respiratory cancer--a meta-analysis of the epidemiologic evidence.

PARTANEN T. Formaldehyde exposure and respiratory cancer - a meta-analysis of the epidemio logic evidence. Scand J Work Environ Health 1993;19:8-15. A recent meta-analysis by Blair and his co-workers stimulated the undertaking of a reanalysis of the epidemiologic evidence of an asso ciation between formaldehyde and respiratory cancer. Emphasis was placed on extracting the maxi mal amount of relevant data from the source studies. In close agreement with the original analysis, the aggregated evidence did not indicate an excess risk or an exposure-response gradient for lung cancer. An exposure-response gradient was seen for both sinonasal and nasopharyngeal cancers. The risk for substantial exposure was significantly elevated (odds ratio 1.7 for sinonasal and 2.7 for nasopharyn geal cancer). It is suggested that at least substantial levels of occupational exposure to formaldehyde are associated with a risk of these cancers. The excess risk would be of the order of 100% or more over background rates.

A recent review in the Scandin avian Journal of Work, Environment & Health by Blair and his coworkers ( I) consolidated the results of over 30 epidemiolo gic studie s on formaldehyde exposure and cancer. A causal role for formaldehyde was considered the most credible for cancers of the nasopharynx and, to a less extent, nose, nasal cavities, and sinuses, but not the lung. The effects on the aggregated results of the substitution and rederivation of some of the data extracted from the studies are scrutinized in the present reanal ysis. The substitutions and rederivations of input were based on considerations of the maximal amount of information and maximal relevance.

Methods
The source studies for the present reanalysis were, with some updating, very much the same as in the original meta-analysis (I) . The overlaps between some of the studies in the data were removed according to reference 1. To retain comp arability, the statistical treatment of input from the studies waswith an important addition pertaining to the upper respiratory sites -kept the same as in the original analysis (l). Thus the observed (0) and expected (E) frequencies? were summed over the studies in the 1 Institute of Occupation al Health, Helsinki, Finland. 2 In studies in which expected freque ncies were not readily available, the expected values were derived by dividing the appropriate observed values by the appropriate risk ratios or odds ratios.
Reprint requests to : Dr T Partanen , Institute of Occupational Health, Departm ent of Epidemiology and Biostatistics, Topeliuksenkatu 41 a A, SF-00 250 Helsinki, Finland. 8 broad category of any formald ehyde exposure exceeding background level and in the graded categories of low-medium and substantial exposure. In several studies, graded categories of exposure were supplied, variably delineated in each study via exposure probability, intensity, duration , cumulated exposure, or some other related indicator of the authors' choice. The aggregate risk ratios (RR) were estimated as aggregated observed-to -expected ratios, and the 95% confidence limits were set for the RR values by the Poisson model (2).
The upper respiratory sites were considered in the subcategories of (i) nasal cavity and sinuses, (ii) nasopharynx, and (iii) others (oropharynx, hypopharynx, lip, tongue , salivary glands, and mouth). The data for these sites derived overwhelmingly from case-referent studies with imprecise expected frequencies of exposed cases. These data were therefore aggregated via the more appropriate log-Gaussian, fixed-effects model for risk ratios (3), in addition to the Poisson format. The procedure, in essence, combined the estimates of the log RR values (or log odds ratios) by weighting each by the inverse of its variance. The confidence limits of an aggregated RR were then computed on the assumption of a Gaussian distributi on of the log of the estimate of the aggregated RR. The required component RR values and the variance s of the log RR values were extracted from the study reports. In some reports, estimated standard errors of the log RR values were readily provided; in others they were recovered from the reported confidence limits. In the few standa rdized mortality ratio (SMR) studies addressing upper respiratory sites, the variance of the RR (of the SMR, that is) was taken as the inverse of the observed frequency. These and other procedures have been de-scribed by Greenland (4). A chi-squared test for homogeneity of the component RR values (3) was applied .

