Methods for small area population forecasts: state-of-the-art and research needs

Abstract

Small area population forecasts are widely used by government and business for a variety of planning, research and policy purposes, and often influence major investment decisions. Yet the toolbox of small area population forecasting methods and techniques is modest relative to that for national and large subnational regional forecasting. In this paper we assess the current state of small area population forecasting, and suggest areas for further research. The paper provides a review of the literature on small area population forecasting methods published over the period 2001-2020. The key themes covered by the review are: extrapolative and comparative methods, simplified cohort-component methods, model averaging and combining, incorporating socio-economic variables and spatial relationships, ‘downscaling’ and disaggregation approaches, linking population with housing, estimating and projecting small area component input data, microsimulation, machine learning, and forecast uncertainty. Several avenues for further research are then suggested, including more work on model averaging and combining, developing new forecasting methods for situations which current models cannot handle, quantifying uncertainty, exploring methodologies such as machine learning and spatial statistics, creating user-friendly tools for practitioners, and understanding more about how forecasts are used.

An Open‐Source Model for Projecting Small Area Demographic and Land‐Use Change

“…The models presented in this paper draw strength from each of these categories, so we consider the broad context here and return to how each have informed our models in the discussion section of the paper. A subsequent review of small area projection methods over the past decade undertaken by Wilson et al (2021) identifies similar broad headings, with the addition of microsimulation and machine-learning methods. We utilize microsimulation, but not machine learning in our models.…”

“…The key criticism leveled at these approaches by Cameron and Cochrane (2017) is that they lack a strong theoretical basis given their deterministic reliance on past trends, especially when used for areas that exhibit less predictable growth. Nonetheless, they have been found to perform well in terms of overall accuracy (Smith, Tayman, and Swanson 2013), often better than more complex methods such as cohort component models at a small area level (Smith and Tayman 2003) and have the distinct advantage of relatively low data requirements (Wilson et al 2021) so can be applied in a broad range of contexts.…”

“…Comparatively, these models have substantial data requirements for each of the demographic components, and while this is seldom a problem at national and regional scale, these requirements are often too great to effectively use cohort component models for small area projections (Swanson, Schlottmann, and Schmidt 2010;Wilson, 2015). Wilson et al (2021) refer to as simplified cohort-component methods, namely cohort change ratios, of which the Hamilton-Perry method (Hamilton and Perry 1962) is widely used, e.g. by Swanson, Schlottmann, and Schmidt (2010) to project census tract populations in the United States.…”

“…While statistical models for demographic projection are attractive because they offer the opportunity to include contextual variables which might be relevant at small area scale, the reviews offered by both Wilson et al (2021) and Cameron and Cochrane (2017) point out that that regression-based models generally do not perform any better in terms of forecast accuracy than simple extrapolative methods. This has been demonstrated by Chi (2009) in a study comparing a regression model (including variables capturing small area characteristics) with simple extrapolative approaches, and Chi and Voss (2011) who additionally incorporated variables capturing the characteristics of neighboring areas.…”

“…see Wu and Birkin 2012). A recurring criticism of using microsimulation approaches for demographic projection is that they are extremely data intensive (Wilson et al 2021) and by extension they are difficult to calibrate and validate .…”