Ann. occup. Hyg., Vol. 46, No. 5, pp. 479-487, 2002
© 2002 British Occupational Hygiene Society
Published by Oxford University Press
Combining Expert Ratings and Exposure Measurements: A Random Effect Paradigm
1 INRS, Department of Epidemiology, BP 23, 54501 Vandoeuvre Cedex; 2 University of Lyon, Lyon; 3 USINOR, Paris, France
Received 30 March 2001; in final form 8 February 2002
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
The aim of this paper is to present a paradigm for combining ordinal expert ratings with exposure measurements while accounting for a between-worker effect when estimating exposure group (EG)-specific means for epidemiological purposes. Expert judgement is used to classify the EGs into a limited number of exposure levels independently of the exposure measurements. The mean exposure of each EG is considered to be a random deviate from a central exposure rating-specific value. Combining this approach with the standard between-worker random effect model, we obtain a nested two-way model. Using Gibbs sampling, we can fit such models incorporating prior information on components of variance and modelling options to the rating-specific means. An approximate formula is presented estimating the mean exposure of each EG as a function of the geometric mean of the measurements in this EG, between and within EG standard deviations and the overall geometric mean, thus borrowing information from other EGs. We apply this paradigm to an actual data set of dust exposure measurements in a steel producing factory. Some EG-specific means are quite different from the estimates including the ratings. Rating-specific means could be estimated under different hypotheses. It is argued that when setting up an expert rating of exposures it is best done independently of existing exposure measurements. The present model is then a convenient framework in which to combine the two sources of information.
Keywords: expert information; exposure measurements; statistical methods
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
A large literature now exists on the use of quantitative exposure assessment in industry-based occupational epidemiology and the use of exposure measurements in this context (Kaupinen, 1994; Seixas and Checkoway, 1995; Heederik and Attfield, 2000). Basically, exposure measurements are either used on an individual basis or after grouping the jobs and/or tasks into exposure groups (EGs). An exposure assessment based on the individual implies that enough exposure measurements are available for all subjects to characterize both present exposure and, in chronic disease epidemiology, past exposure. It is thus rarely applicable, especially insofar as usually only relatively recent measurements are available and it is impossible to measure past exposure. It is therefore usual practice to base exposure assessment on grouping strategies and to base the exposure assignment only on the exposure group to which the individual belongs. A further reason is that attenuation of the dose–response relationship due to lack of precision in the dose assessment is lower in group-based strategies (Kromhout et al., 1994). When a grouping into EGs has been performed, the individual is usually assigned the quantitative exposure corresponding to the mean of the exposure measurements in this EG. This method can be extended by modelling the exposure measurements according to exposure determinants. This approach has been used by, among others, Preller et al. (1995) and Burstyn et al. (2000). In the first case, measurements were taken among Dutch farmers for exposure assessment to endotoxins in the framework of an epidemiological study on respiratory health. As the exposure measurements were done for this study, presupposed exposure determinants could be taken into account while planning the measurements. In the second case, exposure measurements of bitumen and polycyclic aromatic hydrocarbons were pre-existing and were modelled according to the available determinants.
Another method consists of assigning to each EG an essentially dimensionless semi-quantitative rating through the development of an industry-specific job–exposure matrix (JEM). The latter can be based on formal procedures in which the effect of exposure determinants is subjectively estimated and combined in order to obtain quantitative task-specific exposure scores (Cherrie and Schneider, 1999) or on subjective expert opinions who ordinally rate the exposure groups (Moulin et al., 1997, 1998). The former approach presupposes a detailed description of every task and, to be used in the actual estimation of individual cumulative exposure, the percentage of time spent on each task for each member of the different EGs. The latter situation occurs in situations in which no detailed information on tasks and exposure determinants exists or could be abstracted, but in which an attempt at quantifying the exposure was done by obtaining ordinal ratings or exposure ranks from a group of experts. This is the situation we concentrate on in this paper. We want to combine the information coming from the exposure rating with existing exposure measurements. The aim is 2-fold: first, to obtain the best summary for each exposure rating when no measurements exist; secondly, to combine these with the information from the exposure rating for each EG in which measurements have been collected.
This combination works under the assumption that the ratings were obtained independently of the exposure measurements, so that the quality of the ratings can be assessed, and that for each rating a number of EGs were measured.
