Tuesday 12 January 2016

Doubt is their product

The changes to the drinking guidelines last week represented the first review of the evidence since 1995. The Chief Medical Officer, Sally Davies, and the Sheffield alcohol researchers who advised her went out of their way to downplay and dismiss the evidence that moderate consumption reduces heart disease risk and overall mortality. Why?

First, let's familiarise ourselves with what the review found back in 1995:

'For a number of years studies have provided evidence that there is a relationship between moderate alcohol consumption and a reduced risk of death from all causes, and that this benefit was found in people who regularly consumed as little as 1 unit per day.'

'It is now established that the main specific pathology which benefits from alcohol consumption is coronary heart disease.'

'The epidemiological evidence alone linking alcohol and CHD now makes the existence of a protective effect appear very likely. The written and oral evidence received by the Group confirmed that most of the technical epidemiological criteria for a causal association are now fulfilled.'

'A key issue previously complicating the epidemiological data has been the so called “sick quitter” hypothesis outlined by Professor A G Shaper... However, a number of studies since 1987 have controlled for these factors so that we believe Professor Shaper’s reservations cannot be considered as a major explanation of the cardio-protective effect. Other confounding factors such as tobacco use, obesity, diet and age have now been controlled for in enough studies to allow us, on the basis of expert testimony, to be confident that the basic protective effect for CHD by alcohol is scientifically valid.'

'All the evidence we have received confirms that the relationship between all-cause mortality and alcohol consumption follows a J-shaped curve. Non drinkers have higher all-cause mortality than light and moderate drinkers, and heavy drinkers have even higher all cause mortality than either group.'

'All cause mortality is at its lowest at modest drinking levels (at about 1 unit a day for men and women) and does not exceed the mortality level of abstainers until consumption levels which are somewhat higher than the current recommended sensible drinking levels of 14 units per week for women and 21 units for men.'

Pretty clear, no? In addition to finding a clear relationship between moderate consumption and lower mortality risk - which could not be explained by confounding factors - the working group identified several plausible biological pathways which explained how the effect came about.

It is important to stress that the 'sick quitter' hypothesis - which says that unhealthy former drinkers could explain why teetotallers don't live as long - had already been tested extensively by 1995. In 2002, Richard Doll discussed it in an essay titled 'Proof of causality':

One possibility, that the non-drinkers included ex-drinkers who might have given up because of ill health, was excluded by cohort studies in which ex-drinkers and lifelong non-drinkers had been classed separately and by data like our own, which included information on past medical histories and showed the same proportional reduction in risk in drinkers irrespective of any previous history of vascular disease.

Another possibility was that drinkers might have differed from non-drinkers in other ways that would affect the risk of the disease, by, for example, including lower proportions of cigarette smokers, having a lower mean blood pressure and body mass index, a higher level of physical activity, or higher socioeconomic status (Shaper 1995). In fact, an association with smoking has the opposite effect to that postulated, as smoking is more prevalent in drinkers than in non-drinkers (Jarvis 1994), and allowance for it actually increases the evidence of benefit associated with drinking. In many studies it has been taken into account, as have all the other suspect factors in some of the more detailed studies, such as those by Stampfer, et al. (1988), and Thun, et al. (1997). When these factors were all allowed for, the observed differences were hardly altered.

Confounding, often a serious concern when risks vary by less than two-fold, has in this case been tested and found wanting... That the inverse relationship between ischemic heart disease and the consumption of small or moderate amounts of alcohol is, for the most part, causal should, I believe, now be regarded as proved (Doll et al. 1997).

In the last fifteen years, research has continued to show that the J-curve is not the result of 'sick quitters' or other confounding (eg. here, here, here, here and here). A meta-analysis of 34 prospective studies in 2006 showed a clear J-curve that could not be explained by 'sick quitters'. This is how epidemiology is supposed to work. People suggest alternative explanations, new studies test the alternative hypothesis and if the association survives all tests the relationship is assumed to be causal.

In any other field, the debate would have been put to bed years ago. It has been kept alive because many people in 'public health' desperately want to believe that alcohol has no (health) benefits. Almost unbelievably, Sally Davies described the J-curve as an 'old wives' tale' last week. She drew on a report she had commissioned from Sheffield University (the same people who devised the minimum pricing computer model). You can read it here.

The Sheffield report starts the campaign of doubt early in the main text when it mentions the protective effect of moderate drinking and says...

These effects are disputed,(10-13) may be overestimated (14-17) and are probably limited to particular groups within society (18) (see Section 5.4.1.2). If, as appears possible, scientific opinion develops to conclude cardioprotective effects are in fact overestimated and only occur up to very low levels of consumption (e.g. 5 units per week); one of the researchers responsible for the Canadian approach concluded that this would leave the Canadian guideline “in trouble”.(19)

It's worth looking at what all those references refer to:

(10) is an opinion piece in Addiction which discusses the same confounding variables which Doll said had been overcome in 2002.

