environmental exposure-response analysis respiratory ... · 1970-9, 1980-90)*,whichmayberepresented...

Occupational and Environmental Medicine 1997;54:646-652

Exposure-response analysis of risk of respiratorydisease associated with occupational exposure tochrysotile asbestos

Leslie Stayner, Randall Smith, John Bailer, Stephen Gilbert, Kyle Steenland, John Dement,David Brown, Richard Lemen

National Institute forOccupational Safetyand Health (NIOSH),Cincinnati, Ohio45226, USAL StaynerR SmithJ BailerS GilbertK SteenlandR Lemen

Department ofMathematics andStatistics, MiamiUniversity, Oxford,Ohio 45056, USAJ Bailer

Division ofOccupational andEnvironmentalMedicine, DukeUniversity MedicalCenter, Durham,North Carolina 27710,USAJ Dement

National Institute forEnvironmental HealthSciences, ResearchTriangle Park, NorthCarolina, USAD Brown

Correspondence to:Dr L Stayner, Centers forDisease Control, NationalInstitute for OccupationalSafety and Health, Robert ATaff Laboratories, 4676Columbia Parkway,Cincinnati, Ohio45226-1998, USA.

Accepted 26 February 1997

AbstractObjectives-To evaluate alternative mod-els and estimate risk of mortality fromlung cancer and asbestosis after occupa-tional exposure to chrysotile asbestos.Methods-Data were used from a recentupdate of a cohort mortality study ofworkers in a South Carolina textile fac-tory. Alternative exposure-response mod-els were evaluated with Poissonregression. A model designed to evaluateevidence of a threshold response was alsofitted. Lifetime risks of lung cancer andasbestosis were estimated with an actu-arial approach that accounts for compet-ing causes of death.Results-A highly significant exposure-response relation was found for both lungcancer and asbestosis. The exposure-response relation for lung cancer seemedto be linear on a multiplicative scale,which is consistent with previous analysesof lung cancer and exposure to asbestos.In contrast, the exposure-response rela-tion for asbestosis seemed to be non-linear on a multiplicative scale in thisanalysis. There was no significant evi-dence for a threshold in models of eitherthe lung cancer or asbestosis. The excesslifetime risk for white men exposed for 45years at the recently revised OSHA stand-ard of 0.1 fibre/ml was predicted to beabout 5/1000 for lung cancer, and 2/1000for asbestosis.Conclusions-This study confirms thefindings from previous investigations of astrong exposure-response relation be-tween exposure to chrysotile asbestos andmortality from lung cancer, and asbesto-sis. The risk estimates for lung cancerderived from this analysis are higher thanthose derived from other populationsexposed to chrysotile asbestos. Possiblereasons for this discrepancy are dis-cussed.

(Occup Environ Med 1997;54:646-652)

Keywords: chrysotile asbestos; risk assessment; epide-miology

There has been considerable discussion in thescientific literature about the significance oftherisks associated with exposure to chrysotileasbestos.' 2 This debate is of importance topublic health, as chrysotile is the most often

used asbestos fibre in production worldwide3and is also the main source of exposure result-ing from efforts to remove asbestos in theUnited States today.There are only two industrial cohorts that

have relatively pure exposures to chrysotileasbestos and supply sufficiently high qualitydata for exposure-response analysis. These arethe studies of Quebec miners and millers4 anda National Institute for Occupational SafetyHealth (NIOSH) study of South Carolina tex-tile workers.5 Both studies have been used bythe Occupational Safety and Health Adminis-tration (OSHA)6 and the Environmental Pro-tection Agency (EPA)7 in their quantitative riskassessments for asbestos.The NIOSH cohort of chrysotile asbestos

textile workers was recently updated to includean additional 15 years of observation andexpanded to include women and non-whitepeople as well as white male workers.5 Ourpaper presents an exposure-response analysisand risk estimates for lung cancer andnon-malignant mortality from respiratory dis-ease based on the most recent update of themortality of the study of the NIOSH cohort oftextile workers.

