Introduction

Cigarette smoking is a major cause of cancer of the lung, larynx,
oral cavity, and esophagus and is a contributory factor for cancer of
the kidney, urinary bladder, and pancreas (US DHHS 1982). These
cancers will cause 278,700 of the estimated 910,000 new cancer cases
in the United States during 1985 (ACS 1985), or 30.6 percent of the
cancers occurring in the United States other than skin cancer.
Exposures to agents in the workplace other than cigarette smoke
will also cause some of these new cancers, and a number of cancers
will result from the combined effects of cigarette smoking and
carcinogenic exposures in the workplace.

The role that cigarette smoking plays in causing these cancers is
well established and extensively documented (US DHHS 1982). The
role that occupational agents play in the development of these same
cancers continues to emerge as the effects of more agents are
examined both in the laboratory and in the workplace. However,
cigarette smoking by exposed workers makes it difficult to separate
the effects of smoking from the effects of occupational agents for
cancers of sites causally linked to cigarette smoking. For some
agents, such as asbestos, both the large numbers of people exposed
and the magnitude of the increased cancer risk have allowed a
careful examination of the relative contributions of cigarette smok-
ing and the workplace exposure. For most agents, the data are more
limited. Nevertheless, protection of workers requires that regulatory
decisions be made about individual workplace exposures, even in the
face of limited data. In assessing the effects of workplace exposures,
consideration must be given to the interactions of smoking with
agents that increase risk and to the bias introduced into studies of
occupational groups by confounding effects of cigarette smoking.
This chapter discusses the nature and measurement of interactions
between smoking and occupational exposures and the sources and
&ntrol of confounding of smoking and occupational exposures. It is
not intended to be a comprehensive discussion of the epidemiologic
methods used to evaluate workplace exposures, but rather a discus-
sion of how smoking behavior in the workforce can effect the
evaluation of occupational exposures. The data on smoking and
specific occupational exposures are presented in later chapters of
this Report. The discussion of these issues is intended to aid in the
design and interpretation of studies of occupational exposure and not
to criticize those studies in which smoking could not be completely
addressed.

Lung Cancer Death Rates and Smoking
A detailed discussion of the causal relationship between cigarette
smoking and the cancers is provided in an earlier Report in this

101


series (US DHHS 1982) and is not repeated here. However, the
relationship between smoking and lung cancer is briefly described,
as a framework for the discussion of interaction and confounding in
subsequent sections of this chapter. Lung cancer was chosen as an
example because of its strong link to smoking and because it is the
greatest cause of cancer death in both men and women (ACS 1985).

Lung cancer will cause an estimated 125,600 deaths in 1985 (ACS
1985): 87,000 men and 38,600 women. For men, this represents more
than 8 percent of all deaths. Current U.S. age-specific lung cancer
death rates increase with age into the late seventies age range and
then decline. However, when death rates for any given birth cohort
of men are examined (Figure l), there is no decline in death rates at
the older ages. This difference between the cross-sectional mortality
statistics and the cohort data is generally attributed to differences in
the smoking habits of successive birth cohorts of men (and women)
during this century. This Report's chapter on smoking patterns in
the U.S. population also carefully documents that cigarette smoking
is not uniformly distributed in the U.S. population, but rather varies
considerably with both age and occupation. This nonuniform distri-
bution of smoking patterns introduces much of the difficulty in
controlling for smoking in occupational studies.

The relationships among age, lung cancer death rates, and number
of cigarettes smoked per day, derived from the mortality study of
U.S. veterans (Kahn 1966), are presented in Figure 2. The risk
associated with smoking is a function of both the intensity of
smoking, as measured by number of cigarettes smoked per day and
depth of inhalation, and the duration of smoking as measured by age
and age of initiation.

The lung cancer mortality ratios derived from the American
Cancer Society (ACS) study of 1 million men and women (Hammond
1966) for smokers compared with nonsmokers, stratified by age and
by number of cigarettes smoked per day, depth of inhalation, and age
of initiation are presented in Table 1. In general, the mortality ratios
are greater in the older age groups and increase with increasing
dosage measure within each age strata. The data demonstrate that
within the broader category of smokers a substantial variation in
risk (up to fivefold) occurs between the different levels of dose and
duration of smoking. The variation in mortality ratios for each
isolated measure in Table 1 almost certainly overestimates the
independent contribution of that measure to the actual risk, owing to
correlation among the measures of number of cigarettes smoked per
day, depth of inhalation, and age of initiation. For example, those
who begin to smoke at a young age also smoke more cigarettes per
day (Shopland and Brown 1985). However, it is unlikely that this
correlation among dosage and duration measures explains all of the
variation in mortality ratios with the isolated measures; therefore, it

102


1885


1880


1875

FIGURE I.-Age-specific mortality rates for cancer of the
      bronchus and lung, by birth cohort and age at
      death, men, United States, 1959-1975
SOURCE: Data derived from McKay et al. W82).

is reasonable to expect that the accuracy of lung cancer risk
estimates for a population would improve with the inclusion of a

103


FIGURE 2.-Death rates from cancer of the lung and
       bronchus in nonsmokers and smokers of
      various numbers of cigarettes per day
SOURCES Kahn (1966).

measure of smoking prevalence, a measure of smoking intensity, a
measure of smoking duration, and a measure of the duration of
cessation for former smokers.

Interactions Between Cigarette Smoking and Occupational
Exposures

Interactions between cigarette smoking and occupational expo-
sures may be examined in the context of a biological process, as a
statistical phenomenon, or as a problem in public health and
individual decisionmaking (Rothman et al. 1980; Saracci 1980;
Siemiatycki and Thomas 1981). In each of these contexts the

104


E   TABLE l.-Number of lung cancer deaths (men), age-standardized death rates, and mortality ratios, by
;'         current number of cigarettes smoked per day, degree of inhalation, and age began
if         smoking, by age at start of study
0
I                         Age35-54             Age 55.69             Age 7cH34           All ages, 35-E-4
g                       Number                 NUdX?r                 Number                 Number
I     SllKking               of     Death   Mortality     of     Death   Mortality     of     Death   Mortality     of     Death   Mortality
cn     characteristics           deaths rate     ration     deatha    rate     ratios     deaths    rate     ratios     deaths    rate     ratios

Current number of cigarettes a day

l-9                   9
lo-19                 15
20-39                138
240                  26

Degree of inhalation

None or slight            19
Moderate               114
D=P                  56

Age began cigarette .3moking

2%                   5
20-24                 31
15-19                 112
< 15                  36

