Papers published before 2000 describing epidemiological studies involving 100+ lung cancer cases were obtained from Medline and other sources. Studies were classified as principal, or subsidiary where cases overlapped with principal studies. Data were extracted on design, exposures, histological types and confounder adjustment. RRs/ORs and 95% CIs were extracted for ever, current and ex smoking of cigarettes, pipes and cigars and indices of cigarette type and dose–response. Meta-analyses and meta-regressions investigated how relationships varied by study and RR characteristics, mainly for outcomes exactly or closely equivalent to all lung cancer, squamous cell carcinoma (“squamous”) and adenocarcinoma (“adeno”).

287 studies (20 subsidiary) were identified. Although RR estimates were markedly heterogeneous, the meta-analyses demonstrated a relationship of smoking with lung cancer risk, clearly seen for ever smoking (random-effects RR 5.50, CI 5.07-5.96) current smoking (8.43, 7.63-9.31), ex smoking (4.30, 3.93-4.71) and pipe/cigar only smoking (2.92, 2.38-3.57). It was stronger for squamous (current smoking RR 16.91, 13.14-21.76) than adeno (4.21, 3.32-5.34), and evident in both sexes (RRs somewhat higher in males), all continents (RRs highest for North America and lowest for Asia, particularly China), and both study types (RRs higher for prospective studies). Relationships were somewhat stronger in later starting and larger studies. RR estimates were similar in cigarette only and mixed smokers, and similar in smokers of pipes/cigars only, pipes only and cigars only. Exceptionally no increase in adeno risk was seen for pipe/cigar only smokers (0.93, 0.62-1.40). RRs were unrelated to mentholation, and higher for non-filter and handrolled cigarettes. RRs increased with amount smoked, duration, earlier starting age, tar level and fraction smoked and decreased with time quit. Relationships were strongest for small and squamous cell, intermediate for large cell and weakest for adenocarcinoma. Covariate-adjustment little affected RR estimates.

Attention was restricted to epidemiological prospective or case–control studies published up to and including 1999, which involved 100 lung cancers or more, and which provided RR estimates for one or more defined major, cigarette-type or dose-related smoking indices. The “major indices” compare ever, current or ex smoking with never or non-current smoking, and refer to smoking of any product, cigarettes, pipes, cigars and combinations, or of specific types of cigarette. The “cigarette type indices” compare smokers of different types of cigarette – filter with plain, manufactured with handrolled and mentholated with non-mentholated. The “dose-related indices” concern amount smoked, age of starting to smoke, duration of smoking, duration of quitting, tar level, butt length or fraction smoked. Pack-years was not considered as it was felt more important to separate effects of extent and duration of exposure. Uncontrolled case studies were not included. There were no further exclusion criteria.

Between 1997 and 2001 potentially relevant papers were sought from Medline and Emtree searches, from British Library monthly bulletins, from files on smoking and health accumulated over many years by P N Lee Statistics and Computing Ltd, and from references cited in papers obtained, until ultimately no paper examined cited a paper of possible relevance not previously examined.

Master copies of all the papers in the study file were read closely, with relevant information highlighted to facilitate checking. Where multiple papers are available for a study, a principal publication was identified, although details described only in other publications were also recorded. Preliminary calculations and data entry were carried out by one author and checked by another, and automated checks of completeness and consistency were also conducted. RR/CIs underwent validation checks [15].

A pre-planned series of meta-analyses was conducted for various smoking indices for each of the three main outcomes (all lung cancer, squamous cell carcinoma, and adenocarcinoma) and also for some indices for two other outcomes (large cell carcinoma and small cell carcinoma). Nearest equivalent definitions are allowed for the three main outcomes, with the terms “squamous” and “adeno” used subsequently to distinguish these results from those specifically for these cell types. Each meta-analysis was repeated, based on most-adjusted RRs and on least-adjusted RRs. For each meta-analysis conducted, combined estimates were made first for all the RRs selected, then for RRs subdivided by level of various characteristics, testing for heterogeneity between levels.

All meta-analyses are restricted to records with available RR and CI values. The process of selecting RRs for inclusion in a meta-analysis must try to include all relevant data and to avoid double-counting. For a given analysis (e.g. of current cigarette smoking), several definitions of RR may be acceptable (e.g. cigarette smoking, or cigarette only smoking), so, for studies with multiple RRs, the one to be used is determined by a preference order defined for the meta-analysis. Preference orders may be required for smoking status, smoking product, the unexposed base, and extent of confounder adjustment. As the definitions of RR available may differ by sex (e.g. a study may provide RRs for any product smoking for males, but only for cigarette smoking for females), the RRs chosen for each sex may not necessarily have the same definition. Sexes combined results are only considered where sex-specific results are not available. Similarly RRs from a subsidiary study are only used where eligible RRs are unavailable from the principal study. When multiple preference orders are involved, the sequence of implementation may affect the selection, so preferences for the most important aspects, usually concerning smoking, are implemented first.

