Thursday, May 14, 2015

Coffee - a guilt-free pleasure?

A lot of research money has been spent trying to find the bad stuff that coffee does to us.  But Monday's piece by Aaron Carroll in the New York Times reviewing the literature concludes that not only is it not bad, it's protective against a lot of diseases.  If he's right, then something that's actually pleasurable isn't sinful after all!  

The piece had so many comments that the author was invited to do a follow-up, answering some of the questions readers raised.  First, Carroll reports that a large meta-analysis looking at the association between coffee and heart disease found that people who drank 3-5 cups a day were at the lowest risk of disease, while those who drank 5-10 had the same risk as those who drank none.

And, by 'coffee', Carroll notes that he means black coffee, not any of the highly caloric, fat and sugar laden drinks he then describes.  But, it can't be that all 1,270,000 people in the meta analysis drink their coffee black, so it's odd that he brings this up.  Fat and sugar are our current food demons, yes (speaking of pleasurable!), but really, does anyone have "a Large Dunkin’ Donuts frozen caramel coffee Coolatta (670 calories, 8 grams of fat, 144 grams of carbs)" (Carroll's words) 5-10 times a day?

A lifesaving cup of black coffee; Wikipedia

So, two to six cups of coffee a day are associated with lower risk of stroke, 'moderate' consumption is associated with lower risk of cardiovascular disease, in some studies (but not others) coffee seems to be associated with a lower risk of cancers, including lung cancer -- unless you're a smoker, in which case the more coffee you drink, the higher your risk.  The more coffee you drink, the lower your risk of liver disease, or of your current stage of liver disease advancing; coffee is associated with reduced risk of type 2 diabetes. And, Carroll reports, two meta-analyses found that "drinking coffee was associated with a significantly reduced chance of death." Since everyone's chance of dying is 100%, this isn't quite right -- what he means, presumably, is that it's associated with lowered risk of death at a given age, and by implication, longer life (though whether that means longer healthy life or not is unclear).

In the follow-up piece, he was asked whether this all applies to decaffeinated coffee as well.  Decaf isn't often studied, but when it is the results are often the same as for caffeinated coffee, though not always.  And sometimes true for tea as well.  So, is it the caffeine or something else in these drinks?

This is an interesting discussion, and it raises a lot of questions, but to me they have more to do with epidemiological methods than the idea that we now should all now feel free to drink as much coffee as we'd like, guilt-free (though, frankly, among 'guilty pleasures', for me coffee isn't nearly pleasurable enough to rank very high on the list!).  Indeed, after presenting all the results, Carroll notes that most of the studies were not randomized controlled trials, the gold standard of epidemiological research.  The best way to actually determine whether coffee is safe, dangerous, protective would be to compare the disease outcome of two large groups of people randomly assigned to drink 1, 2, 3....10 cups of (black) coffee a day for 20 years.  This obviously can't be done, so researchers do things like ask cancer patients, or people with diabetes or heart disease how much coffee they drank for the last x years, and compare coffee drinkers with non-drinkers.

So right off the bat, there are recall issues.  Though it's probably true that many people routinely have roughly the same number of cups of coffee every day, so the recall issues won't be as serious as, say, asking people how many times they ate broccoli in the last year.  But still, it's an issue.

More importantly, there are confounding issues, and the lung cancer association is the most obvious one.  If the risk of lung cancer goes up with the number of cups of coffee people drink a day, that's most likely because smokers have a cigarette with their coffee.  Or, have coffee with their cigarette.

Or, less obviously, perhaps people who drink a lot of coffee don't drink, I don't know, diet soda, and diet soda happens to be a risk factor for obesity, and obesity is associated with cancer risk (note: I made that up, more or less out of whole cloth, to illustrate the idea of confounding).

Una tazzina di caffè; Wikipedia

And what about the idea that decaffeinated coffee, and black but not green tea, can have the same effect?  If there really is something protective about these drinks, and we're going to get reductionist about it and finger a single component, what about water?  Could drinking 10 cups of water a day protect against liver disease, say?  True, not all the studies yield the same results, and the black but not green tea association suggests it's not the water, but not all studies show a protective affect of coffee, either.  But this idea would be easy to test -- Italians drink espresso by the teaspoon and on the run.  Do Italian studies of coffee drinking show the same protective effect?