Never smoked regularly        11

38       6.17       12      68    3.53        5      134      5.32      26      66      4.60
24       3.90      57     168    8.77        10      243      9.62      82      90      7.48
58       9.37      216     264    13.82        27      446     17.62      381     169     13.14
47       7.67      60     334    17.47        6      754     29.84      82     201      16.61

29       4.75      97     203    10.60        14      193      7.66      120     102      8.42
62       8.48      177     224    11.72        20      401      15.88      311     138     11.45
65       9.00      73     266    13.93        13      638     25.26      141     173     14.31

17       2.77      12
36       6.83      72
54       8.71      176
79      12.80      57

6               27

65
212
250
302

19

3.39
11.11
13.06
15.81

3
7
27
9

11

85      3.38      20      39      3.21
306     12.11      110     118      9.72
490     19.37      315     155     12.81
424     16.76      101     183     15.10

25              49      12

NCYI'E: Mortality ration are baeed on death rates carried cut to one more significant fmre than shown
SOURCE Hammond (1966).


concepts are applied somewhat differently, and confusion results
when a move from one context to another is attempted without
consideration of these differences in application. Biological interac-
tion refers to the presence of one agent influencing the form,
availability, or effect of a second agent, and includes physical
interaction such as the adsorption of carcinogens to particulates in
inspired air, process interactions such as the induction by one agent
of an enzyme system capable of converting a second agent into a
carcinogenic metabolite, and outcome interactions such as the
number of tumors produced by separate and combined exposures in
an animal exposure system. Statistical interaction refers to a
departure from the mathematical model used to assess the effects of
the exposure variables. The model being tested may be additive,
multiplicative, or some other form; the outcome of interest may be
death rates, relative risks, or other outcome measures; the indepen-
dent variables may be intensity of exposure, duration of exposure, a
combination of intensity and duration (e.g., pack-years), or a
logarithmic or other transformation of these measures. Public health
interaction usually refers to the presence or level of one agent
influencing the incidence, prevalence, or extent of disease produced
by a second agent. An exposure to two agents that resulted in a
multiplicative effect on lung cancer death rates might show no
interaction using a multiplicative statistical model, but might show a
profound interaction in terms of public health and a variety of
interactions within the biologic system under consideration (i.e.,
human carcinogenesis).

Biologic Interactions

The transformation of normal lung tissue into a clinically mani-
fest lung cancer is a complex, incompletely understood process that
is generally assumed to require multiple inheritable changes within
the cell (Armitage and Doll 1961; Day and Brown 1980). Although
cellular changes are assumed to be requisite for carcinogenesis,
phenomena taking place outside the cell may influence carcinogene-
sis. Cigarette smoke and occupational agents may potentially
interact by influencing the fraction of inhaled carcinogen deposited
and retained in the lung, the rate of metabolic activation of a
procarcinogen into a carcinogenic metabolite, the transfer of agents
across mucosal and cellular boundaries, the vulnerability of the cell
to carcinogenic change (by increasing the rate of cell replication), or
the transformation of the cellular DNA. In addition, cellular DNA
repair, humoral or metabolic factors influencing tumor growth, and
immunologic recognition or destruction of tumor cells are processes
that may influence tumor manifestation and may be affected by
occupational exposures and cigarette smoke. A detailed discussion of
chemical carcinogenesis is beyond the scope of this chapter and is

106


provided elsewhere (Weinstein 1985; Farber 1982); however, this
chapter explores some potential sites of biological interaction
between occupational exposure and cigarette smoke to illustrate the
biologic interactions that may take place.

Cigarette smoking and occupational exposures may interact
through effects of smoking on the dose of the carcinogen that reaches
the cell. Long-term exposure to cigarette smoke impairs mucociliary
clearance (US DHHS 1982) and could alter the dose of an occupation-
al agent retained. Carcinogens may adsorb to particulates in smoke
or to environmental dusts (Natusch et al. 1974; Mossman et al. 19831,
resulting in a higher fractional retention or different distribution in
the lung. The adsorption to dust may also facilitate or inhibit
transport of carcinogens through th, mucus layer. Cigarette smoke
has been shown to increase epithelial permeability in the tracheo-
bronchial tree (Simani et al. 1974); the effect may increase the
exposure of the underlying cell to an occupational agent.

Another potential site of biologic interaction is the metabolic
activation of a carcinogen. A number of agents, including the
polycyclic aromatic hydrocarbons in cigarette smoke, undergo chem-
ical transformation within the body to met,abolites that are consid-
ered to be active carcinogens (Gelboin and Tso 1978a, b). The
majority of known conversions occur through the mixed function
oxygenase system predominately located in the microsomal fraction
of the cell. A number of constituents of cigarette smoke have been
shown to induce this enzyme system (US DHEW 19791, and its
activation may increase the rate of biologic activation of procarcino-
gens in the worksite. Cigarette smoking also alters the cellular
composition of the lung, increasing the number of neutrophils and
activated macrophages in the lung (US DHHS 1984); these cells may
also play a role in the metabolic transformation of occupational
agents.

Much of the consideration of interactions between smoking and
occupational exposures has centered on interactions that might
influence the response of the cell rather than the "dose" of
carcinogen (Siemiatycki and Thomas 1981; Rothman et al. 1980;
Rothman 1974, 1978; Walter and Holford 1978). In a widely accepted
conceptual model, the process of malignant transformation of a cell
into a cancer is considered to be a multistage process requiring
multiple inheritable changes (Armitage and Doll 1961; Day and
Brown 1980). Individual agents may initiate or promote the process
of carcinogenesis. Initiation is thought to be at least a two-stage
process that requires cell division before becoming irreversible
(Farber 1982). Promotion describes the process by which an agent
encourages an initiated tissue to develop focal proliferation. A tumor
initiator may exert its effect through a brief exposure, whereas a
tumor promoter usually requires repetitive contact with initiated

107


tissue to exert its effect. Cigarette smoke is known to contain a
number of compounds that act as tumor initiators and promoters
(US DHHS 1982); occupational exposures reflect a similar range of
agents. Tumor promoters in smoke may influence the effects of
exposure to tumor initiators in the workplace and thus increase the
number of cancers that occur, and the presence of tumor initiators in
smoke may allow the expression of a tumor promoter in the
worksite.