While full multivariable analysis of the data is considered beyond the scope of this report, meta-regression analyses were also carried out using the sets of RRs selected for the main meta-analyses for ever smoking and for current smoking. Following preliminary meta-regressions (not shown), a “fixed model” was fitted to examine the effect on the results of six different categorical variables (sex, location, start year of study, major study type, number of lung cancer cases and number of adjustment factors). Note that the number of lung cancer cases (in the study as a whole), which is referred to subsequently as “number of cases”, is used as an indicator of study size. The significance of each of these variables was estimated by an F-test based on the increase in deviance resulting from its exclusion from the basic model. A list of secondary variables was also defined (relating to more detailed aspects of location, outcome, study type and confounder adjustment, national cigarette tobacco type, the product smoked, the denominator used in the RR, use of proxy respondents, whether the study required 100% histological confirmation of lung cancer, whether the population studied worked in risky occupations, the age of the subjects, and the derivation of the RR) with the significance of adding each characteristic to the fixed model estimated by an F-test based on the increase in deviance. Fuller details are given in Additional file 1: Methods.

Additional tests of the relationship of lung cancer risk to various characteristics of interest were based on corresponding pairs of RR and CI estimates within the same study for the same definition of outcome and exposure, and deriving the ratio of the two RRs. Where the pairs involved independent sets of subjects, the variance of the ratio was also derived, and meta-analyses of the ratio were conducted. Where the pairs involved non-independent sets of subjects the numbers of ratios greater and less than 1 were compared using the sign test. Tests of independent pairs related to sex (males vs. females), age (oldest vs. youngest age group) and race (white people vs. non-white or black people). Tests of non-independent pairs related to level of adjustment (most-adjusted vs. least-adjusted), and to comparisons of product smoked (mixed smokers vs. cigarette only smokers, and vs. smokers of pipes/cigars only). Tests were always carried out for all lung cancer and ever/current smoking. For sex, additional analyses were conducted for current and for ever smoking, for squamous and adeno, and also within level of amount smoked. For level of adjustment, two sets of analyses were run. The first, relating to RRs for ever/current smoking were based on the most-adjusted/least-adjusted ratio, while the second, for highest vs. lowest RRs for age of starting to smoke, duration, years quit and tar level, compared RRs that were most- or least-adjusted for other aspects of smoking.

If possible, sex-specific results are included in the meta-analyses, with combined sex results included only if not. Though variation by sex was not significant in all the main analyses, risk estimates generally tended to be higher for males than females. This is supported by additional analyses comparing RRs within study for the same outcome and exposure definition. Somewhat higher RRs were found in males even in analyses where comparisons were made within the same levels of daily cigarette consumption (about 5, 20 or 45 cigs/day). Even so, the existence of somewhat higher RRs for males does not necessarily indicate any greater susceptibility, as it may reflect their increased exposure to occupational carcinogens, or other differences in smoking history such as greater duration of smoking or increased use of plain and higher tar cigarettes. It should be noted however that in prospective studies where smoking habits were determined at baseline, the greater tendency of males to quit during follow-up may cause bias in the reverse direction. It should also be noted that comparison of smoker/never smoker RRs for men and women does not take account of possible differences in risk between male and female never smokers, the base groups for these comparisons. A detailed overall assessment of this aspect is beyond the scope of this paper, and ideally would involve direct comparison of risk in male and female smokers, with detailed adjustment for age, smoking characteristics and major potential confounding variables. We note that Bain et al. [18] concluded, based on analysis of two large prospective studies and review of results from six other such studies, that “women do not appear to have a greater susceptibility to lung cancer than men, given equal smoking exposure”.

While it is clear that absolute risk of lung cancer rises markedly with age, both in smokers and never smokers, it is far less clear whether the smoker/never smoker RR also does. Predictions based on the multistage model [19] suggest that there should be a modest rise, but there is difficulty in establishing this, especially when the great majority of the studies do not give results by age. Possible effects of age were investigated in two ways. The first method (see Tables ​Tables66 and ​and9)9) was to compare RRs which were specific to subjects in specific age groups. Data here were limited for squamous and adeno, and for all lung cancer suggested a possible increase in RR with age for current smoking, but not for ever smoking. More reliable are the comparisons (described in results section O), of RRs for the highest and lowest age groups within study for ever/current smoking; between-study differences are automatically controlled for under this approach. These showed a 17% greater risk for the highest age group (95% CI 10% to 25%). Whether or not a RR was adjusted for age was considered as a characteristic in the meta-regression analyses, but it never added significantly to the fixed model for either ever or current smoking for any of the three outcomes.

