Re: Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data
I do not believe that this conclusion drawn by the authors is substantiated by the data analysis: '… reduction of alcohol consumption, even for light to moderate drinkers, is beneficial for cardiovascular health.'
This conclusion has been widely reported in the media e.g. Mail Online 15/07/2014:
'Moderate drinking IS bad for your health: Just two glasses of wine a day can cause problems
As few as 12 units a week still has negative effect on health say scientists
Drinking less reduces the risk of heart disease and lowers blood pressure
Challenges claims moderate drinking could have protective effect on heart' http://www.dailymail.co.uk/health/article-2688161/Moderate-drinking-IS-b...
The issue here is whether there is or is not a U-shaped relationship between alcohol consumption and the various measures of cardiovascular health.
I believe that the authors had an opportunity to examine this hypothesis but failed to do so satisfactorily in this article.
I refer here to: ' Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data Supplementary Appendix'.
'Figure S3 Observational association between alcohol intake and traits' plots the fitted relationship for traits. Six of these are behavioural, potentially causal, factors (Ever smoker, Cigarettes/day, Pack years, Education, Physical activity and Binge drinker). Excluding these leaves 24 measures of health status. Of these 24 traits, there is a U-shaped relationship for 16 i.e. the minimum (least adverse) level of these traits occurs at some point above zero alcohol intake. The intake at which the minimum occurs varies as follows (these are not exact values as they have been read off Figure S3):
2.5 8.4 6.4 1.7 3.3 5.8 5.8 4.4 12.9 3.3 3.3 3.3 8.8 4.4 3.6 5.8
which have range (2.5-12.9), mean 5.2 and standard deviation 2.9.
In order to explore whether the odds ratio for each measure of cardiovascular disease varied by alcohol intake, the samples were divided into three strata: (claimed) non-drinkers, light/medium drinkers light to moderate drinkers (>0 to less than 21 units/week) and heavy drinkers (21+ units/week). The largest odds ratios were observed for non-drinkers and the smallest for heavy drinkers. It is on this basis that the authors concluded that there is a monotonically increasing relationship between intake and cardiovascular disease.
The fault in this analysis is that it omits to explore the region in which the U-shaped relationships are observed. The middle stratum should have been divided further, such as into (>0 to less than 3), (3+ to less than 8) and (8+ to less than 21). The middle of these three substrata may then show larger odds ratios, which in the article have been averaged with the surrounding groups, including the apparently less favourable (10+ to less than 21) group.
Thus it can be seen that the current analysis does not support abandoning the U-shaped hypothesis in respect of cardiovascular disease.
This is my main concern, and it is independent of the use of Mendelian randomisation; however, I have a second concern regarding use of the ADH1B rs1229984 A-allele variant as an instrumental variable.
The authors state:
'Likewise, this stratification by alcohol consumption will also serve to validate the ADH1B rs1229984 A-allele variant as a specific instrument for alcohol consumption, as it is expected that in non-drinkers carriage of the rs1229984 A-allele variant will have no effect on cardiovascular traits or events, or a substantially attenuated effect given the known difficulty in correctly classifying long term non-drinkers from self reported questionnaires.'
This is indeed good justification for claimed non-drinkers to form a separate stratum for analysis, but validation depends on whether there is 'no … or a substantially attenuated effect' for the various measures among non-drinkers. Evidence is presented in Figure 1 of the main article. The 95% confidence interval for the odds ratio includes unity for five of the measures: hypertension, C reactive protein, waist circumference, non-HDL cholesterol and triglycerides; it does not include unity for the other three measures: systolic blood pressure, interleukin 6 and BMI. So the odds ratio is significantly different from unity for these three measures, which is not a negligible proportion of the eight, but is there at least an attenuated effect for each of these? Figure 1 suggests not: for all three, the odds ratio is further from unity for non-drinkers than for light to moderate drinkers. Note that all three measures exhibit U-shaped relationships with intake in Figure S3. The expectation regarding non-drinkers is therefore not fulfilled for three of these eight measures. I am concerned that the measures for three apparently important causal factors for cardiovascular disease differ so significantly between carriers and non-carriers of the rs1229984 A-allele and that this is not investigated further in the article.