The process of carcinogenesis is frequently modeled as a multistep
process in which each succeeding step can occur only in those cells
that have undergone the preceding step (Armitage and Doll 1961;
Day and Brown 1980). In this model, agents may influence one (or
more) of these steps, and therefore may have an effect early or late
in the carcinogenic transition. Because the later steps in the process
can occur only in cells that have undergone the changes of earlier
steps, agents that act at separate steps may have multiplicative
effects. For example, an agent that results in a fourfold increase in
the rate of transition from a hypothetical step 1 to step 2 in the
carcinogenic process would result in a fourfold increase in the
number of malignant transformations by increasing the number of
cells available for step 2 and subsequent steps. Similarly an agent
that tripled the rate of transition from step 2 to step 3 would triple
the number of malignant transformations. However, exposure to
both agents would provide a fourfold (300 percent) increase in the
number of cells available for transition from step 2 to step 3 as well
as a threefold (200 percent) increase of the rate of transition from
step 2 to step 3, with a resultant twelvefold (1,100 percent) increase
in the number of malignant transformations. Therefore, the effect of
the combined exposure on number of malignant transformations
(1,100 percent) would be greater than the sum of the effects of
independent exposures (300 percent plus 200 percent).

A similar phenomenon may occur with cigarette smoke and an
agent that has an independent and additive effect as an initiator of
carcinogenesis. The additive effects on tumor initiation may appear
as a multiplicative effect on tumor occurrence because of the action
of the tumor promoters in cigarette smoke. The tumor promoters in
smoke may act on the cells initiated by an occupational agent, as
well as on the cells initiated by smoke, to increase the number of the
cells that become cancers. The number of tumors produced by a
combined exposure could then be greater than the sum of the
numbers of tumors produced by the individual exp<osures separately.

Two additional mechanisms by which cigarette smoking and
occupational exposures may interact are by alterations in the
immunologic surveillance for cancers and by increasing the frequen-
cy of cell division. Differences in the number, type, and function of
cellular components of the immune system have been demonstrated

108


between smokers and nonsmokers (US DHHS 1984) and among
workers exposed to occupational agents (see other chapters of this
Report). The potential for these differences to influence the rates of
clinically manifest cancers (either positively or negatively) is an
issue of considerable interest. The increase in cell turnover in the
respiratory tract in response to the acute toxic and inflammatory
effects of cigarette smoke, or of occupational exposures, may also
influence cancer rates, as it is believed that cells are more
vulnerable to carcinogenic changes during periods of replication.

This discussion is intended to illustrate the kinds of biologic
interactions that might occur between smoking and occupational
agents and not to be a complete description of either the carcinogenic
process or the sites of potential interaction.

Statistical Interaction

Statistical interaction refers to departure from a mathematical
model in assessing the main effects of independent variables; its
presence is often evaluated by the addition of an interaction term to
the independent variables (Siemiatycki and Thomas 1981; Blot and
Day 1979; Saracci 1980). With this approach, the presence of
interaction is dependent on the model being used (Rothman 1974;
Kupper and Hogan 1978). For example, a multiplicative effect can be
adequately modeled without an interaction term on a log scale, but
requires an interaction term on an additive scale. In this section, an
additive model for the effects of two exposures assumes that the
combined exposure produces an effect equal to the background rate
plus the sum of the increases from the background rate of the two
exposures experienced separately. In a multiplicative model, com-
bined exposure results in an effect equal to the product of the effects
produced by the separate exposures.

The following example illustrates this terminology and demon-
strates the dependence of statistical interaction on the selected
model. Assuming that two agents independently increase the risk of
lung cancer and that the separate exposures result in a fivefold and
tenfold increase in risk, respectively, if exposure to both agents
produces an eightfold increase in risk, there is negative interaction
(protective effect) in the additive and the multiplicative models. A
combined risk of 14 indicates no interaction in an additive model, but
a negative interaction in a multiplicative model; a risk of 30 is a
positive interaction with an additive model and negative with a
multiplicative model; a risk of 50 is a positive interaction with an
additive model and no interaction with a multiplicative model; and a
risk of 60 is a positive interaction with both models.

This example illustrates the critical dependence of tests for
interaction on the mathematical model that is selected. Ideally, the
choice of a model is based on biological considerations and not on

109


statistical convenience. For example, if the potential interaction of
two initiators is being examined, an additive model should be used.
The use of a multiplicative model may result in the demonstration of
a negative interaction.

When applied to the multistage biologic model of carcinogenesis,
independent actions at the same step would yield additive effects and
actions at separate steps would yield multiplicative effects (Siemia-
tycki and Thomas 1981; Walter and Holford 1978). This progression
from the biologic model to the statistical effect is easily defended;
however, it is less clear that the reverse progression is valid,
particularly in epidemiologic studies. The demonstration of an
additive effect on lung cancer death rates does not necessarily imply
that the two agents are acting at the same point in the carcinogenic
process, nor does a multiplicative effect guarantee action at separate
steps. As should be evident from the discussion of biologic interac-
tion, cigarette smoke may interact with occupational agents at
points external to the cell, and smoke consists of a variety of agents
with different carcinogenic effects. The complex biologic processes
that underlie the exposure-disease relationships evaluated in epide-
miological studies limit the inference from the results of statistical
modeling to biological mechanisms.

Rothman (1974) and Hogan and colleagues (19781 described
methods of quantifying the magnitude of statistical interaction, and
Kupper and Hogan (1978) described the detection of interaction in
cohort and case-control studies. This Report's chapter on the
evaluation of chronic lung disease also discusses the concepts of
interaction and its measurement in studies of outcomes that are
continuous (i.e., lung function measures) rather than binary (i.e.,
presence or absence of lung cancer).

In the simplest analytical problem, departure from additivity can
be readily assessed when a population has two exposures, the rates
in the presence of each individual exposure are known, and the rates
in the presence and absence of both are known. If the relative risk
(RR) in the absence of exposure is set equal to 1, then the ratio of the
rate in the population with only one of the exposures to the rate in
the population with neither exposure is the RR associated with the
exposure. Correspondingly, the ratio of the rate in the population
with both exposures over the rate in the population with neither
exposure is the RR associated with combined exposure. The magni-
tude of the interaction can then be estimated by the ratio of the
increase in rate with combined exposure (the RR of combined
exposure minus 1) over the sum of the increases from the unexposed
rate produced by the single exposures ((RR,-l)+(RR~11). The
confidence interval around this estimate of interaction can also be
estimated (Rothman 1974) as a measure of its statistical significance.
More complicated estimates of the magnitude of interaction are

110


necessary when the rate in the unexposed population is unknown,
when the rate of the disease being measured is high in the general
population, and when case-control analyses are being performed
(Rothman 1974; Hogan et al. 1978). In general, the size of the
population needed to test for interaction between two exposures is
considerably larger than the size of the population needed to
establish statistically significant effects for the separate exposures.