Although RRs were entered onto the database, if available, there were few studies that provided such data. For eight studies which provided pairs of comparable RRs for ever/current smoking, there was no indication that RRs for white people differed systematically from those for black people (or non-white people). This, of course, does not rule out the possibility that absolute risks for white people and black people with similar smoking habits may differ. As our concern was only with RRs for smoking, and whether these vary by other characteristics, we have not attempted to collect data comparing absolute risk according to these characteristics, such as white/black RRs within never smokers, or within smokers. Detailed analysis and discussion of racial differences in lung cancer risk between black people and white people is therefore beyond the scope of this paper. Elsewhere Lee [20] points out that in the USA black men have a higher risk of lung cancer than do white men. However, interpretation of this difference in terms of effects of smoking is not straightforward for various reasons. Thus Lee notes that though black people are more often current smokers, are less likely to quit smoking, smoke cigarettes with a higher tar level, and have higher cotinine levels, all characteristics predictive of a higher risk of lung cancer, they are also less likely to have ever smoked, smoke fewer cigarettes a day and start to smoke later, all characteristics predictive of a lower risk. Also little or no difference in lung cancer rate is seen between black and white women. Black people are much more likely than white people to use mentholated cigarettes, but no evidence of a difference in lung cancer risk associated with mentholation was found, either in the present analysis or in other reviews [20,21].

A consistent tendency in our meta-analyses was for RRs to be highest in studies in North America, intermediate in Europe and lowest in Asia, particularly in China. There was no very clear evidence of a difference between European countries, or between other countries in Asia, though some of the analyses suggested relatively lower RRs in Greece and Turkey than in the rest of Europe, and higher RRs in India than in the rest of Asia. In an attempt to study a possible explanation for this difference we divided countries into three groups by national cigarette tobacco type. One was the countries (Australia, Canada, India, South Africa, UK and Zimbabwe) which typically use flue-cured Virginia tobacco, another was the countries (all except those in the other two groups) which typically use blended tobacco, and the third included Taiwan and China (countries which used both types quite commonly or where we lacked confirmed information). Including this variable into the meta-analyses did not consistently improve the prediction of our model, a finding which is consistent with the conclusions of other analyses we have conducted based on national data on lung cancer rates and smoking frequency [22]. There are, of course other possible explanations of the clear differences in lung cancer RRs between continents, including genetic differences, and differences in baseline rates of the disease.

Our meta-regressions generally showed a tendency for RRs to be lower in studies which started earlier. There may be a number of reasons for this, such as changes in the relative use of cigarettes and pipes or cigars, and improvement of study quality, with better standardization of questionnaires and definition of products smoked. However we consider the most plausible reason to be changes in patterns of uptake of smoking, with smokers in earlier born cohorts being less likely to have a lengthy smoking career than smokers in later born cohorts.

Though this was only clearly significant in the analyses of ever smoking for all lung cancer, there was a consistent tendency for RRs to be somewhat higher from prospective studies than from case–control studies. If this is a true effect, the explanation for it is unclear.

In order to limit the considerable amount of work needed, we limited attention to studies involving at least 100 lung cancer cases. Given that smaller studies would have contributed much less weight to the meta-analyses than would the studies that were included, we consider that this restriction unlikely to have any material effect on our conclusions. The meta-regression analyses did show a consistent tendency for RRs to be higher in larger studies, though this was only significant for ever smoking (all lung cancer p<0.001), squamous and adeno p<0.05). This tendency is in the opposite direction to that predicted from publication bias. The explanation is unclear.

There was consistent evidence that risk of lung cancer was higher for cigarette only smokers than for smokers of any product, and substantially higher than for smokers of pipes only, cigars only or pipes/cigars only. For current smokers, for example, RRs were 9.57 (7.90-11.59) for cigarettes only, as compared to 4.76 (3.44-6.59) for pipes/cigars only. Mixed smokers tended to have similar risks to cigarette only smokers. Interpretation of this finding is difficult as mixed smokers and cigarette only smokers may have a different total exposure to tobacco, as well as a different cigarette consumption. Data on the types of cigars or pipes smoked have not been recorded on the database, but the increased risk is evident in each continent. The results for pipes and cigars mainly apply to males and to RRs for all lung cancer. Though there are only limited results by histological type, it is interesting that there is no indication of an increased risk of adenocarcinoma for pipe and cigar smokers.

