American Journal of Epidemiology Voi 141, No 7

Copyright O 1995 by The Johns Hopkins University School of Hygiene and Public Hearth Printed In USA
All rights reserved

Epidemiologic Application of a Dosimetric Model of Dust Overload

Mary A. Ballew,1 David Kriebel,2 and Thomas J. Smith 3

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

Evidence from laboratory animals suggests that certain rates of deposition of dust can slow the normal
clearance of the lungs by macrophages. If this occurs in humans, then estimates of exposure or dose for
epidemiologic models may be improved by explicitly incorporating this "overload" phenomenon into lung dose
estimates. Using a model of dust overload, the authors estimated the lung dust dose for a group of workers
exposed to silicon carbide dust. From a knowledge of underlying biologic mechanisms and using goodness
of fit tests, the authors identified parameters for the dosimetric model that yielded dose estimates which fit the
epidemiologic data well. The dosimetric estimates also compare favorably with cumulative exposure as risk
predictors of radiographic evidence of pulmonary fibrosis. The evidence for overload is not strong, however,
since the data are also consistent with other clearance patterns including very slow linear one-compartment
clearance. Nevertheless, in some data sets, use of a dosimetric model instead of cumulative exposure to dust
may reduce misclassification in exposure and improve fit to epidemiologic data. Furthermore, evidence of a
better fit would provide valuable information about biologic mechanisms or aspects of dose such as the effect
of short-term, high-intensity exposures. These in turn might provide information about prevention strategies
such as standard setting. Am J Epidemiol 1995; 141:690-6

dust; occupational exposure; pulmonary fibrosis; silicon dioxide

Researchers seeking to characterize the relations lungs, the dose can be described by a toxicokinetic
between environmental and occupational hazards and model of air concentration inhaled, deposition at the
disease risk are increasingly able to benefit from re- target tissue, and dust removal by clearance processes.
cent improvements in the quality of the quantitative The parameters of this model can be estimated with
exposure data for their epidemiologic models. Quan- physiologic data from other studies or by fitting to
titative exposure data hold the promise of greater epidemiologic data. This paper examines the process
sensitivity through reduced misclassification of expo- of using epidemiologic data to fit some of the param-
sure. However, it is not always clear how to summa- eters of a dosimetric model.
rize the lifetime exposure of a worker in an epidemi-
ologic model. One solution to this problem is to use Description of the study group
mathematical models to derive estimates of "dose"
We obtained data (also analyzed by Smith and oth-
from exposure data. This approach has been advocated
ers (3-5)) on 170 male workers from a Canadian
previously (1, 2), and in this paper we report an
facility producing silicon carbide. The workers mix
application to a group of silicon carbide workers ex- petroleum coke, silica, and sawdust, bake the mixture
amined cross-sectionally for evidence of pulmonary in large furnaces for 36 hours, let the furnace cool, and
fibrosis. then break the silicon carbide into large lumps with
The approach involves two separate modeling steps: pneumatically powered chisels. Because of the contin-
in the first (dosimetric) step, a mathematical model
estimates the tissue dose separately for each worker. In Received for publication February 14,1994, and In final form July
the second, an epidemiologic model characterizes the 21, 1994.
Abbreviation: ILO, International Labor Organization.
association between the disease outcome and the result 1
Department of Work Environment, University of Massachusetts
of the dosimetric model for each worker. Lowell Current address: US Environmental Protection Agency,
The dosimetric model is based on the assumption Boston, MA
2
Department of Work Environment, University of Massachusetts
that each milligram of material causes an increment in Lowell, Lowell, MA.
effect per unit of time it resides in the target tissue. 3
Department of Environmental Sciences, Harvard School of
Therefore the total effect is proportional to the total PuWic Hearth, Boston, MA
Reprint requests to Dr. David Kriebel, Department of Work Envi-
quantity of material present in the tissue integrated ronment, University of Massachusetts Lowell, One University Ave-
over its residence time in the tissue. For dusts in the nue, Lowell, MA 01854.

