Exploring methods for mapping seasonal population changes using mobile phone data

Abstract Data accurately representing the population distribution at subnational level within countries is critical to policy and decision makers for many applications. Call data records (CDRs) have shown great promise this, providing much higher temporal spatial resolutions compared traditional sources. For CDRs be integrated with other in order effectively inform support making, mobile phone user must distributed from cell tower into administrative units. This can done different ways it often not considered which method produces best representation of underlying distribution. Using anonymised Namibia between 2011 2013, four methods were assessed multiple unit levels. Estimates density per ranked each against corresponding census-derived densities, using Kendall’s tau-b rank tests. Seasonal trend decomposition Loess (STL) multivariate clustering was subsequently used identify patterns seasonal variation investigate how impact these. Results show that accuracy results influenced by level. While marginal differences are displayed “coarser” 1, use ranges provided most accurate finer levels 2 3. The STL helpful recognise on further analysis, degree consensus decreasing as scale increases. Multivariate delivers valuable insights units share a similar behaviour. number prescribed clusters, more obtained differ. However, two major identified across all methods, cluster numbers: (a) 15% decrease August (b) 20–30% increase December. Both likely partially linked school holidays people going vacation and/or visiting relatives friends. study highlights need importance investigating detail before conducting subsequent analysis like decomposition. In particular, investigated both terms their area coverage, well relation appropriate based specific application. inappropriate change observed derived conclusions.