Systematic review of the relation between smokeless tobacco and cancer in Europe and North America

Background Interest is rising in smokeless tobacco as a safer alternative to smoking, but published reviews on smokeless tobacco and cancer are limited. We review North American and European studies and compare effects of smokeless tobacco and smoking. Methods We obtained papers from MEDLINE searches, published reviews and secondary references describing epidemiological cohort and case-control studies relating any form of cancer to smokeless tobacco use. For each study, details were abstracted on design, smokeless tobacco exposure, cancers studied, analysis methods and adjustment for smoking and other factors. For each cancer, relative risks or odds ratios with 95% confidence intervals were tabulated. Overall, and also for USA and Scandinavia separately, meta-analyses were conducted using all available estimates, smoking-adjusted estimates, or estimates for never smokers. For seven cancers, smoking-attributable deaths in US men in 2005 were compared with deaths attributable to introducing smokeless tobacco into a population of never-smoking men. Results Eighty-nine studies were identified; 62 US and 18 Scandinavian. Forty-six (52%) controlled for smoking. Random-effects meta-analysis estimates for most sites showed little association. Smoking-adjusted estimates were only significant for oropharyngeal cancer (1.36, CI 1.04–1.77, n = 19) and prostate cancer (1.29, 1.07–1.55, n = 4). The oropharyngeal association disappeared for estimates published since 1990 (1.00, 0.83–1.20, n = 14), for Scandinavia (0.97, 0.68–1.37, n = 7), and for alcohol-adjusted estimates (1.07, 0.84–1.37, n = 10). Any effect of current US products or Scandinavian snuff seems very limited. The prostate cancer data are inadequate for a clear conclusion. Some meta-analyses suggest a possible effect for oesophagus, pancreas, larynx and kidney cancer, but other cancers show no effect of smokeless tobacco. Any possible effects are not evident in Scandinavia. Of 142,205 smoking-related male US cancer deaths in 2005, 104,737 are smoking-attributable. Smokeless tobacco-attributable deaths would be 1,102 (1.1%) if as many used smokeless tobacco as had smoked, and 2,081 (2.0%) if everyone used smokeless tobacco. Conclusion An increased risk of oropharyngeal cancer is evident most clearly for past smokeless tobacco use in the USA, but not for Scandinavian snuff. Effects of smokeless tobacco use on other cancers are not clearly demonstrated. Risk from modern products is much less than for smoking.


Background
Over the last 10 years, interest in smokeless tobacco (ST) as a possible safer alternative to smoking has risen. Although a number of recent reviews have considered the evidence relating ST to cancer, some have not included metaanalyses [1][2][3], and others have only provided quantitative summaries for specific sites: oropharyngeal cancer [4], pancreatic cancer [5], or oropharyngeal, oesophageal, pancreatic and lung cancer [6]. No formal comparisons have been conducted with the well-known effects of smoking [7,8].
The review described in this paper is restricted to studies in Western populations. In practice this predominantly means studies in the USA and Sweden, the only North American and European countries where the two major types of STchewing tobacco and snuffare commonly used [2]. Although ST is also widely used in developing countries, particularly parts of Central and South-East Asia, the tobacco is often used in combination with other products, such as betel nut quid, slaked lime, areca nut and even snail shells [1,2,9]. This review also does not consider the limited data on nicotine chewing gum.
Our first objective is to carry out a comprehensive review of the available epidemiological evidence in Western countries relating ST to cancer, including meta-analyses for as many cancer types as the data justify. In meeting this objective, we take proper account of the potential confounding role of smoking by distinguishing effect estimates which are unadjusted for smoking and those which take smoking into account (either by adjustment in analyses based on the whole population of smokers and non-smokers combined or by restricting analysis to lifelong never smokers). Our second objective is to provide a quantitative indication of the relative effects of ST and cigarette smoking.

Study identification and selection
All reports had to satisfy the following inclusion criteria: published in a peer reviewed journal or the results publicly available, epidemiological study in humans, of cohort or case-control design, study location specified, any form of cancer as the outcome, and chewing tobacco, oral snuff or unspecified ST as the exposure. They also had to fall outside the exclusion criteria: conducted in an Asian or African population, no control group, or inappropriate design (case report, qualitative study or review/metaanalysis). Relevant papers were sought from a MEDLINE search conducted in May 2008 of "cancer" AND ("smokeless tobacco" OR "chewing tobacco" OR "snuff" OR "snus"), supplemented by citations in recent reviews [1][2][3][4][5][6]10] and in the papers obtained.

Data extraction
Reports were grouped by study, and for each study details were abstracted (see Tables 1 and 2  ) relating to the design, period, location, controls used and size, the exposure (method of assessment, type of ST, exposure doses and durations considered), the outcome (cancer sites studied) and issues relating to analysis (type of effect measure, analysis methods, extent of adjustment for smoking and other factors, and availability of doseresponse data). The extent of adjustment for smoking for a study was categorised into five groups: A. no informationeffect estimates are provided but no details are given of any adjustments made; B. no adjustmenteffect estimates are available for the whole population, but smoking is not taken into account; C. never smokersthe only effect estimates available are for never smokers; D. some adjustmenteffect estimates adjusted for smoking are available, but the adjustment is relatively simple, using two or three level broad groupings (for example, ever/never smoked, current/non-current smoker, current/former/never smoker), and takes no account of daily amount smoked or duration of smoking; and E. more adjustmenteffect estimates are available that take into account daily amount smoked, duration of smoking and/or their product (pack-years). Studies were categorised under D or E if smokingadjusted effect estimates are available, regardless of whether some results for never smokers are also presented. The method used to adjust for smoking is not always clear. Studies where the authors merely report that they 'adjusted for cigarette smoking' are included in category D.
Based on the availability of relevant data, 13 cancer groupings (oropharyngeal, oesophagus, stomach, pancreas, other digestive, larynx and nasal, lung, prostate, bladder, kidney, haematopoietic and lymphoid, other and all), were selected, with results for each grouping tabulated in a standard way, with details given of the source, exposure to ST, smoking group, sex, number of cases and adjustment factors for each effect estimate or indication of association (see tables dealing with individual effects estimates, below). For each study the intent is to extract the relative risk (RR) or odds ratio (OR) adjusted for the most factors, relevant to current, former or ever exposure to chewing tobacco, snuff or overall/undefined ST. Where relevant results for a study are reported in more than one paper, those based on the greatest number of cases are used.
Results are included, where available, for the whole population and for never smokers, and for sexes separately. RR or OR estimates based on zero exposed cases (or controls) are not included as providing too little information and because a valid confidence interval (CI) cannot be calculated. Suitable estimates of effect (RR or OR) and precision (CI) provided by the authors are used if possible, estimates otherwise being calculated from available data  [11] Prostate (149) Kneller et al. 1991 [12] Stomach (75) Zheng et al. 1993 [13] Pancreas (57)  Heineman et al. 1992 [16] Multiple myeloma (582) f Zahm et al. 1992 [17] Soft tissue sarcoma (119), pharynx (55), buccal cavity (74) Heineman et al. 1995 [18] Colon (3,812) [35] All (1,572), smoking-related (493), oral (34) p a Only exposures for which results are available are shown. b Main references. Other references supplying limited data are indicated in footnotes. c Numbers of cases are totals for the sexes specified. Numbers of cases exposed to ST are shown in the tables presenting results by site. Cases are deaths, unless indicated. Oral is used as an abbreviation for oropharnyx. d Some limited additional results for the Lutheran Brotherhood cohort, based on follow-up to 1981, were reported earlier for cancers of the prostate, pancreas and oesophagus in IARC Monograph 37 in 1985 [14]. e Some limited additional results for the US Veterans cohort, based on follow-up from 1954 to 1969 were presented earlier for a range of cancers in an abstract by Winn et al. in 1982 [19]. f Cancers listed are incident cases. g NHANES I = First National Health and Nutrition Examination Survey. h Data on ST use were only collected in 3,847 subjects at baseline in 1971-1975, but were collected for all subjects in follow-up surveys in 1982-1984. 6,805 subjects were considered in the mortality analyses [21] and 7,779 in the incidence analyses [22]. i Numbers of cases not given. j CPS-I = Cancer Prevention Study I. k CPS-II = Cancer Prevention Study II. Some additional results for lung cancer, based on mortality to 2002, comparing 111,952 men who quit cigarette smoking with 4,443 who switched to ST, were presented by Henley et al. in 2007 [25]. l Results for chewing and snuff are also given for all cancers and lung cancers. m Some limited additional results, based on follow-up to 1978, were reported by Heuch et al. in 1983 [27] for pancreatic cancer incidence and in IARC Monograph 37 in 1985 [14] for cancers of the buccal cavity/pharynx, oesophagus, pancreas and prostate. n Cancers listed include incident cases.°I ncludes cutaneous malignant melanoma, melanoma in situ and intraocular malignant melanoma. p Numbers are incident cases. An analysis of overall cancer based on 1,574 deaths was also conducted. ST = smokeless tobacco. presented in the source publication, based on methods [115][116][117][118] summarised elsewhere [4]. Where an effect estimate cannot be calculated, statements made by the authors are summarised into terms such as 'no association' or 'no significant association'. Data are summarised for all types of cancer, except those relating to subdivision by type within site (for example, adenocarcinoma or squamous cell carcinoma of the lung, or t(14; 18)-positive andnegative non-Hodgkin's lymphoma or those relating to combined 'other' groups of cancers, which typically vary in definition from study to study).

