Low cigarette consumption and risk of coronary heart disease and stroke: meta-analysis of 141 cohort studies in 55 study reports

In response to Nord, although our estimated relative risk (~1.5) for cardiovascular disease associated with smoking about 1 cigarette per day (CPD) has been shown by others before our article, there seems to be a view that our results were completely driven by the lowest categories of smoking used in our meta-analysis (i.e. some were wide and could include smokers who consumed up to 19 per day). The lowest category range did vary between studies, which is why we provided Supplementary Table B in our paper focussing only on those where the lowest category was up to 5 CPD (7 CPD for one study). However, the ultimate consideration is whether the analyses and conclusions are materially affected.

Below are the number of studies according to the lowest category of smoking reported in all 49 papers on coronary heart disease, along with the published relative risk (RR):
1-3 CPD: n=1, RR=1.63
1-4 CPD: n=3, median 2.74 (range 1.12-2.94)
1-5 CPD: n=1, RR=1.88
1-7 CPD: n=2, RR=1.24, 1.47
1-8 CPD: n=2, RR=1.01, 2.33
1-9 CPD: n=18, median RR=1.65 (range 1.04-3.19)
1-10 CPD: n=2, RR=1.63, 1.90
1-14 CPD: n=13, median RR=1.69 (range 0.95-2.76)
1-15 CPD: n=1, RR=1.61
1-19 CPD: n= 6, median=1.34 (range 1.16-2.30)

Most RRs are high (32 [65%] of the 49 have RR>1.5, and 12 [24%] have RR>2.0), reflecting that they represent people who smoke several CPD. Therefore, when using regression to estimate the RR at 1 CPD, the result will typically be lower than these reported values.

In our paper, we used a regression between logRR and CPD, for each cohort study, to produce a model from which the RR for 1 CPD can be estimated. These values were then combined across the studies in a meta-analysis. We repeated these analyses now, but for the lowest smoking category we used the upper limit minus 1 instead of the midpoint as the independent x-value for the regressions (e.g. 3 CPD for the category 1-4, 7 CPD for the category 1-8, and 8 CPD for 1-9). By assigning the observed RR to almost the highest consumption in each category, this deliberately biases the results away from a high RR at 1 CPD, and helps address the potential concern that some smokers under-estimate their consumption in this low category. Furthermore, in the re-analysis we only used studies where the lowest smoking category was up to 9 CPD.

Among the subset of 14 studies of men, the estimated RR for 1 CPD is 1.48 (95%CI 1.21-1.80) without the deliberate bias (RR=1.48 using all 26 studies of men), which becomes 1.37 (95%CI 1.11-1.69) with the bias. Similarly, among 10 studies of women, the RR for 1 CPD is 1.59 (95%CI 1.22-2.07) without the bias (RR=1.57 using all 18 studies of women), which becomes 1.44 (95%CI 0.99-2.09) with the bias. Therefore, even if we assume that most smokers in the lowest smoking category consume almost the maximum level of that category in all of the studies included, our results are still consistent with our originally reported findings, and the conclusions remain the same.

We draw attention again to some important points. First, our estimates (above and in the main paper) match those from sophisticated modelling of raw data within a large cohort study, where the RR for 1 CPD was 1.53.[1] This should help address comments about ‘extrapolating’ down to 1 CPD in our analysis (even though all the categories in the studies we used must, by definition, include people who smoke 1 CPD), but unfortunately not acknowledged in the BMJ rapid responses. It is also worth noting another individual study where RR=2.78 (95% CI 1.49-2.18) among people who consistently smoked <1 CPD during their lifetime; an unrealistic high point estimate and wide interval due to small number of events but the lower limit is 1.49.[2] Second, because the RR of heart disease associated with secondhand smoke among lifelong never-smokers is about 1.3 (a key piece of evidence used to justify smoking bans in public places internationally), it seems implausible that regular active smoking of 1 cigarette each day can have a RR that is substantially lower than 1.3, and it should be higher; whatever the exact shape of the dose-response relationship. This is indeed shown using several cohort studies, where secondhand exposure has smaller RRs than light active smoking (1-3 CPD).[3] Third, the RR for cardiovascular disease among smokers is known to be under-estimated because the reference group in nearly all studies contains never-smokers exposed to secondhand smoke.[4] Therefore, the RRs we extracted from the papers, and consequently our modelled ones, would be under-estimated to some extent.