Input values
The main difference betwe en the original analysis (1) and the reanalysis was the selection of the input values. Extraction of the maximal amount of relevant data from the different studi es was pur sued. Thus some individu al realization s of input parameters chosen by Blair et al (I) were repl aced by what were considered slightly more appropriate ones. The genera l principl e was, in the qualitative analysis, to contrast the risk connected with formaldehyde exposure that exceede d background exposure to that associated with background exposure (" unexpose d" subjects). The predetection latency period was accounted for when lagged inputs (RR or OR values referring to a period of more than 10 years after the onset of exposure) could be extracted. In the quantitative setting, similarly, the aggr egated exposure-response analyses were based on lagged input values, if these values could be derived. In additio n, confounder-adjusted values were extracted whenever avai lable.
The substitutions made in the input s for lung cancer appear in table 1, and the inputs that were accepted as such are readi ly found in reference 1. The complete inputs for the cancers of the upper respiratory sites are given in tables 3, 4, and 5 in the Results section. All of the changes are explained in the following text.
In Harrin gton & Shannon 's mortality study of British pathologists (5), the figu res of lung cancer for men in England and Wales were repl aced by those fo r both genders in England, Wa les, and Scotland in order to exploit fully the information provided in the report.
In the report by Harrington & Oakes (6), on a later follow-up of British pathologists, the mortality figures for lung cancer among both genders were preferred instead of those of men only; this was numerically a min or adjustment.
In place of unlagged values for lung cancer, lagged equivalents were extracted from the repo rts of Blair et al (7) on United States (US) workers producing and using for malde hyde , Bond et al (8) on Dow Chemical (Texas) workers, and Stayner et al (9) on US garment workers exposed to formaldehyde . For the exposure-res ponse ana lyses in the large mortality study by Blair et al (7), low-med ium forma ldehyde exposu re referr ed to any nonzero expo sure up to 5.5 ppm-years, and any expos ure exceeding 5.5 ppm-years was considered substantial.
Logue et al (10) contrasted the mortality of US pathologists with that of radiol ogists. reported death rates of 1.37/10 3 years for all cancers and 0.2 1/10 3 years (SMR 0.24) for respiratory cancers, the observe d number of respiratory cance rs (as a proxy for lung ca nce r) was calculated as (0.2 1/ 1.37) · 91 = 14, and the expec ted number as 14/ 0.24 = 58 .
From the mortality study by Acheson et al (11, 12) on British chemica l workers, nil, background, and unkn own exposures to formaldehyde were excluded from the category of exposed workers.
The input for lung cance r from a Swedish national record linkage study by Malker & Weiner (13) in the original meta-analysis (1) presum ably incl uded, amo ng others, all physician s, all biologists, most if not all textile wor kers , garment shop assis tants, and paper and cardboa rd produ ction workers as occupationally exposed to forma ldehyde. In the reana lysis, data for fiber workers in plywood and fiberwood factories only were inc lude d.
The recent updating of the data on Finnish woodworkers by Partanen et al (14) provided, forthe qualitative analysis, OR values for lung cancer, with adjustment for age, smoking, and vital status at the time of data collection.
In the case -referent study of sinonasa l cancer conducted by Hayes et al ( 15) in The Netherlands, independent reconstru ction of individual expos ure histories from interview data was acco mplished by two hygien ists. Blair et al (1) used the assessment of one of the two. In the reanalysis, mean freq uencies were calcul ated betw een the two hygienists in the cells of a three-dimensional arra ngement spanned by wood du st (two categories), maximal form aldehyde level (three categories), and case-refere nt status, from the frequencies in table III in reference 15. Th e OR va lues adjusted for wood dust and their standar d errors (on the log scale) for any, low-med ium, and substantia l exposure to formaldehyde were then calcu- Table 1. Changes in input between the original analysis (1) and the reanalysis with respect to lung cancer." [0 = observed frequency, E = expected frequency, RR = risk ratio , Input values that were not changed can be found in reference 1 and they have not been reproduced here. , Respiratory cance rs.

