Exploring methods for mapping seasonal population changes using mobile phone data

Abstract

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