general provide an extremely thorough investigation of issues that are important for radon risk assessment. Careful attention is given to investigating the homogeneity of effects across cohorts and to determining which models provide the best description of the data.
Nevertheless, additional analyses might be desirable. For example, investigation of age at exposure was limited to consideration of age at first exposure, possibly because the use of "windows" for both age at exposure and time since exposure would have been too cumbersome with the use of Poisson regression. Analyses based on age at exposure would be desirable especially if substantial numbers of workers with very early ages at start of exposure continued to be exposed for many years.
The NCI analyses of exposure rate (or exposure duration) were limited primarily to consideration of average exposure rate and did not take account of detailed exposure histories. However, efforts by Tomásek and others (1994a) and by Thomas and others (1994) to use those histories more precisely yielded results that were similar to those obtained in the NCI analyses and based on the average exposure rate.
The NCI analyses did not take account of errors in the exposure measurements (including estimates of both exposure rate and duration, both of which are needed to estimate exposure). Those errors are generally thought to be largest for early periods of mine operations, when exposure rates were highest. In addition, there is considerable variation in the quality of exposure measurements across cohorts. In general, random error in exposure measurements tends to bias overall risk coefficients downward and might also distort the shape of the exposure-response curve. Because estimates of exposure in early years of mine operations are often subject to greater errors than estimates in more recent years, estimates of the modifying effects of exposure rate and possibly of other time-related factors might be exaggerated. Statistical methods are available for adjusting for exposure-measurement errors but tend to be difficult to use. Furthermore, the application of those methods requires that the error structure be specified, an extremely difficult undertaking, given the complexity of the structure and the lack of adequate data for quantifying many sources of error.
It is difficult to investigate the separate modifying effects of the variables evaluated in the NCI analyses, and this could have important implications for extrapolating to values outside the range of the data. A general difficulty is that we do not have adequate knowledge of the biologic rationale of patterns of risk associated with various factors. The lack of adequate biologic understanding has necessitated the descriptive approach taken in the NCI analyses and in analyses conducted to develop the BEIR VI risk models. However, despite its limitations, the descriptive approach provides extremely valuable information and very likely must serve, with appropriate caution, as the basis for developing risk estimates for both occupational and residential exposure.