Any exposure
Scand J Work Enviro n Health 1993, vo l 19, no I lated with the Mantel-Haen szel procedur e (16), and the expected numbers needed in the Poisson analysis were derived as a /OR , Vaughan et al (17) conducted a case-referent study on sinonasal, nasopharyngeal, and oro-and hypopharyngeal cancers in west Washington in the United States. For the low-medium and substantial exposure categories in the reanalysis, IS-year lagged OR values derived through a job -exposure matrix from telephone interview data were ava ilable. The figure s were readily adjusted for age, gender, cigarette smoking, and alcohol consumption by incorpora tion of these cov ariates into a multiple logistic regression model. The graded exposure level s were based on a multiplicative function of maximal exposure level and number of years expose d. For the qualitative analysis, exposure levels over background were collapsed into "any" exposure to form aldehyde , and no adj ustment or lagging seemed to be possible for the reported figures.
Vaughan et al (18) also studied residential exposures to formaldehyde from mobile homes and other structures. Number s referring to living in mobile homes, lagged by 15 years, were selected for the reanalysis. Figures for "any expos ure" were drawn from the category of " > I year in mobile home," adj usted for cigarette smoking, alcohol consumption, gender, and age for sinonasal cancer and for race and cigarette smoking, and lagged by 15 years, for nasopharyngeal cancer.
From the populati on-based case-referent study of sinonasa l and nasopharyngeal cance rs in Connecticut by Roush et al (19), figures adjusted for age at death, year of death , and quality of occ upational information were extracted. The low-medium category of formaldehyde exposure was taken to equal " probably exposed to some level for most of the worker' s worklife and probably exposed to some level for~20 years prior to death," and the substantial level was "probably exposed to some level for most of the worker's worklife and probably exposed to a high level for~20 years prior to death." The latter category was included in the former one in the 10 present analysis becau se of the way the results were presented in reference 19. The low-med ium category therefore served also as the " any" category . A Danish case-referent study of sinonasal cancer by Olsen et a) (20) pro vided lO-year lagged figure s for formaldehyde exposure after adjustment for wood dust exposure.
Three studies missin g in the original analysis ( 1) were included. From a case-referent study of sinonasal cancer by Brinton et al (2 1), gender-adj usted figures for formaldehyde exposure were available. Another case-referent study by Gallagher et al (22) on cancers of the oral cavity and oropharynx provided adjusted figures for "any," as well as "probable or definite" exposure, the latter being used as input for the "substantial" category of exposure. Merletti et al (23) published their case-referent study on cancer of the oral cavity and oroph arynx after the original meta-analysis (1). They provided OR values adjusted for age, education, area of birth in Italy, tobacco smoking, and alcohol consumption. The "any" exposure catego ry served as both "a ny" and "lowmedium" exposure categories in the present analysis, as no separate "low-medium" catego ry was provided by the authors. "Probable or definite" exposure, (ie, expos ure very probabl y or definitely higher than that of the general popu lation) was selected for "substantial" exposure.
A recent report of a record-linkage study of occupational factors for nasopharyngeal cancer in Sweden by Malker et al (24) addressed, among other issues, formaldehyde. A standardized incidence ratio of 3.9, based on four exposed cases , was computed for fiberboard worke rs. However, no cases were reported among veneer and plywood workers, and an adequate input could not be extracted from the report. The study was therefore not considered in the present reanalysis.

Lun g
The results of the qualitative analysis of lung cancer are shown in table 2. Only the Poisson model was applied.
Practitioners of relevant medical specia lties (pathology, anato my, forens ic medicine) displayed a markedly and significantly depressed risk of lung cancer (aggreg ated RR = 0.3 in the original analysis and upon reanalysis) (table 2). The aggregated RR for funeral directors, embalm ers, and undert akers! was 1,0 in both analyses, and, in close agreement Table 3. Log -Gaussian" and Poisson" meta -analyses lor form ald ehyde exposure and sinonasal can cer. (0 = observed frequency, E = expected frequency, RR = risk ratio, SE = standard error, 95% GI = 95 % conf idence interval)  (1) ; the confidence limits wer e recalculated by th e pres ent autho r. C See refer ence 3.

Upper respiratory sites
The complete input data and results of the log-Gaussian and Poisson meta-analy ses for sinonasal, nasopha ryngeal, and other sites of upper respiratory cancers are shown in tables 3, 4, and 5. In the reanalysis of both sinonasal and nasopharyngeal cancers (tables 3 and 4), a significant increase was associated with the "s ubstantial" exposure category (RR 1.7 for sinonasal cancers and 2.7 for nasopharyn geal cancers). Neither an increased risk nor an exposure-response relation was suggested by the aggregated data for the combined category of oropharynx, hypopharynx, lip, tongue , salivary glands, and mouth (table 5).