690
 Dosimetric Model of Dust Overload 691

ual assembly, disassembly, and heating of the fur- trations of dust that they lose their ability to remove
naces, the workers received exposure to silicon car- particles from the lungs. This is the point of overload.
bide dust, hydrocarbon fumes, SO2, and carbon Researchers have found support for the overload hy-
monoxide throughout the workday. The process re- pothesis in experiments on rats exposed to diesel ex-
mained essentially the same over the 40 years during haust (8, 9). Studies of chronic exposures to dust
which this group was exposed. support the idea that particles may sequester in deep
Smith et al. retrospectively estimated the workers' tissue compartments (such as lymph nodes or intersti-
exposures to dust based on individual work histories tial tissues) in such a way that the dust cannot be
and average air concentrations as measured in each job cleared from the body (8-10). It has been suggested

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

and work area in 1980 (3). They also measured pul- that particulates of high toxicity may cause overload
monary function and took chest radiographs for each sooner and at lower concentrations than do particulates
worker during a lung examination in 1980. Three of low toxicity (9, 10).
National Institute for Occupational Safety and Health-
certified chest radiographic readers (B readers) inter- Smith models of lung deposition
preted each radiographic film according to the Inter-
Smith proposed two alternative dosimetric models,
national Labor Organization (ILO) criteria for
and in this paper their epidemiologic performance was
pneumoconioses (5). Also available were age, height,
compared with each other and with that of simple
and smoking status in 1980 and work histories for each
cumulative exposure. The first model consists of a set
worker from 1941 to 1980.
of simultaneous differential equations designed to rep-
resent the movement of dust through six lung com-
Dust overload partments (figure 1) (4). Smith hypothesized that, in
A variety of dusts are known to have toxic, noncar- the specific case of silica, encapsulated dust would
cinogenic effects on the lungs. Probably the most lead to fibrosis in a dose-dependent fashion. Parameter
common mechanism of action is fibrogenesis, the ba- values for this model (10 in all) were estimated by
sis of most of the classic pneumoconioses. The lungs extrapolation from published studies of laboratory an-
have several defense mechanisms against the fibrogen- imals. Smith showed that the dose of encapsulated
esis resulting from prolonged dust exposure of which dust predicted by this model was well correlated with
the most important is probably macrophage clearance. the presence of radiographic opacities in the same data
In the healthy lungs, macrophages can remove large set used in the present investigation (4).
quantities of deposited particles rapidly and with little In 1992, Smith proposed a second, simplified, one-
apparent damage. Toxic dusts, such as silica and as- compartment model to reflect the essential temporal
bestos fibers, impede macrophage clearance through dynamics involved in dust deposition and overload
their inherent cytotoxicity. There is also some evi- (1). Smith postulated that the overload phenomenon
dence, mostly from laboratory animals, that even could be modeled by Hill's equation, the most familiar
"nontoxic" dusts may interfere with normal clearance example of which is the Michaelis-Menten equation
if exposure is sufficiently intense (6, 7). According to
this hypothesis, dust clearance from the lungs is effi-
AIRBORNE
cient while the quantity in the lungs is low ("low DUST ENCAPSULATED
burden"). However, the clearance process has a lim-
T
EXPOSURE
PARTICLES
ited capacity that can be reached at high lung burdens. (LUNG BURDEN)
The phenomenon of "overload" occurs when the
capacity of the clearance system is exceeded (over-
loaded) and further increases in burden do not increase
clearance (6, 7). During heavy exposure, the burden
FREE
PARTICLES s k< 7
TEMPORARY PARTICLES
IN LYMPH
PARTICLES
increases rapidly because the deposition rate of dust NODES
greatly exceeds the clearance rate of dust leaving the
lungs. At low concentrations, clearance operates more MACRO- MACRO-
PHAGES PHAGES \
efficiently, and lung burden may show an approximate (fait) (flow)
equilibrium with exposure. Under this latter condition,
cumulative exposure—the sum of average exposure
intensities (e.g., annually) over the working lifetime—
will be approximately proportional to the lung dose FIGURE 1. Lung fibrosis model based on work by Smith (4). The
model represents the deposition, clearance, and retention of dust in
(3). According to the overload hypothesis, macro- the human lungs The quantity in the compartment "Encapsulated
phages become so heavily burdened at high concen- Particles" is an estimate of the lung dose.