Data presentation
Study-specific results for the different types of cancer are presented in an essentially identical format, with a standard set of information included for each effect estimate included. Points to note about the entries in the various columns are discussed below.

Source
For the case-control studies, the source reference is shown. For the cohort studies, the source reference is also shown, but the study is also identified by name.

ST usetype
The exposure is identified as chewing tobacco ('chew'), 'snuff' or smokeless tobacco ('ST'). ST implies the results relate to smokeless tobacco unspecified by the author, or to use of either chewing tobacco or snuff or both.
ST useexposure Results are presented for current, former or ever use, or simply for 'use' where the timing of exposure was unspecified by the author. For current, former or ever use, the comparison is with never use; for use, it is with non-use.

Smoking
Results are presented only for any smoking (that is, based on the combined population of ever and never smokers) and for never smokers.

Sex
Results are shown, where available, for the sexes separately, though in some studies results are given only for the sexes combined. Within each table, each effect estimate (RR or OR) is given a unique identification number, so that those which are included in specific meta-analyses can readily be seen.

Cases
The number of ST-exposed cases is shown. Total numbers of cases are given elsewhere. Estimates are not presented unless there is at least one exposed case.

Estimate (95% CI)
This is the RR for cohort studies or the OR for casecontrol studies, together with its 95% CI. For many studies, the estimates are not given directly in the source paper, but were calculated from data provided. This involved one or more of the following: estimating numbers of exposed and unexposed cases and controls from proportions exposed given numerically or graphically and, where appropriate, combining numbers over level of exposure or cancer subtype; calculating estimates from a 2 × 2 table, or multiple independent 2 × 2 tables using standard methods [115], and calculating estimates from non-independent RR/ORs by level of exposure or by cancer type using the method of Hamling et al. 2007 [118]. Fuller details of the method of calculation used for each estimate are available on request. In a limited number of studies, as indicated in the tables, estimates were available separately for chewing tobacco and for snuff, but data were lacking for joint use. Here estimates for combined ST use were calculated assuming that no one used both chewing tobacco and snuff. Where there is a choice of relevant estimates from a study, preference is given to the estimate adjusted for the most potential confounding factors, and, for cohort studies, the estimate from the publication with a longer follow-up period.

Adjustment factors
The adjustment factors used for each estimate are shown. For matched case-control studies, the matching factors are not included unless the estimate specifically took this into account (for example, by conditional logistic regression). The factors used have been simplified into a relatively short consistent list, rather than repeating verbatim the wide variety of variable descriptions given by the original authors. Thus 'res' (area of residence) includes any variable based on the location of the subject and, for example, includes centre in multicentre casecontrol studies. 'Diet' includes any aspect of diet, and 'alc' (alcohol) any aspect of alcohol use. Estimates relevant to never smokers are not listed as being adjusted for smoking ('smok').

Layout
For the five columns, ST usetype, ST useexposure, smoking, sex and adjustment factors, any blank entry for a particular effect estimate is assumed to be the same as in the first previous non-blank entry in that column. This avoids needless repetition and makes the tables easier to read.

Meta-analysis
Estimates with no CI are not included in the metaanalyses. The standard error of the logarithm of estimates of effect size was calculated from its reported or estimated CI, assuming that the effect size was lognormally distributed. The logarithms of the effect sizes and their corresponding standard errors form the data points for fixed-effect and random-effects meta-analysis [116].
For most cancer groupings, results of nine random-effect meta-analyses are presented, subject to availability of data (see tables summarising meta-analysis results, below). In the first set of three, any, there is no restriction of estimates on type of exposure or region. In the second set, any ST use (USA), estimates are restricted to those from studies conducted in the USA (or on occasion in Puerto Rico), while in the third set, snuff (Scandinavia), estimates are restricted to those for snuff and for studies conducted in Scandinavia. Each of the three sets of metaanalyses is divided into overall data, smoking-adjusted and never smokers. In the overall data analyses, estimates are not restricted on smoking status or on adjustment for smoking. The smoking-adjusted analyses only include estimates that are for the whole population and adjusted for smoking or are for never smokers. The never smokers analyses are restricted to estimates for never smokers. For oropharyngeal cancer, where more estimates are available, some additional meta-analysis results are shown, based on estimates that are smoking and alcohol adjusted, and on estimates published since 1990.
To avoid double-counting multiple non-independent estimates from the same study, estimates from each study are selected for inclusion in the meta-analyses using order of preference lists for ST exposure (ever use/ unspecified use/current use/former use), then smoking status (anybased on the combined population of smokers and non-smokers/never smokers), and then ST type (ST/snuff/chew), with each list being in order of most to least preferred. At each step we retain those estimates highest up the list, discarding any estimate lower in the preference order. If the procedure ends up with separate estimates for males and for females, both are included in the analysis. In one study [36], the results available are for males for chewing and for females for snuff (see Table 3). Although the procedure, strictly applied, selects only the snuff estimate, it was decided to include both in the relevant meta-analyses.
The presentation of the meta-analyses shows the number of estimates combined; the identification numbers of these estimates (so that they can be related to the preceding table of individual effect estimates); the combined randomeffects estimate, with its 95% CI [116], the chi-squared and P value of homogeneity [119] and the I 2 statistic [120]. The meta-analyses conducted also include a test for publication bias [121] where five or more estimates are combined. Findings significant at P < 0.1 are indicated.
Forest plots are also included for most of the cancers. These are generally based on the smoking-adjusted analyses, with the estimates split by region and shown with cohort data first, then case-control, presented in order of publication year.