Remember when drinking drip coffee was associated with increased cholesterol levels?  Carroll writes:
[A]s has been reported in The New York Times, two studies have shown that drinking unfiltered coffee, like Turkish coffee, can lead to increases in serum cholesterol and triglycerides. But coffee that’s been through a paper filter seems to have had the cholesterol-raising agent, known as cafestol, removed. 
High blood pressure and high cholesterol would be of concern because they can lead to heart disease or death. Drinking coffee is associated with better outcomes in those areas, and that’s what really matters.
So, high blood pressure and high cholesterol aren't in fact associated with heart disease or death?  Or, only in non-coffee drinkers? A word about methods is in order.  The results Carroll reviews are based on meta-analyses, that is, analysis that combines the results of sets of independently done studies.  As even Carroll said, some individual studies found an association between coffee and cancer at a particular site, but the effect of meta-analysis was to erase these.  That is, what showed up in a single study was no longer found when studies were combined.  In effect, this sort of pooling assumes homogeneity of causation and eliminates heterogeneity that may be present when considering individual studies, for whatever reason.  Meta-analysis allows far greater total sample studies, and for that reason has become the final word in studying causation, but it in fact can introduce its own issues.  It gains size by assuming uniformity, and that is a major assumption (almost always untested or untestable), which can amount to a pragmatic way of wishing away subtleties that may exist in the causal landscape.

I'm not arguing here that coffee is actually bad for us, or that it really isn't protective.  My point is just that these state-of-the-art studies exhibit the same methodological issues that plague epidemiological studies of asthma, or obesity, or schizophrenia, or most everything else.  Epidemiology hasn't yet told us the cause of these, or many diseases, because of confounding, because of heterogeneity of disease, because of multiple pathways to disease, because it's hard to think of all the factors that should be considered, because of biases, confounding of correlated factors, because meta-analysis has its own effects, and so on. 

One should keep in mind that last year's True Story, and many every year before that, had coffee -- a sinful pleasure by our Puritanical standards -- implicated in all sorts of problems, from heart disease to pancreatic cancer, to who knows what else.  Why should we believe today's Latest Big Finding?

Even if drinking coffee is protective to some extent, the effect can't be all that strong, or the results would have been obvious long ago.  And, the protective effects surely can't cancel out the effects of smoking, say, or overeating, or 5-10 coffee Coolatas a day.  The moral, once again, must be moderation in all things, not a reductive approach to longer life.


Holly Dunsworth said...

I doubt that we're ever going to show that coffee's going to kill us enough to make us stop drinking it. It's more likely that climate change and economics will decide our future coffee-drinking behavior for us.

Perhaps we should turn our epidemiological/public health attention ($$$) toward the stuff that's so new it's not yet not ingrained in cultural traditions around the world, like the thousands of new chemical compounds invented over the last few decades that we eat, breathe, and shroud ourselves in daily.

Also, I'm starting to wonder if this never-ending cycle of pros/cons to coffee, broccoli, bananas, eggs, wine, chocolate, beer, red meat,... I'm starting to wonder if we eat all this up for the same reason we eat up evolutionary just-so stories. Stories.

Because although they seem more relevant to our lives than evolutionary tales, I bet these kinds of wishy-washy results don't change very many people's coffee-drinking behavior. Who's studying that? Who's studying what people do in response to these findings that aren't very informative? Same with 23andMe's reports that I have 35% risk of developing X disease. Who's studying what people do with their lives once they're given this information? How do you find such studies? What discipline is that?

Holly Dunsworth said...

My first sentence should end with, "or help us enough to make us drink more of it."

Holly Dunsworth said...

Not to continue spouting off, but, yes, to continue...

What this begs for, at least to me, is proof (by whom, from whom?) that these scientific studies, which I believe are eligible for larger pots of public funding are actually more worthy of it than ones that are seen as "just" for the sake of knowledge, like paleontology. Because at this point, this stuff is "just" for the sake of knowledge too. Isn't it?

