An overview of geospatial methods used in unintentional injury epidemiology

Geospatial methods are valuable for understanding injury outcomes because they can be used to recognise patterns of occurrence, identify priority areas for prevention measures and provide more accurate modelling of clustered data that is inherently correlated (Cromley and McLafferty 2011; Ostfeld et al. 2005). While the benefits of geospatial methods have been widely known in broader public health applications for disease surveillance and data exploration in a spatial context (Auchincloss et al. 2012; Martinez et al. 2016; Rezaeian et al. 2007), this review shows that their use in the context of investigating unintentional injuries has been far less common.

Road traffic injuries were the most common category of injury causes investigated through geospatial methods. A possible reason for this could be the long-standing and well-managed injury surveillance systems for road traffic injuries that routinely collect data on the precise location of injury (e.g. specific road intersections). In addition, because there is a well-recognised and significant public health burden from these injuries, especially for fatal cases, they have long been a high priority for injury data systems development and prevention (Ameratunga et al. 2006). Outside of road traffic injuries, the use of geospatial methods has been more limited, mainly used in research of falls, burns and drowning injuries. This might be because of low counts of these injuries in a spatial context. There were some injury cause categories that were notable for their absence in the published spatial epidemiology injury studies, including injuries associated with sport and recreation, an area that could be expanded through future research.

Over the past 15 years, there appears to have been an increasing application of geospatial methods for investigating unintentional injuries, demonstrated by the growing number of published studies using these methods, particularly since 2008. This is likely due to recent advancements in geospatial methods and the development of GIS, which has now made it possible to capture, store, manipulate, analyse, manage and present all types of spatial or geographical data (Fotheringham and Rogerson 2013). It may also reflect the increased availability of routinely collected injury and determinant data that includes a spatial reference, as is now common from government and private organisations.

This review has demonstrated that mapping has been by far the most common spatial analysis approach adopted in injury epidemiological studies. Maps offer the advantage of presenting a clear visual representation of data showing regional or spatial variation in burden or injury risk (Martinez et al. 2016). Maps of standardised mortality ratios, relative risks or other similar statistical measures presented in the reviewed literature are useful for describing the spatial pattern of injury risk. However, basic mapping approaches may misrepresent spatial patterns because estimated standardised mortality ratios or other similar statistical measures do not take into account varying population sizes resulting in apparently large standardised mortality ratios in areas with small populations (Clayton and Kaldor 1987; Lawson et al. 2000). To some extent, this problem can be addressed by applying smoothing models to the risk estimates that take the overall distribution of rates into account (Rezaeian et al. 2007). Widely accepted models such as the empirical Bayes (Clayton and Kaldor 1987) and BYM (Besag et al. 1991) methods, were applied in very few of the included studies that involved small geographic areas with few cases (de Pina et al. 2008; DiMaggio 2015; Heng et al. 2015; Lassarre and Thomas 2005; Silva et al. 2011; Yiannakoulias et al. 2003).

It is fundamentally important that injury epidemiological studies begin to define spatial patterns statistically to determine whether observed clustering patterns occur by chance, or if there are statistically significant clusters that require further investigation (Pfeiffer et al. 2008). Many clustering/cluster detection methods have been developed over the past two decades based on different statistical approaches such as distance based, nearest neighbour, and scanning local rates for point and aggregated data (Auchincloss et al. 2012; Fritz et al. 2013). Our review identified that, in the context of unintentional injury research, very few clustering methods have been applied. Nonetheless, it is evident that the application of these methods has increased over the last eight years, mostly for road traffic injuries, but also falls and drowning. The statistical method regarded as having the best statistical power Tango’s maximized excess events tests (Pfeiffer et al. 2008) has yet to be applied in the context of unintentional injuries. Compared to their application in broader public health studies, other methods such as K-functions and spatial scan statistics were also not common in injury studies.

There were differences apparent in the choice of geospatial methods for clustering/cluster detection in unintentional injury studies when compared to broader public health research, suggesting that unintentional injuries might be different in terms of their spatial contexts and, hence, need to be treated differently. It is beyond the scope of this particular review to assess this more formally, but it is certainly worthy of future research attention. There are no established guidelines to suggest which method is most appropriate for what type of injury data. Largely, it appears the choice of method is dependent on what has been readily integrated into common GIS packages. Each clustering/cluster detection method will produce a different result for the same dataset and that result will also vary based on parameter settings (Fritz et al. 2013). This means that identifying the appropriate method along with parameter settings for a particular dataset is challenging and requires multiple testing. Further research in this area would be a valuable contribution.

Health outcome data routinely collected by private and government agencies is often only available as aggregated summaries for well-defined geographic areas. In such cases, spatial inferences can be made at the aggregated level in relation to socio-economic and environmental risk factors for clues to aetiology (Beale et al. 2008). The increasing availability of routinely collected injury data in the form of aggregated summaries lends itself to potential opportunities for ecological studies (Beale et al. 2008). Statistical challenges for this type of analysis include taking into account variability and potential error in rates, due to unequal population distributions and spatial autocorrelation (Elliot et al. 2000). The included studies that applied spatial regression techniques demonstrated how these methods can help to address statistical challenges associated with aggregated data by geographical regions (Dai et al. 2013; Erdogan 2009; Shenoi et al. 2015). These studies also analysed a diverse range of factors (e.g. neighbourhood, environmental characteristics) which may not be possible to assess at an individual level.

Geospatial methods play an important role in understanding the influence of complex social environments on injury outcomes that will help to develop population level injury prevention strategies (Bell and Schuurman 2010). In addition, they can help to identify which populations/sub-groups are consistently at greater (or lower) risk to inform the targeting of prevention efforts in those areas. This review has demonstrated that there is a move towards the use of more sophisticated geospatial methods from more traditional perspectives with the increasing availability in health and determinant data and also advances in GIS and other technologies. Continued advancement in this area would be well served by a detailed review of the quality of the geospatial methods currently adopted in injury epidemiological studies.