In the first section we present a data set in which these two conditions were fulfilled. In the second section we present our model and our approach to fitting this model. We go on by presenting an approximate formula which allows compromise exposure estimates between ratings and measurements to be obtained. Finally, we apply the method to our data set and conclude with a discussion on the validity of such models.
|
AN EXAMPLE |
|---|
|
TOPABSTRACTINTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
Dust measurements and expert ratings in a steel producing factory
A cohort mortality study is currently being carried out in a large steel producing factory integrating all steps from coke production and iron ore sintering to hot rolling. In order to be able to obtain semi-quantitative exposure estimates for substances possibly involved in this production a factory-specific JEM was set up by a group of seven experts with expertise in industrial hygiene, occupational medicine and epidemiology. Similar exposure groups (SEG) were identified in this factory for which semi-quantitative exposure indices were obtained for a number of substances, including total inhalable dust. A code for reliability was added to each code. A reliability coded 3 meant consensus among the experts, reliability 2 signalled some doubt as to the code and 1 was coded when the rating was considered guesswork or in the case of disagreement between experts. The methodology for obtaining a consensus is described in detail for a similar study on the risks within the hard metal industry (Moulin et al., 1997). These estimates were obtained blind to the existing exposure measurements. Independently of this JEM, all exposure measurements carried out in this factory by industrial hygienists were collected. We report here on a subfile of this database consisting of all individual 4–8 h measurements of atmospheric samples at a flow rate of 1–2 l/min. A total of 412 such measurements obtained between 1980 and 2000 could be assessed, of which six were below the known detection limit. Fifty SEGs were thus characterized. These measurements were obtained for 318 subjects and for 47 subjects some within-subject replication was available. The corresponding ratings (JEM) varied between 1 (lowest exposure ) and 5 (highest exposure). The number of different categories was defined by the experts themselves and corresponded to the finest distinction they thought they could make on the basis of their knowledge.
|
JOINT MODELLING OF EXPERT RATINGS AND EXPOSURE MEASUREMENTS |
|---|
|
TOPABSTRACTINTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
The basic model
The now classical statistical model of exposure measurements in EGs acknowledging the between-worker effect is summarized in the equation below (see for instance Boleij et al., 1995, pp. 106–7).
A measurement Xij obtained in an EG for the jth worker on the ith day is modelled as follows :
Yij = ln(Xij) = µ + ßj + 
ij (1)
where ßj is the random deviation of the jth worker’s true exposure from the mean of the EG and ij the random deviation on the ith day from the jth worker’s true exposure.
It is further assumed that the two random deviations are normally distributed. More specifically
ßj 
N(0, BW
Y2) (1.1)
where BWY2 is the between-worker variance
N(0, WWY2) (1.2)
where WWY2 is the within-worker variance. These two random variables are further assumed to be statistically independent.
The formulation we propose in order to combine the information from the different EGs with the same exposure rating is to extend this model by considering that the true mean exposure of an EG is a random deviation of the mean exposure of all potential EGs with the same rating.
Equation (1) then becomes, indexing the EGs with k,
Yijk = ln(Xijk) = µk + ßjk + ijk (2)
where ßjk stands for the jth worker within the kth EG with
µk N(
r, BXY2) (2.1)
where BXY2 is the between-exposure group variance and r is the mean of the exposure groups with rating r.
An alternative way to write the model would be to specify the deviate 
kr = µk – r
Yijk = ln(Xijk) = r + kr + ßjk + ijk (3)
k N(0, BXY2) (3.1)
As before, the three random components, i.e. the EG within its rating, the worker within his EG and the day-to-day variability, are assumed to be independent. This is the two-way nested random effect model. It is nested as each worker belongs to only a single EG.
After fitting such a model, we obtain for each rating r an estimate of the typical geometric mean (GM) [i.e. exp(r)] of an EG with this rating in the absence of any actual measurement and, on the other side, for each EG with actual measurements an estimate of the GM [i.e. exp(µk)]. The latter, as will be shown later, is a compromise between the GM of the measurements and information from the other EGs with the same rating.
