induction etiketine sahip kayıtlar gösteriliyor. Tüm kayıtları göster
induction etiketine sahip kayıtlar gösteriliyor. Tüm kayıtları göster

The Blind Men and the Elephant -- a post-modern parable

It's an ancient parable; a group of blind men are lead to an elephant and asked to describe what they feel.  One feels a tusk, another a foot, a third the tail, and so on, and of course they disagree entirely about what it is they are feeling. This tale is usually used as an illustration of the subjectivity of our view of reality, but I think it's more than that.

I heard a talk by Anthropologist Agustin Fuentes here at Penn State the other day, on blurring the boundaries between science and the humanities.  He used the parable to illustrate why science needs the humanities and vice versa; each restricted view of the world is enhanced by the other to become complete.

But, this assumes that the tales that science tells, and the tales that the humanities tell are separate but equally true -- scientists feel the tail, humanities feel the tusk and accurately report what they feel.  Once they listen to each other's tales, they can describe the whole elephant.

"Blind monks examining an elephant" by Hanabusa Itchō (Wikipedia)

But I don't think so.  I don't think that all that scientists are missing is a humanities perspective, and vice versa.  I think in a very real sense we're all blind all of the time, and there's no way to know what we're missing and when it matters.  You feel the tusk, and you might be able to describe it, but you have no clue what it's made of.  Or, you feel the tail but you have no idea what the elephant uses it for, if anything.

Here's my own personal version of the same parable -- some years ago we purchased a new landline with answering machine.  Oddly, we have a lot of power outages here, and it seemed that every time I set the time and day on the answering machine, we'd have another outage and the time and day would disappear, having to be set once again.  I decided that was a nuisance, and I stopped setting time and day.

The next time the machine said we had a message, I listened to it, but it was blank. There was no message!  Naturally enough (I thought), I concluded that the time and day had to be set for the machine to record a message.  Unhappy consumers, we contacted the maker, and they said no, the machine should record the message anyway.  Which of course it would have if the caller had left a message, as was proven the next time someone called on unknown day at unknown time and ... left a message.

My conclusion was reasonable enough for the data I had, right? It just happened not to be based on adequate data (aka reality).  But, we always think we've got enough data to draw a conclusion, no matter how much we're in fact missing.  This is true in epidemiology, genetics, medical testing, the humanities, interpersonal relationships; we think we know enough about our partner to commit to marrying him or her, but half of us turn out to be wrong.  Indeed, if all you've seen are white swans, you'll conclude that all swans are white -- until you see your first black one.

No, you say, we did power tests and we know we've got enough subjects to conclude that gene X causes disease Y.  But, it's possible that all your subjects are from western Europe, or even better, England, say, and what you've done is identify a gene everyone shares because they share a demographic history.  You won't know that until you look at people with the same disease from a different part of the world -- until you collect more data.  Until you see your first black swan.

But, you say, no one would make such an elementary mistake now -- you've drawn you controls from the same population, and they will share the same population-specific allele, so differences between cases and controls will be disease-specific.  But, western Europe is a big area, and even England is heterogeneous, and it's possible that everyone with your disease is more closely related than people without.  So, you really might have identified population structure rather than a disease allele but you can't know, until you collect more data -- you look at additional populations, or more people in the same population.

Even then, say you look at additional populations and you don't find the same supposedly causal allele.  You can't know why -- is it causal in one population and not another?  Is it not causal in any population, and your initial finding merely an artifact of ill-conceived study design?

Without belaboring this particular example any further, I hope the point is clear.  You feel the tail, but that doesn't tell you everything about the tail.  But you can't know what you're missing until you ask more questions, and gather more data.

Darwin explained inheritance with his idea of gemmules.  He was wrong, of course, but he had no way to know how or why, and it wasn't until Mendel's work was rediscovered in 1900 that people could move on.  Everything we know about genetics we've learned since then, but that doesn't mean we know everything about genetics.  But theories of inheritance (and much else) don't include acknowledgement of glaring holes: "My theory is obviously inadequate because, as always, there is a lot we don't yet understand but we don't know what that is so I'm leaving gaps, but I don't know how big or how many."  And, in a related issue that we write about frequently here, it's also true that instead of coming clean, we often claim more than we know (and often we know what we're doing in doing so).

Even very sophisticated theories just 15 or 20 years ago had no way to include, say, epigenetics, or the importance of transcribed but untranslated RNAs (that is, RNA not coding for genes but doing a variety of other things, some of them still unknown), or interfering RNAs, and so on, and we have no idea today what we'll learn tomorrow.  But, like the blind men, we act as though we can draw adequate conclusions from the data we've got.

Science is about pushing into the unknown.  But, because it's unknown, we have no idea how far we need to push.  I think in most cases, there's always further, we're never done, but we often labor under the illusion that we are.  Or, that we're close.