Material and methodsSTUDY POPULATIONA detailed description of the design of theNIOSH cohort of chrysotile asbestos textileworkers may be found in several previouslypublished papers.5 89 Briefly, the plant waslocated in South Carolina and began produc-ing asbestos products in 1896. Chrysotileasbestos received from Quebec, British Colum-bia, and Zimbabwe was the only type of asbes-tos processed as raw fibre. Crocidolite yarn wasused in extremely small quantities from the1950s until 1975 (about 2000 pounds), and theexposures resulting from this process arethought to have been low and limited tospecific jobs.5 8The original analysis was restricted to

include white male workers (n= 1247) em-ployed in the textile production operations forat least one month between 1 January 1940 and31 December 1975. In the most recentpublication,5 this cohort was expanded toinclude non-white men (n=546), and whitewomen (n=1229) who met the same employ-ment requirements. Follow up of this cohortfor vital status was extended up to 31 Decem-ber 1990. The analyses presented in this paper

646

on May 30, 2021 by guest. P

rotected by copyright.http://oem

.bmj.com

/O

ccup Environ M

ed: first published as 10.1136/oem.54.9.646 on 1 S

eptember 1997. D

ownloaded from

http://oem.bmj.com/

Occupational exposure-response analysis of risk of respiratory disease and chrysotile asbestos

include all of the sex and race groups from thisstudy including non-white women (n= 19).One of the strengths of this study is the rela-

tively high quality of information on exposurethat was available for estimating historic occu-pational exposure to asbestos. Exposure levelsto chrysotile (fibres >5 gm/ml) by areas of theplant (department and operations), specificjobs, and calendar years have been previouslydeveloped8 and were used with information onwork history to estimate individual exposuresfor the present analysis. Changes in processesand controls were taken into consideration inderiving historical exposure estimates. Cumu-lative exposure to asbestos, which is theproduct of duration and intensity of exposureto asbestos, was the exposure metric used inthe statistical analyses described below.

STATISTICAL ANALYSESExposure-response analyses were conductedfor cancer of the trachea, bronchus, and lung(henceforth collectively referred to as lungcancer), and for asbestosis and pneumoconio-sis (henceforth collectively referred to as asbes-tosis). The underlying cause of death was usedto define the response for lung cancer (9threvision of the international classification ofdiseases (ICD-9)=162). For asbestosis, a mul-tiple cause of death approach'0 was used inwhich all of the fields of the death certificatewere considered. This approach was usedbecause asbestosis is often not listed as theunderlying cause of death on death certificates.Also, the definition of asbestosis was broad-ened to include deaths from pneumoconiosis(ICD-9=505) as well as from asbestosis (ICD-9=501), as the more general term pneumoco-niosis may have been used instead of asbestosison death certificates. Based on these defini-tions, 126 cases of lung cancer and 45 cases ofasbestosis (only one of which was pneumoco-niosis) were available for this analysis.The person-years and deaths stratified by the

covariates for the Poisson regression analysiswere generated with the NIOSH life tableanalysis system." Person-years for this analysiswere counted from the time when a person metthe study requirements until the time whenthey were either lost to follow up, died, orreached the end of the study. For lung cancer,the person-years and observed deaths wererestricted to only include those with at least 15years since the date of first exposure (latency).The person-years and observed deaths were

partitioned into 20 cumulative asbestos catego-ries, which had roughly equal numbers ofdeaths (all causes). Cumulative exposure wasmodelled as a continuous variable from themidpoints of each exposure category. Theseextensive categories permitted a detailed explo-ration of the shape of the exposure-responserelation.

Poisson regression models'2 were used toanalyse the exposure-response relation be-tween exposure to chrysotile asbestos andmortality from respiratory disease with theobserved deaths and person-years generated bythe NIOSH life table analysis system. Different

potential model forms were evaluated, whichinclude functions that have been previouslyproposed for the analysis of epidemiologicalcohort data.'3 Together these models are capa-ble of reflecting a wide range of possibleexposure-response patterns including linear,sublinear, and supralinear.Models in which the effect of exposure either

multiplied (multiplicative models) or added(additive models) to the background hazardrate were evaluated. These models may be rep-resented mathematically as:

Multiplicative: X = X0xf(E)Additive: X = O + f(E)

(la)(lb)

where X is the predicted incidence rate, f(E) isa function of cumulative exposure to asbestos(E) in fibre-year/ml, and X,, is the backgroundincidence which is a function of age, sex, race,and calendar time.The background incidences were modelled

as a log (ln) linear function of the followingcovariates: age (continuous), sex, race (white vnon-white), and calendar time (1940-69,1970-9, 1980-90)*, which may be representedmathematically as:

ln(ko)= P, + PI I(sex=female) + [2 I(race=white)+ f3,(age) + 04 I(year=1970-9) + 35 I(year=1980-9) (2)

where [0 is the intercept, P[. [2) [3, 34, and P5 arethe parameters associated with the effects ofsex, race, age, and calendar year. The I( )s areindicator variables (0 or 1) for the categoricallevels of sex, race, and calendar year.

