Strength of evidence on passive smoking and lung cancer is overstated

EDITOR—Hackshaw et al estimate a 26% excess risk of lung cancer (95% confidence interval 7% to 47%) in non-smokers who live with a smoker.).1 This estimate may be too high.

Their adjustment for smoking misclassification bias is inadequate for two reasons. Firstly, it is based on the comparatively low misclassification rates seen in American and British populationsand ignores evidence of far higher rates in Asian women (table)2^–4Secondly, it uses a new adjustment technique that takes no account of the varying relative risks from smoking in the 37 studies. Despite the fact that lung cancer has many known causes and the evidence that smoking and passive smoking are associated with higher exposure to many adverse lifestyle risk factors,).5 the authors' consideration of confounding is limited to fruit and vegetable consumption. The authors dismiss publication bias, using an argument that addresses only whether such bias might explain the whole observed association. They do not even mention recall bias, and they do not discuss implications of specific weaknesses of the studies. After their inadequate downward adjustment for bias and confounding, they adjust upwards for effects of non-spousal passive exposure to smoke. Such effects are inferred indirectly from cotinine data, and the direct evidence that exposure to smoke in the workplace, in social situation, and during childhood is not associated with risk of lung cancer is ignored.

Hackshaw et al also estimate an excess risk of 19% from passive smoking, by extrapolating from the risk in current smokers, and this may also be too high. The use of particulate matter rather than cotinine to calculate the relative exposure of passive and active smokers would reduce the excess risk estimate at least tenfold; allowance for a quadratic component to the dose-response wouldreduce it further.

Hackshaw et al also create a false impression of precision by using confidence limits that takeno account of the many uncertainties in the adjustment procedures used, and by emphasising the similarity of their estimates of 26% and 19%. If more appropriate adjustments were used for bias and confounding and different assumptions were made in the dose-response extrapolation, both estimatescould remain similar but be an order of magnitude or more lower. Indeed, bearing in mind the possibility of a zero threshold for carcinogenesis, one might even argue that the authors have not conclusively demonstrated that passive smoking has any effect on risk of lung cancer.

Risk extrapolation overestimates risk

EDITOR—In their article on the accumulated evidence on lung cancer and environmental tobacco smoke, Hackshaw et al state that the direct (26%) and indirect (19%) estimates of the excess risk are similar.).6 The direct risk estimate is based on a meta-analysis of 37 epidemiological studies, whereas the indirect risk estimate results from a linear extrapolation of the risk in smokers, on the basis of the concentrations of nicotine and cotinine in body fluids of smokers and non-smokers.

We believe that the risk extrapolation presented by Hackshaw et al significantly overestimates the excess risk, mainly because they use risk and smoking dose in two different groups of smokers.They assume a relative risk of lung cancer related to smoking of about 20, which applies for heavy smokers. Seven cohort studies performed in five Western countries reported relative risks for male cigarette smokers (as a group) of 7.0 to 14.9.).7 Mean concentrations of nicotine and cotinine in body fluids, however, on which the risk extrapolation is based, are derived from male and female smokers with a wide range of smoking habits, including light smoking. Furthermore, the weighted dose ratio of 1% is not based on independent data: some groups of smokers and non-smokers appear twice or four times in the dataset. With the data listed in table 5,).6 we calculate a weighted average for the dose ratio of 0.8%, instead of the reported 1.0%.

Probably the most accurate data on serum cotinine concentrations of the general population wereobtained during the third national health and nutrition examination survey in the United States.).8 In this study a highly sensitive analytical method with an extremely low threshold of detection for serum cotinine of 0.050 ng/ml was used. From the graphically presented distributions of the cotinine concentrations, we calculated arithmetic means of 1.05-1.15 ng/ml and 180-215 ng/ml for those reporting exposure to tobacco smoke at home or work and tobacco use, respectively, yielding a dose ratio of 0.49-0.64%. With the above mentioned range of the relative lung cancer risk for “average” cigarette smokers, an excess risk of 3-10% for environmental exposure to tobacco can be extrapolated, which is three to nine times lower than the risk estimate (26%) based on epidemiological studies.).6

We therefore believe that there is a discrepancy rather than a similarity between the risk of lung cancer related to environmental tobacco smoke that is derived from epidemiological studies andthat which is extrapolated from the biochemically determined exposure dose. Evaluation of the biomonitoring data on genotoxic substances for non-smokers exposed to environ- mental tobacco smoke results in a similar conclusion.).9

Accumulated evidence on lung cancer and environmental tobacco smoke

EDITOR—Hackshaw et al).10 define the outcome of their study as “relative risk of lung cancer in lifelong non-smokers according to whether the spouse currently smoked or had never smoked.” In fact, for the pooled variable, case-control studies used ever exposure to smoking cohabitant(s), whereas cohort studies mostly used ever exposure at start of follow up. The distinction is important if exposure varies and timing plays a part.

