Exploring methods for mapping seasonal population changes using mobile phone data

Accurate, high-resolution population distribution data is crucial for effective policymaking and resource allocation. Traditional data sources, like censuses, often lack the temporal and spatial detail needed for many applications, including planning, resource allocation, disease control, and disaster management. While censuses provide comprehensive spatial coverage, their typical decadal frequency limits their use for contemporary analyses. Smaller surveys have limited spatial coverage and suffer from recall bias. In recent years, mobile phone data, specifically call detail records (CDRs), have emerged as a valuable alternative data source. CDRs provide high-resolution data on user location and call times, enabling more precise and frequent analysis of population distributions and mobility patterns. This study focuses on the critical step of distributing mobile phone users from cell tower locations into relevant administrative units, recognizing the impact this has on the accuracy of population estimates. Inaccurate distribution can lead to skewed results and impact seasonal mobility analysis, making the selection of appropriate distribution methods crucial. This paper explores the impact of using four different distribution methods on the accuracy of seasonal mobility patterns in Namibia.

Existing research highlights the potential of CDRs for population mapping and mobility studies, but often overlooks or simplifies the critical step of distributing user counts from cell towers into administrative units. Studies have shown CDRs can be used to model population distribution and examine human mobility relationships with infectious disease transmission. However, inconsistencies exist in the distribution methods used, and the impact of these choices on subsequent analysis, such as seasonal trend decomposition, remains largely unexplored. Studies have used methods like Point to Polygon allocation and Voronoi tessellation, but more complex methods based on maximum likelihood estimation and signal propagation models are gaining traction. Koebo (2020), for example, compared several methods and found mixed results, concluding that more sophisticated methods didn't always yield significantly improved results. Zulfira et al. (2018) investigated seasonal mobility in Senegal using individual trajectory matrices, but this study takes a different approach by focusing on distribution methods and their effect on seasonal patterns at different spatial scales.

This study uses anonymized CDR data from MTC, Namibia's leading mobile network provider (76% market share in 2010-2012), covering October 2010 to April 2014. The data includes anonymized user IDs, communication times, and cell tower IDs. MTC provided approximate circular coverage ranges for each tower. Data processing followed the approach outlined in zu Erbach-Schoenberg et al. (2016), assigning users to towers based on their nighttime location. Daily and monthly user counts per tower were derived, and data from 2011-2013 were used for analysis. Four distribution methods were assessed: 1) Point to Polygon, 2) Voronoi, 3) Tower Ranges, and 4) Adjusted Voronoi. Each method's user density per administrative unit was compared against census-derived population densities using Kendall's tau-b rank tests for correlation. Seasonal and trend decomposition using Loess (STL) was applied to analyze seasonal user variation. Multivariate clustering was used to group administrative units with similar seasonal user behavior using a K-means algorithm. Three population datasets (2011 census, WorldPop 2011-2013, GHS-POP 2014) were used to assess MTC network coverage in terms of population.

The study found high population coverage (above 85%) by MTC tower ranges, indicating good representativeness of the population. However, the performance of distribution methods varied significantly depending on the administrative level. At the coarsest level (Level 1 – regions), there were marginal differences between methods, but at finer levels (Level 2 – constituencies and Level 3 – enumeration areas), tower ranges showed the highest correlation with census data. At Level 3, other methods, particularly Point to Polygon, showed poor correlation or almost no relationship with the census data. STL analysis revealed two major seasonal patterns across all methods: a 15% user decrease in August and a 20-30% increase in December, likely linked to school holidays and vacation travel. Multivariate clustering showed that increasing the number of clusters amplified the differences between methods. The 'intersection effect', where small intersections between tower ranges/Voronoi polygons and administrative units resulted in extreme seasonal proportions at Level 3, necessitated data filtering. When comparing only covered units, the relative performance of distribution methods changed at levels 2 and 3. Tower Ranges consistently outperformed other methods, especially at finer administrative levels where spatial resolution and accuracy are crucial.

The findings highlight the critical importance of selecting appropriate distribution methods based on the spatial scale of analysis. Tower ranges, when available, provide more accurate results at finer scales due to their better representation of service areas. The study confirmed that more complex methods don't always translate to better accuracy, especially at coarser levels. The consistency of two major seasonal patterns across all methods and levels supports the validity of using CDR data for analyzing seasonal mobility. The observed discrepancies between methods in multivariate clustering emphasize the sensitivity of the chosen approach on the subsequent findings. The results suggest the need for further research to establish thresholds for tower density and other factors that influence the choice of distribution method.

This study demonstrates the value of CDR data for mapping seasonal population changes but emphasizes the crucial role of appropriate distribution methods. Tower ranges, when available, are advantageous for finer-scale analyses in areas with relatively low tower density. Voronoi methods offer a viable alternative when tower range information is lacking. Point to Polygon can be an effective legacy method for coarser-scale analyses with high tower density. STL and multivariate clustering are useful for identifying seasonal mobility patterns. Future research should focus on validating the findings in various contexts, quantifying thresholds for method selection, and investigating the influence of other factors on method performance.

The study's limitations include using modeled mobile phone user distributions as a proxy for actual population distributions, potential biases in CDR data (multiple devices, varying demographics), and the use of annual baselines in calculating seasonal proportions, which can introduce artificial attenuation in areas with network expansion. The 'intersection effect' at finer spatial scales necessitates data filtering. The assumption that rural towers experience more frequent downtime needs additional validation. Lastly, the study's reliance on data from a single mobile network provider limits the generalizability of the findings.