Document YDQpaLDyQOMOvnB5jV2xYwb68

iom) :darn) .ai ir table :agus1on of 7en the n &om canons 5. for a >.:VUI or ndc, if not :olume :or cor:r than Isdicinc. ion and -served. 1Pyright :ored in . meam awnen. :acts of :to the e publi- -isernenc -isemcnt 2. BMJ London mmmd ion rate t to the Vlnl#d instock x d with :c o m e or wthmd, W 10weva, nquirlea ublishcd . Visa. or xcdit or penod cully ` '.,lCIanS for the e ticld of it 6 SC Application of meta-analysis in reviewing ~ -.-...:-. . occupational cohort studies .A*_--- Otto Wong, Gerhard K Raabe ... _-.-.. . . .ipptied Health Sciences, 181 Second .\venue, Suite 628, PO Ros 2078, San Mateo, (lalifornia94401, USA 1) \S'ong .\lc.dical Department, .\[obi1 Business Resources (:wporation, PO Box !Sv. New Hope, I'cnnsylvania, USA . f I.; b a b e - ., ...bpundence to: :I )rro Wong, Applied . ' -...:h Sciences, 1x1 Second '. . x c . Sum 028, PO Box - - -. San Mnreo, Californu .L, ' I . L`SA. :?red 2Y July IY40 Abstract Meta-analysis has been used increasingly in renewing and summarising epidemiological studies. Reviews incorporating meta-analyses have appeared in medical journals in increasing numbers. .Although there are several methodology papers on meta-analysis, most of these papers have been written primarily for discussion among epidemiologists. The present paper considers some of the basic methodological issues, the more practical aspects of meta-analysis, and tar,wets an audience of mainly non-epidemiologists. Thus, the main objective of this paper is to provide some basic ,@delines for nonepidemiologists to evaluate meta-analysis in occupational cohort studies. In this methodology paper, the limitations and problems of traditional qualitative reviews are pointed out. Some of these problems can be dealt wirh by quantitative meta-analysis. The potential limitations and benefits of quantitative meta-analysis are discussed. Rather than replacing traditional qualitative review, quantitative meta-analysis should be made part of the overall assessment. The term "meta-review" is proposed to emphasise the importance of both qualitative and quantitative components in a comprehensive review process. The basic steps in a meta-review are outlined, with a discussion on how to recognise and avoid some of the problems which are likely to occur at each step. A meta- review is u s e l l in selecting studies, and in organising, presenting, and summaris- ing results from individual studies. A meta-review can also be used to detect heterogeneity among studies. Major benefits of conducting a meta-analysis (the quantitative component in a metareview) include the increase in statistical power and the estimate of a properly weighted summary risk estimate. Keywords: literature review. spiriemrology, occupational cohort studies, meta-m.dvsi~.meta-review Meta-analysis has been used increasingly in epidemiological literature revielvs. Its use has generated both enthusiasm and scepticism. Meta-analysis, when used properly, is a powerful research tool. There are several papers in the epidemiological literature on different aspects of meta-analysis. However, these have been written primarily for discussion among epidemiologists. The present paper considers some of the basic methodological issues and the more practical aspects of meta-analysis. and targets as its audience mainly nonepidemiologists, such as occupational physicians, occupational nurses, industrial hygienists, toxicologists, risk managers, or others who have the need and the interest to read epidemiological literature. Thus, a primary objective of this paper is to provide some basic guidelines for non-epidemiologists IO evaluate papers on mera-analysis in occupational epidemiology. As such, highly statistical or technical issues will not be discussed. Furthermore, the discussion will focus on occupational cohort studies. The term meta-analysis was coined by the British psychologist Glass in 1076 to describe "statistical analysis of a large collection of analysis results from individual studies for the purpose of integrating the findings."! Although the term meta-analysis was not used before that time, statistical integration of results from studies began long before then. In particular, several excellent papers describing statistical techniques for combining data from individual experiments had been written by pioneers such as Cochran, Fisher, and Pearson.` Meta-analysis has been used extensively in psychology, sociology and education, and more recemly, in clinical trials. In epidemiology meta-analysis has not been widely used until recently. Although the term metaanalysis was not used, in the 1964 Surgeon General's report Smoking and healrh, data from seven cohort studies available at that time were combined to provide summary estimares of cancer risk for smokers." Various definitions of meta-analysis h a w been published. For example, similar to the original definition by Glass, Huque defined meta-analysis as ``a statistical analysis which combines or integrates the results of several independent clinical trials, considered by the analyst to be combinable.'" Without any que+ tion, statistical analysis is a major part of m e u - analysis. However, 3s will be discussed below, meta-analysis involves more than