Bayesian space-time models for surveillance and policy evaluation using small area data Seminar

Abstract
In this talk, I will discuss various Bayesian space-time models for small area data that we have developed for the purposes of surveillance and policy evaluation. A common feature of the models and applications we consider is that data (counts of events) are available for a set of small areas over time, and we are interested in identifying areas that exhibit unusual temporal trends in the event rate. Since the data for each area are typically sparse, we adopt a Bayesian hierarchical modelling approach to smooth the raw event rates. We consider two different 'detection' scenarios: (1) Policy Evaluation - here, a policy or intervention has been implemented in a known subset of areas, and we wish to evaluate whether this has had a measureable impact on the event rate in those areas compared to other non-targeted areas in the study region (2) Surveillance - here we do not have an a priori subset of areas of interest, but simply wish to identify any areas whose event rate differs markedly from the general time trend. This can be useful to flag-up areas warranting further investigation (e.g. to identify emerging localized risk factors) or to help target areas for, say, resource allocation purposes. For both scenarios, our general approach is to fit a Bayesian spatio-temporal hierarchical model comprising a 'common trend' component and an 'unusual trend' component. For the surveillance model, these two components are formulated independently, and a Bayesian model choice framework is used to detect unusual areas based on the posterior probability that each area belongs to the common trend component. We use a Bayesian estimate of the False Discovery Rate to calibrate these posterior probabilities. For the policy evaluation model, the common trend component is estimated from a set of control areas, while the model for the policy areas includes the common trend component plus an impact function which only comes into play at time points after the policy was implemented, and captures the relative change in time trend following the policy. This is similar in spirit to interrupted time series models. The posterior probability that the impact function is positive (or negative) can be used to evaluate the success of the policy in each area. I will illustrate these models using various health and social science applications, including surveillance of cancer incidence trends using routinely collected cancer registration data from Thames Cancer Registry, and evaluation of the impact on burglary rates of a scheme to reduce "cold calling" (unsolicited house visits to sell products or services) implemented by Cambridgeshire Constabulary.