Both case-control and cohort data can be analyzed with ap
proaches that involve stratification (Kleinbaum et al. 1982; Rothman
and Boice 1979). The data are separated into strata defined by levels
of the occupational exposure and of cigarette smoking. By combining
the information within the separate strata, summary measures can
then be calculated that estimate the independent effects of the
variables and describe their interaction. Although stratified analysis
can be readily performed, its application is frequently limited by the
number of available subjects, both in the entire study and within
specific strata. For example, if an investigator designates four levels
of exposure to an occupational agent and classifies smokers as
currently smoking, previously smoking, or never smoking, twelve
separate exposure categories are created. If age, sex, and race must
also be considered, stratified analysis may be feasible only if the
number of subjects is extremely large.

Statistical modeling represents an alternative that is less compro-
mised by smaller sample sizes and that provides greater flexibility
for controlling confounding and for testing for interaction. Modeling
refers to the specification of a particular mathematical relationship
between the outcome variable, e.g., the occurrence of lung cancer,
and the variables representing the exposures of interest, e.g.,
cigarette smoking and an occupational agent. Statistical methods
describe the adequacy of the model for the data and provide
estimates of the effects of the exposure variables. Modeling can be
performed with the programs available in most conventional statisti-
cal packages, but some special applications may require customized
software.

In analyzing data on the effects of occupational exposures in
populations with a high prevalence of smoking, modeling facilitates
the control of confounding by smoking; multiple variables that
characterize smoking, such as duration, daily amount, and depth of
inhalation, can be entered simultaneously into the model. Further, if
the cumulative exposures to the occupational agent and to cigarette
smoke are temporally correlated, modeling may more satisfactorily
separate their effects, in comparison with stratified analysis.

A recent report by Whittemore and McMillan (1983) illustrates the
application of modeling to occupational data. These investigators
analyzed data collected in the U.S. Public Health Service study of
Colorado Plateau uranium miners, a prospective cohort study of

111


mortality in relationship to exposure to radon daughters in the
mines. Their analysis assessed exposure to radon daughters and
cigarette smoking as risk factors for lung cancer. To assess the joint
effects of smoking and radiation, they developed and contrasted
additive and multiplicative models. They found that the multiplica-
tive model fit the data better than the additive. Of the alternative
multiplicative models, giving the highest likelihood of the data was a
linear function of the variables for smoking and radon daughter
exposure. Whittemore and McMillan then used this multiplicative
model to assess the effects of age and birth cohort. This analysis
complemented the conventional cohort methods that had been
applied previously to the data (Lundin et al. 1971; Archer et al.
1976).

Most conventional forms of modeling assume either an additive or
a multiplicative relationship between the independent effects of the
variables representing the exposures. Case-control data are most
often analyzed with the multiple logistic model (Breslow and Day
1980; Schlesselman and Stolley 19821, although alternatives have
been described (Walker and Rothman 1982; Breslow and Storer
1985). The multiple logistic model is multiplicative; the risk of
disease from multiple exposures is obtained as the product of the
risks from the individual exposures, in the absence of interaction
among the exposures. A variety of approaches have been described
for the modeling of data from cohort studies (Breslow et al. 1983;
Breslow 1985). These models may be developed as additive or as
multiplicative or on other scales.

In developing a model, confounding is controlled by introducing
variables for the potentially confounding exposures. Statistical
interaction among the variables is tested by entering terms formed
as their product or by running the model within groups of subjects
separated by their classification on one of the exposure variables.
When a product term is entered into a model to test for interaction,
the presence and extent of interaction is indicated by the coefficient
calculated for the product term. Most modeling techniques also
supply a test of statistical significance for the coefficient, under the
null hypothesis that its value is zero. Such a test of statistical
significance may not be very powerful (Greenland 19831, and the
coefficient may suggest an interaction of potentially important
magnitude, although it does not reach statistical significance at
conventional levels.

The presence of statistical interaction between two variables
demonstrates that their effects are interdependent, as assessed by
the specific statistical model (Rothman et al. 1980). Statistical
interaction does not necessarily imply biological interaction. In fact,
the interpretation of interaction hinges on the scale on which it is
measured; the choice of the statistical model may determine whether

112


interaction is present or absent, synergestic or antagonistic (Green-
land 1979; Rothman et al. 1980). If possible, the choice of model
should be based on biological considerations. For malignancy, the
results of modeling may be interpretable within the conceptual
framework supplied by the theory that carcinogenesis is a multistep
process (Armitage and Doll 1961; Day and Brown 1980).

Public Health Interactions

From a public health perspective, an interaction occurs when the
number of individuals injured, or the extent of the injury, with
combined exposure exceeds that expected from the sum of the
background rate and the differences between the background rate
and the rates with the individual exposures. Public health interac-
tions can be considered a case of statistical interaction in which both
the model being tested and the outcome measurement scale being
used are defined by their ability to assess the contribution of a given
agent to the disease burden in society. When a positive interaction
occurs in this definition, the term "synergism" should be used. The
model used to examine interactions is often further specified by the
importance of considering the intensity and duration of exposure in
the risk model being examined. Establishing a dose-response rela-
tionship for an exposure supports a causal association, and the slope
of the exposure-response relationship allows an estimation of the
reduction in disease burden that might occur with a reduction in the
workplace exposure. Both of these issues are important in establish-
ing safe levels of exposure in the working environment.

Estimation of the reduction in disease burden due to an occupa-
tional exposure with the lowering of exposure levels has three
components: How much disease will be prevented in those workers
who begin their work exposure at the new levels? How much disease
will be prevented by reducing the exposure of workers previously
exposed to higher levels to these levels? and How much disease can
be prevented by altering the smoking habits of the exposed workers?
For those exposures for which synergism between smoking and an
occupational exposure exists, the sum of these three estimates may
exceed the total amount of disease that occurs in the population
(Samet and Lerchen 1984; Doll and Peto 1981). If a group of asbestos
workers have a fiftyfold increased risk with combined exposure and
a fivefold risk with exposure only to asbestos and a tenfold risk with
exposure only to cigarettes, then elimination of smoking would
eliminate 90 percent of the risk (from 50 to 5) and elimination of
asbestos would eliminate 80 percent of the risk (from 50 to 10). The
sum of these reductions is greater than 100 percent, and points out
that for prevention efforts, the synergistic effect works to potentiate
the effect of the intervention.