the application of statistical techniques. There is d e b nitely a qualitative component in the selection of studies to be included as well as in the inter- pretation of the results. Meta-review of occupational epidemiology In a traditional qualitative literature review, the selection of studies is usually not subject to pre- determined criteria. This was particularly true in the days before the availability of computerised databases. Furthermore, weights assigned to individual studies were usually subjective. In traditional reviews, no quantitative summaries of risk estimates are provided. Therefore they are primarily narrative, and tend to be subjective. A comprehensive literature re15ew should consist of both a qualitative review and a quantitative analysis. T o be distinguishable from a traditional literature review, we propose the use of the term "meta-review" to describe a comprehensive review that includes a meta-analysis. This term seems appropriate as the Greek word "meta" means ``after".cAccording to Webster's dictionary, "meta" is "used with the name of a discipline to designate a new but related discipline designed to deal critically with the original Table 1 shows the steps in conducting a meta-review. Some of the steps are self esplanatory and will be discussed only briefly. Within the framework of occupational srudies, the research question (step 1) should consider and identify the exposure of interest, the exposed population, and the extent of exposure. On the response (health outcomes) side, the specific health end points and the corresponding diagnostic criteria should be clearly stated. Step 3, a literature search or identification of studies, calls for a comprehensive search strategy; and step 3, criteria for inclusion, is vital to the overall search strategy. One question which has been debated to some extent is whether unpublished studies should be included. Negative studies-that is, studies showing no increased risk-are less likely to be published for various reasons, the most common one being that the investigators or journal editors do not think that the results are exciting or informative. This is particularly m e for negative studies with a small sample size, because such studies individually are not conclusive. Negative studies are also less likely to be published in the absence of positive studies or any controversy of the subject matter. This is usually referred to as the "file drawer" problem or "publication bias." The general consensus is that a study's quality should not be judged according to whether it is unpublished or not."' A more troublesome issue is the identification Tablc I Sicps bi c.oiidticiiiig u iiicra-rcvicz 1 Defining the research question 2 literature search 3 Criteria for inclusion in the study 4 Traditional qualitative review 5 Quantirarivr metr-analvsis 6 Integrating traditional qualitativu renew with quantitative me=-analysis 7 Applying criteria for causation in interpretation of all unpublished studies. There is no good solution to this problem. If the reviewer is also ;: primary investigator in that particular field. identifjhg unpublished studies may not be difficult. Otherwise, the best approach may be inquiring of researchers (identified through published reportsj in that field. There is also the reverse file drawer problem. the repeated publications or overrepresentation of positive studies. There is a tendency for positive studies to be published or presented at conferences more than once. In a meta-review on asbestos and gasmc cancer, it was found that some data sets had been published repeatedly. ! Papers reporting on different subcohons of the same population may also complicate the process. Furthermore, studies may have been updated periodically, and results from different versions of the same study may have been published. The general guideline for inclusion in a study is that the most recent and the most complete version of the study should be included. :Also, any serious discrepancies between pasr and current versions should be identified and explained. Finally, studies based on individuai companies or indi\<dual centres may also bt. part of an indus-xlde or multicenue srudy. and caution must be exercised not to enter the same data more than once in the meta-analysis.I' A similar issue is the presentation bias, which refers to selective presentation of study resulrs. For example, the results for rare cancers in a cohort study consisting of workers exposed tc certain chemicals of interest may nor be presented due to small numbers; yet this is exactly the situation where a meta-analysis is most beneficial. Again, contacdng the original. authors may be the only soludon to this problem. This was done in a recent investigation, in which original investigators were contacted to provide unpublished data for analyses of cell type sprcific leukaemia in a combined cohort of more than 208 000 workers in the petroleum indus- In most of the original studies, cell type specific leukaemia results were not analysed or reported due to smallnumbers. Step 4, the rraditional qualitative review, is a main component of the overall process. It summanses the basic information from individua! studies, such as desi-en, subject selection, expisure patterns, diagnostic criteria, observatioE period, potential confounding factors, statisricai techniques, etc. The quantitative review shod2 also point out the stren-gths and weaknesses of individual studies. Table 2 provides a suggesrc.2 checklist for evaluahg studies. Most of tinitems in table 2 are self explanatory and cornmon to any qualitative literature review, and 2` Toblc 2 Cliccklisr .for t-caluuriirg studies Study design Selecrion of study sub~ects Sample size and srausucal power Statistical analysis Comparison populauon Duration and completeness of folliiu up Latency Diagnostic information Adequacy and spectiinn. of'exposure data Concomitant exposures Confounding factors -- I x :ood ISO a ield, dif-. 