Am J Epidemiol Vol. 141, No. 7, 1995

692 Bailew et al.

for enzyme kinetics (11). Toxicokinetic models of this from published animal and human data, or by an
sort are often formulated as differential equations or as empirical fit to epidemiologic data. Both approaches
their discrete analogs, difference equations. Construct- were used here. Smith estimated &Q to have a value of
ing the model as a difference equation allows one to 0.84/year (based on Bohning's estimate of 0.82 years
solve even quite complex, nonlinear equations analyt- for the normal half-time of clearance in humans (13)).
ically using a desktop computer without the need for However, there was considerable individual variation
elaborate differential equation-solving software. Fol- in clearance among the 11 human subjects used in this
lowing Smith, a single difference equation describes long-term study of inert dust clearance, so that a wide
this model: range of &o values are consistent with the data. By

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

(1) scaling from the burden of diesel particulates that
triggers overload in rats (7), Smith estimated km to be
where: about 184 mg. These parameter values, k$ = 0.84 yr" 1
and J^ = 184 mg, were assumed to be the "best guess"
C, lung burden at the beginning of the time
values in these investigations, while a process of iter-
interval;
ative fitting, described below, was used to identify
dt duration of the time interval, 0.1 years in
these calculations; "best fir" values of these same parameters.
[A] time-weighted average (TWA) air
concentration (mg/m ) during the time
interval; MATERIALS AND METHODS
/ volume of air (m3) inhaled at work during A difference equation was constructed describing
the interval; the simpler Smith model (equation 1) and using a wide
/d fraction of insoluble, respirable particles range of values of the unknown parameters A^ and k^.
deposited in the target area of the lung; and Changing values for km and /^ in the equation allowed
k clearance rate during that time interval us to calculate a dose metric representing different
(yr" 1 ), which is defined as: assumptions about the overload process. The fraction
k k ^ / ^ + Q); and of inhaled dust that deposits in the respiratory tract,/d,
fco = baseline clearance rate (yr *); and
was estimated to be a constant 20 percent, while the
k^ = lung burden producing a 50 percent decline
ventilation rate for each worker, /, was estimated to be
in the clearance rate (mg).
20.8 liter/minute for all subjects over all time. These
Following pharmacologic convention, k equals the assumptions affect the absolute values of the resulting
natural logarithm of two divided by the half time of dose estimates; however, because they were held con-
clearance of dust from the lungs. The difference equa- stant for all study subjects and over the entire time
tion will accurately reflect the continuous change in period, they have no effect on the relative epidemio-
lung burden that would be obtained by solving the logic performance (goodness of fits or magnitude of
equivalent differential equation as long as the change effect estimate) of the dose metrics.
in burden for any dt is less than 10 percent. In practice, For each member of the study group, an estimate of
the equation is solved repeatedly at each increment of the annual mean concentration of respirable dust (mg/
dt, yielding a new value of C, that is essentially a sum m3) was available for all years of employment. Each
of what remains of the burden in the previous interval
worker's exposure history (ranging from 1 to 40 years)
after clearance has acted upon it, plus the contribution
was used to estimate the annual burden of dust in the
to burden from the fraction of exposure in the current
lungs from the start of employment until 1980 (the
interval that reaches the target tissue.
year in which chest radiographic examinations were
Hattis and Silver found that the one-compartment
Smith model fit well to data on rats and humans performed). The annual burdens from the start of em-
exposed to diesel exhaust or cobalt particulates (D. ployment until 1980 were summed for each worker to
Hattis and K. Silver, Ashford Associates, Massachu- estimate the dose of dust in the lungs. If there is no lag
setts Institute of Technology, Cambridge, Massachu- or latency period between exposure and appearance of
setts, unpublished report submitted to the Laborers' radiographic evidence of fibrosis, then this dose
Health and Safety Fund, April 1, 1992). It correctly (gram-years) should be proportional to risk of radio-
predicted instances of overload and fit the data better graphic opacities. However, if one assumes that some
than more complicated models (12). time transpires between exposure and response, then
The one-compartment model has two unknown pa- one might wish to lag the dose. Lags of 5 and 10 years
rameters, &„, and fco, which one can estimate either were investigated.

