Background In disease mapping, fine-resolution spatial health data are routinely aggregated for various reasons, for example to protect privacy. Usually, such aggregation occurs only once, resulting in ‘single-aggregation disease maps’ whose representation of the underlying data depends on the chosen set of aggregation units. This dependence is described by the modifiable areal unit problem (MAUP). Despite an extensive literature, in practice, the MAUP is rarely acknowledged, including in disease mapping. Further, despite single-aggregation disease maps being widely relied upon to guide distribution of healthcare resources, potential inefficiencies arising due to the impact of the MAUP on such maps have not previously been investigated. Results We introduce the overlay aggregation method (OAM) for disease mapping. This method avoids dependence on any single set of aggregate-level mapping units through incorporating information from many different sets. We characterise OAM as a novel smoothing technique and show how its use results in potentially dramatic improvements in resource allocation efficiency over single-aggregation maps. We demonstrate these findings in a simulation context and through applying OAM to a real-world dataset: ischaemic stroke hospital admissions in Perth, Western Australia, in 2016. Conclusions The ongoing, widespread lack of acknowledgement of the MAUP in disease mapping suggests that unawareness of its impact is extensive or that impact is underestimated. Routine implementation of OAM can help avoid resource allocation inefficiencies associated with this phenomenon. Our findings have immediate worldwide implications wherever single-aggregation disease maps are used to guide health policy planning and service delivery.

Background All analyses of spatially aggregated data are vulnerable to the modifiable areal unit problem (MAUP), which describes the sensitivity of analytical results to the arbitrary choice of spatial aggregation unit at which data are measured. The MAUP is a serious problem endemic to analyses of spatially aggregated data in all scientific disciplines. However, the impact of the MAUP is rarely considered, perhaps partly because it is still widely considered to be unsolvable. Results It was originally understood that a solution to the MAUP should constitute a comprehensive statistical framework describing the regularities in estimates of association observed at different combinations of spatial scale and zonation. Additionally, it has been debated how such a solution should incorporate the geographical characteristics of areal units (e.g. shape, size, and configuration), and in particular whether this can be achieved in a purely mathematical framework (i.e. independent of areal units). We argue that the consideration of areal units must form part of a solution to the MAUP, since the MAUP only manifests in their presence. Thus, we present a theoretical and statistical framework that incorporates the characteristics of areal units by combining estimates obtained from different scales and zonations. We show that associations estimated at scales larger than a minimal geographical unit of analysis are systematically biased from a true minimal-level effect, with different zonations generating uniquely biased estimates. Therefore, it is fundamentally erroneous to infer conclusions based on data that are spatially aggregated beyond the minimal level. Instead, researchers should measure and display information, estimate effects, and infer conclusions at the smallest possible meaningful geographical scale. The framework we develop facilitates this. Conclusions The proposed framework represents a new minimum standard in the estimation of associations using spatially aggregated data, and a reference point against which previous findings and misconceptions related to the MAUP can be understood. Electronic supplementary material The online version of this article (10.1186/s12942-019-0170-3) contains supplementary material, which is available to authorized users.

Objective To examine the impact of the modifiable areal unit problem (MAUP) in an investigation of factors associated with ED demand in Perth, Western Australia, in 2016. Furthermore, to advocate a means of avoiding this impact. Methods ED presentations were classified as: urgent medical, non‐urgent medical, urgent trauma or non‐urgent trauma. In each group, sex‐stratified, age‐adjusted multivariate associations with socio‐economic status and distance to the nearest ED and general practitioner (GP) were estimated. Modelling was undertaken using different sets of spatial units: Australian Bureau of Statistics (ABS) Statistical Areas Level 1 (SA1s) and numerous aggregate‐level zonations of SA1s (ABS SA2s and others). Results Estimates obtained using the different units often varied widely: for seven (30%) of 24 strata defined by combinations of sex, ED type and covariate, the smallest and largest effect sizes differed in terms of direction; further, for 11 (65%) of the remaining 17 strata, the largest effect size was at least twice as high as the smallest. This demonstrates the MAUP's impact and that analyses based on a single set of spatial units are unreliable. To resolve the observed variation, we highlight the SA1‐level estimates. Conclusions When formulating interventions targeting reduced ED utilisation, policy planners should be guided by evidence based on analysis of appropriate spatial units. This ideal is undermined by the widespread lack of acknowledgement of the MAUP in studies examining drivers of ED demand using spatially aggregated data. To avoid the MAUP, only estimates obtained through examining a minimal geographic unit should be relied upon.

High rates of dental-related potentially preventable hospitalisations are thought to reflect poor access to non-hospital dental services. The association between accessibility (geographic and financial) to non-hospital dentists and potentially preventable hospitalisations was examined in Western Australia. Areas with persistently high rates of dental-related potentially preventable hospitalisations and emergency department (ED) presentations were mapped. Statistical models examined factors associated with these events. Persistently high rates of dental-related potentially preventable hospitalisations were clustered in metropolitan areas that were socioeconomically advantaged and had more dentists per capita (RR 1.06, 95% CI 1.04-1.08) after adjusting for age, sex, socioeconomics, and Aboriginality. Persistently high rates of ED presentations were clustered in socioeconomically disadvantaged areas near metropolitan EDs and with fewer dentists per capita (RR 0.91, 0.88-0.94). A positive association between dental-related potentially preventable hospitalisations and poor (financial or geographic) access to dentists was not found. Rather, rates of such events were positively associated with socioeconomic advantage, plus greater access to hospitals and non-hospital dental services. Furthermore, ED presentations for dental conditions are inappropriate indicators of poor access to non-hospital dental services because of their relationship with hospital proximity. Health service planners and policymakers should pursue alternative indicators of dental service accessibility.

IntroductionThe Modifiable Areal Unit Problem (MAUP) arises from the aggregation of data organized by spatially defined boundaries. Aggregated values are influenced by the shape (zone effect) and scale of the aggregated units. Aggregations of the same data using different zones or scales can give different analytical results, none reliable. Objectives and ApproachUsing population-level administrative health data in Western Australia, the objectives were to: accurately measure the association between health service utilization and demographic, socio-economic, and service accessibility variables; and develop models to accurately forecast areas of high health service utilization into the future. Multiple zone designs and aggregation scales were used to examine the impact of MAUP in association studies. These zone designs and scales were then used in all-subset model selection processes, combined with repeated k-fold cross-validation, to generate forecast maps of areas having high future rates of health service utilization. ResultsThe impact of the MAUP and methods to reduce this bias in association studies will be presented, for both simple and complex model designs. Maps indicating gradients of predicted probabilities of high rate of health service demand in the future can be used to optimize the placement of services, through the use of catchment areas based on road-network travel distance and population distributions. Conclusion/ImplicationsThe impact of the MAUP on the analysis of spatially-aggregated data has been considered intractable. However, methods to reduce the impact of the MAUP can improve policy and planning decisions based on such studies.