Using data from a Swedish population based case-control study,).11 we simulated possible bias from misspecification of dichotomous variables regarding exposure to tobacco smoke from the spouse or at work, or both, and time since cessation of exposure. The table illustrates the potential bias introduced by using never/ever variables for exposure from the spouse or at work if only exposure from that single source during the past 20 years is aetiologically relevant, with a true relative risk of 2.0. 

After people give up smoking, much of their excess risk disappears within 10-20 years).12 and it has recently been suggested that this is also true forthe risk from passive smoking.).11The simulated bias increases, particularly for women and when a spousal exposure variable is used, if relevant exposure morerealistically occurs from both sources but is defined by never/ever exposure to one source (table (bottom)). Here, exposure to environmental tobacco smoke from work, spouse, or both within the relevant time is considered valid. This assumption is reasonable, as recent data show an increasing importance of exposure to tobacco smoke at work compared with exposure at home with regard to prevalence, duration, and intensity, in many countries.).13).14 Since early studies often investigated populations with a higher prevalence of smoking among men than among women and lower occupational rates among women, increasing problems with alternative sources of environmental tobacco smoke and cessation of spousal exposure for women may partly explain why the relative risks for the pooled variable tend to decrease over time.).10

Our data also suggest that sex specific exposure patterns could contribute to sex differences in observed relative risk for a never/ever variable.).10 Although ever spousal exposure is more common among women in our study, a greater proportion of ever exposed men have been exposed recently, which is also true for exposure at work.).11Important differences in timing and sources of exposure may exist between sexes in other studies.

The pooled variable may thus be more heterogeneous than Hackshaw et al made clear. On the basisof our considerations and supported by their dose-response estimates,).10 true current exposure could well convey a higher risk than the estimate reported by Hackshaw et al.

Authors' reply

EDITOR—We accept the last footnote to Lee's table (the correction increases the adjusted estimate of lung cancer risk from 1.26 to 1.27). We disagree with his other points.

(1) Misclassification rates in the Asian data from Lee's table are implausibly high. They wouldmake the pooled unadjusted estimate of excess risk in the 17 Asian studies).15 two to four times higher than the 20 American and European studies. But they were identical (23%).

(2) The adjustment technique was that previously used,).16 modified by Lee himself (personal communication). We used Western smoking/lung cancer relative risk estimates. Using lower estimates from countries with shorter duration of smoking would reduce the effect of the bias.

(3) Confounding factors must be linked to both lung cancer and smoking(active or passive); thisleaves only diet, which cannot explain the association.

(4) Publication bias (against small negative studies) cannot be excluded completely but does not explain the excess risk because of the clear association in large studies.

(5) Cohort studies are not subject to recall bias; they yielded a similar risk estimate to case-control studies. Study quality is important but not easily characterised. Studies with different strengths and weaknesses yield strikingly consistent overall results.

(6) Categorising people by their reported exposure outside the home only is unsatisfactory; it is invalidated, variable, and likely to reflect a very small average exposure.

(7) Tobacco particulates cannot be measured in body fluids. Cotinine (from nicotine) is a validated marker of exposure in non-smokers married to smokers and shows that they have higher exposure levels both inside and outside the home.

(8) Confidence intervals indicate statistical precision; sensitivity analyses show that bias and confounding could not reasonably explain the association.

Contrary to Scherer and Heller's opinion the 20-fold relative risk between cigarette smoking and lung cancer is reasonable—it was 22.4,).17 and 18.518 (second 20 year follow up). Several studies show that cotinine levels in non-smokers exposed to environmental tobacco smoke are about 1% of those in active smokers. Scherer and Heller's estimate from one study (about 0.5%) still indicates exposure and expected risk).19 it is lower probably because over 40% of “exposed” non-smokers were children.

Lee, Scherer and Heller have not considered the evidence as a whole and ignore important evidence. Non-smokers inhale carcinogens from tobacco smoke. Carcinogens in general have no threshold. Non-smokers exposed to environmental tobacco smoke show an excess risk unexplained by bias, commensurate with the extent of exposure, with a dose-response relation. The only reasonable conclusion is that breathing other people's smoke causes lung cancer.

We agree with Nyberg and Pershagen; some studies included spouses who were former smokers, which tends to underestimate risk of lung cancer.