| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Ann. occup. Hyg., Vol. 46, No. 4, pp. 431-432, 2002
© 2002 British Occupational Hygiene Society
Published by Oxford University Press
Letters to the Editor |
Reply
Division of Environmental and Occupational Health, School of Public Health, University of Minnesota, Minneapolis, MN 55455, USA
Burstyn and Kromhout have provided a thoughtful and substantial critique of my paper (Ramachandran, 2001). They raise several issues that are relevant in a general sense to subjective exposure assessment and several that are specific to my paper. I hope my response clarifies these issues to some extent.
I agree with Burstyn and Kromhout that industrial hygienists need to be calibrated before subjectively assessing exposures. However, the evidence for this is mixed and there is a limited amount of validatory data on this issue (Kromhout et al., 1987; Hawkins and Evans, 1989; Post et al., 1991; Hornung et al., 1994). Hawkins and Evans (1989) have shown that industrial hygienists gave ‘remarkably accurate subjective estimates of the mean exposure without historical data and even better estimates with the data’. Thus, providing assessors with reference exposure levels may help to minimize bias. Kromhout et al. (1987) showed that subjective qualitative exposure classification (from no exposure to high exposure) could be performed reliably by hygienists within a given work situation, but the relative classification could not be extended from one work situation to the next. Post et al. (1991) found that the success of exposure estimation depended on the type of chemical and its usage and that in some instances providing exposure data for calibration purposes did not improve the performance of hygienists in ranking exposures. In all the above-mentioned studies, the hygienists directly estimated exposures after reviewing relevant information. The procedure used in Ramachandran (2001) is different in that it uses a more structured elicitation where the expert assessors provide only the inputs (e.g. ventilation flow rate) to an exposure model. This is somewhat similar to the structured subjective assessment described by Cherrie and Schneider (1999). As a further precaution, the experts are provided with anchoring information (e.g. ventilation flow rate in a specific year) that essentially calibrates them to the specific situation. However, such calibrating information is often available only from more recent periods in the plant’s history. The beneficial effects of such information is diluted as the experts are asked to assess determinants of exposure farther and farther back in time. This is seen in figure 5, where the two experts differ by a factor of 
10 in the earliest years (1958–60), but their differences are not statistically significant in the later years.
One limitation of Ramachandran (2001), acknowledged on p. 656, is that only two experts were used, thus providing little opportunity for in-depth analysis of expert uncertainty. This is clearly an area that warrants further study. However, I disagree with the assertion that the combined error bars of the expert-based distributions would cover similar ranges of values to the measured distributions. As figure 8 shows, the expert judgment distributions (even if we ‘combine’ them) are narrower than the measurement distributions.
Burstyn and Kromhout also assert that the experts did not take into account within- and between-worker variability while it is reflected in the distribution of personal exposures. This is a valid and useful criticism in that the expert judgment distributions do not allow parsing of the total variability into within- and between-worker components. There are two responses to this criticism. First, the experts provide probability distributions of the input parameters to the exposure model (equations 5 and 6). The variability caused by uncertainty in the input parameters to equation 5 (the general ventilation model) may be considered to be within-worker variability (since it arises from general environmental conditions). In contrast, the variability caused by uncertainty in input parameters to equation 6 may lead to both between- and within-worker variability, and the components cannot be separated. Thus, while the total variability (between- and within-worker) is represented in the expert distribution, no parsing of its components is possible. This should certainly be a topic of interest to the exposure modeling community. Secondly, it should be pointed out that the same is also true for the exposure data in this study. The personal exposure measurements and the konimeter and HiVol measurements that were converted into personal exposure measurements did not have enough information for parsing out the within- and between-worker variability. Exposure data were available only as summary statistics for groups of workers and individual data were not available. This situation is common in occupational epidemiological studies.
I agree with Burstyn and Kromhout that we are far from having accurate generic exposure models for all workplaces, and do not claim otherwise in my paper. Certainly, the general ventilation model used in my paper might be problematical in other situations. However, the model choice in the procedure described in Ramachandran (2001) is left to the experts. Indeed, even for the converter aisle the model may not be the most appropriate and so a model uncertainty factor (K) was included to reflect this. I also know from experience that obtaining information on exposure determinants may, at times, be more difficult than obtaining exposure measurements. However, if exposure data are sparse, there is little choice but to seek out such information. It is unclear why Burstyn and Kromhout assert that I ‘virtually forced the uncertainty in determinants of exposures to be low’. The methodology followed does exactly the opposite. The uncertainty estimates for the exposure determinants are provided by the expert assessors. As the description on p. 654 makes clear, the judgments of the experts were constantly challenged by the interviewer to prevent underestimation of uncertainty in the judgments.