Am J Epidemiol Vol. 141, No. 7, 1995

 Dosimetnc Model of Dust Overload 693

Data used in epidemiologic models

10
In the epidemiologic models, smoking status was 78 3 78 6 84.5 83.0 01
dichotomized into "never smoked" (0) and "ever T T T T

smoked" (1), and the scores of radiogTaphic opacities 78.3 78.4 831 83.0
1 - T V V T
were collapsed into "none" (ILO scores of 0/0) or "any 1.0
profusion" (ILO scores of 0/1 or greater). To create an
epidemiologic model for the risk of developing radio- 01 -
graphic opacities, logistic regression was used to esti- 10

mate the change in the odds of radiographic opacities

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

78 3 78 3 78 3 78 3
found by contrasting various levels of exposure or a 100
0 01 - T T T T
dose metric to the lowest category of exposure or dose
metric.
The exposure data were highly skewed, with a large 0 001 p I

number of workers receiving relatively low cumula- 01 10 100 1,000 10,000

tive exposures. When either cumulative exposure or km (mg)
dose metric was used as a continuous variable, good- FIGURE 2. Error log likelihood statistics from logistic regression
ness of fit statistics were poor, and logarithmic trans- models for the risk of radiographic opacities from lung doses esti-
mated in dosimetnc models with various values of the parameters k0
formations did not improve the fits appreciably. This and km. Note that smaller values of log likelihood indicate greater
is a common problem in logistic regression models consistency with the data; differences >6.0 correspond to p < 0.05
with highly right skewed independent variables. To in tests of the null hypothesis of equal goodness of fit. When
cumulative exposure Is used instead of a dose metric, the log
avoid this problem, the exposure or dose variable in likelihood is 83 0 The dashed line indicates the region of relatively
the logistic regression models was categorized into poorer fitting models.
thirds (i.e., cutpoints at the 33.3 and 66.7 percentiles).
When exposures were lagged, workers who began
employment less than 5 or 10 years before the medical
The relative odds of radiographic opacities at the
survey had no exposure. When exposure was lagged
three dose levels were calculated repeatedly with dose
by 10 years, 78 workers (46 percent) had zero expo-
metrics based on different values of k^ and km spread
sure, while 14 (8 percent) had zero exposure with a lag
across a grid of values (figure 2). The resulting surface
of 5 years. For the 10-year lagged data, the division
of values for error log likelihood represented the rela-
into thirds had to be accomplished differently to ac-
tion between ICQ and km and the goodness of fit of the
commodate these no exposure workers. The members
logistic model. We generated such a surface for un-
of the no exposure group (n = 78) became the low
lagged estimates of dose (figure 2) and for doses with
exposure (reference) category, and the remainder were
lags of 5 or 10 years (data not shown). One can
equally divided into the medium and high categories
compare the log likelihoods of nested models: the
(n = 46 in each).
difference in two log likelihoods has a chi-squared
Preliminary analyses including both smokers (n = distribution and, with 2 df (Ic^, k^, the chi-squared
132) and nonsmokers (n = 38) showed that smoking statistic for an a level of 0.95 is ^6.0.
status strongly modifies the effect of exposure or dose.
However, with such a small data set and, in particular,
very few nonsmokers, it was difficult to obtain stable RESULTS
estimates for risk or for the interaction between smok- Of the 170 workers in the study group, 100 (58.8
ing and risk on each of three levels of a trichotomized percent) developed radiographic opacities. Of the 132
exposure or dose variable. As a result, analyses were workers who smoked, 84 (63.6 percent) developed
restricted to the 132 smokers and excluded consider- radiographic opacities. The mean number of years
ations of effect modification. Logistic regression mod- worked was 14.4; the median was 9.5, and the range
els that estimated the risk of radiographic opacities was 1.2-40 years. Ten percent of the group worked 35
among smokers and used only trichotomized cumula- or more years. A large number of workers (approxi-
tive exposure or dose as predictors consistently mately 25 percent) worked between 5 and 10 years,
showed good fits to the data. After controlling for and the rest of the workers were evenly distributed
exposure, age was not found to be an important pre- across time periods. Their cumulative exposure to dust
dictor of risk of radiographic opacities. All study sub- had a mean of 8.7, a median of 7.0, and a range of
jects were male. Goodness of fit was evaluated in the 0.3-36.8 mg/m3 years. Dividing the median cumula-
logistic regression models with the log likelihood sta- tive exposure by the median time worked yielded a
tistic. median population exposure of about 0.7 mg/m3 of