Modelling options
Incorporation of prior information
Much is known about typical geometric standard deviations (GSDs) for exposure measurements within EGs and it therefore appears reasonable to incorporate this prior knowledge in the data description, thus following a Bayesian approach. This corresponds either to WWY2, the within-worker, day-to-day variance, or if the between-worker variance is ignored, to the total within-EG variance WXk2 = WWk2 + BWk2. For instance, a European Working Group (CEN TC137 1999) made the following statement ‘Depending on the activity or branch concerned, the geometric standard deviations are known to vary between 2.0 and 2.7’; other authors (Seixas and Checkoway, 1995; Kromhout et al., 1996) cite somewhat lower GSDs, which correspond to our experience. This depends of course on how homogenous the EG is. If we characterize a workplace which involves very different tasks, our a priori idea of the GSD might be high. On the other hand, if we characterize a well-defined and relatively stable task, our a priori estimation of the GSD is lower.
A mathematically convenient statistical distribution a priori is to assume that the log variance has an inverse 
distribution with parameters a and b. For instance, a = 2 and b = 0.7 would express the prior belief that the GSD is in the interval 1.4–5.3 with 95% probability with a prior median of 1.9.
There is less quantitative literature available on the between-worker variance if one excepts the paper by Kromhout et al. (1994), in which the authors give the range of observed within- and between-worker GSDs. Following this work, non too informative prior information can be written as an inverse distribution with parameters 1 and 0.05, which would express the prior belief that the GSD is in the interval 1.12–4.04 with 95% probability with a prior median of 1.3.
Modelling the rating-specific mean
Several approaches as to how to model the rating-specific means r can be proposed.
1. The simplest model, and the one which is closest to the measurements, is to assume that each rating is independent of the others. The fit of this model would be nearly equivalent to the fit of a series of models, one for each rating.
2. The other extreme is to consider that the r values increase linearly with r, which is equivalent to stating that the ratios between the GMs of contiguous ratings are constant. Mathematically this corresponds to r = 0 + 
r. If our aim is to test whether the (mean log) exposure increases with the expert rating, the statistical significance of the slope could be considered as such a test. Such a linear increase is very hypothetical, however.
3. The following model is intermediate between the first model, which does not consider any relation between consecutive ratings, and the second, which fixes the parametric form of this relation. The model we consider here is obtained from the first one by adding an order constraint, i.e. we fit the model assuming that r + 1 > r for all values of r. Such a model is meaningful if we consider that the information obtained from the experts is more reliable than the information from the measurements. This might be the case when the measurements were taken for unknown reasons or when the circumstances under which the measurements were taken were not adequately documented. The present model would then yield the best quantitative estimate of the rating-specific mean under the hypothesis that the experts were ‘right’ in the sense that the rating-specific means are indeed increasing.
Fitting the model
The models we presented in the first section could, in their simplest form, be estimated by standard methods for random effects (Searle et al., 1992). In this framework the EG-specific estimated means (µk) would be called best linear unbiased predictors (BLUPs), and it has been shown that in this context they are ‘better’ (Robinson, 1991) estimates than the raw means of the log measurements. However, at least some replicates would be necessary at each random level, i.e. every worker in every EG must be measured several times and more than one worker must be measured within each EG. Moreover, the fitted model would make one further assumption, which is unlikely to be realistic in practice. This assumption is that the within- and between-worker variances are constant across the EGs. Also, these methods do not allow for a feature which is quite common in exposure measurements, namely the fact that a non-negligible fraction of measurements are below the detection or quantification limit. Finally, traditional methods and software neither allow prior information on the variance components to be built in nor modelling of the values. However, it is possible to fit such models using Monte Carlo Markov chain methods, like so-called Gibbs sampling. This consists instead of maximizing the likelihood, in sampling repeatedly from the a posteriori distribution, which is the likelihood times the a priori distribution. This technique is presented in detail for the problem of inference with censored log normal exposure measurements in Wild et al. (1996). We fitted it using the BUGS (Bayesian inference using Gibbs sampling) software (Gilks et al., 1996), which is a general purpose as yet freely available program to fit such models. The results of such software are large number of samples (typically at least 10 000) for all parameters (µk, r and the variance parameters), from which means, standard errors, medians and confidence intervals can be derived. It is further possible to include an EG without measurements within each rating, which would yield not only a typical mean of an EG with a given rating (which will be equal to r; see below) but without measurements, but also an associated confidence interval. Note that if we are interested in arithmetic means, which are more useful in epidemiological studies, they can also be obtained with their confidence intervals without any further computation.