Ken Weiss said...

It's all the same story, when it comes to complex causation, which even this is about. It has more to do with maintaining the research establishment and its supporting industries than with actual results or public health science. If we were doing what was really needed, it would include actually figuring out real cures for clearly genetic disease, for example, or real ways to change behavior that leads to serious common disease. Coffee is not the problem, nor the cure (except for putting us in a better mood).

Holly Dunsworth said...

I just realized that the causation problem is a ... problem... for studying how/whether scientific findings affect our dietary habits. Meh. I'm going back to grading now.

Ken Weiss said...

We are, collectively, in my opinion not asking what physicists call well-posed questions. We are assuming physics and chemistry-like replicability, orderly probability distributions reflecting causal factors, and so on. These are unrealistic, and maybe that fact is itself the deeper truth. Maybe there just is no simple causal-effect predictability of the sort people are naturally wishing for.

Anne Buchanan said...

I have an idea. Do you think Nature would publish it? We could do n of 1 studies on coffee drinking.

Ken Weiss said...

YES! I'll drink a cup, and then have a heart attack or not, then drink another cup, ... and so on. If I eventually have a coronary, then we know coffee's the cause. If not, then we know coffee protects!

Simple as, well, simple as a typical story in Science or Nature!

Holly Dunsworth said...

Sounds like you're both on coffee, for the good or bad of it.

Ken Weiss said...

Yes, and no decaf fake stuff for me, either! But we mostly drink tea (until someone shows that that's really, really bad for you!).

Anonymous said...

There were times when I would drink couple of cups per day, other times I would avoid coffee. However what stuck with me is that coffee is bad for you. So it's kind of like when I drink coffee, I feel guilty. This is not a good thing, considering how many studies indicate that it's healthy to drink coffee in moderation. However I guess the health industry did its job properly and I just feel that coffee is unhealthy. Recently I read a blog post stating that 15-20 minutes of a powernap is much more energizing than a cup of coffee. I have mentioned this and got a great response - coffee is like a quirky habit, routine that you love doing. So why give up I ask =)

Ken Weiss said...

Yes, and a few years ago it was wine, then eggs, then butter, then cholesterol, then salt. As Hippocrates said some 3000 years or so ago, moderation in all things! And don't read the 'science' news; it will just upset you---and that really IS bad for your health!

Geoff Dougherty said...

Before attributing the back-and-forth results on coffee to study design, it's probably worth unpacking study design a bit.

First, while randomized trials are inevitably referred to as the "gold standard" of causation, they, too, can yield misleading results. The people who participate in RCTs are invariably not the same people for whom we'd like to make inferences about coffee and stroke. The RCT folks are healthier, wealthier and better educated, so a finding in one direction or another from an RCT may not represent the underlying population value.

Secondly, even though an RCT might direct someone to drink X cups of coffee per day, the reality is that trial participants never do exactly what they're told, and so you're left to either analyze the data based on their actual behavior, which breaks the randomization process and opens the results to confounding, or to analyze based on the exposure they were assigned to, which inevitably makes it more likely that you will miss any real effect that might be present in the data.

While some of the studies mentioned by Carroll might have been case-control studies, many were either cross-sectional or cohort studies, which means people were asked about coffee consumption contemporaneously, before they got sick.

This limits the possibility for recall bias.

That said, measuring exposures is tough to do, particularly when they involve what we eat or drink. But unless the error is somehow associated with the outcome (People on the verge of heart disease systematically report more coffee consumption), this type of thing will mostly bias the study results toward the null.

Confounding is always an issue. I haven't read all of these studies, but I strongly suspect they all controlled for smoking, BMI, statin use, age, gender, and other well-known predictors of disease that could also be associated with coffee. So the main potential for bias arises with unknown/unmeasured confounders

The nice thing about unknown confounders is that they're mostly caused by the same thing as known confounders (i.e., bad health), so adjusting for known confounders often yields an accurate result.

