Warren Jochem , University of Southampton
Andrew J. Tatem, University of Southampton
Accurate and timely data on the population of local areas is vital for policy and decision making and for monitoring progress towards development goals. Yet in many places, population data are out of date and a complete population census is difficult to complete. This work addresses the challenge of producing accurate, high spatial resolution population estimates in the absence of a full census. We develop a marked spatial point process model to jointly model the density of building locations and population per building from samples of georeferenced households. We use a Bayesian framework and make predictions of the population with uncertainty at a 1 km grid cell resolution. We apply our model to a simulated georeferenced census which enables us to test different data sampling scenarios and evaluate predictions against a known population. The initial results suggest that point process models with a shared spatial effect have the potential to support population mapping and estimation; however more work is needed to investigate the sensitivity of the model. Ongoing work is also extending the basic joint model form to include geospatial ancillary data as covariates to improve the predictive performance. Further development of spatial point process models and related statistical techniques can open up opportunities to make fuller use of a wider range of datasets to study population distributions and to make accurate predictions of the population.
Presented in Session 29. Estimation Methods