113


Confounding of Occupational Exposures by Smoking Behavior

By the nature of the employing industries, most occupational
exposures occur to a limited number of individuals who are often
geographically clustered and who are not representative of the U.S.
population. Prospective studies of cancer rates in populations that
are representative of the U.S. population generally contain too few
individuals with specific occupational exposures to allow analysis by
occupational exposure. Therefore, most studies of occupational
exposures involve populations selected on the basis of a specific
exposure. Then either these selected populations of exposed workers
are compared with a control group or individuals with high dose
exposures are compared with individuals with low dose exposures.
Validity depends upon the comparability of the groups being
examined for variables that may influence cancer risk. other than
occupational exposure. Age is one such variable, as rates of most
cancers increase with increasing age. For those cancers linked to
smoking, the comparability of the smoking habits of the various
exposed subjects is a second such variable. This variation may
potentially confound an association between an occupational expo-
sure and a cancer known to be associated with smoking, and control
for this potential confounding may be critical for an unbiased
evaluation of such an association.

Sources of Confounding

Confounding is the distortion of the apparent effect of an exposure
on risk brought about by the association with other factors that can
influence the outcome (Last 1983). Cigarette smoking can be a
confounding factor in occupational studies through an association
(either positive or negative) with the exposure in question. As
described earlier in this chapter, the major determinants of smoking-
related risk in a population include smoking prevalence, intensity of
exposure, and duration of exposure. Each of these measures can
potentially confound an occupational exposure.

Smoking Status

In occupational studies, cancer mortality in the occupational
group is often compared with that in the entire population of a given
geographic area. Age-specific death rates are available for the U.S.
population on an annual basis and can be used to develop an age- and
calendar-year-adjusted overall expected number of deaths, or a
cause-specific expected number of deaths, for the population of
workers being examined. The ratio of the actual number of deaths in
the exposed population compared with the expected number in the
general population, multiplied by 100, is referred to as a standard-
ized mortality ratio (SMR) for the exposed population. The SMR may

114


be based on national mortality data or on data from the geographic
location of the exposure group. In addition to providing a control
population, the use of SMRs also adjusts for differences in age
distribution between the exposed population and the population on
which the SMR is based.

Cigarette smoking behavior is not uniformly distributed through-
out the U.S. population. As demonstrated in the preceding chapter,
there are substantial differences in smoking behavior among men
and women, blacks and whites, different age groups, and different
occupations. It is not surprising, therefore, that the smoking
behavior of selected populations of exposed workers might differ
markedly from the average for the U.S. population, and these
differences would be expected to influence the SMR for smoking-
related cancers,

Axelson (1978) has suggested that the effect on the SMR of
differences in smoking habits could be estimated by dividing the
population being examined into various smoking categories, multi-
plying the proportion of the population in that smoking category by
the relative risk of developing disease produced by that smoking
category, and summing the resultant numbers. The ratio of this
number, calculated for the exposed population and compared with
the number for the population on which the SMR is based, is then a
multiplier that can be used to evaluate the effect on the SMR of the
smoking habits of the exposed population.

In its simplest form this calculation would use only the proportion
of smokers and nonsmokers in the population and a single relative
risk number for the smokers. The effect that differences in smoking
habits might have on the SMR for three different relative risks due
to smoking is shown in Table 2. These different relative risks
correspond approximately to the different relative risks for different
sites of cancer associated with smoking (US DHHS 1982). Blair and
colleagues (1985) have compared the crude and smoking-adjusted
SMRs for different job categories in the population of the U.S.
veterans study. They used four categories: smoker, never smoked, ex-
smoker, and other. In general, adjustment for smoking did not
substantially alter the SMRs for lung cancer (R 0.881, and the
differences were small for most job categories (the largest difference
between crude and adjusted SMR, 68.0).

Measures of Smoking Intensity

The risks due to smoking increase with increasing number of
cigarettes smoked per day and depth of inhalation (Table 1) (US
DHHS 1982). A calculation, similar to the one in the preceding
section, can be performed using separate risk estimates for light
smokers and heavy smokers and for ex-smokers. The magnitude of
the effect on the SMR for lung cancer of a range of different smoking

115


TABLE 2.-Effect of differences in smoking prevalence on
      the relative risk of an occupational group
      compared with a control group

                       Proportion of smokers in exposed group
Assumed risk   Proportion of smokers
due to smoking    in control group      .l       .3       .5       .I        .9

.3
.5
.I
.9

.3

.I
.9

.5

9

1.00     1.18     1.36      1.55     1.73
.85      1.00      1.15      1.31      1.46
.73      .87      1.00      1.13     1.27
.65      .76      .8a      1.00      1.12
.58      6f3      .79      .89      1.00

1.00      1.57      2.14      2.71     3.29
64      1.00      1.36      1.73     2.09
.47      .I3      1.00      1.27     1.53
.37      .58      .79      1.00     1.21
30      48      .65      .83     1.00

1.00      1.95      2.89      3.64     4.79
.51      1.00      1.49      1.97     2.46
.35      .67      1.00      1.33      1.65
.26      .51      .75      1.00      1.25
21      .41      .60      .80      1.00

prevalences and dosages is shown in Table 3, calculated using a
relative risk of 7 for smokers of less than one pack per day, 20 for
smokers of over one pack per day, and 4 for ex-smokers. These
relative risks were drawn from the major prospective mortality
studies on smoking (US DHHS 1982). The proportions of smokers
and ex-smokers in the population and the percentage of smokers who
smoke more than 20 cigarettes per day were drawn from the data
presented in the preceding chapter for the U.S. population between
the ages of 20 and 64. On the basis of the data, the current
differences in smoking patterns between blue-collar men and the
total male population might be expected to result in a 10.2 percent
elevation in the SMR for lung cancer. A hypothetical population
with a prevalence of current smoking of 80 percent might have a 59.9
percent increase in the lung cancer SMR. Correspondingly, a
population with a low smoking prevalence might have a 45.1 percent
reduction in the SMR. These numbers are similar to those calculated
for the Swedish population by Axelson (1978) as outer limits of the
adjustment that might need to be made in lung cancer SMRs,
secondary to differences in smoking patterns in an occupationally
exposed population.