Am J Epidemiol Vol. 141, No. 7, 1995

694 Ballew et al.

silicon-respirable dust. The workers varied in age from Lagging the dose metric by 5 or 10 years had several
20 to 64 years. About 40 percent of the workers fell effects. Goodness of fit improved for all dose metrics
between the ages of 25-35, about 25 percent were over and for cumulative exposure. For example, dose met-
50, and there were relatively few workers between the rics estimated with km = 10 mg, ^ = 0.01 yr" 1
ages of 40-45 years. Thus, the population had a yielded logistic regression models with log likelihoods
bimodal age distribution that perhaps reflects hiring of 76.3 (5-year lag) and 75.5 (10-year lag). However,
practices at different time periods. the lagged exposure data also yielded a much "flatter"
Changing km and ^ produced widely varying esti- goodness of fit surface; with a lag of 10 years, the
mates of lung dose (mg-years) of silicon carbide (table goodness of fit values only varied from 75.7 (the best

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

1). As ICQ decreased (i.e., half-life increased), the mean fit ) to 77.1. Cumulative exposure yielded an interme-
estimated dose increased. Thus, at a low baseline diate goodness of fit.
clearance, a significant amount of dust was estimated Regardless of lag, the pattern of log likelihood val-
to accumulate in the lungs. Also, as / ^ decreased, the ues was consistent: dose metrics with the lowest val-
mean dose increased. A small km implies that the ues of km andfcgfit the data best and larger values of
macrophages quickly reach overload, clearance slows, km and k^ fit worse.
and dust rapidly accumulates in the lungs. The L-shaped region of better fits yielded strong
The goodness of fit statistics resulting from models measures of association between dose and risk of
with varying values of km and k$ can be summarized as radiographic opacities. The goodness of fit resulting
a surface plot: the x and y dimensions of the plot from the one-compartment dosimetric model with bet-
represent alternative values of the two unknown pa- ter fitting values of km and k^ (table 2) is somewhat
rameters, and the values on the surface are the log better than that which results from use of the "best
likelihoods for models based on particular values of km guess" values of the parameters (table 3 - difference in
and /CQ (figure 2). Thus, for example, the value 84.5 in log likelihood = 4.8, p = 0.09), while the odds ratio
the upper right-hand corner of figure 2 is the log for high exposure compared with low exposure is
likelihood that results from fitting a logistic model considerably larger in the former than in the latter
with dose metrics based on the hypothesis that km = model (8.0 vs. 3.4). Cumulative exposure also yielded
184 and Jfcp = 3.5. In these data, this surface was fairly substantially lower risk estimates than the better fitting
"flat"; that is, the best fitting (minimum) and worst model (table 4).
fitting (maximum) values of log likelihood differed by The six-compartment dosimetric model, with pa-
about six (p = 0.05). Best fitting values of log like- rameter values estimated from the published literature,
lihood occurred with low values of km (0.5 mg), re- yielded risk estimates and a goodness of fit similar to
gardless of ICQ, and low values of k0 (0.01 yr" 1 ' tir2 = the one compartment model with better fitting values
70 years), regardless of the values of km. Poorer fits of *o and km (table 5).
occurred at high values of ICQ and km. Thus, few if any Overall, the models with the better fits were also the
values of km and ICQ on this surface were inconsistent models with the highest odds ratio, and the converse
with the data. The log likelihood was also calculated was also true. In other words, models with poorer fits
for a model using cumulative exposure rather than a
dose metric. It was relatively large (83.0) and margin-
ally different from the best fitting values (figure 2). TABLE 2. Relative risk of radiographic opacities for those
with three levels of estimated dose, from a "better fitting"*
dosimetric model: kg = 0.01 yi~ 1 t and km = 10
TABLE 1. Descriptive statistics for alternative dose metrics
and cumulative exposure
95%
Estimated lung dose Odds
confidence
Dose metrics with selected values (gram-years) ratio interval
o f V ' and kmi (gram-years) Cumulative
exposure
<16.2 1.0
it = 3 5 k = 3 5 k = 084 k,= 0.01
and and and and (mcym»-years) 16 2-87 2 15 0 7-3.5
= 10,000 frm=184 *' - = 1 0 >87.2 80 2 5-26