Evaluation of the additive model (lb) waslimited to a simple linear function for model-ling the exposure-response relation, as thesemodels were generally found to fit the data farworse than the multiplicative models. Thissimple function may be represented math-ematically as:

f(E) = PE E (3a)

Different parametric functions were evaluatedfor modelling the exposure-response relationfor the multiplicative models including the fol-lowing forms:

Log-linear:f(E) = exp(PE E) or ln(f(E)) = [E E

Log-quadratic:f(E) = exp(QE, E + f3E2 E2)

Additive relative rate:f(E) = 1 + ,BE E

Power:f(E) = exp(PE ln(E+a))

(3b)

(3c)

(3d)

(3e)

where PE (and PEI + PE2) are the parametersassociated with exposure to asbestos (E), and ais a constant that is added to the exposure forthe power model. The value of a was solved byiteratively fitting the model with differentvalues of a until the deviance of the model was

*A broad category was used for the first period of study(1940-69) because there were relatively few deaths fromlung cancer or asbestosis during the early period of thestudy. This earliest category was used as a controlcategory, which is represented by the intercept.

647



.bmj.com

/O

ccup Environ M


eptember 1997. D

ownloaded from

http://oem.bmj.com/

Stayner, Smith, Bailer, Gilbert, Steenland, Dement, et al

minimised (note: for this model the 1ground hazard rate is .,, x aPE).An informal statistical evaluation of gi

ness of fit was performed by comparingdeviances ofthese models (technically notthese models are nested, and thus, a focomparison was not always possible).models with the smallest deviance were cotered to be the best fitting models. Also, 1models were graphically evaluated by coning the fit of these parametric models wcategorical model, and a cubic spline mo(For the categorical model, the numbs

exposure categories was reduced to 10 froiby simply combining adjacent categories-example, categories 1 and 2, 3 and 4 etcimprove the stability of the estimates ofrates. The categorical exposure functionbe represented mathematically as:Categorical:

10

f(E)= 1((ok I(exposure category=k)))k=2

where: Pk are the parameters and IQindicator variables for the k=9 highest expccategories, and the lowest exposure categcused as the control group.The restricted cubic spline is a model

makes flexible assumptions about the for:the exposure-response relation based on aunknown parameters. Essentially, the appriconsists of fitting cubic polynomials wdefined intervals of the exposure variableare restricted to be smooth at the cut off pa(or knots) which separate the intervals. Forestricted cubic spline model four knotsused at the 5th (pO5), 25th (p25), 75th (jand 95th (p95) percentiles of the cumulexposure to asbestos distribution. This mmay be represented mathematically as:

Restricted cubic spline:f(E)=exp(f3,E, + 02E,, +13,E3)

Table 1 Comparison of results of exposure to chrysotile asbestos from fitting alternatiPoisson regression models to the mortalities for lung cancer

Results for asbestos

Modelform (number in text) Parameter estimate SE Model deviancy

Baseline model(2)* - - 716.8 (df =24Additive model(3a) 4.79e-08 1.24e-08 701.3 (df=24.Multiplicative models:Log linear(3b) 7.21e-03 1.13e-03 685.0 (df=24.Log quadratic(3c) 676.9 (df=24,

P 1.72e-02 3.62e-03P 2 -4.36e-05 1.55e-05

Additive relative rate(3d) 2.19e-02 7.00e-03 679.0 (df=24-Power(3e) 678.1 (df=242a 6.100 4.86e-01 7.64e-02

Spline(3f) 678.5 (df=24,13, 2.68e-02 2.34e-02P 2 -0.0001 0.000103 0.0001 0.0001

Categorical(3g) 673.5 (df=240.81 S X < 1.64 -0.05 0.541.64 S X < 2.74 0.21 0.542.74 S X < 4.93 0.70 0.464.93 X < 8.76 0.70 0.488.76 < X < 17.80 0.73 0.4917.80 < X < 38.33 0.58 0.5138.33 < X < 79.40 1.19 0.4579.40 < X < 136.89 1.42 0.44X ¢ 136.89 2.02 0.45