Sensitivity analysis
For each estimate included, the value of Q 2 is calculated by w (xx ) 2 , where w is the inverse-variance weight, x is the logarithm of the effect size and x its mean. Q 2 is the contribution of the estimate to the heterogeneity chisquared statistic [116]. Where there is significant (P < 0.05) heterogeneity of estimates, sensitivity to potentially outlying estimates is tested by removing that with the largest Q 2 value and rerunning the analyses. This process is continued until there is no longer significant heterogeneity.
Sensitivity to the criterion for including estimates based on ST exposure is also tested by rerunning the metaanalyses with the preference list for ST exposure changed from ever use/unspecified use/current use/former use to current use/ever use/unspecified use/former use.

Meta-regression analysis
For oropharyngeal cancer, fixed-effects regression analysis is used to investigate how the estimates selected for the first set of meta-analyses vary by region (USA; Scandinavia; other), period × study type (cohort; casecontrol published before 1990; case-control published after 1990), sex (male; female; combined), ST exposure (ever or unspecified use; current use), smoking (any, adjusted for smoking; any, unadjusted for smoking; never) and alcohol adjustment (yes; no). For those other cancers where more than five estimates are available and where there was evidence of significant (P < 0.05) heterogeneity, the meta-regression analyses use a more limited variable list: region, sex, and smoking as above, and also study type (cohort; case-control).
Regression analyses are only conducted based on the overall data and smoking-adjusted data. The analyses successively introduce the most significant factor into the model, stopping when no further factor significant at P < 0.05 can be added. Significance is estimated by treating the ratio of the deviance per degree of freedom (d.f.) explained by the factor to the residual deviance per d.f. as an F statistic. For oropharyngeal cancer some additional analyses investigate the drop in deviance resulting from introducing each factor individually, and others are conducted having excluded 'outlying' observations with a very high Q 2 value.
Estimating deaths attributable to smoking RRs for current and former cigarette smokers (compared with never cigarette smokers) for men aged 35+ for seven major cancers caused by smoking (lip/oral cavity/ pharynx, oesophagus, pancreas, larynx, lung, bladder, kidney/other urinary organs) were obtained from the American Cancer Society Cancer Prevention Study II (CPS-II) [122]. Numbers of deaths for these seven cancers occurring in US men aged 35+ in 2005 were obtained from WHO [123]. Estimates of the proportion of current and former cigarette smokers in US men aged 35+ in 2005 were obtained from the National Health Interview Survey [124].
Defining D i as the number of deaths for cancer i (i = 1,..., 7), R ci and R fi as the RRs for current and former cigarette smokers for cancer i, and p c and p f as the proportions of current and former cigarette smokers in the population, the estimated number of deaths, D i * , that would have occurred  had the whole population the risk of never smokers, is then estimated by: The number of deaths avoided from these seven cancers, had the whole population the risk of never smokers (that is, the deaths attributable to smoking) is then estimated by: Estimating deaths attributable to ST in a population of never smokers Let us further define R si as the estimated relative risk from ST for cancer i based on the meta-analyses using smoking-adjusted effect estimates. Where R si is estimated to be less than 1, it is taken to be 1 for the purposes of calculating deaths attributable to ST.
For a population of never smokers, the number of deaths from cancer i that would have occurred had the same proportion of men used ST as had ever smoked is then estimated by: The increase in overall deaths from these seven cancers is then given by: Also for a population of never smokers, the number of deaths from cancer i that would have occurred had all the men used ST, is estimated by:  Table 4, and 'Cases' is the number of cases in ST users as defined. NA = not available. e Abbreviations used: alc = alcohol, asp = aspirin, bmi = body mass index, edu = education, exer = exercise, ins = insurance status, occ = occupation, res = area of residence, resp = respondent, smok = smoking. f The population included < 0.5% females. g RR/OR and/or 95% CI estimated from data provided in the source. h The average ridit duration of chewing did not differ significantly from the controls for any type of oral cancer. i RR/OR and/or 95% CI estimated from data provided in the source assuming that no one both chewed and used snuff. j "Attributable oral cancer risk due to smokeless tobacco use based on a case-control study at Sinai Hospital in Detroit", Perry et al., unpublished. Cited by Gross et al. 1995 [91]. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk.
The increase, compared with E, is then calculated by:

Results
The MEDLINE search identified 690 publications. Two hundred and thirty-eight were rejected as describing studies conducted in Asia or Africa or relating to products typically used there, 96 as not describing epidemiological studies, 112 as not relating to cancer and 163 as being reviews, letters or comments not providing primary data. Seventeen were rejected as having an inappropriate study design and three as not providing relevant results. This left 61 apparently relevant publications. Taking into account also citations in recent reviews [1][2][3][4][5][6]10], and eliminating publications that referred to studies more recently or completely covered in other publications, a total of 104 publications were considered. Twenty-five related to nine cohort studies, and 79 to 80 case-control studies. Fuller details of the search are given in Figure 1, whilst the studies and publications considered are presented in the following two sections.

Cohort studies
Results relating ST use to mortality or incidence have been reported for nine cohort studies, with results provided by multiple publications for some studies. Six studies have been conducted in the USA and are based on the Lutheran Brotherhood cohort [11][12][13][14], the US Veterans cohort [15][16][17][18][19], the Iowa cohort [20], the First National Health and Nutrition Examination Survey (NHANES I) Follow-up cohort [21,22], and the American Cancer Society Cancer Prevention Study I (CPS-I) [23] and Study II (CPS-II) [23][24][25]. One study was based on two Norway cohorts [14,26,27] while the remaining two were conducted in Sweden; one based on construction workers [28][29][30][31][32][33][34], and the other on a cohort in Uppsala County [35]. Fuller details of these studies are given in Table 1

Figure 1
Flow chart for search strategy for review of literature on smokeless tobacco and cancer. The flow chart shows the number of publications identified by the MEDLINE search, and the number excluded by reason. The number of additional publications identified from reviews and secondary references is also indicated, as is the total number of publications considered in the review and meta-analysis, subdivided by study type.
particularly in the CPS-I and CPS-II studies where the analyses of Henley et al. [23] restricted attention to never smokers of cigarettes. The US studies generally present results for combined ST use, the main exception being the analyses of CPS-II [23] where some separate analyses are presented for snuff and chewing tobacco. The results from the Swedish studies relate to snuff use, as do the main results from the Norwegian study [26].

Case-control studies
Results relating ST use to cancer have been reported for 80 case-control studies, with Table 2 providing details for each study, in chronological order of publication, of the location, period and controls, as well as the exposures, cancer types and sexes studied. Eighteen were published between 1920 and 1975 [36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51][52], 30 between 1976 and 1990  and 32 between 1991 and 2007 . In general there was one publication per study, but Bjelke [52] reported results from two studies, while the reference to Gross et al. [91] is to a review, which cites results from an unpublished study by Perry et al. Of the 80 studies, 56 were conducted in the USA, 11 in Sweden, three in each of Canada, Denmark and the UK, and one in each of Brazil, Norway and Puerto Rico, with one study conducted in five countries. Most of the studies involve only one or a small number of cancer types, but one study [55] involves a very wide range. The majority of the studies involve less than 1,000 cancer cases but 10 are larger than this [38,55,56,66,70,74,89,97,99,106]. The numbers of cancers in ST-exposed subjects are typically much lower than this, as will become evident when the results for the individual sites are presented. Of the different cancer sites, oral cancer is by far the most often studied. Of the 56 US studies, 11 provide results only for chewing tobacco, five only for snuff, and 18 only for ST, with the remaining 22 results for more than one type. Seven of the 11 studies in Sweden restricted attention to snuff, with three also considering chewing and one only considering chewing.