Median 1.0 179 11.8 196 70 Log likelihood = 78.3

Mean 1.3 551 37.2 62 3 87 * Many different combinations of parameter values yield similar
SD* 1.1 74.1 53.3 81.4 7.4 goodness of fit statistics (see text). This table shows the results of a
Minimum 0.0 0.2 0.2 0.2 0.3 logistic regression using one of these pairs of parameter values and
Maximum 5.4 391.0 306 0 419 0 36.8 dividing the study subjects into three subgroups of equal size based
on Individual dose estimates.
1
• Baseline clearance rate (yr- ). t Baseline clearance rate. A k0 of 0.01 corresponds to a
t Lung burden estimated to cause a 50% decline in the clearance half-time of 69 years.
clearance rate (mg). t Lung burden estimated to cause a 50% decline in the
$ SD, standard deviation. clearance rate.

Am J Epidemiol Vol. 141, No. 7, 1995

 Dosimetric Model of Dust Overload 695

TABLE 3. Relative risk of radlographlc opacities for those best fits on the response surface for error log likeli-
with three levels of estimated dose, from the "best guess"* hood (figure 2). When the lag was longer, the fit was
dosimetric model: kg = 0.84 y r 1 t and km = 184 mg$
better and the odds ratio was higher. However, since
95% fewer workers have long durations of employment,
Estimated lung dose Odds confidence
(gram-years) ratio Interval
models with long lags were based on very small num-
bers and were correspondingly less precise.
1.0
10-45 12 0 5-2.9
>45 3.4 1.3-8.9 DISCUSSION
Log likelihood = 83.1 We used a previously published one-compartment