Discussion
Some of the central problems of epidemiologic metaanalyses were encount ered in this reanal ysis. Thus the choice of the observed and expected frequencies, or the RR values and their variances, from the various studies may have suffered from subjectivity in a single-author exercise. A weighting of the component inputs by study quality was not formally attempted (but the choice of input s was given some thought), and the source data were weighted by precision only. Some heterogen eity and inconsistency was unavoidable among the input s from the different studies, particularly for lung cancer. All input was based on published -and most of it peer-re viewed -result s. The procedure was double-edged in that some data could have been missing, but, if the missing data were available, their relevance and quality might have been difficult to judge.
Considering the statistical procedures, the Poisson model resulted in narrower confidence limits than did the log-Gaussian risk ratio model. Strictly taken , the Poisson model is incorre ct when the expected frequencies contain error , which is essenti ally the case when the expected values are based on relati vely small numbers of reference subjects. In particular, when aggregating results of a few small-sized casereferent studies, the Pois son formula may create a deceptive illusion of precision, and the log-Gaussian risk ratio model is preferable. The changes in the point estimates of the RR values between the original and the rean alysis were not however materially affected by the model; the changes in the input values were decisive.
Despite a fair number of changes in the input values, the results of the reanalysis agreed generally well with those of the original analysis (l). Thu s, in light of epidemiologic eviden ce, it doe s remain unlikely that workplace exposures to formaldehyde pose any substantial lung cance r hazard among human s. The exposure-response grad ient for lung canc er was slightly inverse, if any. The statistical significance on the Poisson model of RR = 1.2 for indus trial workers derived from large numbers. Some persistent confounding (eg, from smoking) may have accounted for the increa se, but a weak true effect of formald ehyde could not be absolutely excluded. The deficit observed among anatomists, pathologists, and speciali sts in forensic medicine was confounded by social class and had nothing to do with formaldehyde.
The aggregated data for the mixed category of cancers of the oroph arynx, hypopharynx, lip, tongu e, salivary glands, and mouth was not indicative of any association with formaldehyde exposure. The data derived from three studies only, and the inputs were somewhat heterogeneous, the RR values from the component studies ranging from 0.4 to 1.8.
Upon reanalysis, and in slight contrast to the original analysis (1), sinonasal cancers did reveal an exposure-response gradient in the sense that the risk in the category of substantial exposure was significantly elevated (RR =1.7 in the log-Gaussian model) . Such a gradient was observed only for nasopharyngeal cancers in the original analysis (1). The rederivation of the input seemed to sharpen the gradients slightly for both sites, as compared with the results of the original analysis (1) . On this evidence, the two contiguous sites appear to be the most likely targets for the carcinogenic action of formaldehyde in humans.
Confounding by wood dust is a concern, particularly when nasal cancer is considered, and it might be argued that some of the postulated nasal carcinogenicity might be due to some constituent of wood particulates other than formaldehyde. In two (IS , 20) of the four studies that quantitatively contributed the most to the reanalysis of sinonasal cancer, however, exposure to wood dust was controlled for. In the one that addressed exposure-response for formaldehyde (15), a positive gradient was observed, and in the one that did not (20) a significant overall excess risk was found. Hayes et al (15) made a further note that formaldehyde exposure in their study was linked with squamous cell carcinoma, a finding consistent with data on rodents (27), while wood dust was the most strongly associated with nasal adenocarcinoma, as could be expected, particularly with hardwood dust, from extensive epidemiologic data (28,29).
The aggregated exposure-response gradients for nasal and nasopharyngeal cancers were probably understated because the graded categories (low-medium and substantial) reflected the full variety of conceptions and definitions adopted in the different studies. An unknown degree of rnisclassification between exposure categories resulted from this heterogeneity. In some studies, job-exposure-matrix-type translation of job titles into exposure levels probably compounded the misclassification rates. If, on the other hand, publication bias or a file-drawer phenomenon (3) was operating, it would probably have had a positively biasing effect also on the exposure-response gradients. In light of the considerable industrial and economic importance of formaldehyde, however, the assumption may be defendable that also nonpositive high-quality studies on formaldehyde and cancer tend to get submitted and published.
