Rationality Case Study - Ad-36

by Perplexed1 min read22nd Sep 201054 comments

32

Case StudyMedicine
Personal Blog

Edit: Title change - may cause RSS resend.

There is a widespread belief here that we pretty much have epistemic rationality nailed.  Instrumental rationality, ... well, nobody is perfect.  We are still working on that.  But epistemic rationality?  That is a breeze.  Simply a matter of avoiding some well known biases, doing Bayesian updating, and maintaining a certain kind of mutual respect in discussions.  Maybe throw in a bit of Pearl, if you need to address causality in addition to correlation.  We know how to handle evidence.  And what is epistemic rationality, after all, besides the optimal evaluation of evidence?

A recent posting made the suggestion that we ought, as a community, to make our expertise available to society, by tackling important and controversial questions and (presumably) then sharing our consensus with the world.  But will we reach a consensus?  A simple test case might be useful.  The purpose of this posting is to propose a test.  I am providing some links to an ongoing scientific controversy which (judging by press coverage) is of considerable interest to the general public.  Given the attention given to health issues here, I suspect that it will also be of interest to the LW population in general.  So here goes.

There has been some suggestion that a particular strain of the common cold virus - Ad-36 - might be a major contributor to human obesity.  Become infected by the virus, and you are at increased risk of becoming fat.  However, a certain amount of skepticism has been expressed - the evidence may not be as good as it is made out to be.

I'm completely new to this controversy - just found out about it today.  I think that it would be an interesting test case to have LW as a group look into this question (quite a lot of information is available online, and what is not can be found in university libraries to which I assume many of us have access).  So, what do you all think?  Does Adenovirus strain #36 cause human obesity?  If so, how much?  If not, where exactly have the doctors and scientists who believe otherwise gone wrong in their thinking?

32

54 comments, sorted by Highlighting new comments since Today at 10:26 AM
New Comment

EDIT: I have retracted the conclusion of this comment.

Figuring this out is basically a matter of doing a literature survey, judging a few of the studies for basic sanity, and reporting the results. Being able to do this is a very important skill, and everyone should practice it at least once so that if they suddenly develop a medical condition, they know how to double check their doctor's work. So here's what I found in the literature:

So that's correlation in both controlled animal tests and observational human tests, correlation in cultured human tissue, and observed mechanisms. A few names show up in a bunch of the studies, but there are also replications by unrelated research groups. There were even more studies to sort through, but I stopped there, because that's more than enough: being infected with AD-36 makes you fatter, proven beyond a reasonable doubt. Any commenters nitpicking the statistics and sample sizes of individual studies are missing the point, because the evidence is the sum of all the studies, and they are not all flawed, or at least not flawed badly enough to seriously undermine the conclusion.

That said, there are conclusions which it is tempting to leap to, but which I could not find support for. First, it is tempting to conclude that AD-36 infection is responsible for a large fraction of the rise in obesity in the United States. However, I could not find any data on the overall prevalence of AD-36 in the US or world population, nor on factors affecting AD-36's transmissibility; the studies I found seem to involve participants drawn from narrower geographic regions, whose exposure would be correlated and therefore not representative of the US as a whole. I also couldn't find any sort of historical data to indicate that AD-36 prevalence has risen, rather than remain constant. Not that I would expect to find historical data in any case, since that's very difficult to collect for recently discovered pathogens - but there are several inferential steps still not covered. The most likely conclusion, therefore, is that AD-36 accounts for some of the rise in obesity in the US, but an unknown amount, and not all. I also saw references to several other pathogens indicated to cause obesity in humans and animals, but did not research them in depth.

I stopped there, because that's more than enough: being infected with AD-36 makes you fatter, proven beyond a reasonable doubt. Any commenters nitpicking the statistics and sample sizes of individual studies are missing the point, because the evidence is the sum of all the studies [emphasis added], and they are not all flawed, or at least not flawed badly enough to seriously undermine the conclusion.

You are the first person to actually take a firm position. I am enough of a Popper fan to applaud you, even if you end up being proven wrong. Thx.

I'm still unconvinced/neutral myself, but for the sake of argument I'll make a few criticisms of your conclusions. I wonder how you will respond.

