinds of issues)

Again, I don’t think this is too different from smoking. You can imagine that smoking is just a proxy for a more satisfying cause of lung cancer. For example, smoking depresses immune function, letting a virus get in and cause the cancer. But in fact we’ve studied this pathway and our best guess is that this isn’t true. So we just say “smoking causes lung cancer”.

Argument 7: Heritability numbers rely on assumptions that can be violated

All numbers are like this.

You can never 100% prove a number with no assumptions. You can just check as hard as you can, try to falsify it, and find that it’s stood the test of time.

So for example, Dr. Aftab points out that twin studies could break down if a small difference in the intrauterine environment could have vast effects on future outcomes. But when people have tried to investigate this, they mostly find that it doesn’t, aside from a couple of perinatal outcomes you’d expect. Or: twin studies could break down if identical twins have much more correlated environments than fraternal twins. But when people try to investigate this, they mostly find that they don’t.

The results of twin studies have been mostly corroborated by adoption studies, a separate methodology. A third methodology, relatedness disequilibrium studies, gets slightly different numbers, but not enough to change the predominantly genetic nature of schizophrenia (also, some people here object to them; I’m not smart enough to follow this conversation but would love if someone explained it to me).

There are also reasons to expect twin studies to underestimate heredity; for example, any measurement error shows up as environmental variation. If a construct is hard to measure (schizophrenia is so famously hard to define that we’re not even sure it exists) then twin studies will underestimate its heredity. And “heredity” probably underestimates the level of genetic effect, since some of the variants causing schizophrenia might be mutations during embryonic development.

So the question isn’t whether you can think of a possible bias that a study can have, but whether you honestly believe the bias is big enough that you should discount the study - and if so, what prior you retreat back to. My impression is that the biases in twin studies are relatively small, and that the prior from other studies is similar enough that you can’t justify retreating to a more environmentalist position in particular.

To keep to our theme, did you know that the phrase “correlation is not causation” was popularized by pro-tobacco statisticians to prove that the original smoking-causes-cancer studies might be biased?

Argument 8: Heritability numbers produces counterintuitive results about risk at an individual level.

We discussed this in the context of twin pairs earlier. Instead of rejecting heritability numbers, I think people should get better intuitions.

Also, this is equally true of everything else. Even though smoking causes 80% of lung cancer, if Alice gets lung cancer and Bob smokes exactly the same as Alice, Bob will have a much lower than 80% chance of getting lung cancer. The solution isn’t to stop describing smoking as causing lung cancer, it’s to get better intuitions.

Argument 9: Epilepsy is also ~80% heritable, but we don’t call epilepsy genetic

Fine. Epilepsy is also genetic.

Why am I insisting on this so hard?

I think if E. Fuller Torrey had discovered that something fun and interesting like toxoplasma or social defeat explained 80% of the variance in schizophrenia, everyone would say “Oh! That causes schizophrenia!” and forget all the nitpicking. This would happen even though toxoplasma can cause other things, even though it might not explain the exact causal pathway by which toxoplasma causes schizophrenia, etc. I think people really want things not to be genetic, so when they do turn out to be genetic, they apply higher standards for whether you can call that “the cause”. Then people underestimate how much genes matter.

This is well-intentioned: people want to fight back against a disease, so they want to exhaust all hope of finding environmental causes (which they think they can change) before giving up and attributing it to genetic causes (which they think they can’t). But as we often discuss here, this is backwards - society is fixed and biology is mutable.

A story: once I was seeing a young man who became psychotic for a few days every time he used cannabis. Along with medical treatment, I gave him the obvious suggestion: stop using cannabis! This was both because even a few days of psychosis can be pretty bad, and because I was worried he was at high risk for schizophrenia and cannabis might eventually push him permanently over the edge. His parents attended the appointments and very strongly reinforced the “don’t use cannabis” message. Every few months, he would use cannabis, become briefly psychotic, and need me to help get him out of it. Every few months, I would use all the tools I had - contracts, motivational interviewing, tearful lectures from his parents, etc - to try to convince him not to use cannabis. Every few months, he would swear he definitely wouldn’t use it again. Every few months, he would come back, psychotic again, after using more cannabis. Until one time it pushed him over the edge and he became schizophrenic and as far as I know he still is.