One of the basic assumptions made in the risk adjustment
calculations described is that differences in smoking behavior (and
the resultant risk) can be described by simple prevalence numbers
(percentage of smokers, never smokers, and ex-smokers) or by using

116


TABLE S.-Effect of differences in smoking prevalence on
      the standardized mortality ratio for lung cancer

Smoking status

Group         Total

  Current                         SMR
c 20      220      Former     Never   multiplier

u s.
population


White collar


Blue collar

Hypothetd
1OW

40 9       29.4       70 6       40 0       19.1      1.0



39.9       27.8       72.2       40 8       19.7      0.994


47 1       28 2       71.8       34 8       18.1      1.102


20 0       29 4       70.6       20.0       600      0.549

Hypothetical
high

800       294       70.6       10.0       10 0      1.599

a division of current smoking prevalence into heavy smokers or light
smokers. Other characteristics of smoking behavior have also been
shown to influence lung cancer risk, including depth of inhalation,
age of initiation (duration), and tar and nicotine yield of the cigarette
smoked (US DHHS 1981, 1982). The differences in lung cancer
relative risks among male smokers in the ACS study of 1 million
men and women resulting from differences in depth of inhalation
and age of initiation are presented in Table 1. It is apparent that
substantial differences in lung cancer mortality ratios (up to fivefold)
can occur within the broad category of smokers because of differ-
ences in the various dosage measures. It also appears that, in
general, the difference in mortality ratios between the highest and
lowest exposure categories was greater in the older age group than in
the younger age group.

When the SMR is based on the general population, in which
smoking behavior is in the middle range of the dosage measures in
Table 1, it is unlikely that differences in behaviors between an
exposed population and the general population would equal the
differences between the highest and lowest dosage categories.
However, sizable differences may occur, and the values shown in
Table 1 can be used to estimate the impact of these differences. If the
lowest age of initiation (under 15 years) were used as the risk for the
exposed population, and the risk for an age of initiation of age 20 to
24 were used for the control population, there would be a 30 percent
increase (using one risk value for all current smokers) in the SMRs
listed in Table 3. This would increase the SMR for the hypothetical
high smoking prevalence population to 207.4. A corresponding
adjustment for a difference in depth of inhalation could increase

117


hese numbers even further. However, because there is almost
:ertainly some correlation among the various dosage measures
smokers of higher numbers of cigarettes per day are more likely to
nhale and to have begun smoking at an earlier age), it is not valid to
rest these numbers as independent measures of risk. It does seem
:lear, however, that substantial variations can occur in the "expect-
?d SMR" for a population, based on differences in smoking preva-
lence, differences in number of cigarettes smoked per day, and
probably differences in age of initiation. These adjustments suggest
that SMRs in excess of 200 may occur owing to differences in
smoking patterns and differences in depth of inhalation. The use of
high tar and nicotine cigarettes might increase the SMR even
further.

In the description of differences in smoking patterns by occupation
presented in the preceding chapter, only modest differences between
blue-collar workers and white-collar workers were found for age of
initiation and number of cigarettes smoked per day. However, larger
differences in these dosage measures are present among some of the
subcategories of blue-collar and white-collar workers. Substantial
variation from national norms in the various dosage measures may
also occur because of sampling and selection bias in the small
population samples that are often a real limitation in occupational
studies. Even in larger studies, such as the study of 17,800 asbestos
insulation workers (Hammond et al. 1979), substantial differences
between the asbestos-exposed workers and the general population in
number of cigarettes smoked per day are demonstrable (82.8 percent
of the asbestos workers smoked more than 20 cigarettes per day in
contrast with 68.5 percent of the men in the general population).

Failure to control for differences in smoking behavior may lead to
a spurious impression of interaction. A spurious interaction pro-
duced by differences in smoke dose has a greater public health
significance when the outcome is an apparent antagonism rather
than a synergism. If the workers who smoke and are exposed to a
given agent smoke fewer cigarettes per day, or began smoking later
in life than the control population, an apparent protective effect (i.e.,
a less than additive effect) of the occupational exposure may result.
In this setting, if the population of nonsmokers is too small to
evaluate the effects of the occupational agent, only the biased
estimate of the agent's effect on smokers will be available; the
spurious antagonism may mask the effect of an occupational
carcinogen by lowering the rate of lung cancer in the workers with
combined exposure. A lower number of cigarettes smoked per day
may be a relatively frequent confounder in worksites where smoking
is not allowed during working hours, and a later age of initiation
may exist in workforces with higher education levels. Thus, lack of
information on smoking may lead to biased estimates of the effect of

118


an occupational agent, and even to the impression that the agent has
no effect. This potential for missing the effects of an occupational
carcinogen makes the incorporation of dosage data a critical part of
the consideration of statistical interactions.

This discussion has used examples in which differences in smoking
dosage measures resulted in spurious interactions between smoking
and occupational exposures. However, the same potential exists for
differences in occupational exposure dose between smokers and
nonsmokers in the exposed population. If the smokers in the exposed
population have a greater exposure to an occupational carcinogen
than the nonsmokers, then the effect of combined exposure might be
expected to appear to be greater than additive.

A companion question of "dosage" measurement among the
smokers in occupational studies is how to classify pipe and cigar
smokers and former smokers. Pipe and cigar smokers have a lower
risk of developing lung cancer (but not oral cancer) than cigarette
smokers and are distributed differently by age, reflecting the greater
use of pipes and cigars by older men (US DHEW 1979). To the extent
that differences in the use of pipes and cigars exist among exposed
groups and control populations, the effects of smoking may be
confounded if pipe and cigar smokers are classified in the study as
smokers. Pipe and cigar smokers should be either analyzed as a
separate category, or if the number of subjects is too small for
separate analysis, they may be combined with light smokers as part
of a dose-response relationship. A similar problem arises with
former smokers. The lung cancer risk in former smokers declines
with the increasing duration of cessation. Few people begin to smoke
after age 25, and the percentage of the population who have quit
smoking increases with increasing age. Many occupational settings
have been the focus of intensive cessation efforts, particularly those
worksites where an increased lung cancer risk has been established
or suspected. These efforts, as well as the other previously described
reasons for differences in smoking patterns, may make the preva-
lence and age distribution of former smokers in an occupationally
exposed population different from that in a control population;
therefore, former smokers should not be included with current
smokers in an analysis of occupational exposures but should be
treated as a separate category.