(11) is a study that doesn't really 'dispute' the J-curve.

(12) is a letter from Tim Stockwell and friends responding to a meta-analysis in the British Medical Journal which found that the link between good cardiovascular health and moderate alcohol consumption was 'beyond question'. That meta-analysis is not cited in the Sheffield report at all, nor is the authors' reply which accuses Stockwell et al. of adopting 'an extreme methodological position, proposing to dismiss an entire body of literature on the basis of the presence of predictable limitations in individual studies. This dogmatic and dichotomous approach to the evaluation of epidemiological studies is counterproductive to scientific epistemology.' If you read Stockwell's letter, you will see that they have a point.

(13) is an opinion piece by Tim Stockwell and friends casting doubt on the alcohol-health link.

(14) is a study that clearly found lower heart disease risk among moderate drinkers.

(15) is a comment piece by Tim Stockwell and friends.

(16) is a revised meta-analysis by Tim Stockwell and friends which claimed that the protective effect of alcohol disappears when adjustments are made to the data and some studies are excluded.

(17) is a study which found that a gene associated with not drinking was also associated with better cardiovascular health.

(18) is a meta-analysis which concluded that 'some form of a cardioprotective association was confirmed in all strata' but that this 'cannot be assumed for all drinkers'. Its authors have since published a meta-analysis which found clear evidence of a J-curve, even when only lifetime abstainers were included (ie. no 'sick quitters').

(19) is a short opinion piece by (you guessed it!) Tim Stockwell.

There you have it - the case for the prosecution. Not quite a one man crusade by Tim Stockwell but not far off. Needless to say, you can find dissenting voices and contrary evidence in any field of research, but on the basis of these minority reports the Sheffield crew not only portray the science as highly contentious but also predict that the science will eventually agree with Tim Stockwell.

Towards the end of the report, the Sheffield boys and gals address the J-curve at more length. They start by acknowledging (but not referencing) the evidence...

An extensive literature including well-executed meta-analyses of high quality primary studies have found an association between moderate drinking and reduced risk of cardiovascular disease and particularly ischaemic heart disease.

There is a 'but' coming, of course...

This literature has attracted substantial debate regarding whether evidence is sufficient to conclude that low levels of alcohol consumption have a causal relationship with improved cardiovascular health. The debate includes detailed critique of both observational and meta-analytic studies, exploration of potential biological mechanisms explaining observed cardioprotective effects and arguments regarding the public health relevance of establishing the veracity of cardioprotective effects given alcohol’s undisputed risk for other diseases and the various alternative options for reducing cardiovascular risk.(10,12,13,17,94)

Those references are some of the articles and studies mentioned above plus a comment piece in Alcohol from 2007 which said 'In view of the potential risks of alcohol, a more cautious view about the beneficial effects of alcoholic beverages is warranted.' It is notable that all the citations given to prove the existence of a 'debate' here are from people on one particular side of the debate. This is a bias that permeates the Sheffield report. For example, no fewer than twelve Tim Stockwell publications are cited in the report, along with eight publications by John Holmes of Sheffield University. By comparison, there are zero citations of studies by people like Richard Doll who produced several studies showing (and testing) the J-curve, nor is there any mention of people like Di Castelnuovo who conducted the 2006 meta-analysis.

The Sheffield crew then proceed to give four possible reasons why this 'extensive literature' is wrong...

First, there is evidence that participants in epidemiological cohort studies may differ with regard to their underlying health status compared to the general population. One reason for this is such studies often recruit participants with no underlying health conditions at baseline. The resulting potential for bias was demonstrated in a major European prospective cohort study which included at baseline people with chronic disease. Risk estimates were calculated for both the whole cohort and for a subsample of the cohort who were free of chronic disease at enrolment. Relative risk of cardiovascular mortality was lowest in those with light to moderate alcohol use; however, this was only the case among the subsample free from chronic disease at enrolment. This suggests sample selection processes for typical cohort studies may disproportionately exclude those at cardiovascular risk from moderate drinking leading to overestimation of any cardioprotective effect.

This is a reference to this study which found that moderate drinkers had a 26-48 per cent lower risk of heart disease mortality. The study included graphs showing a clear J-curve for all-cause mortality (the one below is for men).


The protective effect for heart disease was only statistically significant for people who did not have a disease at the time of enrolment, but even among those who had a disease upon enrolment, heart disease risk was not higher.

Second, estimates of risk relationships between alcohol consumption and health conditions are commonly quantified by calculating the risk of a given level of consumption relative to the risk of zero consumption (i.e. abstention). In practice, this means assuming that, after controlling for a range of confounding factors such as age and gender, drinkers and abstainers only meaningfully differ in terms of their alcohol consumption and a narrow set of other factors.

That's what epidemiologists are supposed to do. Smokers are different to nonsmokers in ways that have nothing to do with tobacco. Promiscuous people are different to virgins. People who drink large quantities of sugary drinks are different to people who don't. This is why epidemiologists adjust for factors that might bias the results. Alcohol research is no different to any other type of observational research in that respect - and, as Doll pointed out, non-drinkers tend to lead healthier lifestyles than drinkers, therefore any bias likely leads in the opposite direction to what the Sheffield authors are implying.