7 be 'Ugh lem, ition Josiconv on that ily." f the the 3een :rent mbina omcd." past and dual be udy, :the ialy- hich ults. ina d to pre3Ctly henhOrs This hich wide spenore dus- type :d or . is a ium:dud xpolaon ;tical odd es of sted i the :omId as y u ~ i r w-TSGU~JC clornJoliL~-- --__m i -_. ____J - '7 have special relevance in a rneta-analysis. The can be measured by a single score. In any principal objective of this review step is to iden- event, quality scoring does not improve objec- tify studies of acceptable quality for inclusion in tivity, as any scoring system can be arbitrary subsequent meta-analyses and to identlfy fac- and subjective. tors contributing to heterogeneity between Whether a particular study should be srudies, which can then be used to guide further included or excluded in a rneta-analysis should stratification of data (stratified meta-analyses). be based on a set of criteria such as those sug- Study design is used to group individual gested in the checklist in table 3. Only those studies for rneta-analysis. Generally, separate studies which meet certain minimum require- meta-analyses should be performed on studies ments should be considered further for inclu- of different designs. It is not usually advisable sion. An additional assessment has to be made to combine studies of different designs in a to find whether all such eligible studies can be single meta-analysis unless it can be deter- included. mined that study design has little or no influ- T h e first consideration in a quantitative ence on study characteristics such as quality of meta-analysis (step 5 ) is the ability of individual data, specificity of exposure, and uniformity of studies to be combined. T o some extent, the diagnoses. In reality, study design is usually definition of a study can sometimes be artifi- one of the most important determinants of cial or arbitrary. For example, a study of data quality, exposure specificity, and diagnos- petroleum refinery workers from an oil com- tic criteria. Similarly, studies with very differ- pany can consist of two or more refineries. .in ent statistical techniques, different comparison industrywide study obviously consists of more populations, or different diagnostic categories than one company. Data from individual should generally not be lumped into a single refineries or companies could have been pub- analysis. For example, in the literature on lished separately. Thus, the issue ofwhether or occupational epidemiology, some studies may not studies can be combined may a!ready exist report proportional mortality ratios (PhLRs) at the level of individual studies. and others standardised mortality ratios . The primary question to ask is: do studies (SICLRS). It is generally agreed that SMR stud- to be included in the meta-analysis have suffi- ies are far better than P%LR studies, and that cient common ground to consider the scien- P%Rs usually overestimate cancer risks in tific issue in question? Table 3 provides a list industrial populations due KO the healthy of items to be considered in evaluating worker effect. Therefore, ShLR and PMR whether studies can be combined. In occupa- studies should not be combined in a single tional studies, the question often asked is: is meta-analysis. Similarly, if one study examines there a relauon between exposure to a certain total digestive cancer in general and another chemical and a specific disease? Therefore, stomach cancer in particular, it would be inap- both specificity and patterns of exposures are propriate KO combine these studies as different obviously important determinants in deciding diagnostic categories are involved. whether it would be meaningful to combine The selection of study subjects, adequacy certain studies. As discussed already, separate and specificity of exposure data, concomitant rneta-analyses should be performed for studies exposures, and confounding factors should be of different design due to the heterogeneity in examined at the individual study level; includ- data collection and data quality. ing studies which are significantly deficient in Other study characteristics should also be any of t5ese areas would incresse the hetero- examined. The fundamental question is "Are geneity in the resultant meta-analysis. For studies so different that they should not be example, if different concomitant exposures combined?" Studies should be carefully exam- are present in individual studies, hererogeneiry ined for similarity or differences of characteris- in the combined analysis will