One of the methods that has been used to control for the
differences in smoking between control groups and exposed popula-
tions, or between cases and controls (Liddell et al. 1984), is to
examine the dose-response relationships of smoking and occupation-
al exposure for lung cancer. An example of such an analysis
performed on a group of asbestos miners using a case-control
approach is presented in Table 4. The risk of developing lung cancer
is shown to increase with increasing cumulative asbestos exposure in

119


TABLE 4.-Risks of lung cancer, by cigarette smoking and
      asbestos exposure, relative to all 223 cases and
      715 referents for whom smoking histories were
      reliable; unmatched analysis

Exposure accumulated up to 9 years before death of case!
            (f/mL)q

Pack-years `

Law       Medium
(<loo)     (< um

High and
very high
(2 1,ooo)

All

0       Number of cases            6          7          10        23
      Number of referents         103         61          37        201
      Relative risk             0.19        0.37         0.87       0.37

1, 140   Number of cases            29         27          34        90
      Number of referents         123         93          63       279
      Relative risk             0.76        0.93         1.73       1.03

240     Number of cases            40         35          35        110
      Number of referents         117         79          39       235
      Relative risk             1.10        1.42         2.88       1.50


All     Number of cases            75         69          79       223
      Number of referents         343        233         139       715
      Relative risk             0.70        0.95         1.82       1.00

s Sumber ofcqprettes a day:20 x duration in year6
SOURCE LIddell et al r19@4)

all three categories of smoking dose. Stratification is useful for
examining exposure-response relationships, an important element
in establishing a causal association between a given exposure and
lung cancer.

If stratification is used to control the confounding between
smoking and an occupational exposure, careful consideration must
be given to the relative magnitudes of the effects of smoking and
occupational exposure on lung cancer risks when determining the
number of smoking dose categories compared with the number of
occupational exposure dose categories. As discussed elsewhere in this
Report, the prevalence of smoking has been higher among men born
between 1910 and 1930 than among men born in later decades. This
cohort of men represents the older workers in many occupationally
exposed populations, and it is these same workers who were
previously exposed to levels of occupational agents that substantially
exceeded the levels currently experienced. Thus, populations of older
workers have had higher cumulative exposures to occupational
agents than their younger peers at the same age, and have also had
higher cumulative exposure to cigarette smoke than their younger
peers at the same age. The result may be a residual confounding
between cumulative occupational exposure and cumulative smoke
exposure in assessing the effects of these two exposures. If bhe

120


magnitude of the effect of smoking is large compared with the
magnitude of the effect of the occupational exposure, and few broad
categories of smoking status are used with a greater number of
categories of occupational exposure, then higher levels of smoking
dose may occur with increasing occupational exposure dose category,
generating a spurious dose-response relationship. Correspondingly,
too few occupational exposure categories may result in a spurious
strengthening of the dose-response relationship present for smoking.
The total number of categories that can be used in this kind of
analysis is usually limited by the number of lung cancer patients
available for analysis; therefore, the distribution of the dosage
categories to smoking and to the occupational exposure should
reflect the relative magnitude of the effects of the separate expo-
sures on lung cancer risk.

Duration of Exposure

In models of lung cancer risk due to smoking behavior, separate
terms for intensity of smoking and duration are commonly included.
In a risk model developed by Doll and Peto (19781 for the study of
British physicians, the term for intensity of exposure was raised to
the second power and the term for duration of exposure was raised to
the power of 4.5.

Confounding may arise because of correlation between age and
duration of exposure. Because of the importance of duration of
exposure (and its covariate age) on lung cancer risk, the majority of
the lung cancer cases will develop in the older members of a
population. Correspondingly it is the smoking prevalence and dosage
among these older workers that will largely determine the lung
cancer risk for the population. The mean prevalence or mean dosage
measures for the population do not take into account the effect of
duration of exposure on the lung cancer risk. In a comparison of
populations with different age distributions of smoking prevalence,
or of the prevalence of heavy smokers, the population with the
higher prevalence in the older age ranges will have the higher risk.

A final source of concern in examining the relationship between
occupational exposure and lung cancer in cigarette smokers is
generated by the lag time between the exposure to a carcinogen and
the clinical manifestation of lung cancer. This lag time is a
combination of the induction period (the time from exposure to
disease initiation) and the latent period (the time from disease
initiation to clinical manifestation) (Rothman 1981). This lag period
is not fixed, but rather has a broad distribution over perhaps 50 or
more years (Nicholson et al. 1982).

Epidemiologically, the shortest lag times are identified by the
interval between the age of onset of exposure and the age when an
increased relative risk can first be demonstrated secondary to the

121


exposure. For some exposures, once the exposure period has exceed-
ed the shortest lag time, the relative risk often increases rapidly
with increasing duration of exposure (Nicholson et al. 1982),
resulting in a dramatic increase in disease rates with increasing age.
It appears that the shortest lag period for smoking-induced lung
cancers is in the range of 15 to 20 years, as demonstrated by the rise
in lung cancer death rates that begins after age 30 to 35. The lag
period for occupational carcinogens in lung cancer is not well
characterized, but some agents have lag times similar to that found
with smoking (Nicholson et al.  1982; Selikoff and Lee 1978).
However, the onset of exposure to cigarettes and occupational
carcinogens may occur at substantially different ages. Any such
difference needs to be considered when examining the interactions of
occupational exposures and smoking.

Ideally, the study of an occupationally exposed cohort would follow
the entire cohort until the last survivor had died, so that late effects
of exposures would not be missed. The reality of examining working
populations and the need for timely assessment of existing risks
makes the examination of workers at a variety of ages the norm in
epidemiologic studies. In this setting, careful consideration of the
differences in age of onset of smoking and of occupational exposures
is necessary if the effects of occupational exposure are not to be
missed or underestimated. For example, assuming that the average
age of onset of smoking is 15 and the average age of onset of a
particular occupational exposure is 25, the combined exposure effect
is one of equal and additive risks of lung cancer and the lag time for
both agents is 20 years. The lung cancer risk due to smoking would
begin to increase at age 35, but because of the lo-year difference in
age of onset of exposure, the risk due to the occupational exposure
would not begin to be expressed until age 45, and even then would
appear to be much smaller than the risk due to smoking because of
the effects of the longer duration of exposure to cigarettes. If the
cohort of workers with these two exposures is relatively young, with
few older workers, then the effect of an occupational exposure may
be missed or substantially underestimated. A similar concern exists
when examining an agent that was introduced into the workplace 20
to 30 years ago. The cohort of exposed workers would represent a
cross-section of ages, and therefore a cross-section of smoking habit
durations. An additive risk effect of the occupational exposure would
be small in comparison with the cumulative risk secondary to
smoking in the older workers, and the number of cases of lung
cancer in young workers (where the risk effects might be more equal)
would be small. Again, the effect of an occupational carcinogen could
easily be missed in this setting.