So, if we can't blame confounding or measurement error, what can we blame? The list is long:

* Small study size. Smaller studies are more likely to generate extreme results, so it's no surprise to see a series of studies w/ low N ping-ponging back and forth across the risk/no risk line

* Different populations. If I study coffee in white women under 40 and you study it in black men over 40, we may get different results. The resulting headline will likely read "Coffee is dangerous/healthy" regardless of the specifics of the population

* Different exposures. I decide to study what happens to people who drink more than four cups of coffee a day. You decide to study the effect of any coffee consumption vs. none. Unsurprisingly, we get different results

* Different outcomes. I happen to have data on ischemic stroke lying around, and I decide to do a study on association with coffee. You have data on people with hemorrhagic stroke. Our analyses point in different directions, and again result in duellng headlines: "Coffee prevents/causes stroke!"

The big surprise here, to me anyway, is how consistent this is with the way things go in almost all scientific endeavors. Someone does a study. Someone else thinks that if they address the same question in a different way, or with different data, they might get a better/different answer. They do it, and their colleagues try to reconcile the competing results. They come up with an explanation and test it, and thus the cycle continues.

Anne Buchanan said...

Geoff, thanks so much for your detailed comments. Absolutely. Or, one population drinks Dunkin Donuts coffee Coolatas and the other drinks espresso, but they are both called 'coffee'; your definition of heart disease is different from mine; when you started to drink coffee is in fact the crucial factor, but people don't remember that accurately; older people used to drink instant coffee (say), and that was in fact the cause of the findings of coffee being a risk factor. And so on.

I don't find it surprising that this is just the way things go. Our methods are good when risk factors are strong (cholera, HIV, Ebola), but not so good when everyone is unique and risk factors have weak effects. Science is better when all observations are replicable, and that just doesn't happen in epidemiology or genetics. Throw on top of that all the other issues that we've mentioned, and, well, should I have that after dinner cup of coffee or shouldn't I?

Ken Weiss said...

It isn't always exactly as you very clearly outline are issues in many life and health sciences. I recently learned of, and have read, a book by Milton Rothman called Discovering the Natural Laws. It's not a new book, but it shows how the 'laws' of physics were tested and developed. There, a basic principle is 'triangulated' by various experimental designs and so on, and the underlying principle is homed in on with asymptotically increased precision. With a good theory to test, an a prior theory one might say, we can refine our understanding and see if the theory really is correct, etc.

There are issues and uncertainties in physics, certainly, but this basic process is not, in areas we're discussing, how things work. Here we don't have adequate theory to test. We compare internally (cases vs controls, drinkers vs non-drinkers). There seems little interest even in developing an adequate theoretical basis, and we borrow methods developed for truly replicable phenomena, as we often say here. Or, our theory that things are not replicable is not being understood, accepted, or used adequately. Therein lies many of our problems and issues, I think.

When, how, or whether we can develop an adequate theory to test is an open question. Why we persist in what we're doing is a separate question that, in my view again, is about the sociopolitics of current science as much as it is about the science itself.

Geoff Dougherty said...

I don't know that the way things work in epidemiology is terribly different, Ken. Someone comes up with an animal study suggesting exposure X is carcinogenic. Another group does a case-control study. Someone else does a cohort study. Group #4 tries to find a situation where X has been withdrawn or increased in a community due to quasi-random factors, and dubs this a natural experiment. Group 5 does some bench science and comes up with a plausible pathway by which X causes cancer. If three or four of these things all point in the same direction, we begin to think it would be smart to avoid X.

I completely agree with you that we need theory, and that some of the things we test are more about socio-politics than science. It's clear, to me at least, that residential segregation has far more to do with cardiovascular disease than coffee does (or doesn't). But good luck untangling that web of causality. Or, on second thought, wish me luck; it's my dissertation.

Ken Weiss said...

We need thoughtful wrestle with the problem, rather than just caving in to 'normal science' (as Kuhn put it), which is what statistical epidemiology and genetics are largely about these days. I hope you've found a tolerant, thoughtful advisor!

Anonymous said...

The individual's reaction to coffee and long-term effects is probably genetically variable, and heritable.

Like what this meta-analysis of twin studies tells us; non-additive effects of genes and shared environment matters very little. For all traits, actually.