An approximate formula
In this section we propose an approximate formula for the simplified model in which we ignore the between-worker effect. This would be appropriate if only one exposure measurement per subject was available, although the between- and within-worker variances would then not be distinguishable. This formula is not shown as a replacement for a full statistical analysis, but as a way to obtain an intuitive idea as to what influences the estimates. It should therefore not be seen as a means to compute these estimates.
For a given EG k with expert rating r, its log geometric mean µk can be estimated from formula (4) (see Searle et al., 1992, section 9.3; Gelman et al., 1995, Ch. 3) under the assumption that we know 
k and r, as a weighted average of the mean 
of the nk log transformed measurements (this would be the natural estimator ignoring the expert rating) and r the mean of all potential µk values sharing the same expert rating r.
(4)
where
k = WXk2/BXr2 (5)
and WXk2 = WWk2 + BWk2 is the total variance of the log measurements in EG k (WX stands for within exposure group), which is the sum of the within- and between-worker variance BXr2 denoting the variance of all potential µk values sharing the same expert rating r.
In this formula one can see that if for EG k the number nk of exposure measurements increases, the rating information will, as it should, have less and less influence on estimation of the EG-specific mean. k, the ratio of the within- and between-EG variance, can be interpreted as the number of measurements equivalent to the information from the expert rating. This information becomes larger relative to the information from the measurements when the within-EG variance of this EG becomes larger (in this case the measurements carry less information). This may be due in practice to EGs which should perhaps be split into different EGs, possibly with different ratings. In contrast, the information from the rating becomes smaller when the between-EG variance becomes larger, i.e. when EGs sharing a rating are more dispersed. This might happen when the number of possible ratings is not large enough so that very different EGs are given the same rating. Unfortunately, this formula cannot be mathematically extended to account for the between-worker variance. However, even in its present form it is not operational, as it presupposes a knowledge of the variance parameters.
A result which can be derived from equation (3) and which is still true for a complete fit of the model, including within-worker effects, is that when no measurement exists (nk = 0) the best estimate of the EG-specific mean µk is r.
|
APPLICATION |
|---|
|
TOPABSTRACTINTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
Figures 1–3 show graphical displays of the results of fitting the three models to our example. Table 1 presents the corresponding numerical results for selected EGs and predicted GMs for EGs with no measurements. The results are all based on a full statistical analysis based on Gibbs sampling, carried out with the BUGS software and incorporating all sources of variation mentioned in the section presenting the model. The approximate formula of the preceding section does not allow for between-worker variance and does not allow the flexible modelling options shown in Figs 2 and 3 and therefore was not used. The triangles show the GMs of the measurements exp(
) of the EGs with which the estimated GMs exp(µk) (according to the different models) represented by the square dots can be compared. Furthermore, the rating-specific means r are represented by crosses with their (Bayesian) confidence intervals.
|
|
|
The striking feature of the first model (Fig. 1) is the increasing trend in the mean over ratings 1–4 and the sharp drop-off in rating 5, which was totally unpredicted. The largest difference between the mean log measurements and the predicted mean µk occurred in EG 4. This is the EG with the highest empirical geometric mean [exp
= 16.8]; the corresponding estimate in model 1 was exp(µ4) = 7.4. Conversely, EG 14, the lowest point in rating 2, whose empirical GM was 0.26, was estimated at 0.63. On a log scale this change is even larger than for EG 4. On the other hand, the GM of EG 22, the highest point in rating 4, despite being markedly larger than the rating-specific mean 4, is close to the empirical GM of the measurements, as it is based on a large number of measurements. The estimated log mean µk of the EGs, for which the empirical mean of log measurements 
is close to the corresponding r, are of course close to both r and 
. The within-worker GSDs vary from 1.59 to 2.95 and the between-worker GSDs vary from 1.23 to 1.49, with the exception of two very heterogeneous EGs in which the GSDs were 3.29 and 3.83.
|
In Figure 2 we show the results of the fit of a linear trend (model 2). Despite the strong non-linearity (the drop-off for rating 5), this trend fits reasonably well and, seen as a test, it is highly significant: the estimated slope is 0.278 with a 95% Bayesian confidence interval (0.096–0.460) and a 0.092 standard error.