Adjustment for smoking
ST use is not a major subject for many of the publications from which results have been extracted. While reference is made to ST in the title of one or more papers relating to six of the nine cohort studies (NHANES I, CPS-I, CPS-II, Norway Cohorts, Swedish Construction Workers and Uppsala County), the same is true for only 15 of the 80 case-control studies. For many of the other studies [39,42,59,61,66,70,89,91,102,104,108,109,111,113,114], the reports only provide limited information about ST use in the text, simply giving percentages of users in the cases and controls or even saying there was an association or no association, but without giving supportive data. Many papers consider ST independently of smoking, with no attempt to adjust ST effect estimates for smoking, even though for many of the cancers considered smoking is known to be a cause, and often a major cause.
To summarise the extent to which the available effect estimates were adjusted for smoking, the studies were divided into five groups (A = no information, B = no adjustment, C = never smokers, D = some adjustment, E = more adjustment) as described more fully in the methods.
Of the nine cohort studies, the numbers in the five categories were, respectively, 0, 1, 3, 3 and 2. The Iowa study [20] failed to take smoking into account at all, while the CPS-I and CPS-II studies [23] and the main results from NHANES I [22] were restricted to never smokers. In the remaining five cohort studies, the extent of smoking adjustment varied from publication to publication, but amount smoked or duration of smoking were never taken into account in the US Veterans, Norway cohorts and Uppsala County studies so they are classified as group D.
In the Lutheran Brotherhood study, amount smoked was taken into account in the analyses of pancreatic cancer [13] and stomach cancer [12], and in the Swedish Construction Workers study, amount smoked was adjusted for in the analyses of stomach and oesophageal cancer [34], and cutaneous squamous cell carcinoma [29], and they are therefore classified as group E.
Of the 80 case-control studies considered, details of the adjustment factors used are not provided in either of the studies reported by Bjelke [52] or in two other studies [93,109] Table 3 presents individual effect estimates from six cohort and 34 case-control studies, with 36 of the 40 studies providing estimates with CI that could be used in meta-analyses, the other four [40,71,86,92] finding no significant relationship. Thirty-eight of the 41 estimates included in the first meta-analysis (see Table 4) are those given in our earlier review of ST and oral cancer [4], three recently published studies [32,35,113] Table 3 was included if available, otherwise that for chewing tobacco or snuff was used. b Smoking-adjusted includes estimates for smokers and non-smokers combined, adjusted for smoking if available, and estimates for never smokers otherwise. c The actual estimates included are identified by their RR/OR identification numbers as given in Table 3. d Includes estimates 24 and 25 from a study in Puerto Rico [49]. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk. When the analyses are restricted to US studies, the pattern is similar to that for the overall data, with the effect estimates reduced when attention is limited to those that are smoking-adjusted, and close to 1.0 when estimates that are adjusted both for smoking and alcohol are considered. The effect estimate for never smokers is significantly increased (3.33, 1.76-6.32), based on five small studies, in total involving 19 ST-exposed oropharyngeal cancer cases.

Oropharyngeal cancer
No real evidence of a relationship with snuff use is seen in studies conducted in Scandinavia, where seven estimates, all adjusted for smoking, and four additionally adjusted for alcohol, give a combined estimate of 0.97 (0.68-1.37). However some heterogeneity should be noted, a high RR of 3.1 (1.5-6.6) in the Uppsala County study [35] conflicting with six other estimates ranging from 0.67 to 1.10.
Many of the higher estimates seen in Table 4 come from older studies which often did not adjust for smoking. If attention is limited to studies published since 1990, which generally did adjust, no association is seen. Indeed, the combined estimate from the 14 smokingadjusted studies published since 1990 is 1.00 (0. 83-1.20), and shows no significant heterogeneity.
While the choice of 1990 as the cut-point was not defined a priori, the change in estimates about that time is very clear. As shown in Figure 2 The findings are very similar to those in an earlier review [4]. That review provides additional meta-analyses of the slightly smaller data set, further investigating variation by type of ST, sex, study design, study location and study period. It also provides full details of the various types of cancer that have been considered in the source papers.
The evidence presented suggests that snuff as used in Scandinavia has no effect on oropharyngeal cancer risk. Products used in the past in the USA may have increased the risk but any effect that exists now seems likely to be quite small. Table 5 summarises the data from four cohort and 10 casecontrol studies. For five of these studies effect estimates with CI are not available, one of these [52] reporting a 'synergistic effect of tobacco chewing and alcohol', another [19] presenting a RR of 2.28, but not whether it was significant, and the others [14,40,60] showing no significant relationship. Of the remaining nine studies, six provide smokingadjusted estimates, three of which are also adjusted for alcohol. Though estimates are generally somewhat above 1.0 in these nine studies, they are rarely significant, exceptions being the estimate of 1.92 (1.00-3.68) for snuff in never smokers in the Swedish Construction Workers study [34] and that for chewing of 2.39 (1.23-4.64) in the Wynder and Bross case-control study [44].

Oesophageal cancer
The meta-analyses (see Table 6 and Figure 3) show some indication of an association, though this is not always statistically significant. Based on all available smokingadjusted data, the combined estimate for any ST use is 1.13 (0.95-1.36, n = 7), somewhat lower than when there is no restriction to smoking-adjusted data (1.25, 1.03-1.51, n = 10). The corresponding analyses show no real indication of an effect for snuff in Scandinavia, but are more suggestive for the USA. Even here, the smoking-adjusted estimate is not significant (1.89, 0. 84-4.25), though this is based on only three small studies, involving a total of 11 cases using ST. The estimates based on all the available smokingadjusted data include an any smoking RR of 1.00 (0.79-1.27) from the study with the largest weight, the Swedish Construction Workers study [34], this RR being derived by combining the findings for adenocarcinoma and squamous cell carcinoma. The meta-analyses for never smokers give a higher combined estimate of 1.91 (1.15-3.17, n = 4) for any ST use, mainly because they use a higher (combined adeno/squamous) estimate of 1.92 (1.00-3.68) for the Swedish Construction Workers study [34].
Overall, the data must be regarded as providing suggestive evidence of a possible weak relationship between ST use and oesophageal cancer. Table 7 presents results from 12 studies, eight of which provide a total of 17 estimates which could be used in meta-analyses. Although the Swedish construction workers study [34] shows a significant increase in risk of stomach cancer associated with snuff use for never smokers (RR 1.33, 95% CI 1.03-1.72), no other significant associations are reported, and the metaanalyses conducted (see Table 8 and Figure 4)

Figure 2
Smokeless tobacco and oropharyngeal cancer by study type and period of publication (smoking-adjusted data). The 19 individual smoking-adjusted relative risk (RR) and 95% confidence interval (CI) estimates separated by study type, and for case-control studies by period of publication, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication. In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 3 for further details relating to the estimates, and Table 4 for fuller details of the meta-analyses.
detailed data. No association with stomach cancer was reported by Weinberg et al. [67] or for the US data considered by Bjelke [52]. However, Bjelke did report an "Association ... with tobacco chewing" for the Norwegian data, and a standardised mortality ratio of 1.51 was given for the US Veterans' Study [19], but not whether this was statistically significant.
The combined evidence does not indicate an effect of ST use on the risk of stomach cancer. Table 9 presents results from four cohort and seven casecontrol studies. For four of the studies effect estimates that can be included in meta-analyses are not available; two [75,84] of these studies merely reported finding no association, one [19] reported an elevated RR of 1.65 with no CI, and another [82] a reduced RR of 0.80, also with no CI. Of the other seven studies, significant increases have been reported in two. The Norway cohorts study [26] reports an increase in ever users of snuff in a  Table 6, and 'Cases' is the number of cases in ST users as defined. NA = not available. f Abbreviations used: alc = alcohol, bmi = body mass index, edu = education, exer = exercise, incm = incidence or mortality, res = area of residence, rflx = reflux symptoms, smok = smoking, NA = not available. g The population included < 0.5% females. h RRs for adenocarcinoma (1.0, 95% CI 0.6-1.5) and squamous cell carcinoma (1.0, 0.8-1.4) combined. i RRs for adenocarcinoma (0.2, 95% CI 0.0-1.9) and squamous cell carcinoma (3.5, 1.6-7.6) combined. j The average ridit duration of chewing was non-significantly lower in the oesophageal cancer cases. k RR/OR and/or 95% CI estimated from data provided in the source. l The abstract noted a "synergistic effect of tobacco chewing and alcohol". m RR/OR and/or 95% CI estimated from data provided in the source assuming that no one both chewed and used snuff. n The authors noted the percentage of ever users was "slightly higher" in the controls than in the cases for chewing but not for snuff. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk.