The other environmental risk factors for schizophrenia are equally hard to change. Poverty? Okay, don’t be poor, thanks for the important life advice. Social defeat? “Doctor, are you saying I have to never let anyone defeat me?” “Yes, it’s my official medical recommendation that you become invincible.” The only thing in this category I’m really excited about is fish oil supplementation, and even that might or might not replicate.

Even this is overestimating the tractability, because half of the environmental variance in schizophrenia is non-shared, which usually means measurement error and embryological random noise we can’t control. The other half - about 10% of total variance - probably is interesting environmental stuff we can control, but at this point the crumbs are so small that they’re hard for studies to pick up and they’re all just guesses. If cannabis really is involved in schizophrenia, it’s probably 1-3% of the variance - but nobody is entirely sure it’s involved (1, 2), because picking up things that cause 1-3% of variance in a noisy world is really hard. It’s not just that social factors explain a small percent of the variance and are hard to change, it’s that we only have tentative guesses about what they are.

By comparison, you can very clearly halve your children’s risk of schizophrenia through polygenic selection, which costs only a few hundred dollars if you’re already doing IVF. You don’t need to worry about whether your teenager will ignore your recommendation not to use marijuana, you don’t have to fiddle around with shaving a few points off the variance, and you don’t have to worry that you’re chasing correlational phantoms. Just pay a few hundred bucks and you’re done. And polygenic screening gets better every year. In a decade or two you can probably eliminate the risk entirely.

Still, if you look at the resources on how to avoid schizophrenia, the ones doctors are supposed to give people from high-risk families when they’re considering having kids, they never mention polygenic screening. It’s all just “don’t do drugs” and “avoid getting socially defeated”.

It’s even worse than that, because people keep trying to sabotage polygenic screening! The psychiatric genetics teams are trying to prevent screening companies from using their data! Sometimes it’s because this completely voluntary process vaguely reminds of them of eugenics. Other times it’s because they somehow try to pretend the amount of variance involved doesn’t matter or isn’t worth it, even though it’s a million times more than the drug abuse and social defeat issues people constantly obsess over. But other times it’s even weirder - a bioethicist in this article and a geneticist in this one both say variants of “health care should be about treating schizophrenia, not preventing it”. This is both totally antithetical to the spirit of real clinical medicine as it’s practiced, and ethically odious to anyone who has witnessed the side effects of antipsychotics first-hand. I cannot wait until this kind of thinking ends up in the shameful dustbin of history, alongside all those 18th-century people who tried to ban vaccines because diseases were a divine punishment and humans shouldn’t interfere.

And one step to getting them in that dustbin is acknowledging that this mostly genetic disease is, in fact, mostly genetic.

EDIT AFTER THINKING ABOUT THIS MORE: One way to think about this is that if we knew for sure that all of the genes involved in schizophrenia were for an autoimmune condition, we might say “schizophrenia is an autoimmune disease”. If we knew for sure that all the genes involved in schizophrenia involved susceptibility to a certain virus, we might say “schizophrenia is a viral disease”. In both of these cases, it would still be genetic in some sense, but we would be more interested in the downstream mediator. But if all of the genes in schizophrenia exert their effect in different ways with different mediators, and there’s no story we can tell beyond the accumulated damage analogy above or the dynamical systems approach, maybe we might as well call it genetic, since the genome is the furthest-downstream place where it has a clear and comprehensible unified cause. I think this is the most likely outcome, but Torrey and others disagree. I don’t know if this is the crux of our larger disagreement.

  1. Actually it’s more complicated than this: if you look at the simulation argument from last time, it’s hopefully obvious that every case is caused both by genetic and by environmental factors, or that it’s not really meaningful to try to disentangle them. I don’t know whether real schizophrenia is like this.

  2. Except the causation here is more like genes → some general high-psychopathology-risk state → schizophrenia. So the exact right analogy isn’t smoking/lung cancer, but something more like obesity/heart attacks. We know that obesity → some general metabolically dysregulated state → heart attacks, but we still colloquially say that obesity can cause heart attacks.

  3. Would we really? If schizophrenia was caused by an omnipresent virus, but the only people who got it, got it entirely because of genetic differences from the rest of the population, why would we call the virus “the cause” rather than the genes? Is it just because we’re used to calling pathogens the causes of things from our long history with infectious disease? It might be equally reasonable in this case to think of schizophrenia as a genetic immunodeficiency. I don’t know exactly how to think about things like this.