To begin, if you are going to sum all of the studies, you should probably include the negative ones, such as this European study or this US military study. It seems to me that rationality requires that the negative evidence somehow be explained away before the positive evidence is taken at face value. I have trouble explaining away the European study, though, of course, it might be argued that a different substrain of the virus is endemic on the European continent.

Another reason for uneasiness is the fact that the positive research is coming from so few research groups, and that the leaders of some of those groups have applied for a patent - as you yourself pointed out.

And finally, maybe it shouldn't, but the fact that the most recent study presented such patently misleading statistics gives me a bad feeling about the whole hypothesis.

I actually missed those two negative studies - I only went through the first few pages of Google Scholar and PubMed results; I guess I should've been looking more closely at the preexisting discussion here. My confidence is lowered. Now for those two studies themselves -

The European one has a line in the abstract that seems to explain it: "No adenoviral DNA could be found using PCR on visceral adipose tissue." That suggests either a significantly different strain or a significantly different (and better) immune response. I haven't read the text of the military study (that one costs money), but the most likely confounding factor there would be military recruitment filtering out a lot of obese people; the child exposure test weakly suggests that the effect is largest on children whose adipose tissues are still growing, while the obese in the military are probably mostly those who started out fit and gained weight as they aged.

I have the text of the military study if you want to PM an email address.

Since I posted this, two studies that I missed with negative results have been pointed out, and it's also been pointed out that several of the studies that I thought were independent replications may actually have been subject to editorial influence by Dhurandhar and Atkinson. Additionally, serious questions have been raised about their integrity which I am not able to evaluate. Therefore, I am retracting my statement that is is proven that AD-36 causes substantial weight gain. I now assign this statement a probability of only 0.4.

Rationality as the measure of all things.

I would like to suggest a more specific title, such as "Rationality Case Study: Adenovirus strain #36" or something similar.

That way, not only do you satisfy the general principle that titles should be informative as possible, but you also avoid giving the impression that this post is purely rhetorical (with an implied thesis that you supplied earlier).

Ok with me. But since this is my first top-level posting, is there anything I need to know before I change the title?

I have heard people say that changing the title causes the post to be re-sent in the RSS feed (I wouldn't know, since I don't use RSS myself). Perhaps that bug has been fixed by now, but if not, then it is a good idea to make any title changes as soon as possible after posting.

I think that happens because if you change the title, it'll get a different title and URL sent in the RSS feed, and the reader is going to think it's a different item. There's an RSS tag, , which is meant to work around this by allowing the feed to specify a unique identifier for each item, but the LW RSS feeds use it incorrectly; it uses the post's URL as its value, and that changes when the post's title changes. I think it would work correctly if it chopped off the post title part and just sent (e.g.) "http://lesswrong.com/lw/2qm/" in the tag.

I think we've still got the bug.

There is a widespread belief here that we pretty much have epistemic rationality nailed here.

Really? I don't have that sense. Take a look at A Request for Open Problems and note that the top 4 suggestions (by comment rating) are all about epistemic rationality.

Or maybe we're so far behind on instrumental rationality we don't even know the right questions yet! Or maybe the request for "well-defined" questions creates a bias toward epistemic rationality.

We don't have epistemic rationality nailed well enough to create an oracle. So, we certainly don't have it very well "nailed".

Thx for the link. I hadn't seen that discussion.

Nikhil V Dhurandhar and Richard L Atkinson, two researchers behind much of the research on Ad-36 and obesity, filed for a patent on using AD-36 antibody presence as a predictor of obesity. The initial filing was on Sep 9, 2005, after all the research I can find by those two was published, so it does not appear to constitute a conflict of interest.

(Edit: Actually they published more after that, but the conclusion was well proven before the filing and they seem to have moved on to details and mechanisms)

The blog "Junkfood Science" had a posting a little over a year ago that was highly critical of the Ad-36 obesity hypothesis. Based on this, I'm going to shift from a neutral stance to strong skepticism.

I just don't think there is any credible evidence that this virus is a cause of human obesity.

ETA: Ca-ching!

Blog post noted - I think the criticism of Dhurandhar human correlation studies sound correct - jimrandomh cites AD-36 correlated with greater BMI and body fat mass, but not with greater insulin resistance which should be examined as well, but that's the only study left of those promoted here showing a positive effect in humans that hasn't already been subject to clear criticisms.