Figure 3
Smokeless tobacco and oesophageal cancer by region (smoking-adjusted data). The seven individual smokingadjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 5 for further details relating to the estimates, and Table 6 for fuller details of the meta-analyses. smoking-adjusted analysis based on the whole population (1.67, 95% CI 1.12-2.50) but not in an analysis based on never smokers (0.85, 0.24-3.07). Conversely, the Swedish construction workers study shows no increase in a smoking-adjusted analysis based on the whole population (0.9, 0.7-1.2), but an increase in never smokers (2.0, 1.2-3.3). None of the three metaanalyses presented in Table 10 (see also At most, the overall data weakly suggest a possible effect of ST on pancreatic cancer risk. A fuller discussion of these data is available elsewhere [5].
Other cancers of the digestive system Table 11 summarises evidence relating to cancers of the digestive system other than those considered already in Tables 5, 7 and 9. Nine studies are considered, four

Figure 4
Smokeless tobacco and stomach cancer by region (smoking-adjusted data). The eight individual smoking-adjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control).
In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 7 for further details relating to the estimates, and Table 8 for fuller details of the meta-analyses.
cohort and five case-control, with one or two studies providing data for colon cancer, rectal cancer, colorectal cancer, small intestine cancer, liver cancer, gall bladder and bile duct cancer. These data, which are insufficient for meta-analysis, include two statistically significant effect estimates: an RR of 1 There are rather more data for the combined category of all cancers of the digestive system. Of the four studies providing data, all conducted in the USA, NHANES I

Figure 5
Smokeless tobacco and pancreatic cancer by region (smoking-adjusted data). The seven individual smokingadjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 9 for further details relating to the estimates, and Table 10 for fuller details of the meta-analyses.
[22] and CPS-II [23] show no relationship, CPS-I [23] a weak, but significant, positive relationship, and the case-control study of Sterling et al. [89] a significant negative relationship. Overall, the combined estimate (see Table 12 and Figure 6), all based on smokingadjusted data, is 0.86 (0.59-1.25, n = 5), with significant evidence of heterogeneity (P = 0.002). The analysis for never smokers removes the case-control study and eliminates the heterogeneity. However the combined estimate of 1.14 (0.99-1.33, n = 4) remains nonsignificant.
More data are needed before any conclusion can be drawn for these cancers. Larynx and nasal cancer The data shown in Table 13 are quite limited. The evidence for nasal cancer is based on only three studies, none reporting a significant association with ST use. Seven studies investigated the relationship of ST to larynx cancer, two providing no effect estimates and merely reporting a lack of association. Control for confounding variables is very limited, with only two studies providing estimates adjusted for smoking, only one adjusting for alcohol and no study presenting any results for never smokers. The only study to adjust for smoking and alcohol [102], which shows no relationship of snuff to risk of larynx cancer, is the only study conducted in Scandinavia. Two US studies [55,56] report a significant relationship, however, and, as shown in Table 14 (see also Figure 7), an association is seen in the overall data (1.43, 1.08-1.89, n = 5).
Given the independent role of smoking and alcohol in larynx cancer [7,8], and the lack of association in the one study that has adjusted for both these factors [102], any independent association of ST use with larynx cancer risk has not been established. More data are needed before any conclusion can be drawn on the role of ST in larynx and nasal cancers.

Figure 6
Smokeless tobacco and overall digestive cancer (USA smoking-adjusted data). The five individual relative risk (RR) and 95% confidence interval (CI) estimates, all smoking-adjusted and for the USA, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown is the combined estimate, derived by random-effects meta-analysis. This is represented by a diamond of standard height, with the width indicating the 95% CI. See Table 11 for further details relating to the estimates, and Table 12 for fuller details of the meta-analysis. Table 15 summarises data from six cohort and three casecontrol studies. The case-control studies provide only estimates for smokers and non-smokers combined, and only one of these is adjusted for smoking. The cohort studies all provide estimates for never smokers, with two also giving smoking-adjusted results for the overall population. The meta-analyses (see Table 16 and Figure 8) show no evidence that ST use increases risk of lung cancer, with the combined estimate for smoking-adjusted data 0.99 (95% CI 0.71-1.37). However, there is considerable heterogeneity (P < 0.001), the major contributors to this being the high RR of 6.80 (1.60-28.5) in never smokers in NHANES I [22], the significant increase of 1.77 (1.14-2.74) from CPS-II [23], and the low RR of 0.70 (0.60-0.70) for the Swedish construction workers study [32]. While the combined estimate for never smokers for any ST use is greater than 1.0 (1.34, 0.80-2.23, n = 5), it is not statistically significant.

Lung cancer
While the data have unexplained heterogeneity, they do not provide any clear indication of a relationship of lung cancer to ST use.   Table 14, and 'Cases' is the number of cases in ST users as defined. NA = not available. e Abbreviations used: alc = alcohol, res = area of residence, smok = smoking. f The average ridit duration of chewing was non-significantly lower in the larynx cancer cases. g RR/OR and/or 95% CI estimated from data provided in the source. h RR/OR and/or 95% CI estimated from data provided in the source assuming that no one both chewed and used snuff. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk.

Figure 7
Smokeless tobacco and larynx cancer by region (overall data). The five individual relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 13 for further details relating to the estimates, and Table 14 for fuller details of the meta-analyses. Only estimates 3 and 16 are smoking adjusted.
marked differences between switchers and quitters in a range of characteristics, with adjustment substantially reducing the RR estimate from the age-adjusted estimate of 1.92 (1.63-3.26). Table 17 presents data from five cohort and two case-control studies, all conducted in the USA. No significant association between ST and prostate cancer is evident in five studies, but significant increases are seen in the Lutheran Brotherhood Study [11] and, for current snuff users only, in the casecontrol study by Hayes et al. [96]. Based on the five studies which provide usable data, the overall estimate (see Table 18 and Figure 9) is 1.20 (95% CI 1.03-1.40).

Prostate cancer
Prostate cancer is not considered smoking related [7,8], and more information on its relationship with ST is needed before any clear conclusion can be drawn.

Figure 8
Smokeless tobacco and lung cancer by region (smoking-adjusted data). The six individual smoking-adjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by randomeffects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 15 for further details relating to the estimates, and Table 16 for fuller details of the meta-analyses. Bladder cancer Table 19 summarises data from the Norway cohorts study [26] and from 12 case-control studies. None of the casecontrol studies were conducted after 1990, and with the exception of two studies in Denmark [43,62], all were carried out in the USA or Canada. The great majority of the estimates are non-significant, and based on 10 smokingadjusted estimates the overall estimate (see Table 20 and Figure 10) is 0.95 (95% CI 0.71-1.29). However, there is significant heterogeneity due mainly to estimates 8, 12 and 22, which show a positive association, the last two of which are significant, and estimate 31 which shows a significant negative association.
Considered together, the data provide no real evidence of an association between ST and bladder cancer. Table 21 summarises evidence from one cohort and nine case-control studies, none conducted in Sweden. The estimates are generally based on small numbers of cases  Table 18, and 'Cases' is the number of cases in ST users as defined. NA = not available. e Abbreviations used: pov = poverty, res = area of residence, smok = smoking. f RR/OR and/or 95% CI estimated from data provided in the source. g RR/OR and/or 95% CI estimated from data provided in the source assuming that no one both chewed and used snuff. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk.  Table 22 (see also Figure 11) are elevated, with some evidence of heterogeneity, but none are statistically significant. Based on five smoking-adjusted estimates the overall estimate for any ST use is 1.09 (0.69-1.71).