In the paper, we explicitly used the word ‘about’ 1 CPD in the Abstract and Summary box, given what we say at the start of the Discussion section. However, our estimates for 1 CPD, as part of light smoking, are well-supported by other evidence (as discussed above and in the full article).

The non-linear shape (steep at low levels) between exposure and risk of heart disease has been accepted by expert groups, including the US Surgeon General’s reports on tobacco and health,[4,5] and shown by others for stroke at low exposure when examining secondhand smoke.[6] Because Nord and his colleagues believe that this shape might be incorrect, it would seem reasonable for them to use raw/individual data from large datasets to investigate it further for themselves, as there is no analysis using summary results from the cohort studies that can address their concerns to their satisfaction.

Allan Hackshaw
Dusan Milenkovic

1. Pope et al. Lung cancer and cardiovascular disease mortality associated with ambient air pollution and cigarette smoke: shape of the exposure-response relationships. Environ Health Perspect. 2011;119(11):1616-21

2. Inoue-Choi et al. Association of Long-term, Low-Intensity Smoking With All-Cause and Cause-Specific Mortality in the National Institutes of Health-AARP Diet and Health Study. JAMA Intern Med. 2017;177(1):87-95
3. Pope et al. Cardiovascular mortality and exposure to airborne fine particulate matter and cigarette smoke: shape of the exposure-response relationship. Circulation. 2009;120(11):941-8

4. The Health Consequences of Smoking: 50 years of progress. A Report of the Surgeon General. Atlanta: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, Coordinating Center for Health Promotion, National Center for Chronic Disease Prevention & Health Promotion, Office on Smoking and Health. U.S. Department of Health & Human Services, 2014.

5. How tobacco smoke causes disease. The biology and behavioural basis for smoking-attributable disease. A Report of the Surgeon General. Atlanta: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, Coordinating Center for Health Promotion, National Center for Chronic Disease Prevention and Health Promotion, Office on Smoking and Health. U.S. Department of Health and Human Services, 2010

6. Oono et al. Meta-analysis of the association between secondhand smoke exposure and stroke. J Public Health (Oxf). 2011;33(4):496-502

I was not expecting to continue this exchange, but two claims by professor Cole call for correction.

Let me first say that when the purpose of the analysis is to document a disproportionately steep increase in risk from smoking 1 CPD rather than being a non-smoker, it is a priori strange to exclude non-smokers from the regression.

Cole seems to confuse ‘is’ and ‘ought’. It IS a fact that Hackshaw et al truncated the regression analysis at 1 CPD. It does not follow that I am wrong when I say that they SHOULD not have done so. If they had included zero, any number of regression lines from 3 CPD down to 0 CPD (all starting at 3 CPD with the slope observed between 20 CPD and 3 CPD) would have fitted equally well mathematically, given that there are no observations between 0 and 3. The various possible regression lines – having different slopes from zero - would have yielded a variety of estimates of risk at 1 CPD, including some much lower than the estimate published by Hackshaw et al. It would have become entirely clear that the specific choice of regression line made by Hackshaw et al did not follow from their data and was no more plausible (and to many people probably less plausible) than other candidate lines. Does Cole deny all this?

Cole claims that ‘non-smokers cannot be plotted on the dose-response curve for smokers’. I wonder what he means by this. If the regression had included zero, non-smokers’ risk would have been the starting point of the regression line. If he means that the dose-response curve is not linear all the way from zero (i.e. that risk is not proportional to consumption), this is not a ‘novelty of Hackshaw et al’s analysis’. Bjartveit and Tverdal showed that many years ago, and I obviously did not assume such proportionality when I in my previous response explicitly accepted to build on the ‘disproportionately high risk at 3 CPD’.

Cole earlier devoted a whole response to defending extrapolation from 2 CPD after I had documented that there were no observations below 3 CPD (in terms of means in subgroups). I was puzzled then, and am no less puzzled now by the fallacies above and at Cole’s complete unwillingness to acknowledge that the ‘one cigarette per day’ issue should not have been the theme of the paper. Is the BMJ showing sufficient care and objectivity when considering my criticisms?

I thank Professor Cole for a clarifying, albeit unconvincing response.