Figure 3 shows the fit of the order-constrained model 3. As built in the model, the rating-specific mean increases, but slower than the linear trend. It is noticeable that the estimates based on this model are within the confidence intervals of the first model. The result of this order restriction was mainly a lower estimate for rating 4 and to a lesser degree for rating 3, whereas the estimate for rating 5 was of course higher than in the first model. It can be seen that the model-based estimates of µk seem closer to the empirical GM 
than in model 1.
All the influences predicted by the approximate formula can be seen from the results given in Table 1. For instance, EG 17 (with rating 3) has a very large GSD and is therefore highly influenced by the rating-specific means. On the other hand, EG 22, characterized by 48 measurements on 13 subjects, is only marginally influenced by the rating.
We fitted these models again, excluding all situations for which the experts had expressed doubt as to their classifications. Four EGs, three of which had been rated 2 and one of which had been rated 3, were thus excluded. This led to a slight increase in the estimates for rating 2, as the two lowest EGs were thus excluded.
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
Even in industry-based occupational epidemiological investigations there are often relatively few exposure measurements and when such measurements exist, they were often taken for compliance assessment rather than to characterize typical situations. In these contexts existing exposure measurements cannot be the sole basis for exposure assessment for epidemiological purposes and the different methods for quantitative exposure assessment presented by Heederik and Attfield (2000) cannot always be applied. Exposure measurements are therefore sometimes replaced by ordinal expert ratings for which one substitutes in the analysis of the data some assumed central values based on existing exposure measurements. Our proposed method is a formalization of this procedure. It is, however, different in two aspects from this ad hoc method.
In estimation of the ratings means the different EGs are weighted according to the precision with which the EG means are known; this precision depends not only on the number of measurements but also on their dispersion.
When measurements exist for a given EG, the estimated mean for this EG is estimated from the measurements combined with information obtained from the measurements of all the other EGs with the same expert rating. The relative weight of the two types of information depends both on the quality of the rating as measured by the between-EG standard deviation and on the homogeneity of the exposure in the current EG as measured by its within-EG standard deviation.
A sine qua non condition of our model is that the expert ratings were obtained independent of the measurements. Thus the central point to be discussed is: does a procedure which does not make use of exposure measurements in the expert ratings make sense at all? And if yes, when?
A first point we want to emphasize is that no information is lost when ignoring measurements in the rating of the EGs. The expert ratings are only used to obtain rough ranks, which are calibrated by the existing measurements. The estimates of the EG-specific means are dominated by the measurements as soon there are enough of them and borrows information from the other EGs if not. This has the consequence that when the experts classifications ‘disagree’ with the exposure measurements, the resulting estimate is still mainly based on the measurements, although corrected by the knowledge from the experts. The rating-specific mean, on the other hand, is more or less a weighted mean of the EG-specific means and uses all the information.
In contrast, had we used the results of the measurements to rate the EGs, the between-EG variance within the ratings would be considerably underestimated, so that its influence when combining it with future exposure measurements would be exaggerated. In general terms, it is better to have two independent sources of information: measurements on the one hand and expert ratings on the other, so that when both coexist for some EGs, one can validate the other. In some way the expert rating is a synthesis, based on the raters’ experience, of all the determinants of the exposure and, seen like that, it is clear that the measurements themselves cannot be considered as determinants of the exposure and should therefore not interfere with the rater’s rating.
This brings us to the second question as to when we should apply an expert-based approach. It is clear that if all the objective determinants of exposure are documented for each measurement, it is better left to a regression model to estimate the effect of each determinant (Burstyn et al., 2000) than to rely on a subjective synthesis by whatever experienced experts. However, in an epidemiological context one still must quantify each EG by its ‘typical’ determinants to be able to use the results of this regression in an historical exposure assessment for subjects for which work histories are the only source of information. But a between-EG variance exists, as well as the between-worker variance, and should be taken into account in the statistical exposure modelling when the exposure determinants have been obtained for EGs and not for each individual worker. For instance, even if the industrial process and all other determinants are quite similar, the mean exposure in the different EGs can be markedly different. Ignoring this between-EG variance has the same effect as ignoring the between-worker variance, i.e. biasing the results towards the EG with the largest number of measurements.