The paper you mention appears in the journal together with a comment by R. L. Atkinson, whose own research and interest conflict was criticized in the blog. Furthermore, the same Italian researchers who produce that paper in the liver journal, also published in Atkinson's obesity journal at about the same time, presumably based on the same data.

I haven't read either Italian paper, only the abstracts, but my impression is that the same error - failure to properly account for confounding variables, particularly age - occurs in this Italian data as occurred in the American data.

It seems to me that when you eliminate the research conducted by Atkinson and Dhurandhar, or published in the journal they control, you are left only with the negative results from Europe.

Negative results from the U. S. Navy appeared in Atkinson's journal, which does tend to alleviate my paranoia arising from the fact that most of the positive results appeared in that same journal.

The ideal test, of course, would be to piggyback on a longterm study like Framingham which presumably has blood samples and weight/BMI data on a large number of people taken every two years. Then it could be determined whether weight gain preceeded or roughly coincided in time with or followed AD-36 infection. I don't know what the cost of AD-36 testing would be. Of course, it might be that patent royalties would have to be paid to Atkinson's company, just to do the testing. :)

It may be time to wrap this case study experiment up, and do a "lessons-learned" post-mortem. RobinZ, would you care to do the honors of commenting first on this?

In my opinion, there's no good evidence that Ad-36 is associated with weight gain. Even though it appears plausible from the results observed in animals, the facts that (1) a clear conflict of interest has been shown for the persons responsible for the greatest bulk of papers on the subject, (2) the studies in humans appear to suffer from severe sampling flaws which may be expected to obscure the proposed effect, and (3) the least flawed study (so far as I can determine) is the US military study suggest that the conclusion is, at best, speculative. I don't have a proper reference class to calibrate my estimate by, but I would tentatively say that "Ad-36 causes obesity in humans" has a surprisal of 10 bits.

Edit: This estimate is intentionally higher than my prior for the raw hypothesis but still very small.

I would like to see jimrandomh discuss this before I call it settled among those interested enough to study it - I'll send him a PM.

Criticisms of the recent Gabbert et al AD36<->obesity paper:

The data indicated that for the AD36 Negative group, the ages were 8-11 yrs (18%), 12-15 yrs (65%), and 15-18 yrs (17%). For the AD36 Positive group, the ages were 8-11 yrs (5%), 12-15 yrs (32%), and 15-18 yrs (63%). This is as we surmised above: more older kids are in the AD36 group. The weight average for the AD36 Negative group was 69 kg (+/- 24 kg SD); for the AD36 Positive group is was 93 kg (+/- 24 kg SD).

Thus, one sure finding is that older kids are heavier: in fact, they were about 24 kg heavier, which translated to about 50 lbs. It is at least good to see that the press release got this figure correct.

Another finding is that older kids are more likely to have been previously infected by AD36, also as we surmised (in two years, you have have plenty of colds).

There is no modeling of the expected distribution of biometric properties of the two populations (AD36-negative and -positive) given the other statistics (age,sex,race) reported. This is extremely surprising to me.

However, it's not just that older kids (adults, really) weigh more - the size-relative metrics BMI, and waist/height ratio were also higher in the AD36 antibody group. But still, those should be adjusted for age as well; their distribution (and mean) will surely change with age, and definitely will change for the worse with age in the destined-to-be-obese (it takes time to blossom into full adult obesity). The fact that no such adjustment was made means that the study contributes almost no additional evidence, but this could be corrected with a proper analysis. And, supposing the raw data is available, this can happen.

The two populations also have a significant sex difference: the AD36-negative group is 58% male, and the AD36-positive is only 47%. However, it looks like adult men and women have similar recommended BMI and waist/height ratios (actually, women are recommended to have slightly lower) - I don't know what the actual averages are. Racially, there's a 13% shift from "non-hispanic white" to "hispanic" in the AD36-positive population. While I don't know how different those groups are re: AD36-positivity or BMI etc., this should be considered as well. The age mismatch is definitely the most severe problem.

The "discussion" section lists much other work which seems to provide better evidence of a AD36-obesity link (I assume the authors are leaving out any negative results that don't support their views). For example:

[In] a small substudy of adult twins with discordant AD36-specific antibody status, 13 Twins with antibodies to AD36 were noted to have higher BMI values and greater proportions of body fat then their respective antibody-negative twins.