result, as this tics such as eligibility of study subjects, cohort would reflect the influence of different con- definition, exposure definition and classifica- comitant e-xposures. Finally, although sample tion, disease diagnostic criteria, data collec- size or statistical power are important Criteria tion, etc. However, because many statistical for evaluating a study's quality on an individ- tests are available to evaluate heterogeneity of ual basis, it should not be part of the basis for study results or effect estimates, in many including or excluding a study in a meta- meta-analysis papers the emphasis on assess- analysis. As will be discussed later, one of the ment of heterogeneity has been incorrectly main benefits of a meta-analysis is the pooling placed on the results of studies rather than on. of small studies to improve sfatistical power. the basic characteristics of studies. In other Some authors have suggested the use of "quality scoring" to determine whether a study should be included. This suggestion has Tuble 3 Factors detcnriirring zultctlter studies cdn bc met with strong opposition. For example, in a corrrbirird in LI metn-~rraiysu recent commentary quality scoring was characterised as "the most insidious form of subjectivity masquerading as objectivity."l & We do not recommend the use of quality scoring for the simple reason that it would be impossible to treat dikyerent study char3cteristics (size, exposure information, completeness of follow up, ctc) that are related to quality 3s if they are Study design Speciticity of exposures PlttemS Of CXptJSUreS Concomitant exposures Specificlry of hralrh end points Uniformity ut'diagorric cntem Statistical procedures Informution available from individual studies for rneta-mulvsib C~~mporisupnopulations Heteropeneic?.of ciyect, nsk usiimatus words, heterogeneity of studies is more a scientific than a statistical issue. Nevertheless, statistical tests for heterogenein of study results can sometimes provide clues to sources of disagreement among studies." If the test for heterogeneity is significant, the underlying studies must be scrutinised closely. Once the source for heterogeneity is identified, the decision whether to combine the studies must be carefully considered. From the practical point of view, statistical procedures used in individual studies play an important part in the ability of studies to be combined, as do the type and amount of information presented in individual studies. Only studies presenting adequate information can be included in the meta-analysis. In occupational cohort studies, as will be discussed later, statistical procedures are usually similar and information needed for a meta-analysis is usually presented. Although several study designs are appropriate for investigating occupational health, for chronic diseases such as cancer, most studies use the historical cohort design. iMost occupational cohort studies are based on documented employment or exposure histories, and those included in meta-analyses usually consist of workers at similar places of work or from the same industry. As such, information on exposure is relatively reliable and exposure pattern is homogeneous. When linked to industrial hygiene data, such information can be relatively specific. Another epidemiological study design commonly used is the case-control study. Typically in a case-control study occupational information is based on recalls of the subjects themselves or family members, whir$h may not be reliable or specific. Other populption based case-connol studies may use population regismes to identify the industries and jobs of cases and controls. These different approaches to exposure classification berween cohort and case-control studies can result in significant heterogeneity. Diagnostic criteria can also be different between cohort and casecontrol studies. Therefore, in most situations we do not recommend combining cohort and case-control studies in a single meta-analysis. The meta-analysis should at least be stratified by study design. As the title suggests, this discussion will focus on the statistical procedure used in occupational cohort studies-the standardised mortality ratio (SMR). In a meta-analysis, once it has been decided which studies will be combined, one of the goals is to calculate a quantitative summary estimate of the effect or risk of exposure, as well as the corresponding precision, usually in the form of a 95% confidence interval (95% CI). The summary estimate should make maximum use of the information available and preserve the properties of the original estimates in individual studies (such as adjusunent for confounding). In general, the summary estimate is some type of weighted average of the effect measures or risks from the original studies, with weights inversely proportional to the variance of the individual estimates-that is, a combination of stratum specific effect measures, with individual studies as the strata. For cohort mortality studies, the summary risk estimate is the summary SMR or meta-SMR, which is defined as the ratio of the total number of observed deaths (summed over all studies) to the total number of expected deaths. That is, meta-S,MR = Z,O,E,E,, where the subscript i denotes the