Massive meta-analysis of twin studies: The shared environment and non-additive genetic variation have little impact.

Ken Weiss said...

Thanks for pointing out this new paper, which we had not seen if it was even online when we wrote this post. There are a mix of semantic, epistemological, and conceptual issues here.

Of course everything is 'genetic' in many fundamental ways. That environment is important cannot be denied, and there are shared aspects not all of which can be adequately accounted for. And the molecular way that genes (and other functional genomic elements) work is inherently by interaction; there is no reason to think that genetic variation is all additive at the actual molecular level. There are many issues here, too many for a response to a Comment.

If the coffee responses are genetic as you note, but also complex, to the extent that people are mainly different in their genetic aspects of the trait, then saying that the trait is 'genetic' may be true but at the same time rather misleading or even useless.

Anonymous said...

From my reading of the paper, twin resemblance for a majority (69%) of traits appears to be solely due to additive genetic variation. This vindicates Visscher's and Plomin's work into finding the genetic instatntiation by assuming the additive model for the heritability of traits, which is incontrovertibly "there" considering the decades of twin studies.

Anne Buchanan said...


It behooves the reader of a scientific paper to think, first, about how it might be wrong. Words like "vindicated" and "incontrovertibly" suggest that's not your approach. Science should not be ideology.

Anonymous said...

It seems to me, Prof. Buchanan, that those who rail against additive effects are doing so on the basis of ideology.

Visscher demonstrated that in this paper, I think it's clear the importance of additive effects:

Anne Buchanan said...


I think it's time for us to agree to disagree.

Anonymous said...

Sure, but are you disagreeing with my reading of Visscher's paper and that meta-analysis, or with the papers themselves? That's what I'd like to know, because I do not understand what the disagreement is exactly.

Anonymous said...

Here's a defense of the additive model I agree with. If I have not articulated my view properly, this is it:

I don't know why you'd disagree with it.

Anne Buchanan said...

There isn't going to be a single explanation for all traits. Some will be due to gene X environment interaction, some gene X gene interaction, some essentially single gene, some primarily environment (though even then of course genes are involved) and so on. So, the insistence that the additive model is the right one is just wrong. It might explain some traits. It certainly won't explain them all.

Ken Weiss said...

These are old, and legitimate arguments for the additive components of complex traits. But they first of all state the approximate additivity, and this does in now way vitiate the likelihood that among many of the pairwise (or more-wise) interactions at the molecular and cellular level that things are additive.

Secondly, the idea that evolution favors additivity is that the interactive effects can be transmitted if they are at a single locus, for example, because only one allele, not the combination, is transmitted to an offspring, breaking up the interaction effect. But the same sort of thing is true for complex multilocus traits even for additive combinations, and if the alleles are common enough (because they're good enough fitness-wise), then the good combinations will be common in the population. This applies, with its own dynamics, whether the effects are additive or include important aspects of interactions.

Some have argued for the importance of epistasis because they want to wave away the problem of 'missing' heritability. That argument hasn't convinced very many people so far, I think, and even the advocates of the view basically concede that testing for epistasis is devilishly hard in practice.

All complex analytic studies, like the twin study, are meta-analyses and make all sorts of assumptions, including about LD patterns, and interpretations hang on assumptions to a considerable extent--even if, overall, nobody can deny the importance of additive effects or their relative ease of detection in heritability studies.

Others note that at the actual molecular level, there is no reason for pure additivity. Evolution screens on (1) luck, where additivity is irrelevant, and (2) selection which for reasons your link points out, which are not at all newly discovered!, make evolutionary sense. The central issue is the extent to which any of this makes traits predictable from genotypes. The answer is that will be partly so, depending on the genotype, and the extent to which they are so is not fixed or predictable. There is also no prior theory for the degree to which environment and genes interact additively. To some degree they will, to some approximation they will, but in any given instance they may or may not.

But there is no need to argue about whether the methods such as are used are detecting additive effects in the statistical estimation sense. They give collective pictures of the role of genes under particular circumstances.

Now, this is taking things far enough, because it is a case advocacy that never ends, whatever your motives are. So you should start your own blog and explain your point of view fully. We have to move on.