(evidence of correlation, not causality - would be quite strong evidence if the number of twins and BMI/body fat mismatch were large enough)

infection of nonhuman primates, rodents, and chickens with AD36 increased total body fat independent of energy intake

(causality in animal models, although i would hope for exposure rather than infection as the trigger - perhaps p(infection|exposure) is high enough that it doesn't matter)

adipose-derived stem/stromal cells ... infected with AD36 showed increased differentiation and higher levels of lipid accumulation than noninfected control cells

(causal, but in vitro: doesn't guarantee net fattening in the context of a human body)

An excellent analysis. I agree that the Gabbert study is seriously flawed. However, the animal experiment model evidence is definitely suggestive in the other direction. But besides the obvious objection that these animals, even the primates, aren't human, there is also the fact that these were caged animals.

On the other hand, the negative evidence from the U.S. Army study is not very convincing either. Unless things have changed dramatically since my day, obesity is not exactly tolerated there and exercise is not considered optional.

What U.S. Army study?

This one. Sorry about that.

Hmm ... having read the study, now, I have to disagree with your earlier remark:

On the other hand, the negative evidence from the U.S. Army study is not very convincing either. Unless things have changed dramatically since my day, obesity is not exactly tolerated there and exercise is not considered optional.

These factors would seem to create the opposite implication: that those found to be obese in the U.S. Army would be disproportionately those who are obese for non-behavioral reasons.

Hmmm. Having read it myself now, I am now very confused. I strongly recommend that everyone participating in this case study read this US military study before making up their mind. PM me (or RobinZ?) with your email address and I will email you a copy.

Elsewhere in this thread, Wei Dai took me to task for suggesting that epistemic rationality was a "solved problem". He was right. It occurs to me that as I recently tried to slog my way through Pearl, I learned that there was still controversy as to the proper way to handle confounding variables. ERV's criticisms definitely showed that the Ad-36 proponents weren't properly handling the confounding variable of age. This US military study convinces me that they are not properly handling the confounding variables of sex and race/ethnicity. I'm almost tempted to conclude that all of the human subject evidence of the Ad-36/obesity link is practically worthless for this reason.

But there is another reason for looking at the military study. They find that past the age of 27 or so, the number of Ad-36 positives among their population falls. That suggests that it may be the case that widespread exposure of humans to Ad-36 may be a recent thing. Which is interesting because the worldwide "obesity epidemic" also seems to be a recent thing.

You have the text (pdf) of that study? Could you PM it to me?

I guess I was making the assumption that even Ad-36 positives would be able to combat obesity if sufficiently motivated to exercise. But, having reread the abstract now, I regret my criticism - when I made the criticism I was thinking that they were comparing non-obese individuals of varying BMI.

I didn't think you could PM files - if you want to PM an email address, I can get it for you that way.

infection of nonhuman primates, rodents, and chickens with AD36 increased total body fat independent of energy intake

(causality in animal models, although i would hope for exposure rather than infection as the trigger - perhaps p(infection|exposure) is high enough that it doesn't matter)

This wasn't clear to me - isn't infection more specific than exposure? I mean presumably only infection would hae metabolic effects, and exposure doesn't always result in infection

The causal intervention is really an act of exposure.

If you expose once, or expose until infected, or expose but exclude those not infected, then the difference between infected and not-infected populations is obscured; the reason some individuals were infected (or not) from a single exposure needs to be explained. If it isn't, then I can say that part of any difference between the infected and not-infected populations is due to whatever factor made some of them fall prey to the infection on one exposure.

Ah right, I see.

Cursory initial Internet searches identify Dr. Nikhil V. Dhurandhar as a prominent researcher associated with studies demonstrating weight gain associated with Ad-36. Two relevant papers: one showing correlation between Ad-36 infection and obesity in humans and one showing causal linkage between Ad-36 infection and weight gain in marmoset monkeys. Whether marmosets are a good human analogue for this purpose is an interesting question I am not qualified to discuss.

The study criticized by ERV in the "certain amount of skepticism" link does not appear to be associated with Dhurandhar. That said, I would like to see reliable studies not associated with Dhurandhar before I ascribe significant probability to the theory.

