Model-supported procedures and “big (spatial) data” in regional health services research I: Spatial reference and accessibility

The regional disparities in the provision of health services have come to the fore, particularly against the background of the discussion about demographic change. The focus here is primarily on existing and expected supply bottlenecks in rural areas (Greß/Stegmüller 2011, Lessing/Herr 2015, Ried 2016). The system of statutory health needs planning has been criticized in the past, particularly with regard to its spatial control function, with the planning areas considered too large for both analysis and planning purposes and the failure to take into account the actual accessibility of health facilities (Greß/Stegmüller 2011: 22ff.). In the course of the reform of demand planning in 2013, this problem was already addressed through smaller-scale planning levels and the possibility of regional adjustments to demand (KBV 2013, Schulz et al. 2015).

>> From a geographical perspective, however, this falls short: The sub-discipline of regional (or geographical) health services research examines the provision of health services to the population with particular attention to spatial disparities. The perspective here is expressly very small-scale, although the focus is often on the accessibility of these facilities (Augustin et al. 2015, Neumeier 2013, 2016, Pieper/Schweikart 2009, Stentzel et al. 2015). In a sense, it represents the spatial science counterpart to medical or nursing science-oriented health services research. Recently, a “digital agenda” for health services research has been called for: Since ever larger amounts of data are 1.) available in the healthcare system and 2.) also acceptable due to the IT-related possibilities “Big data” offers unimagined possibilities for upgrading healthcare research – but only if theoretical-conceptual considerations do not give way to “seeking for correlations” (Schrappe 2016).
The theses mentioned are to be expressly welcomed from a quantitative-geographical perspective, as comparable discussions are being held here with reference to “big spatial data” (Graham/Shelton 2013, Kitchin 2013) and small-scale supply analyzes require extensive data availability. In this article, based on a project on small-scale health care in southern Lower Saxony (Wieland/Dittrich 2016), the basic concept of geographical health care research is explained using examples. The focus here is on making the need for a small-scale view visible and on explaining some general aspects of methodology and data availability. Since, as shown, the conventional approaches have theoretical-conceptual limitations, in the second part of the article two model-based approaches for analyzing and optimizing small-scale care with regard to general practitioners in the above-mentioned study area are presented.
Small space and transport accessibility as key variables
On the relevance of the spatial aggregation level
Requirements planning operates in the needs analysis and the determination of doctor's offices with a total of four levels of planning areas, with the 883 middle areas for family doctor care representing the smallest level. The next highest level for general specialist medical care is provided by the 372 districts and independent cities, while specialized specialist medical care is based on the level of the 97 German spatial planning regions and separate specialist medical care is based on a total of 17 regions of the statutory health insurance associations (KBV 2013: 7). The middle areas usually consist of several municipalities and come from the central places concept of spatial planning, which postulates a spatial hierarchy of supply locations (regarding private and public infrastructure, e.g. retail, postal and financial services, health and educational institutions). The middle areas here form the central local areas, i.e. the areas that are also served by centers beyond their own municipal area, by middle centers (Einig/Zaspel-Heisters 2016: 12ff.). Demand planning therefore already takes into account supra-regional demand from medical practices, which is tailored to a more or less large area depending on the doctor's specialization.
Small areas represent the most important central basis for analysis in regional health services research, which is why very local spatial levels are usually used. The following example uses the smallest scale for which at least rudimentary demographic information (ie at least approximate population numbers) is available over a wide area, namely the city or districts of politically defined local authorities (cities and municipalities, joint municipalities). In a research project at the University of Göttingen, all health facilities (doctors' practices in all departments, pharmacies, nursing services, hospitals, etc.) in southern Lower Saxony (city and district of Göttingen, districts of Osterode am Harz and Northeim) were recorded and assigned to this regional level. Without exception, the facilities were researched from public sources (e.g. doctor search from the responsible association of statutory health insurance physicians), since a complete list of all doctor locations, as with most other types of facilities, is not made publicly available for data protection reasons. At the sub-area level (a total of 420 city and district districts), supply indicators (each facility per 1,000 inhabitants) were created (Wieland/Dittrich 2016: 11ff.).
Figure 1 shows the levels of care for general practitioners in the study area, divided into five levels in color. In addition, the thin black lines show the municipal boundaries (as of 2015) and the gray lines show the boundaries of the central areas (as of 2010). First of all, it should be noted that in many of the sub-areas (303 or 72.1 %) there is no general practitioner (dots colored red). Furthermore, particularly in the regional center of Göttingen, which is divided into 69 statistical districts, there are many sub-areas with medium levels of coverage (orange and yellow dot signature). There is an enormous heterogeneity in the number of general practitioners available, particularly in centrally located communities such as Einbeck, Northeim or Duderstadt. The representation also shows how heterogeneous the middle areas are at whose level the demand planning for general practitioners is carried out, as can be seen, for example, in the Einbeck middle area (consisting of the municipalities of Dassel and Einbeck) in the north-west of the study area: the general medical offer is concentrated in a few sub-areas . Interestingly, particularly high levels of care (green dots) are achieved in individual small districts of mostly large communities: However, the care here is only “good” in relation to the size of the population, as there is usually only a single doctor there, but the population is only a few hundred inhabitants lies. This already reveals a weakening of the informative value of this approach, at least if only small-scale density values are taken into account.
The absolute or relative levels of care in themselves have no significance with regard to the small-scale disparities in the number of general practitioners. For this purpose, two indicators of concentration or dispersion are used, both of which are regularly used in the analysis of spatial and socio-economic disparities, the Gini coefficient and the coefficient of variation (Berthold/Müller 2010: 591f.; Leßmann 2005: 28f.). The Gini coefficient is an absolute concentration measure that shows the distribution or concentration of characteristics (e.g. income, sales) among characteristics (e.g. households, companies); it is standardized to a value range of zero (no concentration) and one (complete concentration). The coefficient of variation, on the other hand, is a standardized measure of dispersion (unit-independent normalization of the standard deviation), the minimum value of which is also zero (no dispersion at all), but has no upper limit. In the present case, Gini coefficients (GK) were calculated for the distribution of the facilities across the sub-areas as well as coefficients of variation (VK) for the spread of the levels of care. The same indicators were calculated again for spatial levels above (municipalities, middle areas and rural districts). Table 1 shows the results of these dispersion and concentration indicators for general practitioners, psychological psychotherapists and pharmacies. The calculation in R (R Core Team) was carried out using the REAT package (Wieland 2017).
It can be clearly seen that the larger the aggregation level under consideration, the lower the values of the dispersion or concentration indicators (that is, they indicate smaller spatial differences): With regard to general practitioners, the smallest is at the district level represent the spatial level considered here, a Gini coefficient of 0.839 was determined for the concentration of the absolute number of facilities in all districts. Considering the normalization of this coefficient to values between zero and one, this represents a strong spatial concentration. The coefficient of variation of the equipment levels at this spatial level is 2.407. At the municipality level, both the concentration indicator and the dispersion indicator decrease to 0.665 and 0.821, respectively. The middle area level above represents the demand planning level relevant for general practitioners, at which levels of care are determined and doctor's offices are allocated: the two indicators here are only 0.477 and 0.265, respectively. Accordingly, even lower disparity values are achieved at the district level. A comparable result can be seen for the other two facility types.
The problem at hand is called the Modifiable Area Unit Problem (MAUP) in geography and other disciplines with quantitative-spatial questions. This possible distortion due to the use of different spatial aggregation levels can result in a scaling effect (the homogeneity of the spatial units increases with the size of the spatial aggregation levels) and/or in a zoning effect (assignment of small spatial units to differently tailored aggregation levels, e.g. assignment of districts to communities, area code regions or postal code areas) (Manley 2014: 1158ff.). In the present case, the spatial levels are strictly hierarchical (each district is assigned to exactly one municipality, each municipality to exactly one district, etc.), so that a scaling effect can be observed here: the supply disparities - measured in the form of concentration and dispersion measures - are increasingly smaller, the coarser the spatial resolution.
If one compares the planning level (central areas or districts) with the lowest level (city and district areas) with regard to doctors and psychotherapists, it becomes clear that these spatial levels are obviously only able to a very limited extent to reflect the actual spatial differences in outpatient care to record equipment. Now it is certainly debatable which spatial unit represents the best alternative and whether such a small-scale level is necessary or justified, since in many areas a practice would simply not be economically viable. In the present case, however, it should be noted that in 89 sub-areas (21.2 %) the population is above 1,671 inhabitants, which is used in demand planning as the resident-doctor ratio (and thus also as the average minimum demand) for general practitioners (KBV 2013 : 5).
In any case, these results show that a small-scale level is necessary for analyzing supply structures. Exactly how this is designed must be carefully discussed and decided: apart from a community or district level, the aggregation level of the census grid cells introduced in Germany in the context of the 2011 census could also be used (Neutze 2015: 64ff.). However, this form of analysis fundamentally does not provide any information about the actual accessibility of the relevant facilities. This shortcoming has a two-fold effect: on the one hand, real traffic conditions are not taken into account, and on the other hand, the facilities are clearly assigned to a sub-area, with additional supply effects being ignored. A district in which there are no doctors can still have “good” care if, for example, there is a wide range of services in several neighboring towns and this is easily accessible.