Kidney cancer
While there is a suggestion of a possible relationship, more data are needed before any firm conclusions can be reached. Table 23 summarises evidence from three cohort and seven case-control studies for overall haematopoietic cancer and for specific types. The only report of a significant association is the OR of 4.0 (1.3-12.0) for non-Hodgkin's lymphoma in the case-control study of Bracci and Holly [112]. However, the combined evidence from the five studies (see Table 24 and Figure 12)

Figure 9
Smokeless tobacco and prostate cancer (USA overall data). The five individual relative risk (RR) and 95% confidence interval (CI) estimates, all for the USA, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 17 for further details relating to the estimates, and Table 18 for fuller details of the meta-analyses.

Figure 10
Smokeless tobacco and bladder cancer by region (smoking-adjusted data). The 10 individual smoking-adjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by randomeffects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 19 for further details relating to the estimates, and Table 20 for fuller details of the meta-analyses.  Table 22, and 'Cases' is the number of cases in ST users as defined. NA = not available. e Abbreviations used: bmi = body mass index, edu = education, hosp = hospital, res = residence, smok = smoking, tadm = time of admission. f RR/OR and/or 95% CI estimated from data provided in the source. g Estimated assuming ORs for chewing and snuff are independent. h The authors also report the results of an analysis adjusting for the effects of the matching factors, body mass index, decaffeinated coffee use and continuous pack-years of cigarette smoking. The authors estimated an OR (95% CI) of 0.87 (0.15-5.14) for the effect of chewing among never smokers of cigarettes, and of 26.00 (4.41-153.00) for the joint effect of pack-years cigarette smoking and chewing tobacco use. These results could not readily be incorporated into the meta-analyses as no overall estimate for chewing tobacco use adjusted for cigarette smoking was available. i Analysis uses hospital controls. j Analysis uses population controls. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk. smokers, but not in a smoking-adjusted analysis for smoker and non-smokers combined (1.1, 0.8-1.4). Finally, based on the Swedish construction workers study, Odenbro et al. [29,33] report that snuff use is associated with a reduced smoking-adjusted risk of cutaneous squamous cell carcinoma (0.64, 0.44-0.95) and, in never smokers, with a reduced risk of melanoma (0.65, 0.52-0.82). These isolated reports need confirmation in other studies before any effect of ST can reliably be inferred. A study in Cherokee women [125,126] which shows no association of breast cancer with ever ST use, with an odds ratio adjusted for age at diagnosis estimated as 1.24 (0.26-6.02), is not considered in Table 25 as the study is of cross-sectional design. It contributes little to the evidence.
Overall cancer risk As shown in Table 26, ST use has been related to overall cancer risk in five cohort studies and one case-control study. Two of the 12 estimates shown are smoking-adjusted estimates for smokers and non-smokers combined, one (estimate 10) showing no association at all (RR = 1.00) and the other (estimate 12, based on the case-control study [89]) a reduced OR of 0.64 (95% CI 0.53-0.78). The remaining 10 estimates, all from cohort studies, and all adjusted for age and various other potential confounders, are for never smokers. As shown in Table 27 and Figure 13, the combined estimate for all the smoking-adjusted data is not elevated (0.98, 0.84-1.15, n = 7). However, the combined estimate for never smokers, which excludes the low estimate from the case-control study, is a significant 1.10 (1.02-1.19, n = 6). The estimate for never smokers is similar for the US data (1.10, 1.01-1.20, n = 4) and the Scandinavian snuff data (1.10, 0.94-1.29, n = 2). The data are consistent with any excess risk of cancer in ST users being small.

Publication bias
There are 49 meta-analyses presented that combine five or more effect estimates. The test of publication bias [121]

Figure 11
Smokeless tobacco and kidney cancer by region (smoking-adjusted data). The five individual smoking-adjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by randomeffects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 21 for further details relating to the estimates, and Table 22 for fuller details of the meta-analyses.
shows none to be significant at P < 0.01, and two significant at P < 0.05, similar to the numbers one would expect by chance. Both the significant cases (see Tables  22 and 24) arise due to a single high effect estimate, with the other estimates included in the analysis relatively close to 1.0. Table 28 shows the effect on the smoking-adjusted analyses of successively removing those RR/OR estimates with the largest Q 2 values. Results are only shown for those cancers where significant (P < 0.05) heterogeneity was evident, and removal continues until no significant  Tables 1 and 2. b ST implies smokeless tobacco unspecified, or combined snuff use or chewing. c Ever, former and current ST use were compared with never ST. Use indicates timing not given and comparison is with non use. d 'Id.' is the RR/OR identification number used in Table 24, and 'Cases' is the number of cases in ST users as defined. NA = not available. e Abbreviations used: alc = alcohol, asp = aspirin, bmi = body mass index, edu = education, exer = exercise, occ = occupation, smok = smoking, tadm = time of admission, yriv = year of interview. f The population included < 0.5% females. g Estimated from data on limited number of exposed cases for eight sub-types of haemopoietic cancer. h RR/OR and/or 95% CI estimated from data provided in the source. i Data for six subtypes of leukaemia were also provided, but none were statistically significant. j Data for five subtypes of non-Hodgkin's lymphoma were also provided, but none were statistically significant. k Estimated from data for t (14,18)-positive and t (14,18)-negative cases. l Estimated from the results for chew and snuff, assuming that no one both chewed and used snuff. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk.

Figure 12
Smokeless tobacco and non-Hodgkin's lymphoma by region (overall data). The five individual relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inversevariance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 23 for further details relating to the estimates, and Table 24 for fuller details of the meta-analyses. Only estimates 5, 13 and 19 are smoking-adjusted.   Table 27, and 'Cases' is the number of cases in ST users as defined. e Abbreviations used: alc = alcohol, asp = aspirin, bmi = body mass index, edu = education, exer = exercise, occ = occupation, pov = poverty, res = area of residence, smok = smoking. f RR/OR and/or 95% CI estimated from data provided in the source. g Number of cases estimated from data provided in the source. CI = confidence interval; ST = smokeless tobacco; OR = odds ratio; RR = relative risk.  Table 3 id. numbers 1, 15, 21, 22, 34 and 35).

Figure 13
Smokeless tobacco and overall cancer by region (smoking-adjusted data). The seven individual smoking-adjusted relative risk (RR) and 95% confidence interval (CI) estimates, separated by region, are shown numerically and also graphically on a logarithmic scale. They are sorted in order of year of publication within study type (cohort, case-control). In the graphical representation individual RR estimates are indicated by a solid square, with the area of the square proportional to the weight (inverse-variance) of the estimate. Also shown are the combined estimates, for the subgroups and overall, derived by random-effects meta-analysis. These are represented by a diamond of standard height, with the width indicating the 95% CI. See Table 26 for further details relating to the estimates, and Table 27 for fuller details of the meta-analyses.
For 8 of the 12 cancers, the change to the meta-analysis estimate from the altered preference is very small, by ± 0.02 at most. For oropharyngeal cancer it increases by 0.06, for larynx cancer by 0.11, for lung cancer by 0.12 and for pancreatic cancer 0.15. None of these changes materially affect the significance or the interpretation. Although there is perhaps a slight indication that associations may be stronger for current use, the tendency of most studies to report results only for ever or unspecified ST use limits the extent to which this can be investigated. Changing preferences did not materially affect the heterogeneity of the estimates. The effect of similarly changing the preference on the other meta-analyses shown earlier (for example, for never smokers or by country) also did not materially affect the results obtained (data not shown).  Figure 2).

