Document YDZQrkrxk7LV8q24JLgpg248O

l 368 1. T. Hodgson and A. Oarmnn recently (Vainio and Bofetta, 1994) who conclude that the overall evidence indicates an interaction in the multiplicative region. This implies that the rela lation. and this was the only study where an explicit adjustment for smoking was made. Unadjusted data was used for all other studies. tive risk of lung cancer due to asbestos exposure will be the same for smokers and non-smokers alike. Thus SMRs for lung cancer based on a reference popu lation with the same smoking habits as the cohort members should only reflect the effect on mortality due to asbestos exposure. An earlier review by Berry et al. (1985) estimated that the effect of asbestos exposure was about 1.8 times greater in non-smokers than in smokers (though with confidence limits which Fibre type and industry process For the purpose of summarising the information given in the studies, each cohort was given a fibre type classification of I, 2 or 3 letters according to the type of fibre used, with the letters y, a and o rep resenting chrysolite, amosite and crocidolite exposures respectively. For example: did not exclude a simple multiplicative interaction). If this is the case the observed effect of asbestos on lung cancer rates will be greater in populations with lower smoking prevalence. However, given the rela tive lung cancer risks typical of smoking (about 15fold) and asbestos exposure (about 2-fotd) together with the generally high prevalence of smoking in the observed populations, the scope for bias--if there is indeed a differential effect of the scale suggested-- is limited. In either case, a problem arises when the smoking habits of the cohort members differ from those of the reference population, which is the case 'yao' 'yo' 'a' means all three commercial asbestos types were used in the cohort means chrysotile and crocidolite were used means only amosite was used The order of the letters indicates the relative itnportanceof the fibres used. Very small quantities of fibre were ignored in some cohorts (Carolina, New Orleans plant I, Connecticut), the reasoning for this in each cose is set out in Appendix A (Table 14). In a similar way, for display in tabular and graphical data sum maries, industry process was coded as follows. for some of the cohorts reviewed. For this reason, any information about smoking given in the studies was summonsed. The amount of information given was very variable, and could be categorised os follows: M C T I Mines Cement Textiles Insulation Products 1. No information given, (Ferodo, US Insulators. Pat erson, South Africa, Johns Manville, Albin). 2. The percentage of the cohort that smoked, usually F L O Friction Products Lagging and work with insulation Other based on a cross sectional survey conducted in a particular year, (Connecticut, Balangero, Quebec, Mela-analytic issues Pennsylvania. Rochdale, Wittenoom). 3. Comparison of the prevalence of smoking in the The aim of a meta-analysis is to identify where evi dence from different studies is discrepant; ideally, to cohort and the reference population. (New Orle explain the reasons for the discrepancies; and where ans, Massachusetts, Carolina). data from different studies are coherent to combine 4. Estimation of the effect of any differences in them into a common summary which will be more prevalence--for example calculation of smoker precise and soundly based than the estimate from any adjusted lung cancer SMRs, (Vocklabruck) single study. For this review the coherence of esti 5. Data on prevalence of smoking within exposure- mates of Rl and J?M from different studies has been categories--but with no external comparison assessed in a Poisson regression framework, fitting a (Ontario).. common value of the parameter of interest across a group of studies and testing the residual deviance Most studies fell within the first two of the above between the observed and predicted numbers of categories. In these cases only subjective judgements events (mesothelioma or lung cancer deaths) in the could be made by the authors about the smoking hab studies in the group. Confidence limits around the its of the cohort members. Also, cross sectional stud group estimates were calculated by profile likelihood ies were often based on a small proportion of the cohort and may not be very representative. For most studies which addressed the issue the authors con methods. Confidence limits are not shown for the means of groups which show very significant hetero geneity, since such limits have no ready interpret cluded that there was no major difference in smoking ation. Indeed, in this situation it is not clear that the prevalence or that the slight differences in prevalence mean' itself has any natural meaning. Faced with were not likely to change the expected number of clearly discrepant data, purely statistical criteria can lung cancer deaths in a substantial way. Of the studies not be used to decide on a `correct' summary or . where comparative smoking data were given, the compromise estimate. Vocklabruck cohort showed the largest difference in The statistical analyses in this report only take cohort smoking habits and those of the general popu account of the statistical variability of the mortality \ ' '!