* The baseline model includes the effect ofage, calendar year, race, and sex. The other modeinclude these terms as well as terms representing asbestos exposure.

tack- where: E,=cumulative exposure to asbestos,and E, and E, are functions of cumulative

;ood- exposure as described by Herndon andthe Harrel.'4

all of From the statistical and graphical evalua-trmal tions, a final functional form was chosen forThe modelling the relation between exposure tonsid- asbestos and the response variables. Furtherthese evaluation ofpotential interactions between theipar- exposure and the other covariates, and ofith a higher order exposure terms (quadratic anddel.'4 cubic) were evaluated before arriving at a finaler of model for risk assessment purposes.m 20 Finally, a "threshold" model"5 was consid--for ered to assess whether there was evidence that-to exposures below a certain level were equivalentf the to 0 exposure-that is, a threshold was present.may This model may be represented mathemati-

cally as:Threshold:f(E) = exp(p3, (E-@)) if E > 0 f(E) = 1 if E S 0

(3h)where 0 is a threshold parameter that wasare solved by iteratively fitting the model and

)sure setting the parameter to the midpoint of each)ry is ofthe 20 exposure categories until the deviance

was minimised.that All ofthe models were fitted with the Epicurem of program.X fewoach PREDICTION OF WORKING LIFETIME RISKSwithin Estimates of excess lifetime risk of dying fromthat lung cancer and asbestosis were developed foroints varying levels ofexposure to chrysotile asbestos,r the based upon an actuarial method that waswere developed in a risk analysis of radon exposures?75), (BEIR IV 1988), which accounts for the influ-lative ence of competing risks. It was assumed for thisiodel estimation procedure that workers were ex-

posed to a constant asbestos concentration for45 years between the ages of 20 and 65. Theannual risks were accumulated up to age 90.

(3g) Age specific background rates for lung cancer've and asbestosis were estimated from the final

Poisson regression models developed for theseoutcomes. Age specific background rates forcompeting causes of death were estimated byapplying life table methods to the study cohort.

L29) Results28) POISSON REGRESSION ANALYSES

28) Lung cancer27) Table 1 and figure 1 shows the results from fit-

ting the various Poisson regression models28) described in the methods section. Exposure!7) was a highly significant predictor (P < 0.00 1) of

lung cancer mortality in all of the models26) evaluated. The simple linear model (model 3a)

provided a poor fit to the data when contrastedwith the multiplicative models in table 1.

20) Between the two multiplicative models (model3b, and 3d) that used only one parameter forexposure to asbestos, the additive relative ratemodel (model 3d) gave the best fit to the databased on the criteria of minimum deviance.The deviance of the models was not appreci-ably improved by the models with additionalparameters for exposure to asbestos such as the

is also quadratic model (model 3c) or the powermodel (model 3e).

648



.bmj.com

/O

ccup Environ M


eptember 1997. D

ownloaded from

http://oem.bmj.com/


Categorical-- Additive RR----- PIAi r

Log-linear- -Spline

04 ~~~Log-quadratic0.

0

2-

N

0 50 100 150Asbestos exposure (f-y/ml)

Figure 1 Lung cancer mortalities as a function of cumulative exposure to asbestospredicted by alternative models for white men aged 50 in 1940-69.

Examination of figure 1 essentially confirmsthe impressions based on examination of devi-ances. The curve for the additive relative ratemodel provides similar estimates of the rate asthe spline model, and is reasonably consistentwith the rate estimates from the categoricalmodel. The quadratic and power models alsoseem to provide similar estimates, whereas thelog linear model seems to produce lowestimates of the hazard rate.Based on this evaluation the additive relative

rate model (model 3c) was chosen as the basisfor further analysis. There was no indication ofa significant interaction between any of thecovariates (age, race, sex, or year) and exposureto asbestos, or of a need for higher order terms(quadratic or cubic) to represent exposure. Anevaluation of interaction with time since firstexposure (latency) and exposure to asbestoswas performed by fitting a model with separateslopes for exposure with 15 to < 30, 30 to < 40and > 40 years of latency. This model wasfound to fit the data significantly (X'= 6.5,df=2, P=0.04) better than the simpler additiverelative rate model and was chosen as the finalmodel for predicting lifetime risks. Table 2show the parameter estimates and SEs fromthis final model. The goodness of fit of thismodel was judged to be good based on the factthat the model deviance was much smaller thanthe numbers of degrees of freedom. Based onthis model, the relative rate per unit of cumula-tive exposure to asbestos (X) from this modelwould be (1 + 0.022(X)) with 15-29 years oflatency, (1 + 0.037(X)) with 30-39 years oflatency, and (1 + 0.011(X)) with 40 years of