Some further epidemiologic data, independent of the input used in this report, may enhance the nasal-nasopharyngeal hypothesis . Matanoski (30) has reported a 4.7-fold increase in nasopharyngeal and pharyngeal cancers (and also an excess risk of pancreatic and brain cancers and leukemia) among pathologists, as compared with psychiatrists. The author, however, suspended judgment as to the role of Scand J Work Environ Health 1993, vol 19. no 1 formaldehyde, since there were other chemical exposures encountered by the pathologists. This study was not included in the formal reanalysis. Similarly, the data suggesting an elevated risk of nasopharyngeal cancer associated with formaldehyde exposure , reported in Sweden (24), were excluded, as the assessment of formaldehyde exposure was somewhat unclear in that particular study.
If the nose-nasopharynx hypothesis gains credibility, then rats and humans will come closer to each other in that the carcinogenic action of formaldehyde would affect both around an immediate target , which for rodents is the anterior mucosa of the nasal cavity and which for humans, for anatomic and physiological reasons, is both the anterior mucosa of the nasal cavity and contiguous sites such as the nasopharynx. It is also worthwhile to acknowledge that -in contrast to the animal cancer tests in which chemicals are administered at the maximal tolerated dose (31) -some of the levels of formaldehyde exposure scheduled in the animal carcinogenicity tests (27,(32)(33)(34) that induced tumors in nasal passages of rats were in the range of realistic workplace exposures. Extrapolation between species poses problems , however, as usual. Both the disposition of formaldehyde and , possibly, the tumorigenic response seem to be species dependent among rodents alone, as no carcinogenicity was demonstrated in hamsters (35) in an exposure scenario comparable with that of the rat experiments; and the effects on mice were much less pronounced than those on rats (33). The latter finding has been explained simply by the fact that, when exposed to formaldehyde, the mouse reduces its breathing volume substantially more than the rat, and therefore also lowers the amount of exposure (36).
The biological plausibility of the nasal-nasopharyngeal hypothesis is enhanced by the fact that formaldehyde, as a reactive compound, is rapidly and extensively metabolized to formate. Formaldehyde has genotoxic, typically cross-linking properties (37,38). It is notable that reactions of formaldehyde with cellular deoxyribonucleic acid take place in the turbinates and anterior nasal mucosa in rodents, and in monkeys also in the nasopharynx, larynx, trachea, and carina (39). The sinuses however seem to remain unaffected in monkeys (39), a finding which may be relevant with respect to humans. In recent studies, micronuclei were found to occur more frequently in the epithelial cells of the buccal cavity (40) and nasal mucosa (41) in workers exposed to formaldehyde than in unexposed subjects.
With some relevance to respiratory cancers, the epidemiologic evidence on formaldehyde and nonrespiratory cancers is insufficient and inconsistent (1). Elevated risk for cancers of some nonrespiratory sites remains a possibility. After the original metaanaly sis, a short report (42) was published showing significant excesses of follicular non-Hodgkin's lymphoma and acute myeloid leukemia among Iowa and 13 Scand J Work Environ Health 1993, vol 19, no I Minnesota embalmers and funeral directors. In an animal bioassay (43), leukemia and gastroint estinal tumors were induced by the admini stration of formaldehyde in drinking water , a finding suggesting that the ingestion of formaldehyde, for example, in drinking water, might be associated with a carcinogenic hazard.
In summary, the current epidemiologic evidence is about sufficient for the conclusion that at least substantial occupational formaldehyde exposures are carcinogenic, and the risk is the most likely to concentrate on the nasal cavities and nasopharynx, but probabl y not the lung. The increase in the risk would be, considering the predom inant downward confounding in the epidemiologic studies, 100% or more over the background incidence. The excess absolute canc er burden may remain somewhat limited, however, if the bronchi and the lungs -for which the background incidence is much highe r -remain unaffected. This possibility seems likel y, cons idering the current biological, experimental, and epidemiologic evidence,