Burstyn and Kromhout assert that it is not necessary to accurately quantify variability since only a group-level exposure assessment is being proposed, for which a Berkson error model will suffice. The assumption (which they also state) is that valid measures of central tendency are available. Of course, the validity has to be established through a sensitivity analysis for which estimates of variability are necessary. Additionally, as Cherrie and Schneider (1999) point out, for a case–control study it is essential to estimate exposure by job title for individuals. The next step after the procedure described in Ramachandran (2001) would be estimation of individual exposure estimates based on employment history. In such a situation a mixed Berkson and classical error model (Zeger et al., 2000) is appropriate, and then a reduction in the variability of exposure estimates becomes a useful exercise.
Lastly, regarding the usefulness of the exposure measurements, I do not intend to diminish their importance, and certainly the Bayes formalism needs both measurements and expert input priors. At the same time, I was by no means as data rich as Burstyn and Kromhout suggest. The confusion regarding this is completely my fault. For the 19 yr period there were 251 measurements and not 828. Table 3, for instance, gives the impression that there were 118 measurements of workroom inhalable dust each year from 1976 to 1979 and 50 personal inhalable measurements each year for the same period. Actually, there were 118 and 50 measurements of each type over the entire 4 yr period, not each year. Even this information was available only in summary form. For instance, the 50 individual personal inhalable measurements were not available; instead, only the summary description (the average and 95% confidence interval) was available. In this situation the more detailed analysis that Burstyn and Kromhout call for (and that is indeed desirable) is not feasible. They also cite several references which show that, even in the absence of historical measurements, current measurements are useful for retrospective exposure assessment. While this may have been the case for the studies cited by them, it is difficult to generalize this to all situations. In addition, for measurements made more than a few decades ago there is little contextual information available and so systematic biases are impossible to estimate. The method proposed by Shlyakhter (1994) and described at length in Ramachandran (2001) is the only one, to my knowledge, that confronts this problem quantitatively. The approach proposes widening the error bars by an empirical factor of 2 and is based on Shlyakhter’s analysis of several historical environmental data sets. However, even if this factor of 2 is not applied, the variability in the measurement data is so much greater than the variability in the expert priors that it does not make any significant difference to the posteriors.
Burstyn and Kromhout are entirely correct that when quantitative measurements are available we should use them instead of expert judgment. The Bayesian methodology described in my paper addresses situations where such measurements are sparse. Then the only practicable approach is to supplement this objective information with subjective judgments using information about the work environment and practices to assess exposure levels and their associated uncertainties.
REFERENCES
Cherrie JW, Schneider T. (1999) Validation of a new method for structured subjective assessment of past concentrations. Ann Occup Hyg; 43: 235–45.
Hawkins NC, Evans JS. (1989) Subjective estimation of toluene exposures: a calibration study of industrial hygienists. Appl Ind Hyg J; 4: 61–8.
Hornung RW, Greife AL, Stayner LT et al. (1994) Statistical model for prediction of retrospective exposure to ethylene oxide in an occupational study. Am J Ind Med; 25: 825–36.
Kromhout H, Oostendorp Y, Heederik D, Boleij JSM. (1987) Agreement between qualitative exposure estimates and quantitative exposure measurements. Am J Ind Med; 12: 551–62.
Post W, Kromhout H, Heederik D, Noy D, Duijzentkunst RS. (1991) Semiquantitative estimates of exposure to methylene chloride and styrene: the influence of quantitative exposure data. Appl Occup Environ Hyg; 6: 197–204.
Ramachandran G. (2001) Retrospective exposure assessment using Bayesian methods. Ann Occup Hyg; 45: 651–67.
Shlyakhter A. (1994) An improved framework for uncertainty analysis: accounting for unsuspected errors. Risk Anal; 14: 441–7.
Zeger SL, Thomas D, Dominici F et al. (2000) Exposure measurement error in time-series studies of air pollution: concepts and consequences. Environ Health Perspect; 108: 419–26.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  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]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||