Table 2 Parameter estimates and SEsfrom the bestfittingmodelfor lung cancer mortality

Model parameters Parameter estimates SE

Intercept -16.51 0.56Sex (female) -0.95 0.20Race (non-white) -1.05 0.29Year (1970-9) -0.06 0.30Year (1980-90) 0.47 0.30Age 0.07 0.01Asbestos x latency (15-29) 0.022 0.012Asbestos x latency (30-39) 0.037 0.012Asbestos x latency (>40) 0.011 0.006

Model deviance=672.5; df=2426.

latency. For example, for 45 years of exposureto 0.1 fibre/ml the predicted relative rate wouldbe 1.10 for workers with 15-29 years of latency.The deviance of the threshold model (model3h), relative to a model without the thresholdparameter (model 3b), was not reducedregardless of what value of 0 was chosen.Hence the results from this model did not pro-vide any support for the existence of athreshold type response for lung cancer.

ASBESTOSISTable 3 shows the results from fitting the vari-ous Poisson regression models described in themethods section for asbestosis (fig 2). Theexposure-response relation was found to behighly significant (P100 fibre/ml).Based on this analysis, the power model was

selected as the most appropriate model for fur-ther evaluation. There was no evidence of asignificant interaction in the power modelbetween exposure to asbestos and any of theother covariates included in the baseline func-tion. Table 4 shows the parameter estimatesand SEs from the final power model. Thegoodness of fit of this model was judged to begood based on the fact that the model deviancewas much smaller than the numbers of degreesof freedom. Based on this model, the relativerate for cumulative exposure (X) would beequal to ((X+0.5)"13/(0.5)'"3). For example, for45 years of exposure to 0.1 fibre/ml the relativerate would be 19.95.The fit of the threshold model (model 3h),

relative to a model without the thresholdparameter, was not improved regardless ofwhat value of 0 was chosen. Hence the resultsfrom this model did not provide any supportfor the existence of a threshold type responsefor this outcome.

PREDICTION OF LIFETIME RISKSTable 5 shows the predicted lifetime excessrisks oflung cancer and asbestosis assuming 45years of exposure to varying exposures of

649



.bmj.com

/O

ccup Environ M


eptember 1997. D

ownloaded from

http://oem.bmj.com/


Table 3 Comparison of results of exposure to chrysotile asbestos from fitting alternativePoisson regression models to the mortalities for asbestosis

Results for asbestos

Modelform (numberffrom text) Parameter estimates SE Model deviance

Baseline model(2)* - - 293.61 (df= 13Additive model(3a) 5.46e-08 8.13e-09 229.60 (df=13Multiplicative models:Log linear(3b) 1.54e-02 1.50e-03 207.07 (df=13Log quadratic(3c) 182.54 (df=13

3, 4.59e-02 6.78e-032 -1.IOe-04 2.29e-05

Additive relative rate(3d)t -Power(3e) 179.84 (df=13a 0.5P 1.30 1.70e-01

Spline(3f) 179.71 (df=13Pl 1.08e-01 6.76e-021 2 -2.57e-04 3.12e-0413 2.64e-04 3.22e-04

Categorical(3g)4 181.66 (df=130 X


Table S Predicted excess lifetime risks oflung cancer and asbestosis assuming 45 years ofvarying time weighted average (TWA) exposure levels of chrysotile asbestos

Lifetime excess risk estimates *

Disease TWA (fibreslm3) White men White women Non-white men

Lung cancer 0.1 5 e-03 3 e-03 2 e-030.3 1 e-02 1 e-02 5 e-020.5 2 e-02 2 e-02 9 e-020.7 3 e-02 2 e-02 1 e-020.9 4 e-02 3 e-02 2 e-021.0 4 e-02 3 e-02 2 e-022.0 8 e-02 6 e-02 4 e-023.0 1 e-01 9 e-02 5 e-02