I do not believe that this conclusion drawn by the authors is substantiated by the data analysis: '… reduction of alcohol consumption, even for light to moderate drinkers, is beneficial for cardiovascular health.'
This conclusion has been widely reported in the media e.g. Mail Online 15/07/2014:
'Moderate drinking IS bad for your health: Just two glasses of wine a day can cause problems
As few as 12 units a week still has negative effect on health say scientists
Drinking less reduces the risk of heart disease and lowers blood pressure
Challenges claims moderate drinking could have protective effect on heart'
http://www.dailymail.co.uk/health/article-2688161/Moderate-drinking-IS-b...
The issue here is whether there is or is not a U-shaped relationship between alcohol consumption and the various measures of cardiovascular health.
I believe that the authors had an opportunity to examine this hypothesis but failed to do so satisfactorily in this article.
I refer here to: ' Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data Supplementary Appendix'.
'Figure S3 Observational association between alcohol intake and traits' plots the fitted relationship for traits. Six of these are behavioural, potentially causal, factors (Ever smoker, Cigarettes/day, Pack years, Education, Physical activity and Binge drinker). Excluding these leaves 24 measures of health status. Of these 24 traits, there is a U-shaped relationship for 16 i.e. the minimum (least adverse) level of these traits occurs at some point above zero alcohol intake. The intake at which the minimum occurs varies as follows (these are not exact values as they have been read off Figure S3):
2.5 8.4 6.4 1.7 3.3 5.8 5.8 4.4 12.9 3.3 3.3 3.3 8.8 4.4 3.6 5.8
which have range (2.5-12.9), mean 5.2 and standard deviation 2.9.
In order to explore whether the odds ratio for each measure of cardiovascular disease varied by alcohol intake, the samples were divided into three strata: (claimed) non-drinkers, light/medium drinkers light to moderate drinkers (>0 to less than 21 units/week) and heavy drinkers (21+ units/week). The largest odds ratios were observed for non-drinkers and the smallest for heavy drinkers. It is on this basis that the authors concluded that there is a monotonically increasing relationship between intake and cardiovascular disease.
The fault in this analysis is that it omits to explore the region in which the U-shaped relationships are observed. The middle stratum should have been divided further, such as into (>0 to less than 3), (3+ to less than 8) and (8+ to less than 21). The middle of these three substrata may then show larger odds ratios, which in the article have been averaged with the surrounding groups, including the apparently less favourable (10+ to less than 21) group.
Thus it can be seen that the current analysis does not support abandoning the U-shaped hypothesis in respect of cardiovascular disease.
This is my main concern, and it is independent of the use of Mendelian randomisation; however, I have a second concern regarding use of the ADH1B rs1229984 A-allele variant as an instrumental variable.
The authors state:
'Likewise, this stratification by alcohol consumption will also serve to validate the ADH1B rs1229984 A-allele variant as a specific instrument for alcohol consumption, as it is expected that in non-drinkers carriage of the rs1229984 A-allele variant will have no effect on cardiovascular traits or events, or a substantially attenuated effect given the known difficulty in correctly classifying long term non-drinkers from self reported questionnaires.'
This is indeed good justification for claimed non-drinkers to form a separate stratum for analysis, but validation depends on whether there is 'no … or a substantially attenuated effect' for the various measures among non-drinkers. Evidence is presented in Figure 1 of the main article. The 95% confidence interval for the odds ratio includes unity for five of the measures: hypertension, C reactive protein, waist circumference, non-HDL cholesterol and triglycerides; it does not include unity for the other three measures: systolic blood pressure, interleukin 6 and BMI. So the odds ratio is significantly different from unity for these three measures, which is not a negligible proportion of the eight, but is there at least an attenuated effect for each of these? Figure 1 suggests not: for all three, the odds ratio is further from unity for non-drinkers than for light to moderate drinkers. Note that all three measures exhibit U-shaped relationships with intake in Figure S3. The expectation regarding non-drinkers is therefore not fulfilled for three of these eight measures. I am concerned that the measures for three apparently important causal factors for cardiovascular disease differ so significantly between carriers and non-carriers of the rs1229984 A-allele and that this is not investigated further in the article.
Trevor Sharot B.Sc., M.Sc., C.Stat.
Competing interests: No competing interests