We have investigated the relationship of lung cancer risk to various indices of the dose–response relationship. We did not record data on our database for pack-years, as we wished to investigate the separate roles of daily amount smoked and duration of smoking. Indeed, previous work (e.g. [19,28]) has in fact suggested that pack-years is not a valid measure, as for example, smokers of 20 cigs/day for 40 years and smokers of 40 cigs/day for 20 years have very different smoking RRs despite their identical pack-years. For those indices that we did consider where there were substantial amounts of data – daily amount smoked, duration, age of starting to smoke, and time of quit (relative both to current smoking and to never smoking) – there was very clear evidence that greater exposure leads to greater risk, not only for all lung cancer, but also for squamous and adeno. The results by time of quit extend the observation that RRs in ex smokers are intermediate between those of never smokers and current smokers. Because dose–response results are expressed in categories of exposure which vary from study to study, there are difficulties in combining the evidence over studies. We have used two approaches. One is to consider the RR for the highest vs. lowest level of exposure (where highest and lowest refer to expected risk, so that early ages of starting, for example, are considered highest). The other is the key value approach where we consider categories including a specified level of exposure and not including another specified level. Both approaches have limitations. The highest vs. lowest approach will vary between study in the ratio of exposures considered, while the key value approach, although combining results relating to different exposures in different studies to a lesser extent, necessarily omits results from studies with broader categories while somewhat arbitrarily selecting or discarding RRs from studies with narrow categories. Work is ongoing on a third approach to fit a dose–response curve to the RRs and estimated dose mid-points of the categories for each study. This approach is complex, and was considered outside the scope of the current paper, which was more intended to summarize major features of the data. However, a future paper is planned which will describe the shape of these dose–response relationships including characteristics of the curves, such as the estimated time after quitting by which half the excess risk associated with continued smoking has disappeared. We note that, when considering RR for time of quitting, the problem of “reverse causation” needs to be taken into account, as evidenced by the data in Table ​Table1919 showing no decrease in risk compared to current smokers for quitters of about 3 years. Our analyses also showed that for all lung cancer, risk increased with increasing tar level and with increasing fraction smoked (or equivalently short butt length), data here being more limited and non existent by histological type. As noted earlier, when discussing cigarette type, the relationship with tar level is not an artefact of inappropriate adjustment for amount smoked [27], as has been claimed [26].

Almost a third of RRs used in meta-analyses were not directly available from the source or calculated directly from cross-tables of exposure by outcome, and required more complex methods to derive the required RR. It was reassuring that whether or not the RR was derived did not (with one minor exception) add predictive power to the main meta-regression models, suggesting that our extensive use of derived RRs caused no material bias.

The statistical analyses investigated the role of various characteristics on the estimated risk of all lung cancer, squamous and adeno in relation to ever and current smoking, but generally did not formally test the effect of exclusion of specific studies with extreme RRs or large weights. An exception was the case of study LIU4 for ever smoking and all lung cancer, this study not giving data for current smoking or by histological type. The two sex-specific RRs for this study together contributed 50.9% of the weight for the 328 available RRs from all the studies, and its exclusion increased the overall fixed-effect RR from 4.22 (95% CI 4.16-4.28) to 6.47 (95% 6.34-6.60). However there was little difference in the random-effects estimates, and in the meta-regression analysis the two LIU4 RRs did not produce unusual standardized residuals, suggesting that the relatively low RRs from this study (2.76, 2.69-2.83 for males, and 2.86, 2.77-2.95 for females), were due to the characteristics of the study included in the model (in particular that it was conducted in China) and not due to its unusual results. While there are other large studies, none involved nearly as many lung cancer cases as LIU4, and we feel it unlikely that excluding other specific studies would have had a major effect on our meta-analysis estimates or on our conclusions as to how RRs varied by exposure, outcome and study and RR characteristics.

We did not exclude studies on the basis of the population studied. However, most studies include subjects broadly representative of the general population. A small number of studies were conducted in miners or in other occupations with a known or suspected lung cancer risk, such as welding or foundry working. Risky occupation was considered as a characteristic in the meta-regression models but was never found to be an independent predictor of RRs associated with ever or current smoking.

Another source of bias is misclassification of smoking status. Random misclassification would dilute the association, as would any tendency for cases to deny or understate their smoking more than for the general population. Any tendency for current smokers to claim to be ex-smokers, as might happen in a study conducted in a clinical setting or where patients have been advised to stop smoking, would tend to inflate the risk for ex smoking. Adjustment for misclassification would be difficult, as denial rates are likely to vary by aspects of the study design, the way questions are asked, and also by sex, age, location and other demographic variables.