Asbestosis 0.1 2 e-03 1 e-03 1 e-030.3 9 e-03 4 e-03 3 e-030.5 2e-02 8e-03 6e-030.7 3 e-02 1 e-02 9 e-030.9 4 e-02 2 e-02 1 e-021.0 4 e-02 2 e-02 1 e-022.0 1 e-01 5 e-02 4e-023.0 2e-01 8e-02 6e-02

* The excess risk estimates are expressed in scientific notation where e represents the power to thebase 10 that the number should be multiplied by. For example, the excess lifetime risk of lungcancer for white men at 0.1 fibres/M3 is 5 e-03, which is equivalent to 5/1000 workers.

contrast, the exposure-response relation forasbestosis seemed to be non-linear on a multi-plicative scale in this analysis. This relation wasin fact sublinear, which implies that the risk ofasbestosis drops off more rapidly with reduc-tions in exposure than does the risk of lungcancer.There was absolutely no significant evidence

for a threshold in either the lung cancer orasbestosis models. The fit of these models wasin fact found to be maximised when thethreshold parameter was set to zero. Thus theresults from this analysis fail to provide anysupport for arguments that have been made fora threshold for the effects of chrysotile asbestoson risks of lung cancer and asbestosis.2'Based on this analysis, the predictions of risk

for lung cancer are somewhat higher than thepredictions for asbestosis at current exposurelevels. The excess lifetime risk for white menexposed for 45 years at the recently revisedOSHA standard of 0.1 fibre/ml was predictedto be about 5/1000 for lung cancer, and 2/1000for asbestosis. It was not possible to modelrates for mesothelioma based on this cohort,because there were too few cases. However,given the fact that there were over 60 excesscases of lung cancers and only three ofmesothelioma, it is obvious that the risk ofmesothelioma is far less than that of lung can-cer for this population. Overall, these risk esti-mates indicate that it may be appropriate tocontrol exposure to chrysotile asbestos evenbelow the current OSHA standard if techni-cally feasible.There are several assumptions and sources of

uncertainty underlying the predictions of riskmade in this paper that must be recognised.Firstly, these epidemiological observations arebased on relatively high exposure levels com-pared with current conditions and thus somedegree of extrapolation beyond the range of thedata was made to predict risks for currentexposure conditions. However, this extrapola-tion was not as extreme as is often the case inquantitative risk assessments. The averageexposure intensity (cumulative exposure/duration) of this cohort was about 6 fibre/ml

and predictions were only made for exposuresas low as 0.1 fibre/ml.

Secondly, as with nearly all epidemiologicalinvestigations of this nature, questions may beasked about the accuracy of exposure estimatesthat were used in this analysis. The quality ofthis information was unusually high comparedwith most retrospective cohort mortality stud-ies. The exposure classifications in this studywere based on over 5900 measurements andexposure conditions did not change appreci-ably over the time course of the study.8 Therewas a need to convert measurements that werebased on a method that estimates millions offibres per cubic foot (mfpcf) to the currentmethod of fibre/ml that are >5 ,um in length. Ithas been suggested that these conversions mayintroduce large errors into the risk assessmentprocess.2' Also, it has been argued that analysesbased on cumulative exposure are an oversim-plification which ignore the separate effects ofintensity and duration of exposure.4 Unfortu-nately it is difficult, if not impossible, toseparate these effects in studies such as this onebecause of a lack of information on variationsin intensity, and the ever changing exposurepatterns of workers.

Thirdly, the absence of individual infor-mation on cigarette smoking habits for thisentire cohort introduces some degree of uncer-tainty into this analysis. Information on smok-ing was available for a sample of the cohortwhich suggests that smoking habits amongblack men were lower, white men were similar,and white women were lower compared withthe general sex and race specific population ofthe United States.' However, the fact that thisanalysis was restricted to comparison of rateswithin the cohort reduces the possibility of biasdue to confounding by cigarette smoking.Confounding would only be possible if ciga-rette smoking was associated with the potentialfor exposure to asbestos in this cohort, whichseems unlikely. Of greater possible concern isthe lack of consideration of the potential inter-action between cigarette smoking and exposureto asbestos in the induction of lung cancer.22Berry et al in a review of studies on this issuereported that non-smokers exposed to asbestoshave a zero to fivefold greater relative risk oflung cancer than smokers who have anexpected value of 1.8.2" These results suggestthat the interaction between smoking andasbestos may be greater than additive but lessthan multiplicative. In any case, the resultsfrom this analysis may be viewed as valid for apopulation with a similar distribution of smok-ing habits, but may either over or underesti-mate risk for other populations depending ontheir distribution of smoking habits.

