Using Land-Use Modelling to Statistically Downscale Population Projections to Small Areas

Local government planners, property developers, large businesses and other stakeholders typically require good quality projections of the spatial distribution of the future population at the small-area level. Many approaches are available to project future populations, but all suffer from limitations due to their strict underlying assumptions or limited availability of data. In this paper we apply a novel approach to small-area population projection that combines cohort-component projections at the district level with grid-based land use projections at a fine (four-hectare) geographical scale. In our approach, residential population is directly estimated in the land use model, while a separate statistical model is used to link non-residential population to non-residential land use (by type). The model can then be used to project future small-area populations using projections of future land use from the land use model. We compare four data and model specifications for the statistical modelling, using either absolute land use area or principal components as explanatory variables, and using either OLS or Spatial Durbin model specifications. All four model combinations perform reasonably well for the Waikato Region of New Zealand, with good in-sample (2006) and out-of-sample (2013) properties. However, a naïve model based on constant shares of growth outperforms all four of our models in terms of forecast accuracy and bias. Notwithstanding the underperformance relative to a naïve model, our results suggest that land use modelling may still be useful, because the model is understandable by local authority planners and elected officials, and generates greater stakeholder ‘buy-in’ than black-box or naïve approaches.