This discussion uses a simple statistical model of independent
additive effects in concert with a biological concept of lag time.

122


Interpretation based on this kind of biologic extrapolation of
statistical concepts is hazardous at best; nevertheless, some consider-
ation of the differences in the age of onset of exposure should be part
of both the biologic and the statistical considerations of the
interactions between smoking and occupational exposures.

Control of Confounding

The examination of the risk associated with an occupational
exposure generally requires a comparison group. Prospective mortal-
ity studies of the general population generally have too few
individuals with the exposures of interest to allow analysis. There-
fore, cohort and case-control formats have commonly been used. The
control groups in either of these formats may be external (i.e.,
separate population) or internal (i.e., workers with high exposure
compared with workers with lower exposure). A variety of methods
have been used to deal with the confounding of occupational
exposure by cigarette smoking.

Comparisons Using External Control Populations

Common external control populations are the national or regional
populations. Death rates in these populations can be used to
generate age- and time-adjusted expected numbers of deaths for the
exposed population, with the ratio of actual deaths to expected
deaths as the SMR. The large numbers of deaths in these large
control populations results in relatively stable death rates over time
for the common causes of death, and the smoking habits of these
populations are often available from national or regional survey
data. However, the smoking habits of the population are not known
in relation to the cause of death, which limits the use of this data to
control the confounding of occupational exposure by smoking in
occupational cohorts. If the smoking habits of the workforce are also
known, then the magnitude of the effect that the differences in
smoking habits might have on the SMR can be estimated by
assigning risk values to the proportions of the populations in
different smoking categories (as described in the section on sources of
confounding) (Axelson 1978). This adjustment for differences in
smoking prevalence ignores trends over time as well as a variety of
other potential sources of confounding. However, when this ap-
proach is used, the smoking-adjusted SMR alters the expected value
of the SMR from the value of 100 that was expected prior to
adjustment for smoking.

An alternative approach is to use an external control population
for whom the smoking habits are known in relation to the causes of
death. The use of a control population with known smoking habits
allows the direct comparison of populations of smokers and non-

123


smokers with and without the exposure being investigated. These
direct comparisons allow an examination of the risk of the occupa-
tional exposure in the absence of smoking (i.e., in never smokers) and
also the examination of potential interactions between smoking and
occupational exposures. A study may be constructed to prospectively
or retrospectively examine the lung cancer death rates in a cohort of
occupationally exposed workers compared with a control population,
or a group of patients with lung cancer may be identified and
matched with a set of controls without lung cancer in order to
examine the frequency of a given occupational exposure in the two
groups. In examining lung cancer risk, it is important that the
control population be similar to the exposed population in age,
ethnicity, socioeconomic status, and geographic location.

In general, studies are designed to be able to identify levels of lung
cancer risk due to occupational carcinogens that are lower than the
level of risk due to smoking. This potential difference in magnitude
of effect needs to be assessed carefully when considering the level of
detail with which the smoking data are obtained and examined.

The selection of a control group for an occupational study is often
influenced by the ease with which data can be collected as well as by
the comparability of the control group with the exposed workers.
Control groups can be selected from unexposed workers in the same
plant, from workers in different plants where no exposure occurred,
from populations selected from the same geographic locations as the
workers, and from populations being followed as part of other
epidemiologic investigations. Some of these control groups may have
substantial differences in smoking behavior from the exposed group.
For example, if management and administrative employees are
included in the control group, the prevalence of smoking in the
control population or in comparison with a blue-collar exposed group
may be reduced. Similarly, controls selected from different worksites
may have different smoking patterns owing to differences in work
rules, age of employees, or other demographic factors, or simply by
chance. Populations drawn from other epidemiologic studies may
also have different smoking patterns, and the mode of determination
and definition of smoking status may be different from that used in
the exposed group.

A common method of controlling for the confounding due to
smoking is to separately examine smokers, nonsmokers, and former
smokers. This allows examination of the independent effects as well
as of the interactions; however, the examination of smoking patterns
represents slightly different challenges in each of these groups.

Lung cancer risks may be examined in nonsmoking populations of
occupationally exposed and nonexposed individuals for two separate
reasons. First, such analyses can establish whether a risk due to
occupational exposure occurs in the absence of cigarette smoking or

124


whether exposure only modifies the effect of smoking. Second,
nonsmokers represent the lowest dosage category in examining the
dose-response relationship for smoking. The demonstration of an
effect of an occupational exposure in the absence of cigarette
smoking requires a population of lifelong nonsmokers who have
neither smoked cigarettes or cigars or used a pipe. In contrast, when
a dose-response relationship is being examined, it would not be
unreasonable to combine never smokers with pipe and cigar smok-
ers, or even with light smokers, as a low dose group for lung cancer
risk (pipe and cigar smokers should not be included in the low dose
group for oral cancer risk). For exposures with modest increases in
lung cancer risk, the low prevalence of never smoking status,
coupled with the low expected risk of lung cancer in this group,
means that large populations of workers must be examined in order
to define the risk of exposure in the absence of smoking. Most
occupational studies are limited by the size of the workforce being
examined, and therefore, it is often necessary to combine never
smokers with low smoking risk groups in order to have an adequate
sample size. Once this combination has taken place, the study can
examine only the effect of low smoke exposure coupled with
occupational exposure, rather than the effects of occupational
exposure in the absence of smoke exposure.
The low prevalence in many current workforces of people who
have never smoked and the low risk of lung cancer in this group
generally means that only a very few lung cancer deaths occur in
this group, limiting the number of deaths for which to perform an
analysis of the effects of an occupational exposure in the absence of
smoking. For example, in the large study of asbestos insulation
workers (Hammond et al. 1979), only 5 lung cancer deaths were
recorded in nonsmokers out of more than 8,000 asbestos-exposed
workers (smokers and nonsmokers included) whose smoking habits
were known. Drawing inferences from small numbers of lung cancer
cases is necessary in occupational studies, but two important caveats
should be considered. First, it is essential that lung cancer patients
placed in the never smoking category are actually individuals who
have never smoked. The inclusion of even modest numbers of
misclassified smokers or light smokers may increase the number of
lung cancers over that expected on the basis of the risks in the never
smoker, nonexposed control population. For this reason it is critical
that the data on smoking habits be accurate and obtained in the
same way in the exposed population as in the control population.
When the level of monetary compensation for occupational disability
may be influenced by smoking status, workers may be motivated to
define themselves as never having smoked, regardless of their actual
smoking status. In many studies the determination of smoking
status is made for the living subjects by questionnaire or interview

125