However, returning to the original question as to when an approach based on expert rating is appropriate, the correct answer is probably: as soon as a modelling approach is not feasible. This is the case when the exposure determinants are not sufficiently characterized or if the measurements did not cover all possible determinants, for instance if a past industrial process had never been measured but could be subjectively assessed by experts.
The parametric assumptions underlying the model are another critical point. While the log normality of exposure measurements within exposure groups is a standard assumption, we added the further assumption that the geometric means of the EGs are log normally distributed within the ratings. This is the same hypothesis as that classically made for between-worker effects and is equivalent in some sense to assuming that the rater’s error is multiplicative, which is a reasonable assumption. Except for this rather weak argument, the only argument is probably that it makes the statistical computations feasible. Furthermore, in our example the plots do not show any gross violation of this hypothesis. A formal statistical test does not, however, seem possible, as even under the null hypothesis the empirical distribution of the EG-specific means can be quite varied, depending on the between-worker effects and the numbers of measurements for each worker. The parametric assumption is the price to pay for the possibility of extrapolation of the measurements from the set of EGs for which measurements have been obtained and should be critically examined in each practical situation.
In the example we used to illustrate our paradigm we showed that the two aims we cited at the beginning, namely to estimate the rating-specific mean and to combine the ratings with the measurements, could be accomplished meaningfully. The first of the three models presented provided a description for each rating, the second showed that there was a highly statistically significant increase in the mean with the rating, whereas the third provided estimates under the assumption of increasing GMs. If we want to use these data in order to quantify the exposure in a cohort study for which the ratings were coded, it is probably not reasonable to base a quantitative exposure estimation procedure on the highly parametric second model with the present information available, which points at a less than exponential increase. On the other hand, it is hard to believe that the drop-off in rating 5 is a common feature of all the EGs which were rated 5; the third model would therefore be a natural choice. However, such a decision must take in the circumstances of the study and is beyond the scope of this paper.
An aspect we want to stress is that these rather complex models involving two separate random effects and highly unbalanced and sometimes censored data could only be fitted using modern statistical techniques based on Markov chain Monte Carlo methods. A first introduction to these techniques in the context of industrial hygiene can be found in Wild et al. (1996) and in references therein.
The precision of the estimates obtained by the present methodology also depends on the number of categories defined a priori. If this number is small, the expert assessment can only be coarse, as quite different situations must then be grouped, thus increasing the between-EG variance within each rating. Thus one could conclude that the larger the number of possible ratings, the better one could base them on a semi-quantitative procedure like that developed by Cherrie and Schneider (1999). Increasing the number of categories has, however, two limitations. First, if expert knowledge is scarce or if the overall exposure variance is small (i.e. the EGs are not very different), increasing the number of categories will only add noise. Secondly, if this number is high, the number of EGs characterized by actual measurements must also be high if one wants to be able to assess the validity of the expert assessment without relying on a parametric assumption (as in model 2) or on an a priori assumption of increasing exposure (model 3). Given these considerations, a reasonable strategy would be to set the number of categories high enough to allow the experts to code the finest difference they can possibly make, possibly increasing this number if the EGs are very different. On the other hand, when combining ratings with measurements, it may be necessary to re-group contiguous categories if the number of EGs is too small within each rating and if one wants to assess the agreement between measurements and ratings. No such a posteriori grouping is, however, necessary if one is willing to accept the supplementary hypotheses of either model 2 or 3.
Acknowledgements—We acknowledge the contributions of two anonymous reviewers and of the associate editor Dr van Tongeren who helped to improve the paper significantly. The work of the second author (E.A.S.) was done within the framework of his Ph.D. work, funded by Rhône-Poulenc.
|
FOOTNOTES |
|---|
* Author to whom correspondence should be addressed. Fax: +33-383-50-20-15; e-mail: wild@inrs.fr
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONAN EXAMPLEJOINT MODELLING OF EXPERT...APPLICATIONDISCUSSIONREFERENCES |
|---|
Boleij J, Buringh E, Heederick D, Kromhout H. (1995) Occupational hygiene of chemical and biological agents. Amsterdam: Elsevier.
Burstyn I, Kromhout H. Kaupinen T. Heikkila P. Boffetta P. (2000) Statistical modelling of historical exposure to bitumen and polycyclic aromatic hydrocarbons among paving workers. Ann Occup Hyg; 44: 43–56.
