Small area estimation of incidence of cancer around a known source of exposure with fine resolution data

Objectives-To describe the small area system developed in Finland. To illus trate the use of the system with analyses of incidence of lung cancer aroun d an asbestos mine. To compare the performance of different spatial statist ical models when applied to sparse data. Methods-In the small area system, cancer and population data are available by sex, age, and socioeconomic status in adjacent "pixels", squares of size 0.5 km x 0.5 km. The study area was partitioned into sub-areas based on es timated exposure. The original data at the pixel level were used in a spati al random field model. For comparison, standardised incidence ratios were e stimated, and full bayesian and empirical bayesian models were fitted to ag gregated data. Incidence of lung cancer around a former asbestos mine was u sed as an illustration. Results-The spatial random field model, which has been used in former small area studies, did not converge with present fine resolution data. The numb er of neighbouring pixels used in smoothing had to be enlarged, and informa tive distributions for hyperparameters were used to stabilise the unobserve d random field. The ordered spatial random field model gave lower estimates than the Poisson model. When one of the three effects of area were fixed, the model gave similar estimates with a narrower interval than the Poisson model. Conclusions-The use of fine resolution data and socioeconomic status as a m eans of controlling for confounding related to lifestyle is useful when est imating risk of cancer around point sources. However, better statistical me thods are needed for spatial modelling of fine resolution data.