Frank Dunstan University of Wales College of Medicine

1. Sources and effects of bias in investigating links between adverse health outcomes and environmental hazards The results are only as good as the data! Frank Dunstan University of Wales College of Medicine

2. Outline of talk • Why do so many spatial studies fail to find evidence of the effect of risk factors? • Is it because exposure is usually measured inadequately? • Consider a point source of risk and the effects of Distance as a surrogate for exposure Migration • More generally in looking at association between the spatial variation of disease incidence and risk factors, what is the effect of measurement error?

3. How do we measure exposure? • Distance from focus is often used as a surrogate • Algorithms of Stone, Bithell, Tango etc use different models under the alternative hypothesis – but the usual assumption is that risk decreases monotonically with distance. • It is implicit that it is the same in all directions • ‘Circles’ approach similar • Models of transmission of risk make this implausible

4. ‘Circles’ method

5. Example on congenital anomalies • Data on all births in Wales in 15 year period, linked to records of congenital anomalies • Locations obtained using a GIS • Data on landfill sites which changed significantly in the period – 24 in all • Individual data on maternal age, birthweight • Census data on deprivation • Is the opening of a site associated with an increased risk of anomalies?

6. Modelling • Rate varied significantly between hospitals and by year of birth – adjustment for these needed • Risk modelled as function of Age of mother Hospital Gender Year of birth Deprivation • Calculate observed and expected for each square of side 250m (for example), then smooth standardised differences using kernel smoothing

7. Smoothed risks around 2 landfill sites, before and after opening Astbury Quarry Standard

8. Nantygwyddon Trecatti

9. Interpretation • Need to consider the change in risk pattern, comparing before and after opening. • Different sites have different risk patterns. • Pooled results across sites must be interpreted carefully. • Possibly due to geographical differences. • Risk does not seem isotropic – possibly affected by wind, water flow, topography of site. • What is the effect on tests and estimates of risk?

10. Simulation exercise • Assume that a certain amount of pollutant is spread from a point source • Consider different patterns of spread • isotropic • Concentration on direction of prevailing wind • Non-monotonic • Based on scenario of births to provide detailed data – but interested in relative magnitudes of power, etc, rather than absolute • Does geography matter?

11. Results • Show power – Stone’s method for simplicity (patterns the same for others) • Mean estimated odds ratio if using ‘circles’ with correct threshold • These vary between sites because of the distribution of the population

12. Results on power – 3 sites Isotropic Not isotropic Not isotropic Not monotonic

13. Results on odds ratio – 3 sites Isotropic Not isotropic Not isotropic Not monotonic

14. Migration • Large numbers of people move house each year • Many diseases associated with environmental risks are believed to be due to long term exposure • Taking place of residence at diagnosis as representing exposure is potentially misleading – exposure may have arisen from previous locations • Effect will be to weaken the apparent risk • We planned to use the NHSAR to identify appropriate models – but the data are not yet available

15. Migration model • Based on the population around a site, divided into census EDs (between 150 and 200, depending on the site) • Assume a fixed probability of moving each year • Probability of destination of move decreases with distance • Assume the background rate varies across EDs according to a log-normal distribution • Assume a monotonically-decreasing risk from the source

16. Monitor total number of exposure-years on each individual • Assume a logistic model for the risk of a case as a function of exposure-years • Use ‘circles’ method for simplicity to assess effect • Estimate effect on odds ratio and power

17. Typical results from a site • Odds ratio and power decrease markedly as migration increases. • Absolute values depend on parameters – pattern seems to be preserved but local geography matters

18. Errors in variables • Ecological studies by administrative area • Take area-based disease rates (mortality, incident cancer cases etc.) • Risk factors also defined at area level • Often from census data • Also from irregularly measured factors • So these are unlikely to be reported at correct levels

19. Typical problem – leukaemia incidence against ionising radiation

20. Simulation • Poisson regression & spatial models – only Poisson results shown for brevity • Measure of deprivation used as covariate • Spatial correlation induced • Classical measurement error model • Interested in bias and in the estimate of the SD of the regression coefficient

21. Typical simulation results – effect on bias • Based on all-Wales (908 wards) and a sub-region (111 wards) • Bias increases with error SD as in other contexts • Effect of correlation more on the estimated SD • Spatial model (BYM) gives similar parameter estimates but with better estimate of SE

22. Conclusion • In investigating the risk around a source we need a proper measure of exposure; distance is not enough. • Methods which assume the risk decreases monotonically with distance lack power. • Effects will vary with geographical location and account must be taken of local conditions. • Migration can have a considerable effect on the extent of exposure. This is particular important when distance is used a surrogate for exposure. More work is needed on better models. • A proper investigation requires detailed studies at individual level, of locations and people to assess exposure accurately.