ith srudy, 0 the observed deaths, and E the expected deaths. Under the null hypothesis (SMR = l), the approximate variance of SMR in the ith study is lE,.Furthermore: meta-SMR = X:{W', x SI'VIR,I where W, = EjX.E: is the weight. That is, the weights for the meta-SMR are propomonal to the inverse of the variance of individual studies. In other words, more weight is given to larger studies, and \ice versa. In calculating a meta-SMR, each srudy is treated as a separate stratum, and the data pooling is therefore similar to the indirect standardisation in individual studies. T h e total data from individual cohorts can be considered as a "megacohort." Figure 1 shows the concept graphically. It should be evident from both the statistical expression meta-SMR = Z,{W, x S M q : and fig 1 that a meta-SMR is based on a series of comparisons within each study, although the computational formula meta-SMR = X,O,T.E, shows that it is the ratio of the total observed deaths KO the total expected deaths. A meta-SMR is simply a summary risk index based on all the data, with Figure 1 Relutiom betweeti subcohons. cohorts, aird nrcgacohort. iised lysis, 111 be ate a Ct or 1dmg onfiesti. the >perjtud- wme sures ~1ghtS f the an of !kidtaliry sum:d as med total the la1 to studs to dY data direct total nsids the From viR = AR is each mula j the total dy a with individual studies as a stratifying variable. The 95% CI of the meta-SMR can be calculated with the same procedure as SMRs for individ- ual studies (for example, treating it as a Poisson variable1").Data needed to calculate a meta-SMR are the observed and expected deaths for the causes of interest from individual studies, which are usually reported. Once the meta-analysis has been completed, the result should be integrated with the qualitative component of the meta-review (step 6), especially the discussion on the strengths and weaknesses of the individual studies. It should be emphasised that a meta- analysis will not compensate for any deficiencies that the individual studies may have, other than low statistical power. As such, it is important to discuss the impact of suspect individual studies on the meta-analysis. Analysis of sensititiq- can be helpful for this. In such an analysis, studies of marginal quality, with slightly dissimilar exposures or slightly different from other studies in some aspects, are removed, and the result is then compared with the one that includes them. The results of a meta-analysis are directly applicable to two important criteria of causa- tion (step 7 ) :strenprh and significance of asso- ciation and consistency. The strength of association is given directly by the magnitude of the meta-SMR, and the statistical significance is measured by the 95% CI. Risk cstimates from individual studies, especially for rare diseases, can be unstable and inadequate in statistical power. If the results from individual studies are in general agreement, however, a summary risk estimate such as the meta-SMR takes the consistency of results into consideration. The meta-SMR, based on the combined database, is more stable and is less vulnerable to flucruations due to small numbers than the original studies. Thus, a meta-analysis takes into consideration not only the size of risk estimates from individual studies but also their consistency. So far our discussion of statistical proce- dures in meta-analysis of cohort studies has been based on the summation of observed and expected deaths. This method is commonly used because the information needed is usually available. Depending on the information presented in the underlying studies, other statistical techniques may be used.'- Is In general, the weights for the summary risk estimate are inversely proportional to the vanance of the individual study. The choice of a particular statistical technique depends on information presented in individual studies. A n y valid statistical procedure of rnera-analysis should be based on an objective weighting system. This implies that exposed and non-exposed groups are compared within each study. In other words, each study is treated as a stratum in the meta-analysis. The separate identity of each study is preserved, and it is the estimates of the effect from each study that are combined according to appropriate weights."' AS already discussed, the method commonlv used in combining cohort studies clearly satisfies this criterion. On the other hand, simply pooling resuns mom individual studies, ignoring study as a stratifying variable, is not a valid procedure.:" For completeness, we also mention another statistical issue of meta-analysis: fked effects versus random effects models. Fixed effects models assume a single effect estimate for all studies, whereas random effects models take variability between studies into consideration. Most methods of meta-analysis commonly used (including the one discussed here) are k e d effects. There is an ongoing debate among epidemiologists about the choice of models. In this paper it suffices to say that the choice of models does not change the risk cstimate, but may affect the width of the 95'X CI. It should also be noted that, as already discussed, the definition of a study can be arbitrary and the concept of variability between