According to Hackshaw et al, the reviewed studies suggest a linear dose-response relationship between cigarette consumption and cardiovascular risk in the interval from 3 to 20 CPD. Assuming this is is correct (see alternative assumption below), I agree with Cole that when extrapolating risk downwards from 3 CPD, it is reasonable to start with extending the straight line between 20 and 3 CPD. But given the disproportionately high risk at 3 CPD, the extrapolation cannot continue to be linear all the way down to zero. At some point the straight line has to begin to curve downwards (in order for RR to be 1 when consumption is zero). For estimation of risk at 1 CPD it is decisive where the departure from linearity starts.

Close inspection of table 1 shows that in most subgroups Hackshaw et al for some reason applied a linear extension downwards from 5 CPD with less steepness than the one between 5 CPD and 20 CPD. This is counterintuitive, the opposite of ‘well-evidenced’ (Cole’s term) and contributed to a high estimate at 1 CPD.

Hackshaw et al furthermore chose not to force the regression line through the origin. This allowed them to stay with linearity all the way down to 1. Had they instead chosen to depart from linearity somewhere between 2 and 3 CPD the resulting estimated risk ratio at 1 CDP might have been half of what they found or even less. A priori the latter alternative seems at least as plausible as the former. Hackshaw et al did not consider it, and unlike what readers of the article are led to believe in figures 1, 3 and 4, their data do not at all speak to this question.

Additionally, should a slightly curved regression line (convex upwards) between 3 CPD and 20 CPD yield a better fit than the straight line, the extension could reach 1 CPD at an even lower risk level than what I indicate as possible above.

Cole gives three reasons for accepting a linear extension all the way to 1 CPD. First, he notes that the extrapolation range is less than a tenth of the total data range. This argument must be weighed against the fact that linear extension to 1 CPD forces an intuitively strange sharp fall in the curve at 1 CPD. Second, Cole argues that ‘the lowest consumption categories include 1 CPD, so they are already represented’. This argument is irrelevant given that all analysis is based on the means in the various intervals. Third, Cole claims that ‘it is biologically implausible that the dose-response curve should be linear from 20+ down to 3 CPD, but veer materially from the line between 3 and 1 CPD’. This argument is countered by my first point: At some level between 3 and 0 CPD there has to be a downward bending of the curve, and no level between 3 and 1 is more biologically plausible than others.

I conclude that the value of Hackshaw et al’s study lies in the evidence it provides for estimating (a) the relative risk at 3 CPD, (b) the dose-response relationship between 3 and 20 CPD and thus (c) the slope of a plausible extension line at the point where this line starts, i.e. at 3 CPD. The severe mistake of Hackshaw et al was to push the ‘one cigarette per day’ issue and make this the main message. Their claims in figures 1, 3 and 4 as to what each of 37 studies of risk at moderate and high consumption levels imply for risk at 1 CPD are absurd. When focussing on risk at 1 CPD they went from evidence based estimation to non-evidence based choice between a number of possible alternatives, some of which are more alarming and sensational than others. They chose the highest conceivable estimate. This is deeply unsound scientific practice. It seems from the last paragraph in their introduction that they were ‘on a mission’. It is reasonable to think that this affected their choice and explains their failure to discuss more or at least equally plausible alternatives. If they had given a detailed account of the observed consumption levels in their included studies it seems unlikely that the reviewers would have allowed the ‘one cigarette per day’ issue to become the main theme of the article.

After the publication of Hackshaw et al’s paper we must expect many claims in tobacco control in the future to the effect that surprisingly high cardiovascular risk at only one cigarette per day is ‘well documented’. From a scientific point of view it is sad to know that the BMJ allows this to happen.

In a response to my report on Hackshaw et al's data, Cole et al, on behalf of the BMJ, wrote that the estimation of risk at 1 cigarette per day was an 'extrapolation' from observations at 2 CPD. Zahl corrects Cole et al by saying that the extrapolation was from 3 rather than 2 CPD. The correction is highly incorrect and misleading. In 30 out of 34 studies, the observations of 'low consumption' pertained to consumption in the order of 8-14 CPD. In each of these 30 studies a risk at 1 CPD was estimated. I am saying that this is absurd and that this property of the evidence should have been reported by the authors themselves and thus made clear to the reviewers. Could Cole et al please comment on this rather than on non-existent data at 2 CPD?