But, should ductal cancer in situ, a form of breast cancer, be treated?  And how will we know for sure?  Systems biology sounds like a great idea, but how will we ever know we've taken enough of a given system into account to explain what we're trying to explain?  Will physicists ever know whether the multiverse, or the symmetry theory is correct (whatever those elusive ideas actually mean!)?

Phlogiston was once real, as were miasma and phrenology, the four humors, and the health benefits of smoking.  It's not that we don't make progress -- we do now know that smoking is bad for our health (even if only 10% of smokers get lung cancer; ok, smoking is associated with a lot of other diseases as well, so better not to smoke) -- but we've always got the modern equivalent of phlogiston and phrenology.  We just don't know which they are.  We're still groping the elephant in the dark.

How do we know what we think we know?

Two stories collided yesterday to make me wonder, yet again, how we know what we think we know.  The first was from the latest BBC Radio 4 program The Inquiry, an episode called "Can we learn to live with nuclear power?" which discusses the repercussions of the 2011 disaster in the Fukushima nuclear power plant in Japan. It seems that some of us can live with nuclear power and some of us can't, even when we're looking at the same events and the same facts.  So, for example, Germans were convinced by the disaster that nuclear power isn't reliably safe and so they are abandoning it, but in France, nuclear power is still an acceptable option.  Indeed most of the electricity in France comes from nuclear power.

Why didn't the disaster convince everyone that nuclear power is unsafe?  Indeed, some saw the fact that there were no confirmed deaths attributable to the disaster as proof that nuclear power is safe, while others saw the whole event as confirmation that nuclear power is a disaster waiting to happen.  According to The Inquiry, a nation's history has a lot to do with how it reads the facts.  Germany's history is one of division and war, and nuclear power associated with bombs, but French researchers and engineers have long been involved in the development of nuclear power, so there's a certain amount of national pride in this form of energy.  It may not be an unrelated point that therefore many people in France have vested interests in nuclear power.  Still, same picture, different reading of it.

Cattenom nuclear power plant, France; Wikipedia


Reading ability is entirely genetic
And, I was alerted to yet another paper reporting that intelligence is genetic (h/t Mel Bartley); this time it's reading ability, for which no environmental effect was found (or acknowledged).  (This idea of little to no environmental effect is an interesting one, though, given that the authors, who are Dutch, report that heritability of dyslexia and reading fluency is higher among Dutch readers -- 80% compared with 45-70% elsewhere -- they suggest because Dutch orthography is simpler than that of English.  This sounds like an environmental effect to me.)

The authors assessed reading scores for twins, parents and siblings, and used these to evaluate additive and non-additive genetic effects, and family environmental factors.  As far as I can tell, subjects were asked to read aloud from a list of Dutch words, and the number they read correctly within a minute constituted their score.  And again, as far as I can tell, they did not test for nor select for children or parents with dyslexia, but they seem to be reporting results as though they apply to dyslexia.

The authors report a high correlation in reading ability between monozygotic twins, a lower correlation between dizygotic twins, and between twins and siblings, and a higher correlation between spouses, which to the authors is evidence of assortative mating (choice of mate based on traits associated with reading ability).  They conclude:
Such a pattern of correlation among family members is consistent with a model that attributes resemblance to additive genetic factors, these are the factors that contribute to resemblance among all biological relatives, and to non-additive genetic factors. Non-additive genetic factors, or genetic dominance, contributes to resemblance among siblings, but not to the resemblance of parents and offspring.  Maximum likelihood estimates for the additive genetic factors were 28% (CI: 0–43%) and for dominant genetic factors 36% (CI: 18–65%), resulting in a broad-sense heritability estimate of 64%. The remainder of the variance is attributed to unique environmental factors and measurement error (35%, CI: 29–44%).
Despite this evidence for environmental effect (right?), the authors conclude, "Our results suggest that the precursors for reading disability observed in familial risk studies are caused by genetic, not environ- mental, liability from parents. That is, having family risk does not reflect experiencing a less favorable literacy environment, but receiving less favorable genetic variants."

The ideas about additivity are technical and subtle.  Dominant effects, that is, non-additive interactions among alleles within a gene in the diploid copies of an individual, are not inherited as additive ones are (if you are a Dd and that determines your trait, only one of those alleles, and hence not enough to determine the trait, is transmitted to any of your offspring).  Likewise, interactions (between loci), called epistasis, is also not directly transmitted.

There are many practical as well as political reasons to believe that interactions can be ignored.  In a practical sense, even multiple 2-way interactions make impossible sample size and structure demands.  But in a political sense, additive effects mean that traits can be reliably predicted from genotype data (meaning, even at birth): you estimate the effects of each allele at each place in the genome, and add them to get the predicted phenotype.  There is money to be made by that, so to speak.  But it doesn't really work with complex interactions.  Strong incentives, indeed, to report additive effects and very understandable!