Cherrie JW, Schneider T. (1999) Validation of a new method for structured subjective assessment of past concentrations. Ann Occup Hyg; 43: 235–45.
Gelman A, Carlin JB, Stern HS, Rubin DB. (1995) Bayesian data analysis. London: Chapman & Hall.
Gilks WR, Thomas A, Spiegelhalter DJ. (1996) A language and program for complex Bayesian modelling. Statistician; 7: 247–59.
Heederik D, Attfield M. (2000) Characterization of dust exposure for the study of chronic occupational lung disease—a comparison of different exposure assessment strategies. Am J Epidemiol; 151: 982–90.
Kaupinen T. (1994) Assessment of exposure in occupational epidemiology. Scand J Occup Environ Health; 20 (special issue): 19–29.
Kromhout H, Symanski E, Rappaport SM. (1994) A comprehensive evaluation of within- and between-worker components of occupational exposure to chemical agents. Ann Occup Hyg; 37: 253–70.
Kromhout H, Tielemans E, Preller L, Heederik D. (1996) Estimates of individual dose from current measurements of exposure. Occup Hyg; 3: 23–39.
Moulin J-J, Romazzini S, Lasfargues G et al. (1997) Elaboration d’une matrice emplois-exposition dans l’industrie productrice de métaux durs. Rev Epidemiol Santé Publique; 45: 41–51.[Web of Science][Medline]
Moulin J-J, Wild P, Romazzini S et al. (1998) Lung cancer risk in hard metal workers. Am J Epidemiol; 148: 241–8.
Preller L, Kromhout H, Heederik D, Tielen MJM. (1995) Modelling long-term average exposure in occupational exposure–response anaysis. Scand J Occup Environ Health; 21: 504–12.
Robinson GK. (1991) That BLUP is a good thing—the estimation of random effects. Statist Sci; 6: 15–51.
Searle SR, Casella G, McCulloch CE. (1992) Variance components. New York, NY: John Wiley & Sons.
Seixas NS, Checkoway H. (1995) Exposure assessment in industry specific retrospective occupational epidemiology studies. Occup Environ Med; 52: 625–33.
Wild P, Hordan R, Leplay A, Vincent R. (1996) Confidence Intervals for probabilities of exceeding TLV with censored log-normal data. Environmetrics; 7: 247–59.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  P. Logan, G. Ramachandran, J. Mulhausen, and P. Hewett Occupational Exposure Decisions: Can Limited Data Interpretation Training Help Improve Accuracy? Ann. Hyg., June 1, 2009; 53(4): 311 - 324. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 E Bourgkard, P Wild, B Courcot, M Diss, J Ettlinger, P Goutet, D Hemon, N Marquis, J-M Mur, C Rigal, et al. Lung cancer mortality and iron oxide exposure in a French steel-producing factory Occup. Environ. Med., March 1, 2009; 66(3): 175 - 181. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
G. Ramachandran Toward Better Exposure Assessment Strategies--The New NIOSH Initiative Ann. Hyg., July 1, 2008; 52(5): 297 - 301. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
T. OGDEN Annals of Occupational Hygiene at Volume 50: Many Achievements, a Few Mistakes, and an Interesting Future Ann. Hyg., November 1, 2006; 50(8): 751 - 764. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
A Leclerc Exposure assessment in ergonomic epidemiology: is there something specific to the assessment of biomechanical exposures? Occup. Environ. Med., March 1, 2005; 62(3): 143 - 144. [Full Text] [PDF]
|
|
|
|
|
|
A. BURDORF and M. V. TONGEREN Commentary: Variability in Workplace Exposures and the Design of Efficient Measurement and Control Strategies Ann. Hyg., March 1, 2003; 47(2): 95 - 99. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
E. A. SAULEAU, P. WILD, M. HOURS, A. LEPLAY, and A. BERGERET Comparison of Measurement Strategies for Prospective Occupational Epidemiology Ann. Hyg., March 1, 2003; 47(2): 101 - 110. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
I. BURSTYN and H. KROMHOUT Who Qualifies to be an Expert? Ann. Hyg., January 1, 2003; 47(1): 89 - 89. [Full Text] [PDF]
|
|
|
|
|
|
I. BURSTYN and H. KROMHOUT The Babel of Multicenter Exposure Assessment Ann. Hyg., September 1, 2002; 46(8): 649 - 652. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||