Fourthly, there is a serious potential for dis-ease misclassification in this study particularlyfor asbestosis. Death certificates are not gener-ally regarded as a reliable source ofinformationfor asbestosis.24 This is primarily becauseasbestosis is often not recognised as the under-lying cause of death. We have tried to minimisethis problem by using a multiple cause of deathapproach in this analysis. However, it is likelythat this approach has failed to detect all of the

651



.bmj.com

/O

ccup Environ M


eptember 1997. D

ownloaded from

http://oem.bmj.com/


cases of asbestosis in this cohort and thus therisk of asbestosis is likely to have been underes-timated. Lung cancer is generally recognised asthe underlying cause of death, and thus a mul-tiple cause of death approach was not neces-sary for this outcome.

Fifthly, the selection of an appropriate modelis (as always) a major source of uncertainty fora risk analysis. We have evaluated many modelsin this paper, rather than simply assuming alinear model as in previous analyses. None theless, the choice of models was based ongoodness of fit and not on knowledge of theunderlying mechanism.

Probably the largest source of uncertaintyrelates to the suitability of these findings to begeneralised to current exposure to asbestos inthe workplace or elsewhere. The predictionsfrom these analyses on risks of lung cancerwere higher than previous OSHA estimates forall forms of asbestos, and substantially higherthan the risk predictions based on analysis ofQuebec miners. The reasons for these widelyvarying results are not known. Initially, it wassuspected that they might be attributed to dif-ferences in tremolite contamination or errors inassessment of exposure. However, these theo-ries were ruled out by subsequent pathologystudies.25Another hypothesis that has been advanced

is that the higher risks of lung cancer in thetextile plant may be related to exposures tomineral oil.25 This hypothesis is inconsistentwith the finding that mineral oils have not beenshown to induce lung cancer in studies ofworkers exposed to machining fluids.26 Fur-thermore, the relation between chrysotileasbestosis and risk of lung cancer was notaltered when exposure for mineral oil was con-trolled for in a nested case-control study of theNIOSH asbestos cohort.5A viable hypothesis that might explain these

discrepant findings is that the percentage oflong fibres was higher in the asbestos textileindustry in South Carolina than in the Quebecmining industries.5 It is known that long thinfibres were preferred for use in the textileindustry. Also, the carding process used in thetextile industry sheared the asbestos into longthin fibres. There is also substantial toxicologi-cal evidence that long thin fibres are more car-cinogenic than short thick ones.27 If fibredimensions are the explanation for thesediscrepant findings then it would be importantto know whether the distribution of chrysotilefibre lengths and widths in current operationsare more similar to those experienced histori-cally in the NIOSH textile cohort or in theQuebec miners and millers. Until this issue isresolved, it would seem prudent to consider theestimates of risk from the NIOSH textilecohort, as well as those based on the Quebecmining and milling cohort, as relevant for pre-dicting a range of potential risks for currentindustrial and remediation operations thatinvolve chrysotile asbestos.

We acknowledge and thank Drs Richard Hornung, Jim Dedden,David Umbach, and Patricia Sullivan for their helpful reviews ofthis paper.

1 Mossman BT, Bigman J, Corn M, Seaton A, Gee JBL.Asbestos: scientific developments, and implications forpublic policy. Science 1990;24:294-301.

2 Stayner LT, Dankovic DA, Lemen RA. Occupational expo-sure to chrysotile asbestos and cancer risk: a review of theamphibole hypothesis. Am J Public Health 1996;86: 179-86.

3 Pigg BJ. The uses of chrysotile asbestos. Ann Occup Hyg1994;38:453-8.

4 McDonald JC, Liddell FDK, Dufresne A, McDonald AD.The 1891-1920 birth cohort of Quebec chrysotile minersand millers: mortality 1976-88. Br Jf Ind Med 1993;50:1072-81.

5 Dement JM, Brown DP, Okun A. Follow-up study of chrys-otile asbestos textile workers: cohort mortality andcase-control analyses. Am J Ind Med 1994;26:431-47.

6 OSHA (1986). Occupational exposure to asbestos, tremo-lite, anthophylite, and actinolite. Federal Register 1986;51:22612-747.