Going over the literature, I did see Dhurandhar's name a lot. But I also saw a few without it, and their results weren't any different. And some of them have him last in the coauthor list, which suggests that he wasn't closely involved - rather, the researchers called him up for a consultation.

Some initial questions, before I even read into any of the controversy:

  • Would there be an expected evolutionary benefit for Ad-36 to increase obesity?
  • How can we distinguish correlation and causation in this case? One possible confounding variable - high vitamin D levels are positively correlated with viral resistance and negatively correlated with body fat.
  • Have we or can we try this on experimental animal models? A controlled labortory test is the most effective tool for testing the biological effects of isolated factors (as in koch's postulates)

Would there be an expected evolutionary benefit for Ad-36 to increase obesity?

I wouldn't worry too much about this. A lot of pathogen effects have nothing to do with evolution directly, just with the human body being very complicated and it being very hard to do one thing without doing twenty other things at the same time.

For example, some scientists think a Coxsackie virus indirectly causes diabetes. The virus infects, the immune system does its usual job of mobilizing to fight the virus, but something about the virus "looks" enough like beta cells in the pancreas that the immune system gets confused and destroys them too. This doesn't provide any benefit to the virus afaik, it's just something that happens when you have a finite number of ways biological molecules can be configured.

Likewise, chlamydia can make people infertile, not because there's some evolutionary benefit to chlamydia in making people infertile, but just because fertility's a complicated process and in the process of invading the uterus chlamydia ruins all those fine-tuned organs.

I don't know anything about this particular virus, but if it does increase obesity I would expect it to be more like the Coxsackie virus and less like some weird evolutionary adaptation (although those aren't unheard of - see eg toxoplasma)

Would there be an expected evolutionary benefit for Ad-36 to increase obesity?

For people or for the virus?

The "for Ad-36" means: "for the virus", I figure.

Would there be an expected evolutionary benefit for Ad-36 to increase obesity?

I don't know. But the benefit, if it exists, might be to the virus or to the human.

How can we distinguish correlation and causation in this case?

It would be easy, if medical ethics permitted the experiment of exposing half the test subjects to the virus, without telling the test subjects that this is an obesity experiment.

Have we or can we try this on experimental animal models?

I believe there are animal model test results out there. You can easily find them by Googling, etc.

This is similar to the notion of polymath projects (apparently quite successful): http://polymathprojects.org/general-polymath-rules/

This paper is primarily a report on the results of infecting rats with Ad-36. Other people, jimrandomh in particular, have already mentioned it. I mention it here, not because of the results it reports, but rather because of this fascinating introduction:

Although obesity is associated with several health risks, not all obese individuals are adversely affected by it. To better utilize health care resources, it would help immensely to selectively target prevention and treatment efforts by identifying “at risk” individuals from those who are relatively protected from the adverse effects of obesity. Recent emerging evidence from animal models and human studies indicates that some forms of obesity may indeed carry relatively low health risk. We reported that human adenovirus Ad36 infection appears to be one such marker of “low risk” obesity. [snip summary of evidence] Thus, Ad36 induces obesity with metabolically favorable profile in animals and is associated with such obesity in humans.

I'm not sure whether this makes the Ad-36 hypothesis less important, or more important. If a sizable fraction of obese people can stop worrying about the long term health effects of their condition, isn't that a Good Thing?

I'm not sure whether this makes the Ad-36 hypothesis less important, or more important. If a sizable fraction of obese people can stop worrying about the long term health effects of their condition, isn't that a Good Thing?

Knees are still a problem, I'm given to understand, but yes.

Ceteris paribus AD-36 increasing obesity is a trivial information if true.

We have plenty of other means of making rats, monkeys, and humans fat - like feeding them vegetable oil and sugar. And if you haven't noticed this is exactly what we have been doing to ourselves - modern American diets are 36% vegetable oil and sugar, and everyone else is only a decade behind.

Epidemiologically AD-36 is entirely wrong, weight increase is broad process affecting most people, and viral origin would predict one subpopulation being strongly affected, while everyone else wasn't at all.

Broad increases in weight are what you'd expect from widespread environmental factors, like sugar, high trans-fat, and high-omega-6 vegetable oils being added to very wind range of food.

Lack of correlation with insulin resistance also seems to be rather damning.