studies can be ambiguous. Random effects models have been criticised for assigning more cven weight to all studies, thus incorrectly overemphasising the results of small studies.'" Interested readers are referred to some recent publications on the issue of fixed and random effects models. ' ' I'' Discussion and an example Meta-analysis is a powerful tool in reviewing, summarising, and integrating scientific data from multiple studies. As stated by Glass dr a1 "the findings of multiple studies should be regarded as a complex data set, no more comprehensible without statistical analysis than would hundreds of data points in one study."" .4 meta-analysis not only provides a summary risk ~Stimatewhich is properly weighted, but also takes consistency of results of individual srudies into consideration. An additional bene- fit of meta-analysis is that it comprehends the multiple comparison problem exhibited in individual cohort studies. Typically in an occupational cohort study more than 40 causes of deaths are examined. Statistically significant results for as many as two causes could have occurred by chance alone given a type 1 error of 5% (or a 95% confidence level). By combining data across studies, chance excesses or deficits for a parricular cause in individual studies are less likely to influence the metaanalysis. The meta-SMR discussed has several desirable attributes. It is the natural extension of the indirect standardisation of summarising results from age, sex, and race specific strata in the original studies. Being an extension of the indirect standardisation, the meta-analysis preserves the statistical properties of the original estimates as well as the original interpretation of the effect measure. In particular, by treating each study as a stratum, the summary ct'fect measure preserves the adjustment for confounding made in the original studies. The number of expected deaths in each study is proportional to the size and age of the individual cohort as well as the duration of observation. From the practical point of view, the data required for a calculation of a rneta-SMR are almost always available in the individual reports. . Tabk 4 Summaty ojnronal~rfronrbrain cancer b prroleu?rrworkers': Wong 1480' &plan 1986. Schottenfeld cr a1 1981. -Nelson 1985 Uonp cr a/1986 Hams si ul 1982. Hams cr ul 1985 Wen dr ul1983 Morgan and Wong 1984 Morgan and Wong1985 Enterline and Henderson 1985 Di\me o ul 1985 Divine and B m n 1987 Rushton and Aldmon 1981 Rushton a d hldcnon 1981 Thcriault and Gouler 1979. .Theriaulrand Provencher 1987 Christie 1987 Total (mew-analysis) placf 17 US refineries 17 US refineries US refineries 10 US refineries 2 California refineries 1 Louisiana refinery 3 US reiinerics 1 Texas refinery 1 Texas refinery 1 Pennsylvania refinery 1 California refinen. 13 US retinenes Production and pipeline in US 8 UK refineries Distribution centers in UK 1 Canada refinery 1 Canada refinery Australia refineries 20163 20163 55007 10763 14179 8666 21698 I 5095 6139 4263 1621 19077 11098 34781 23306 1205 1207 10489 'Escluded from h e calculation of rhe meta-SMR because of overlap of data. Follmu up 1962-71 1962-80 1977-79 1970-82 1950-80 1970-77 1970-77 1935-78 19-15-79 1046-79 1959-78 19-17-77 1946-80 1950-75 1950-75 1928-75 1928-8 1 1981-86 8 22 8 .1.835 15 23 9 7 3 31 11 36 39 3 4 2 210 Expcczcd dah 1129 21.63 4.90 9.60 17.48 4.90 13.00 22.85 8.26 6.64 2.09 2840 15.90 44.77 3634 0.77 1.87 2.70 209.70 SMH (95% CI) iw0.71 p.31 tu 0.89 (0.56to 1.33) o m1.63 (0.70to 3.21) (0.3610 1.631 1.25 (0.78to 1.89) 1.02 (0.33 to 2.38) 1.15 (0.64to 1.90) 1.01 (0.64IO 2.37' 1.09 (050 to 2.071 1.05 (0.4210 2.161 1.43 (0.30to 4.18'1 I .11 (0.76 to 1 43 0.69 (0.35to 1.58) 0.80 (0.56to 1.1 1) 1.07 (0.76 to 1-46) 3-89 (0.80 to 11.37) 2.13 (05810 5.451 0.74 (0.09 to 2.67) 1.00 (0.8710 1.15) Table 4 provides an example of meta-analysis of data on brain cancer in petroleum workers originally reported by Wong and Raabe in 1989.': As discussed in the original reporc, both published and unpublished (but otheruise meeting the eligibility criteria) cohort studies of petroleum workers were eligible for inclusion in the meta-analysis. The PhLR studies were not included because of their inherent limitations. Case-control studies were also not included because of the inadequacy of exposure mformazion (based primarily on recalls), but they were included in the discussion. An examinanon was made of all cohort studies for evidence for heterogeneity of cohort identification, data collection, and stadsdcal analysis. Based on this examination, two cohort studies were excluded because of methodological or data deficiencies. For the remaining cohort studies, it was decided that no significant heterogeneity was present and the data could be combined in a meta-analysis. All these studies were historical cohort studies based primariiy on employment records. Exposures in these studies were relatively homogeneous in the sense that a11 study subjects were peaoleurn workers. Methods of follow up and companson populations were also similar. In the metaanalysis, duplicate data (such as indusuyviac studies or data based on earlier follow ups' were excluded. Table 4 and figure 2 show that the brair. Nelson 1985 - -Wong et a1 1986 !E Hank et a1 1985 - Wen eta1 1983 - 7 r- -2 u, 2Morgan and Wong 1984 - 0 Morgan and Wong 1985 - -Enterline and Henderson 1985 -Divine et all985 -Divine and Brown 1987 II f 2 5 5 m 0) Q Rushton and Alderson 1981 - 2 2 1 0 Rushton and Aldenon 1981 - 2 N I Therialt and Provencher 1987 - tn -=Christie 1986 s! In - stn u) Total (Meta-analysis) - 0.1 Fiptrre 2 M m d i n ~ f m n rbrabr iair;c.r in perrolcuar workers. I - 'i -1I I T 8I' i 1 -- i I I--& 1.o - I .