Meta-regression analyses
Based on the 41 overall estimates (whether smokingadjusted or not) for oropharyngeal cancer, where the deviance is 335.6 (P < 0.001), the most significant factor is sex (P = 0.004, drop 83.4 on 2 d.f.). Though drops in deviance of 20 or more are also seen for region, period by study type and smoking status, with estimates high for females, USA, old case-control studies and data unadjusted for smoking, no other factor is significant at P < 0.05 after adjustment for sex. The high deviance of 335.6 is clearly due to very high Q 2 values for some estimates, and further analyses were run excluding these estimates (ids 1, 7, 21, 22 and 34 in Table 3). This reduces the deviance considerably, to 84.4, though it is still highly significant (P < 0.001). However, again sex was the most significant factor (P = 0.02), with no further factor significant at P < 0.05 after adjusting for sex.
Meta-regression analyses were not attempted for larynx, nasal or prostate cancer or for overall digestive cancer or  also in the sensitivity analysis shown in Table 28. For lung cancer, a tendency was noted for never-smoking estimates to be high, significant for both the smokingadjusted data (P = 0.025) and the overall data (P = 0.029). This difference reflected the two high estimates already noted in the sensitivity analysis.
Summary of meta-analyses for ST use in Western populations Table 30 brings together all the meta-analysis results for ST use in Western populations. Based on smoking-adjusted data, significant increases (P < 0.05) are seen for oropharyngeal cancer, though not based on studies published since 1990, and for prostate cancer, but not for any other cancer considered. For never smokers, significant increases are seen for oropharyngeal cancer (again not when based on studies published since 1990), for oesophageal cancer and also for overall cancer. Compared with the smoking-adjusted estimates, the estimates for never smokers tend to be more variable, due to smaller numbers of ST-exposed cases studied, though they consistently exceed 1.0. Table 31 similarly brings together the results for ST use in the USA. With the exception of oesophageal cancer in never smokers, significant increases seen in Table 28 are again significant here, with an increase additionally seen in the smoking-adjusted estimate for larynx cancer (although based on only a single study).

Summary of meta-analyses for snuff use in Scandinavia
As shown in Table 32, the meta-analyses of results provide overall effect estimates that, with one exception, are never significantly increased and generally are close to 1.00. The exception is for oesophageal cancer, where the marginally significant increased RR seen in relation to snuff use for never smokers (1.92, 1.00-3.68) derives solely from the Swedish Construction Workers study [34]. In that study, no increase was seen in smoking-adjusted analyses for the whole population (1.00, 0.79-1.27). Unlike the corresponding results for the USA, where meta-analysis estimates are predominantly greater than 1.0, the estimates for snuff as used in Scandinavia are as often below 1.0 as above 1.0. Generally, the results do not suggest that snuff as used in Scandinavia has any adverse effect on cancer risk.

Dose response data
Results relating the various cancers to dose of exposure to ST are only reported in a few studies and are not presented in detail here.
For oropharyngeal cancer, eight studies were identified that related risk to extent and/or duration of exposure. In seven of these studies, which all show no overall relationship of ST with risk in Table 3 [32,55,[89][90][91]104,113], no significant dose-response relationships are seen. It was only in one study [61], that did show a clear overall relationship, that a significant (P < 0.001) trend in risk with increasing duration of exposure is seen, though only for cancers of the gum and buccal mucosa, and not for other mouth and pharynx cancers.
For other cancer sites relatively few studies report doseresponse data. In the CPS-II study [23] no trends with duration or frequency are seen for either total or lung cancer, while in the Swedish Construction Workers study no trend is seen for cutaneous squamous cell carcinoma with years of snuff dipping [29] or for oral cancer or lung cancer with daily amount of snuff consumed [32]. A significant trend (P < 0.01) is reported with daily amount of snuff consumed Comparison of the effects of smoking and of ST use Table 33 summarises the results of analyses comparing the effects of smoking and of ST use, for seven smokingrelated cancers [127]. Overall in US men aged 35+ a total of 142,205 deaths were seen from these cancers in 2005, with lung cancer (63.4%) by far the most common. Based on RRs from CPS-II for current and former smoking [122] and estimates of the frequency of current and former smoking [124] for US men of this age group, the total number of deaths that would have occurred if the men had the mortality rates of never smokers can be estimated as 37,468, a reduction (E) of 104,737 deaths. This reduction is proportionately largest for the cancers most strongly associated with smoking (lung and oropharynx), and least for those most weakly associated (pancreas, kidney and bladder).   (Table 14) 1 0.90 (0.50-1.50) 0 -Lung (Table 16) 2 0.71 (0.66-0.76) 2 0.82 (0.52-1.28) Bladder (Table 20) 1 0.83 (0.62-1.11) 0 -Kidney (Table 22) 1 0.72 (0.44-1.18) 0 -Non-Hodgkin's lymphoma ( The smoking-adjusted relative risks for any ST use taken from Table 30 are then used to estimate the number of deaths that would have occurred if the population were never smokers, with ST use either at the same frequency as for current and former smoking combined, 53%, or at 100%. In the first situation, the number of cancer deaths rises from 37,468 to 38,570, an increase of 1,102; in the second situation, it rises to 39,548, an increase of 2,081. These numbers of cancers associated with ST use form, respectively, 1.1% and 2.0%, of E, the number associated with smoking.