Sum of all evidence clearly indicates that AD-36 is not the primary factor behind obesity epidemics. It might or might not be significant independent contributing factor, but remember that there might be plenty of other such independent factors each influencing weight in positive or negative direction, without them all necessarily resulting in any net effect. This requires little imagination, as a lot of pathogens are known for making people lose weight.

tl;dr If epidemiology doesn't fit, you must acquit.

Epidemiologically AD-36 is entirely wrong, weight increase is broad process affecting most people, and viral origin would predict one subpopulation being strongly affected, while everyone else wasn't at all.

Since this subpopulation would be based on genetics, which are largely invisible, how do you know this isn't the case?

Weight distribution is not bimodal. Unless nearly everybody got infected by AD-36, but some people just a bit and some a lot more, you would get bimodal distribution.

Or to put it more drastically - AD-36 story predicts median weight barely changing in spite of skyrocketing mean weight.

My impression is that mean weight has changed very little-- 7 to 10 pounds, with the actual increase being about 25 to 30 pounds at the high end.

The definition of obesity was changed in the 90s.

More details

Now is your chance to be delighted by discovering truth! Here's one summary:

Results The age-adjusted prevalence of obesity was 30.5% in 1999-2000 compared with 22.9% in NHANES III (1988-1994; P<.001). The prevalence of overweight also increased during this period from 55.9% to 64.5% (P<.001). Extreme obesity (BMI >=40) also increased significantly in the population, from 2.9% to 4.7% (P = .002). Although not all changes were statistically significant, increases occurred for both men and women in all age groups and for non-Hispanic whites, non-Hispanic blacks, and Mexican Americans. Racial/ethnic groups did not differ significantly in the prevalence of obesity or overweight for men. Among women, obesity and overweight prevalences were highest among non-Hispanic black women. More than half of non-Hispanic black women aged 40 years or older were obese and more than 80% were overweight.

Every part of weight distribution was affected.

Your comment doesn't address Nancy's factual claims. What do the thresholds in your quote tell you about average values? Interestingly, her link is all about how awful thresholds are! The male graph in your source matches the graph in hers, which, by my eyeballing, matches her numbers. The female graph looks different because females are shorter, but I think the pounds are about the same.

As to your original claim, yes, the graph shows that the distribution is not bimodal. But the mean is moving faster than the median.

Your comment doesn't address Nancy's factual claims. What do the thresholds in your quote tell you about average values? Interestingly, her link is all about how awful thresholds are!

You see increases for every threshold - that is whole distribution moving. And this is just a few years and one country of process that is world-wide and lasts for at least four decades now.

Part of your original claim was that mean weight is skyrocketing. How much has it actually increased?

Oh yes, because we all live in fantasy world where quality historical data is easily available... What is wrong with you people...

Anyway, I dug through to what are pretty much punched card readouts from NHANES I dataset that somehow nobody ever bothered to convert to anything more reasonable.

Here are 10th to 90th percentile of weight in kg, height in cm, and BMI (including children):

  • 18.37kg, 37.76kg, 51.14kg, 56.81kg, 61.46kg, 66.45kg, 71.78kg, 78.13kg, 86.75kg
  • 109.3cm, 144.9cm, 155.5cm, 159.1cm, 162.2cm, 165.2cm, 168.4cm, 172.4cm, 177.4cm
  • BMI: 15.7, 17.6, 19.6, 21.1, 22.5, 24.0, 25.6, 27.5, 30.5

Age >= 18 only:

  • 52.16kg, 56.81kg, 60.55kg, 64.3kg, 68.15kg, 72.23kg, 77.0kg, 82.33kg, 90.26kg
  • 155.3 cm, 158.4 cm, 160.9 cm, 163.3 cm, 165.7 cm, 168.2 cm, 171.1 cm, 174.6 cm, 179.0 cm
  • BMI: 19.7, 21.1, 22.3, 23.4, 24.5, 25.8, 27.1, 28.9, 31.8

NHANES I was 1971-75, already quite a bit into obesity epidemic. These are not anything like pre-epidemic numbers, but good luck finding data much earlier than that.

Data for 2007 says 26th percentile has BMI 25, 73th has BMI 30, so I'll linearly extrapolate median adult American BMI to be 27.3 - i.e. median person weights 13% more than if obesity epidemic stayed at early 1970s' level (and let's guess 20% more than if it never happened).