7 Nicholson WJ. Airborne asbestos health assessment update.Springfield, VA: US Department of Commerce, NationalTechnical Information Service. Final Report EPA-600-8-84-003F, June 1986.

8 Dement JM, Harris RL, Symons Mj, Shy CM. Exposuresand mortality among chrysotile asbestos workers. Part I:exposure estimates. AmJ IndMed 1983;4:399-419.

9 Dement JM, Harris RL, Symons MJ, Shy CM. Exposuresand mortality among chrysotile asbestos workers. Part II:mortality. AmJ IndMed 1983;4:421-33.

10 Steenland K, Nowlin S, Ryan B, Adams S. Use of multiple-causes mortality data in epidemiologic analyses, Am JT Epi-demiol 1992;136:855-62.

11 Steenland KJ, Beaumont J, Spaeth S, Brown D, Okun A,Jurcenko L, et al. New developments in the NIOSH life-table system. J Occup Med 1990;32:1091-8.

12 Frome EL, Checkoway H. Use ofPoisson regression modelsin estimating incidence rates and ratios. Am Y Epidemiol1985;121:309-22.

13 Breslow NE, Day NE. Statistical methods in cancer research.Vol 2. The design and analysis of cohort studies. Lyon: Inter-national Agencyfor Research on Cancer, 1987.

14 Herndon JE, Harrell FE. The restricted cubic spline asbaseline hazard in the proportional hazards model withstep function time-dependent covariables. Stat Med 1995;14:2119-29.

15 Ulm KW. Threshold models in occupational epidemiology.Mathematical Computer Modeling 1990;14:649-52.

16 Committee on the Biological Effects of Ionizing Radiation,Board of Radiation Effects Research, Commision on LifeSciences, National Research Council. Biological effects ofionizing radiation (BEIR) IV Health risks of radon and otherinternally deposited alpha-emitters. Washington, DC: Na-tional Academy Press, 1988.

17 McDonald JC, Liddell FDK, Gibbs GW, Eyssen GE,McDonald AD. Dust exposure and mortality in chrysotilemining, 1910-75. BrJ Ind Med 1980;37:11 -24.

18 Dunnigan J. Linking chrysotile asbestos with mesothelioma.AmJfInd Med 1988;14:205-9.

19 Peto J. Fibre carcinogenesis and environmental hazards. In:Bigon J, Peto J, Saracci R, eds. Non-occupational exposure tomineral fibres. Lyon: International Agency for Research onCancer, 1989:457-70.

20 Browne K. A threshold for asbestos related lung cancer. BrJ Ind Med 1986;43:556-8.

21 Peto J, Doll R, Hermon C, Binns W, Clayton R, Goffe T.Relationship of mortality to measures of environmentalasbestos pollution in an asbestos textile factory. Ann OccupHyg 1985;29:1985.

22 Hammond EC, Selikoff IJ, Seidman H. Asbestos exposure,cigarette smoking and death rates. Ann N YAcad Sci 1979;330;473-90.

23 Berry G, Newhouse ML, Antonis P. Combined effects ofasbestos and smoking on mortality from lung cancer andmesothelioma. BrJ IndMed 1985;42: 12-8.

24 Selikoff J. Use of death certificates in epidemiological stud-ies, including occupational hazards: discordance with clini-cal and autopsy findings. AmJ IndMed 22:469-80.

25 Sebastien P, McDonald JC, McDonald AD, Case B, HarleyR. Respiratory cancer in chrysotile textile and miningindustries: exposure inferences from lung analysis. BrJ7 IndMed 1989;46:180-7.

26 Tolbert PE, Eisen EA, Pottier L, Monson RR, Hallock MF,Smith TJ. Mortality studies of machining fluid exposure inthe automobile industry II. Risks associated with specificfluid types. ScandJ Work Environ Health 1992;18:351-60.

27 Stanton MF, Layard M, Tegeris A, Miller E, May M, Mor-gan E, Smith A. Relation of particle dimension to carcino-genecity in amphibole asbestos and other fibrous minerals.J Nad Cancer Inst 1981;67:965-75.

652



.bmj.com

/O

ccup Environ M


eptember 1997. D

ownloaded from
http://oem.bmj.com/

environmental exposure-response analysis respiratory ... · 1970-9, 1980-90)*,whichmayberepresented...

Documents