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

* These values of the dosimetric model parameters were chosen model to estimate the dose received by the lungs from
based on a review of published studies on experimental human and a given exposure to respirable dust in a silicon carbide
animal dust exposure studies (see text). This table shows the results
of a logistic regression using these parameter values and dividing
processing facility. By varying two parameters that
the study subjects into three subgroups of equal size based on alter the assumed rate and behavior of lung clearance
individual dose estimates. in this model, we were able to generate and compare a
t Baseline clearance rate. A k0 of 0 84 corresponds to a
clearance half-time of 300 days.
variety of dose estimates. Although these differences
t Lung burden estimated to cause a 50% decline in the were not striking, models for the risk of radiographic
clearance rate. opacities varied somewhat in their goodness of fit
depending on the values of the parameters chosen.
TABLE 4. Relative risk of radlographlc opacities for those The possibility of comparing many alternative mod-
with three levels of estimated cumulative exposure* els, each the result of a different set of biologic as-
95% sumptions, raises certain methodologic issues: How
Estimated cumulative exposure Odds
(mg/m»-years) ratio
confidence should the "best" model be chosen? Is the best model
Interval
the one with the largest relative risk, the best statistical
<56 1 0 agreement with the data, or the one selected on the
5.6-11.5 1.6 0 7-3.6
>11 5 35 1.3-9.2 basis of prior knowledge? In a recent paper, one of us
(D. K.) argued that goodness of fit along with other,
Log likelihood = 83 0
nonquantitative criteria such as biologic plausibility
* Cumulative exposure was estimated as the sum of the annual and generalizability should be used in choosing the
exposure estimates for each worker in each year of exposure (see
text). This table shows the results of a logistic regression model best model (2). Salvan has also supported the use of a
estimating risks for three subgroups of equal size based on goodness of fit criterion and provided examples where
individual cumulative exposure estimates. the model with the better fit to the data does not yield
the highest risk estimate (14). In the present study,
TABLE 5. Relative risk of radlographlc opacities for those biologic considerations were used to guide the selec-
with three levels of estimated dose based on Smith's six tion of a model structure, and goodness of fit was used
compartment dust model* to choose among alternative values of the model pa-
Estimated encapsulated dust Odds
95% rameters.
confidence
(mg) ratio
Interval In our reanalysis of data for the silicon carbide
<15 1.0
workers, we found that models assuming large values
15-60 1.1 0.5-2 5 of km and fcg (e.g., km of 104 mg and k^ of 3.5 yr" 1 )
>60 35 1.4-8.9 consistently had the same behavior as cumulative ex-
Log likelihood = 845 posure (i.e., the same values for odds ratios and error
log likelihoods). However, models with lower values
* Dose estimates from a previously published model of dust
deposition (4; see text). This table shows the results of a logistic of km and ICQ had better fits to the radiographic data and
regression using three subgroups of equal size based on individual yielded higher risk estimates. Thus, with certain data
dose estimates. sets, one may be able to distinguish between dose
estimates that behave statistically much as cumulative
showed lower odds ratios. At a very large km and short exposure and other doses that behave differently.
half-life (1 year or less), the models estimated dose For the better fitting models, the risks for radio-
metrics that were very similar to cumulative exposure. graphic opacities generated by the one-compartment
However, these models and cumulative exposure do model were similar to those generated by a much more
not fit the data as well as do other combinations of km complex six-compartment model.
and ICQ. The model that was similar to cumulative These data, as modeled by Smith and ourselves, do
exposure (in terms of magnitude of effect estimators not provide strong evidence for the existence of an
and goodness of fit) was far from the models with the overload phenomenon. Actually, the better fitting

Am J Epidemiol Vol. 141, No. 7, 1995

696 Ballew et al.

models, while spread over a wide range of values of reduce misclassification by exposure. This in turn may
and km, all yielded quite similar dose metrics and were lead to better strategies for prevention of disease.
consistent with a linear rate of clearance from a single
compartment with a very long half-time (perhaps de-
cades). This becomes clear when one considers the ACKNOWLEDGMENTS
pattern in better fitting models (figure 2): They have
This work was supported by grant R01-ES-04202 from
low values of k^, regardless of km, or low values of km,
the National Institute of Environmental Health Sciences.
regardless of ICQ. A low value of either parameter
produced a similar result: If ICQ, the background or

