Introduction Bias occurs in particular when estimates are based on sparse or inadequate data. We have estimated the burden of cancer in Great Britain attributable to occupation using an attributable fraction (AF) methodology, and present an adaptation of Greenland's1 Monte-Carlo sensitivity analysis (MCSA) to account for bias uncertainty.

Methods Sources of bias in burden estimation include using Levin's estimator with adjusted RR, unknown cancer latency, unknown proportions exposed and inadequate estimates of employment turnover. Each source of bias operates on a component of the AF estimator, which is represented by a factor for which a prior distribution is estimated from independent sources. Monte-Carlo repeated sampling from these distributions is then used, recalculating the AF each time.

Results Results are presented graphically for a hierarchy of bias sources that contribute to an overall credibility interval for the AF. For sinonasal cancer and wood dust the intervals for bias due to the variables contributing to the proportions exposed are narrower than the interval for RR random error only, and bias from incorrect use of Levin's estimator makes the least contribution.

Conclusion The method presented illustrates the use of credibility intervals to indicate important sources of uncertainty and facilitates identification of data gaps and future research needs.