*_ . idies i1 or lhort hetd be idies .arily kese the eum pariietawide UPS) >rain J 10.0 cancer meta-analysis consisted of listing the SlMRs and the corresponding 95% CIS from individual studies as well as the meta-SMR and its 95% CI. The presentation of risk estimates as well as other key study variables from all individual studies in a single table is obviously very informative. Similarly, the graphical Presentation is a helpful visual aid. Unfortunately, in some recent meta-analyses only graphical presentations are available. T h e lack of a presentation of numerical data often makes it difficult for the readers to determine what data from the original studies were actu- ally used in the meta-analyses. Thus, we rec- ommend presenting data both numerically and ,graphically. In our example of brain cancer in petroleum workers, the meta-SMR was 1.00 (95% CI 0.87 to 1.15). Therefore, the combined data had reasonable power to rule out an increase in risk of brain cancer as small as 15%. The improvement in statistical power through the use of meta-analysis is evident from the narrow 95% CI of the meta-SMR and the small increase in risk that the data could detect. A meta-review should not stop at the quantita- tive meta-analysis. In the example of brain cancer in petroleum workers, the result of the meta-analysis based on cohort studies was discussed in conjunction with results from casecontrol studies. It was concluded that petroleum workers did not experience any increased risk of brain cancer." It should be pointed out that the application of meta-analysis is not without criticisms. Some of the common criticisms are: the inclu- sion of both good and poor studies; publication bias resulting from the tendency of positive studies being more likely to be published; the reliance on published studies only; overrepresentation of data from studies published repeatedly; and most seriously, the oversimplification of a complicated issue by focusing on a singlle summary risk estimate. Lklost of these criticisms arise from improper application of meta-analysis. T h e general ,guidelines provided earlier should help to minimise these problems. It should also be noted that all of these criticisms, except for oversimplification, are equally applicable to a traditional qualitative review. The last criticism of oversimplification of study results is equally applicable to summaries of data specific to strata-for example, age groups-in an individual study. Several decades ago, the pros and cons of summarising data in an individual study were vigorously debated.'+:' The issue at that time was whether risk estimates from different age groups in a study should be summarised with a single index such as SMR, yet today it would be inconceivable that SMRS or some other summary indices are not calculated in a cohort study. As already discussed, a reasonable and practical approach is to examine data for evidence of heterogeneity, and if necessary provide stratified meta-SMRs or sensitivity analyses. The issue of heterogeneity is more scientific than statistical. Although a statistical test for heterogeneity of study results is help- ful, other aspects of the original studies must be investigated before studies can be combined (table 3). The usefulness of a meta-analysis is obviously dictated by the quality of the input data of the original studies and the amount of information available to the meta-analyst. With regard to the amount of information available, hopefully in the future authors will report their investigations in a uniform manner and provide sufficient details to facilitate any future meta-analysis, and journal editors will include as part of their decisions the potential contribution of a manuscript to a future meta-analysis of similar studies. summary ikleta-analysis is useful in selecting studies, and in organising, presenting, and summarising results fiom individual studies. The benefits of conducting a meta-analysis in a review include enhancing statistical power (particularly when original studies are small), providing a summary risk estimate, taking consistency of study results into consideration, minimising the problem of multiple comparisons, and examining data for heterogeneity. Compared with a traditional qualitative review, a review incorporating a meta-analysis tends to be less subjective. Rather than replacing the traditional qualitative review, a metaanalysis should be made part of the overall assessment. We propose the use of the term meta-review to emphasise the importance of both qualitative review and quantitative metaanalysis. 1 Glass GV. Primary, secondary and meta-analysis of research. Eduurimu' Rrseurch 1976;5:3-5. 2 Cochrm WG. Problems arising in the analysis o f a series of similar cxperiments.Journul of the Royal Starisrid Soci>n, Supplement B 1937;1:102-1 8. 