Is drinking a moderate amount of alcohol bad for you? Answering questions about potential causal effects in humans is difficult. If you want to know what factors affect the lifespan of a lightbulb, your task is simple: get a few hundred lightbulbs, and start experimenting. Lightbulbs are cheap, available, and (for the large part) homogeneous – they all behave the same way. Humans are none of those things. And even if we could experiment on humans with no ethical limitations, we would have to recruit individuals into a randomized trial, divide volunteers at random into two groups, assign one group to drinking a moderate amount of alcohol, assign the other group to drinking no alcohol, and wait for 50 years to compare disease rates between the groups. By the time that we have our results, the generation for whom those results hold would already be elderly, and we would have to assume that our findings hold for the current generation, whom we can finally advise whether to drink alcohol in moderation or not.
Correlation, not causation
An alternative would be to compare disease rates between those who drink alcohol versus those who do not drink alcohol. But this comparison falls short for other reasons. Alcohol drinkers differ from those who do not drink in numerous ways not limited to alcohol consumption. For instance, they are more likely to smoke tobacco, they are more likely to be physically inactive, and they are more likely to have a university education. While we can measure and account for some of these factors, we can never know if we have measured all relevant factors. Hence it is unclear if differences in disease risk between alcohol drinkers and non-drinkers arise due to alcohol consumption, or if there is an alternative explanation.
Exploiting natural variation
Mendelian randomization is an example of a natural experiment. Rather than setting up a randomized trial ourselves, we can rely on factors that behave somewhat like randomization, in that they divide the population into groups that differ as to their average alcohol consumption levels, but are similar with respect to all other factors. In Mendelian randomization, we exploit genetic variation as a form of natural experiment. Some individuals are born with a natural predisposition to drink more alcohol, others are not. In Western populations, these differences are small, on average accounting for differences in the order of an additional unit of alcohol per week. In East Asian populations, there is a common genetic variant linked with the “alcohol flush reaction”, which is carried by around 30-50% of East Asians and is associated with a much larger difference in alcohol consumption.
We can construct a natural experiment by comparing disease rates in those with higher versus lower genetic predisposition to alcohol drinking. If we can find genetic variants that influence alcohol consumption, but do not affect other factors, then we can use these variants to construct subgroups of the population which differ with respect to average alcohol consumption levels, but not with respect to other factors. Finding such variants is plausible as each gene has a specific function, and Mendel’s laws of inheritance state that unrelated traits are inherited independently. Under these conditions, any difference in disease rates between these groups must therefore be due to alcohol consumption.
Mendelian randomization analyses have indicated harmful effects of alcohol consumption on overall mortality rates, as well as on cardiovascular disease risk (including coronary heart disease and stroke) [see review here]. However, the evidence is less strong for cancer, with several studies reporting no consistent evidence for an effect on the majority of cancer types.
Comparisons at low consumption levels
Additionally, some Mendelian randomization studies have compared low alcohol consumers. At very low levels of alcohol consumption, there is no clear evidence that genetic predictors of alcohol consumption associate with increased disease risk. However, there is no sign of a beneficial effect either. It would be unsurprising that the effect of very low alcohol consumption is too weak to be reliably detected, even in very large samples. In one recent study, associations indicating a harmful effect were only evident at a consumption level of around 8-10 grams of alcohol per day (around 1 UK unit per day). Below this level, we cannot be sure – there may be a harmful effect that is too small to be detected, or else there may be no effect.
Limitations and caveats
An important caveat for understanding Mendelian randomization studies is the plausibility of the genetic variants as influencing alcohol consumption in a specific way. If the genetic variants increase average alcohol levels, but they also increase average rates of smoking, or other risk factors, then it would not be reasonable to claim that alcohol consumption is the only possible explanation for increased disease risk. It may be that the true causal risk factor is something else.
For alcohol consumption, we are fortunate. We understand many of the biological processes involved in the body’s processing of alcohol, and we can look at naturally-occurring variation in the genes regulating these processes. By restricting our attention to genetic variants in these regulatory genes, our analysis has the best chances of truly reflecting the causal effect of alcohol consumption.
For other important risk factors, we may not be able to find such genes. Or else there may not be variants associated with large enough differences in the risk factor to allow small (but potentially relevant) effects to be detected. For example, we do not understand which biological mechanisms influence sleep duration.
In the case of alcohol consumption, an alternative approach would be to include all genetic predictors of alcohol consumption in our analysis, regardless of mechanism. However, such an analysis would not provide as reliable evidence for a causal effect as an analysis that only considered variants in alcohol related gene regions, as we are less certain that alcohol is the only factor that differs between the genetically-defined groups. For example, an analysis including 57 genetic variants from various gene regions suggested a protective effect of alcohol consumption on endometrial cancer. This provides lower quality evidence than an investigation based on variants in genes whose functional relationship to alcohol is understood.
Conclusion
Mendelian randomization is becoming a ubiquitous tool in epidemiological research. In some cases, where we can find genetic variants that shift the distribution of a risk factor in a specific way, it can provide reliable evidence about potential causal effects. In other cases, it is not an appropriate tool. And in other cases, it provides some level of evidence, but that evidence is subject to limitations and caveats. Answering complex questions about causal factors influencing disease risk typically requires bringing together imperfect evidence from many sources. Mendelian randomization is often part of that picture.
Written by Dr Stephen Burgess, MRC Biostatistics Unit, University of Cambridge.
All IAS Blogposts are published with the permission of the author. The views expressed are solely the author’s own and do not necessarily represent the views of the Institute of Alcohol Studies.