Discussion
Estimating the effects of ST use We have analysed data relating cancer risk to the consumption of chewing tobacco and snuff as used in Western countries. We have identified 12 cancers (or combined categories) where, as shown in Table 30, it is possible to derive a (random-effects) meta-analysis estimate based on at least five individual independent estimates.
It is notable that no strong association at all is evident and that few of the associations are significant at P < 0.05. Indeed, based on smoking-adjusted data, which might be argued to provide a good compromise between avoidance of bias and loss of power, only the estimates for oropharyngeal and prostate cancer are significant, with that for oropharyngeal cancer not evident in more recently published studies. However, it should be noted that while many of the estimates in Table 30 for never smokers have wide confidence limits, and only those for oropharyngeal and oesophageal cancer and for overall cancer are significant, all the estimates are in fact greater than 1.00. Although publication bias may be relevant, and more data are clearly needed, the consistency of these findings suggests that ST may increase the risk of cancer, though any effect is likely to be quite weak. The results in Table 32 suggest, however, that whether smoking-adjusted data or data for never smokers are considered, there is little or no evidence of an effect of snuff as used in Scandinavia.
There are a number of difficulties in interpreting the results of these meta-analyses. The studies are of varying design, size and quality. Many of the individual study reports have limitations and present less information Increase in deaths in a population of never smokers j Same % become ST users as were smokers (I 1 ) 1,102 100% of population become ST users (I 2 ) 2,081 a ICD 10 th revision codes [127] used are oropharynx (C00-C14), oesophagus (C15), pancreas (C25), larynx (C32), lung (C33, C34), bladder (C67) and kidney (C64-C66, C68). b Numbers of deaths in 2005 from WHO [123]. Here and for other results, the entries in brackets correspond to the notation used in the methods section. c Relative risks from US Surgeon General's Report 1989 [122] [124]. e E = Σ (D i -D i *). f Relative risks for any ST use, based on smoking-adjusted data, as given in Tables 4, 6, 8, 10, 16, 20 and 22 for the seven cancers shown. g Actually 0.99, but taken as 1.00 for the purposes of estimation. h Actually 0.95, but taken as 1.00 for the purposes of estimation. i D i ** and D i *** are calculated as shown in the methods section, assuming that 53.0% or 100.0% respectively of the population use ST. j I 1 = Σ (D i ** -D i *), I 2 = Σ (D i *** -D i *) than is ideal for a meta-analysis. Shortcomings include small numbers of cases, and in particular of cases exposed to ST, lack of histological confirmation, lack of division by cancer site, as well as an unclear description of inclusion and exclusion criteria, details of case and control selection, and methods of exposure assessment. Furthermore, details such as the type of ST used, and duration and frequency of use, are often not considered. The products used vary by country and over time, and increased risks seen in older studies for some cancers may not reflect the risks of more modern products, with reduced nitrosamine levels [128]. For most cancers, the number of effect estimates available is really too limited to allow a very detailed examination of variation in risk by such factors as type of product used, current or former use, country and sex. Though metaregressions have been attempted for a number of cancers, they have not added materially to the interpretation, partly because of the limited amount of data for some cancers, and partly because of the number of apparently outlying estimates, notably for oropharyngeal cancer.
A major problem is that many of the studies fail to adjust for smoking and other important potential confounding variables. Although recent major reviews [7,8] consider that all the cancers considered in Table 30, with the exception of prostate cancer and non-Hodgkin's lymphoma, are caused by smoking, it is evident that a number of the studies do not provide estimates that are either for never smokers or for smokers and non-smokers combined with adjustment for smoking. Even where adjustment for smoking is carried out, this is often by a relatively simple approach, with no account taken of number of cigarettes smoked or duration of smoking. Smokers who also use ST may smoke fewer cigarettes a day than smokers who do not. Failure to adjust for smoking is particularly common for studies of oropharyngeal cancer, with many of the older studies not taking smoking into account at all when considering ST. The potential importance of this is illustrated by the overall estimate for oropharyngeal cancer being substantially reduced, from 1.79 to 1.36, when attention is restricted to smoking-adjusted data.
Adjustment for other risk factors is also important, as shown by the case of oropharyngeal cancer where the smoking-adjusted estimate of 1.36 (1.04-1.77, n = 19) can be compared with the estimate adjusted for smoking and alcohol of 1.07 (0.84-1.37, n = 10). Restricting attention to estimates adjusted for both factors also eliminated the highly significant (P < 0.001) heterogeneity seen in the smoking-adjusted data. Alcohol is also an important factor in the aetiology of oesophageal, larynx and liver cancer [8], but the number of ST effect estimates adjusted both for smoking and alcohol for these three cancers is very low indeed, respectively 2, 1 and 0. Other factors considered rarely, or not at all, include, for example, Helicobacter pylori infection for stomach cancer and diet for digestive cancer.
Another difficulty in interpreting the overall results is the variability of the findings. Heterogeneity significant at least at P < 0.05 is evident in the smoking-adjusted estimates for cancers of the oropharynx (though not in the more recent data), pancreas, larynx, lung and bladder, as well as for overall cancer and overall digestive cancer. As noted above, the evidence is too limited for most of the cancers to allow a proper investigation of the sources of this heterogeneity.
Based on the data analysed, there is little or no evidence of publication bias. However, it should be noted that the number of studies reporting results in a form that cannot be included in the meta-analyses is fairly high, representing up to about 30% for some cancers (see Tables 5,7,9,13 and 17).
We are aware that the smoking-adjusted meta-analysis estimates we report for oropharyngeal cancer ( . Reasons for this, based on a detailed analysis of this review, will be presented in a separate publication in BMC Cancer.

Comparison of the effects of smoking and ST use
In 2005 in US men aged 35 or over, there were a total of 142,205 deaths from seven cancers considered to be caused by smoking. Based on relative risks from CPS-II for current and former smoking [122] and estimates of the frequency of current and former smoking [124] for US men of this age group, we estimate that, had the population at risk the mortality rates of never smokers, the numbers would have reduced by 104,737, with the reduction in lung cancer deaths, 79,195, a major contributor. Any increase in risk resulting from the introduction of ST to a population of never smokers would be very much less than this. Even assuming that the smoking-adjusted meta-analysis estimates for the seven cancers all reflect a true effect of ST, the increase in deaths among a never-smoker population would be by 1,102 if 53% of the population used ST (the same proportion as had ever smoked) or by 2,081 if the whole population did. These increases represent, respectively, only 1.1% and 2.0% of the 104,737 deaths attributed to cigarette smoking.
There are a number of objections that can be made in respect of this comparison. These include the following: 1. The RRs for current and former smoking are based on CPS-II, conducted in the 1980s, and may not reflect those appropriate for 2005, given inter alia changes in cigarettes that have occurred since then. However, CPS-II is widely used as a source of data for calculating deaths attributed to smoking (for example, [8,129]).
2. The RR estimates used for ST use are not specifically for the USA, or for males. However, 62 of the 89 studies considered in this review were conducted in the USA, and 41 of the 58 estimates used in the smoking-adjusted meta-analyses for the seven cancers are for males (with 12 for sexes combined and five for females).
3. The RR estimates used for ST are for any ST use, and do not separate current and former use, due to most studies not providing such data.
4. The calculations are limited to those seven cancers which the US Surgeon General, in his 1989 report [122] considered to be caused by smoking and for which RRs were provided for CPS-II. A more recent report [8] includes stomach cancer and leukaemia as caused by smoking. For stomach cancer, the meta-analyses in Table 6 showed virtually no association with ST use (1.03, 0.88-1.20, n = 8), while the more limited data for leukaemia also showed no clear evidence of a relationship.

5.
It is theoretically possible that ST use might increase the risk of some cancers not increased by smoking. Here one should note the significant association for prostate cancer (1.29, 1.07-1.55).
6. The calculations do not take into account the fact that a proportion of US males aged 35+ already use ST. Given the relatively weak association between cancer and ST use, any attempt to do this would have had relatively little effect.
7. The calculations also do not take pipe and cigar smoking into account.
8. The approach used is somewhat simplistic, and a more realistic (but more complex) calculation might be to compare predicted cancer deaths over a long-term period in a population continuing to smoke as at present, with the predicted number in a population switching from cigarettes to ST.
Despite all these points, it is clear that any effect of ST on risk of cancer, if it exists at all, is quantitatively very much smaller than the known effects of smoking. This is in any case apparent from a simple comparison of the RRs for cigarette smoking and for ST use.

Conclusion
The available data relating to ST use have a number of weaknesses, including inadequate control for smoking in many, and limited data for never smokers. Nevertheless, it is possible to conduct meta-analyses based on smokingadjusted estimates for a relatively wide range of cancers. These show no indication of an increased risk of cancer for snuff, as used in Scandinavia. The overall data for oropharyngeal cancer shows a significant increase in risk associated with ST use, but this is not evident for estimates adjusted for smoking and alcohol, or for studies published since 1990. Any effect of ST may relate mainly to products used in the past in the USA. A weak but significant association with prostate cancer, based on limited data from US studies, requires more confirmatory evidence. Reports of significant associations with pancreatic and oesophageal cancer in an earlier review [6] are not confirmed, and reasons for this will be discussed in a later publication. Risk from ST products as used in North America and Europe is clearly very much less than that from smoking, and is not evident at all in Scandinavia.