Secondly, all these various effects are estimated from samples, not derived from basic theory about molecular-level physiology, and often they are hardly informed by the latter at all.  This means that replication is not to be expected in any rigorous sense.  For example, dominance is estimated by the deviation of average traits in AA, Aa, and aa individuals from being in 0, 1, 2 proportions if (say) the 'a' allele contributed 1-unit of trait measure.  Dominance deviations are thoroughly sample-dependent.  It is not easy to interpret those results when samples cannot be replicated (the concepts are very useful in agricultural and experimental breeding contexts, but far less so in natural human populations). And this conveniently overlooks the environmental effects.

This study is of a small sample, especially since for many traits it now seems de rigueur to have samples of hundreds of thousands to get reliable mapping results, not to mention a confusingly defined trait, so it's difficult, at least for me, to make sense of the results.  In theory, it wouldn't be terribly surprising to find a genetic component to risk of reading disability, but it would be surprising, particularly since disability is defined only by test score in this study, if none of that ability was  substantially affected by environment.  In the extreme, if a child hasn't been to school or otherwise learned to read, that inability would be largely determined by environmental factors, right?  Even if an entire family couldn't read, it's not possible to know whether it's because no one ever had the chance to learn, or they share some genetic risk allele.

In people, unlike in other animals, assortative mating has a huge cultural component, so, again, it wouldn't be surprising if two illiterate adults married, or if they then had no books in the house, and didn't teach their children that reading was valuable.  But this doesn't mean either reading or their mate-choice necessarily has any genetic component.  

So, again, same data, different interpretations  
But why?  Indeed, what makes some Americans hear Donald Trump and resonate with his message, while others cringe?  Why do we need 9 Supreme Court justices if the idea is that evidence for determination of the constitutionality of a law is to be found in the Constitution?  Why doesn't just one justice suffice?  And, why do they look at the same evidence and reliably and predictably vote along political lines?

Or, more uncomfortably for scientists, why did some people consider it good news when it was announced that only 34% of replicated psychology experiments agreed with the original results, while others considered this unfortunate?  Again, same facts, different conclusions.

Why do our beliefs determine our opinions, even in science, which is supposed to be based on the scientific method, and sober, unbiased assessment of the data?  Statistics, like anything, can be manipulated, but done properly they at least don't lie.  But, is IQ real or isn't it?  Are behavioral traits genetically determined or aren't they?  Have genome wide association studies been successful or not?

As Ken often writes, much of how we view these things is certainly determined by vested interest and careerism, not to mention the emotional positions we inevitably take on human affairs.  If your lab spends its time and money on GWAS, you're more likely to see them as successful.  That's undeniable if you are candid.  But, I think it's more than that.  I think we're too often prisoners of induction, based on our experience, training, predilections of what observations we make or count as significant; our conclusions are often underdetermined, but we don't know it.  Underdetermined systems are those that are accounted for with not enough evidence.  It's the all-swans-are-white problem; they're all white until we see a black one. At which point we either conclude we were wrong, or give the black swan a different species name.  But, we never know if or when we're going to see a black one.  Or a purple one.

John Snow determined to his own satisfaction during the cholera epidemic in London in 1854 that cholera was transmitted by a contagion in the water.  But in fact he didn't prove it.  The miasmatists, who believed cholera was caused by bad air, had stacks of evidence of their own -- e.g., infection was more common in smoggy, smelly cities, and in fact in the dirtier sections of cities.  But both Snow and the miasmatists had only circumstantial evidence, correlations, not enough data to definitively prove their were right.  Both arguments were underdetermined.  As it happened, John Snow was right, but that wasn't to be widely known for another few decades when vibrio cholerae was identified under Robert Koch's microscope.

"The scent lies strong here; do you see anything?"; Wikipedia

Both sides strongly (emotionally!) believed they were right, believed they had the evidence to support their argument. They weren't cherry-picking the data to better support their side, they were looking at the same data and drawing different conclusions.  They based their conclusions on the data they had, but they had no idea it wasn't enough.  

But it's not just that, either.  It's also that we're predisposed by our beliefs to form our opinions.  And that's when we're likely to cherry pick the evidence that supports our beliefs.  Who's right about immigrants to the US, Donald Trump or Bernie Sanders?  Who's right about whether corporations are people or not?  Who's right about genetically modified organisms?  Or climate change?  Who's right about behavior and genetic determinism?  

And it's even more than that! If genetics and evolutionary biology have taught us anything, they've taught us about complexity.  Even simple traits turn out to be complex.  There are multiple pathways to most traits, most traits are due to interacting polygenes and environmental factors, and so on. Simple explanations are less likely to be correct than explanations that acknowledge complexity, and that's because evolution doesn't follow rules, except that what works works, and to an important degree that's what is here to be examined today.  

Simplistic explanations are probably wrong.   But they are so appealing. 

Rare Disease Day and the promises of personalized medicine

O ur daughter Ellen wrote the post that I republish below 3 years ago, and we've reposted it in commemoration of Rare Disease Day, Febru...