3 Cuchran WG. The combination of estimates kom different experiments. Biomerncs 1954;lO:101-79. 4 Fisher RA. Scurisncul merhods Jor research -J.orkm. 4th cd. London; Oliver and Boyd, 1932. 5 Peanon ES. The probability integral transformation for testing goodness of fit and combining independent tests of significance.Biommika 1938;30:134-48. 6 Surgeon General. Smoking and healh. Washingcon; US Government Prinung Office, 1964. 7 Huque MF. Expcnence with meta-analysis in NDA sub- missions. h c e e d i n p of rhr Bwphumucologrcul Secnon of the American Stuturicul Association 1988;2:28-33. 8 Wrbsrer's npw %odd Jicrionury of Americun Enqlish. New Y.o~~~ rk: Prcntice Hall. ~~ 1994. 9 Wzbsrer's new colleqrarz Jictionury. Springneld, Massachusetts: Memam, 1990. 0 Cook DJ, Guyatt GH, Ryan G, Clifron J, B u c h g h a m L, Willan A, a d.Should unpublished data be included in a meta-analysis?J!.LCM 1993;169:1749-53. 1 Morgan RW, Foliart DE, Wung 0. Asbestos 2nd gastnc cancer, 3 review of the literature. IVcstY .l.lsd 1Y85;143: h0-5. 2 Wong 0, b a b e GK. Critical review of cancer cpidemiol- ogy in petroleum industry employees, with a quantitative mem-analysis by cancer site. A m btd .bred 1989;13: 283-310. 3 Wung 0,Raabr GK. Cell-type-specific leukemia analyses in a combined cohort of more than 20s 000 petroleum workers in the United Stares and the United i(ingdom, 1937-89. Reyul Toxiul P ~ I I T I I U19~9U5;~21:307-21. 14 Greenland S.Invited commentary: a cntical look some popular meta-analytic methods. Am 7 Epidcnrrol 19Y4; 1Jo:390- h. 15 Greenland S. Quantitative methods in the remew of cpi- demiologic literature. Eptdmrrol RLV I987;9:1-30. 16 Bailar JC, Ederer F. Significance factors for the ratio of a Poisson variable to its expectation. Biumecncs 1964;20: 63943. 17 Armitage P, Berry G. ScurisrILd ritc.tlruif~III mrdicul ri*scun.lr. 3rd ed. Oxford: Blackwell, 1994. ~~ 18 Morgan RW, Wong 0. A review of cpidcminlop;al studies on anihcial sweetencn and bladder cancer. F.(oi&rii .-Ipp/ Toxicol1985;23:529-33. 19 Pocock SJ,Thompson SG. Assessing the widence: the role ofmeto-analysis. In: Marmot M, Elliott P,eds. Conwan hcan d i s u w cpidc7nio&&y, .honi uctidogv 10 p t A k lrsultli. Oxford: Oxford Univmiry I'ress, 1991. 20 Dickersin K,Berlin JA. Meta-analysis: state-of-the-science. Epi&mial Rrv 1992;14:154-75. 11 Peto R Why do we need systematic reviews of randomized Uials? Stat Mcd 1987;6:241. 22 Meier P. Cornment;uy. Sur Akd 1987;6:329-31. 23 Bailey KR Inter-study ditrercnccr: how should they inRuence the interpretation and analysis of results? Stor .\lsd 1987;6:351-8. 24 Glass G, McGaw B. Smith ML. Mcru-urr+w 111 SII.XJ! mcadi. Beverly HiU, California: Sage, lQfil. 15 Yule GU.On some points rcladnp to vital statlsticr. m o x cspeaatly stamtics of occupational monality. .7oimioi o/' the R q d SIUUSZICS UoIc l 1~9~74;97:1-72. 26 Kitagawa EM.Component of a difference beween nvo rates. ~ o of rhcd American Smlutical Associarron 1053; 5 0 1168-94. 27 Kitapwa EM. Standardized compansons in populauon mearch. h o p a p h y 1964;1:196-315. 18 Kitapwa EM. Theoretical considerations in the selection of a monality index, and some empirical comparisons. Hum BIOI 1966;38293-308. 29 Elvebacl;LR. Discussion of "Indices of mortality and tests of their sratistical significance." Hunt Biol1966;38:3224. I I 1 I ! j Vancouver style All manuscripts submined to Occup Environ Med should conform to the uniform requirements for manuscripts submitted to biomedical journals (known as the Vancouver style.) Occup Environ Med, together m5th many other international biomedical journals, has agreed to accept articles prepared in accordance with the Vancouver style. The style (described in full in the BM3, 24 February 1979, p 532) is intended to standardise requirements for authors. References should be numbered consecutively in the order in which they are firsr mentioned in the text by Arabic numerals above the line on each occasion the reference is cited (Manson` confirmed other reports' ' . . .). In future references to papers submitted to Occup Eimirori Med should include: the names of all authors if there are seven or less or, if there are more, the first six followed by et al; the title of journal articles or book chapters; the titles of journals abbrellated according to the style of Index hdedicus; and the first and final page numbers of the article or chapter. Titles not in Index Medicus should be given in full. Examples of common forms of refer- ences are: 1 International Steering Committee of Medical Editon. Uniform reqummenrs for manuscnpts submitted IO biomedical journals. BbfJ 1979;1:532-5. 2 Sorer NA, Wasserman SI, Austen KF. Cold urucana: release into the circulaoon of histamine and eosint': phi1 chemotactic factor of anaphylaxis during cold challenge. S EtiglJ Mcd 1976;294:667-90. 3 Weinstein L. Swam W .Pathogenic properties LT: invading micro-organisms. In: Sodeman U A l r . Sodeman UA,eds. Puthulofic plrysrdo~,nicc1~uiir.w of discard. Philadelphia: W B Saunden, 1974:457- 71 M cC DeP Occ: Fact 3450 Suit, QuC` A DL cM i P Lc, Sen Pnei Hosr 1'Un Sher ! Sher Call: R Be. I S M3 i Insti TenraSv;. hbo oues &isa Monr Cana G Per ,I Come. DrA 1 Unive: Occur Faculp Univ,, Montr H3A : AeCCPl