Downloaded from https://academic.oup.com/aje/article/141/7/690/88942 by University of Adelaide user on 17 February 2024

initial clearance rate, was low, then the effective clear-
ance rate was very slow, regardless of where km was REFERENCES
set. If km, the burden necessary to produce overload, 1. Smith TJ. Occupational exposure and dose over time: limita-
was small, this implied that a worker's lungs would tions of cumulative exposure. Am J Ind Med 1992;21:35-51.
2. Kriebel D. The dosimetric model in occupational and envi-
have rapidly become overloaded and effective clear- ronmental epidemiology. Occup Hygiene, 1994;l-55—68.
ance would again become quite slow. Thus, slow 3. Smith TJ, Hammond SK, Laidlaw F, et al. Respiratory expo-
clearance, rather than overload, appears to be the more sures associated with silicon carbide production: estimation of
cumulative exposures for an epidemiological study. Br J Ind
likely mechanism supported by these data. Med 1984;41.100-8.
The support these particular data provide for slow 4. Smith TJ. Development and application of a model for esti-
clearance rather than either overload or a cumulative mating alveolar and interstitial dust levels. Ann Occup Hy-
giene 1985;29:495-516.
exposure metric cannot be considered strong because 5. Peters JM, Smith TJ, Bernstein L, et al. Pulmonary effects of
the differences in fit between alternative models was exposures in silicon carbide manufacturing. Br J Ind Med
small. Only the best and worst fitting models differed 1984;41:109-15.
6. Bolton RE, Vincent JH, Jones AD, et al. An overload hypoth-
importantly in fit from each other. Perhaps this is not esis for pulmonary clearance of UICC amosite fibers inhaled
surprising given the very small size of the study pop- by rats. Br J Ind Med 1983;40:264-72.
ulation. However, the deposition models considered 7. Morrow PE. Possible mechanisms to explain dust overloading
of the lungs. Fund Appl Toxicol 1989;10:369-84.
here may be fundamentally difficult to distinguish 8. Vostal JJ, Schrenk RM, Lee PS, et al. Deposition and clear-
from one another as they are all essentially cumulative ance of diesel particles from the lung. In: Toxicological effects
processes with duration of exposure a common and of emissions from diesel engines. Lewtas J, ed. New York,
NY: Elsevier, 1982:143-58.
important contributor to all three. Recent simulation 9 Wolff RK, Griffith WC, Cuddihy RG, et al Modeling accu-
studies suggest that 1) certain underlying biologic mulations of particles in lung during chronic inhalation expo-
mechanisms may be more readily distinguished with sures that lead to impaired clearance. Health Phys 1989;
57(suppl l):61-8.
epidemiologic techniques than others; and 2) cumula- 10. Vincent JH, Donaldson K. A dosimetric approach for relating
tive processes that are well correlated with duration of the biological response of the lung to the accumulation of
exposure are difficult, in particular, to distinguish with inhaled mineral dust. Br J Ind Med 1990;47:302-7.
11. Riggs D. Mathematical approach to physiological problems.
these methods (15). Baltimore, MD- Williams & Wilkins, 1963.
We combined statistical and biologic theory to iden- 12. Yu CP, Chen YK, Morrow PE. An analysis of alveolar mac-
tify better dosimetric models. Although the evidence is rophage mobility kinetics at dust overloading of the lungs.
Fundam Appl Toxicol 1989;13:452-9.
not strong, these toxicokinetic models may improve 13. Bohning DE, Atkins HL, Cohn SH. Long-term particle clear-
upon cumulative exposure as an estimate of risk. This ance in man: normal and impaired. Ann Occup Hygiene
study used cross-sectional data only. It would be in- 1982;26:259-71.
14. Salvan A. Is the model with the largest exposure effect the
teresting to test out the fit of a dose metric on longi- best model? Epidemiology (in press).
tudinal data and to experiment with nonlinear dose- 15. Kriebel D. Cumulative exposure: conditions for its use as a
response relations. In certain cases, use of a dose robust estimator of exposure in occupational epidemiology.
Presented at the Conference on Retrospective Assessment of
metric instead of cumulative exposure may predict Occupational Exposures in Epidemiology, Lyon, France,
individual and population risks more accurately and April 1994.

Am J Epidemiol Vol. 141, No. 7, 1995