Article Text

O15-1 Bayesian analysis of silica exposure and lung cancer, incorporating prior information from animal studies and a model for measurement error
  1. Scott Bartell1,
  2. Ghassan Hamra2,
  3. Kyle Steenland3
  1. 1UC Irvine, Irvine, USA
  2. 2Drexel U, Philadelphia, USA
  3. 3Emory U, Atlanta, USA


Introduction Bayesian methods can be used to combine human and animal data for exposure-response analyses. We apply a framework recommended in a recent National Academies report on the US EPA risk assessment process, extending the method to adjust for exposure measurement error.

Methods We used epidemiological data from a pooled mortality analysis of silica and lung cancer (n = 65,980), and animal data from pooled analysis of chronic silica inhalation studies of rats. Bayesian analyses were conducted with either diffuse or informative priors based on the animal data, several different cross-species extrapolation factors, and human exposure measurement error corrections in the absence of a gold standard, assuming Berkson-type error that increased with increasing exposure (which can induce bias in Cox models). Analyses were conducted using untransformed and log transformed cumulative exposure.

Results With 3-fold or 10-fold uncertainty in the cross-species extrapolation factor, the animal prior had little effect on results for pooled epidemiological analyses and only modest effects for some individual epidemiological studies. In contrast, assuming 1-fold uncertainty produced markedly different results for both pooled and individual epidemiological studies. Measurement error correction had little effect in pooled analyses using log exposure. Using untransformed exposure, measurement error correction caused a 5% decrease in the exposure-response coefficient for the pooled analysis and more marked changes for some individual studies. Measurement error correction had little effect on the exposure-response coefficient when exposure was log transformed or when the sample size was large.

Conclusions The Bayesian framework is a principled method for combining human and animal data, with posterior effect estimates reflecting a weighted average of the animal and human results. It may be particularly useful when human data are sparse.

Statistics from

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.