Rather than explaining what it is about the characteristics of non-drinkers that invalidate alcohol research when the characteristics of smokers do not invalidate tobacco research, the Sheffield crew ignore all the studies that have tested for confounding variables and return the subject of sick quitters:

This assumption has been questioned and the characteristics of abstainers and their similarity to the general population have been closely scrutinised. Most significantly, the classification of former drinkers as abstainers has raised particular concerns, particularly where those former drinkers have stopped drinking due to health problems. Meta-analyses which disaggregate abstainers (e.g. never drinkers, former drinkers, occasional drinkers) have concluded that using a single abstainer category leads to overestimation of the cardioprotective effect of alcohol.(15,16,18)

Reference 15 and 16 are two similar studies from the pen of Tim Stockwell and friends. One is a revised meta-analysis which controversially claimed that the protective effect of alcohol disappears when studies are excluded and adjustments are made to the data. Reference 18 is a meta-analysis which concluded that 'some form of a cardioprotective association was confirmed in all strata' but that this 'cannot be assumed for all drinkers'.

It is certainly true that including 'sick quitters' biases the results and makes the J-curve look steeper than it is. That is why most studies in the last 25 years have excluded former drinkers and still find strong evidence of a J-curve. The Sheffield authors choose not to mention this or cite those studies, preferring instead to imply that this is an unanswered question.

Third, alcohol consumption is typically measured in epidemiological studies of long-term health risks as average daily consumption. However, recent evidence incorporating data on frequency of heavy drinking occasions (defined as more than 7.5 units on a single day) has shown an elevated ischaemic heart disease risk for moderate drinkers who have heavy drinking occasions at least once per month when compared to moderate drinkers with fewer heavy drinking occasions. Further analyses suggest any cardioprotective effect from moderate drinking may be attenuated or no longer present among those who have heavy drinking occasions at least monthly.

This was acknowledged in the government's 1995 report. It is why people are advised to drink moderately, not heavily. It does not mean that there is 'no safe level of drinking', nor does it mean that the protective effect is an 'old wives' tale'.

Fourth, alcohol consumption is only one of many variables which have a positive or negative association with an individual’s cardiovascular risk and it has been argued that “groups with different drinking habits differ in several other ways than their drinking, making it difficult to separate the effects of drinking habits from other factors”. For example, both increasing age and smoking status increase individual risk of ischaemic heart disease and estimated risk relationships for alcohol consumption can be adjusted to account for these confounding factors. However, a recent major meta-analysis noted substantial unexplained heterogeneity in risk estimates suggesting other important confounding factors were not controlled for.(18)

Reference 18 is the same study they have cited twice already. This point really just reiterates what was said earlier about there being differences between teetotallers and drinkers. In effect, they are saying that there might be some other factor that might make teetotallers die younger than drinkers but we don't know what it is. Well, there might be, just as there might be something about smokers which has nothing to do with smoking that makes them more likely to get lung cancer, but the temperance lobby has been hunting pretty hard for this mysterious explanation for a long time and come up empty-handed.

This is the sum total of their objections to the J-curve. Essentially, they are just saying 'confounding factors' over and over again without acknowledging that that question has been emphatically answered. It is difficult to see how any amount of evidence could satisfy them and it is notable that they do not raise the same objections to the epidemiological evidence linking moderate alcohol consumption to various cancers, most of which is weaker and all of which is susceptible to the same potential biases and confounding. This point was well made by two scientists when Tim Stockwell demanded an almost impossible burden of proof in 2013...

... it seems that some researchers in the field may be using different standards in assessing the cardioprotective effect of alcohol vs. its detrimental effect. Consider two examples. One is the effects of alcohol on colon cancer. Would the same arguments used to judge the relationship between alcohol and ischaemic heart disease not hold for this relation as well? The other example is the more than 200 other risk relations between alcohol and disease and injury outcomes. Of course, this is not a good argument against scrutinizing the cardioprotective effect of alcohol, but we sense a desire by some in the field to apply tough standards on protective effects and more lenient standards on other effects, where sometimes the responses to very simple survey questions such as ‘Did your partner’s alcohol consumption contribute to your marriage problems?’ are accepted as causal evidence.

Nevertheless, the Sheffield authors believe that they have thrown enough mud to make some of it stick and conclude:

Given these critiques, there is little consensus in the scientific community regarding the presence or size of any cardioprotective effect.

There is little consensus in the public health community about the presence of a cardioprotective effect, but the public health community should not be mistaken for the scientific community. There is little consensus in the anti-vaxxer community about the protective effect of vaccines and there will never be agreement from people like Tim Stockwell about the benefits of alcohol. Rather than waiting for a consensus amongst people who hate alcohol, let's trust the evidence.

No comments: