Last week, a study published in the British Medical Journal (BMJ) linked alcohol consumption with cancer risk, and duly attracted extensive international media coverage. News outlets around the world keenly reported on the carcinogenic properties of alcohol. This description from BBC News was pretty typical: “A large Europe-wide study in the British Medical Journal found that one-in-10 of all cancers in men and one-in-33 of all cancers in women are caused by past or current alcohol intake”. Several agencies took an alarmist stance: many cancers are “caused by drinking” proclaimed the Press Association; alcohol was “a major cause” of cancer according to Agence France-Presse (AFP); “Are you aware of the cancer risks of drinking too much alcohol?” challenged Canadian broadcaster, CBC. Others tried to tease out implications. The Irish Times concluded that “reducing alcohol intake would reduce the number of cancer cases”. The Guardian were more directive, noting that “thousands of cancers could be prevented if men had the equivalent of no more than two drinks a day and women had no more than one”. Over at the Daily Mail, however, people were less optimistic: “Alcohol causes cancer“, declared its headline, “…and giving up won’t help“.
The media were virtually unanimous in their interpretation of the research: the study had established tangible evidence for a causal link between alcohol consumption and cancer onset. Unfortunately for the consensus-makers, the problem with this interpretation is that it’s wholly untrue. The BMJ researchers were not even looking for evidence of causality, and most certainly did not establish it. In fact, in one sense, they did the exact opposite.
Rather than establishing a causal link between alcohol consumption and cancer occurrence, the researchers conducted a study that relied on the premise that such a link exists. In other words, a causal link between alcohol and cancer is not a conclusion of the BMJ study — it is one of its assumptions.
The researchers could hardly have been clearer about this. Throughout their report they refer to the fact that they are “assuming causality”. Both the statement of results in the Abstract and the opening paragraph of their main Discussion section begin with the phrase “If we assume causality…” A string of similar disclaimers are embedded throughout the main text. Including such language enabled the researchers to be very explicit about two key points: (a) they were clearly signalling the possibility that there is actually no causal link between alcohol and cancer; and (b) they were highlighting the very fact that their research methods are not designed to shed light on the basic question of causality at all.
Instead, this was very much a “What if…?” kind of study. The basic research question was, “If alcohol consumption causes cancer, then how much cancer might it be causing?” To address this question, the researchers looked at datasets drawn from eight different European countries, focusing on an accumulated total of 360,000 men and women across an eight-year period. Notably, the data about current and previous alcohol consumption were based on self-reported information (rather than, say, on biochemical tests of blood alcohol levels) while data about cancer incidence were derived second-hand from general medical records (rather than, say, from diagnostic biopsies specifically conducted by the research team). The researchers then crunched the numbers and concluded that the portion of the population who consumed large amounts of alcohol tended to include those who experienced greater incidences of cancer. None of these procedures are problematic from a research perspective; in fact they are well suited to the stated study objectives and the opportunities created by such a large-scale data sweep. However, notably, the methods did not examine causal linkages between alcohol consumption and carcinogenesis.
Nor can the results be interpreted in that way, principally because the study was both correlational and reductionist. While the researchers attempted to statistically rule out alternative explanations (such as the likelihood that alcohol consumption would be highest among people whose cancer risk is elevated by factors such as tobacco smoking, poor diet, and low exercise levels), it is impossible to know for sure whether all extraneous factors have been accounted for. And, statistically, such study designs cannot model real-life complexities in the way actual disease risk factors interact with one another in a combinative fashion. For example, when considered combinatively, the effects of alcohol coupled with those of age, gender, smoking, diet, exercise, pollution, drugs, and/or genetics would generate over 500 separate interactions, each of which would require a separate statistical inference. This is simply beyond the scope of statistical hypothesis testing (even when you have 360,000 cases in your dataset).
It is also worth noting that this kind of study fails to allow for the fact that links between alcohol and cancer may be the indirect consequence of different biological events. For example, a number of studies have shown that consuming moderate amounts of alcohol on a regular basis can actually benefit human health, most particularly by limiting the risk of cardiovascular disease. If alcohol consumption were to contribute to longer life expectancies among some consumers, then this should lead to corresponding increases in rates of cancers normally seen in older adults (because more people would be reaching old age). This isn’t to say that there is strong direct evidence to support this interpretation; the point is simply that, as a cross-sectional statistical exercise, the BMJ study was not equipped to rule out such possibilities.
It is certainly well known that frequent heavy alcohol consumption is extremely detrimental for human health. Furthermore, there are decent biochemical reasons to suspect that alcohol plays a role in some cancers (although the physical mechanisms involved remain mysterious). The BMJ study is notable because it suggests that alcohol could account for up to 10% of cancer in males, which is certainly useful information. However, as the BMJ researchers did not attempt discover causality, but just went ahead and assumed it, the bottom line is that this 10% figure remains a statistical hypothesis. The association it represents may well turn out to be a statistical side-effect of other issues.
In implying that causality was established, most media coverage simply missed the point of the research. The study aimed to establish the possible scope of alcohol-related cancers, rather than the fact that such cancers exist. In an ideal world, news coverage of this study would have emphasized the authors’ repeated statements that causality was being assumed. However, that would have made for less exciting coverage.
After all, who wants to run a story with the headline “If causality is assumed, then alcohol causes cancer”? Readers might think that the sub-editor was drunk.
Brian Hughes is an academic psychologist and university professor in Galway, Ireland, specialising in stress, health, and the application of psychology to social issues. He writes widely on the psychology of empiricism and of empirically disputable claims, especially as they pertain to science, health, medicine, and politics.