Modeling of small-scale accessibility
Modeling the small-scale transport accessibility of facilities has now firmly established itself as a form of analysis in regional healthcare research. The focus here is usually on the accessibility of the closest provider (Augustin et al. 2015, Neumeier 2013, 2016, Pieper/Schweikart 2009, Stentzel et al. 2015). The accessibility of other health or education-related public services is also operationalized in spatial planning based on the travel time to the next location, for example in the federal spatial planning report with regard to hospitals (BBSR 2012: 49). By taking accessibility into account, the problem of assigning a fixed location to a doctor's office, for example, mentioned in the previous chapter can be avoided, while at the same time taking real traffic conditions into account.
Accessibility analyzes have become possible with acceptable effort through the use of geographic information systems (GIS): With the help of a GIS-supported network analysis, it is possible to calculate road distances and/or travel times for various means of transport based on a real route network. This creates a link between starting and destination points and the network of paths connecting them in a network, whereby all individual objects can be attributed in this system (e.g. speed on a stretch of road, stopping at a traffic light). Travel times are usually determined in the form of the shortest route, possibly under additional conditions such as certain stopping points. Typical forms of network analyzes are the determination of optimal routes or the determination of (potential) catchment areas of locations (De Lange 2013: 372ff.).
For network analysis, it is necessary to record all relevant facilities and their locations in georeferenced form (ie with spatial coordinates). In addition, a transport network that is as complete as possible is necessary. For this purpose, data from the OpenStreetMap project (OSM) is increasingly being used in accessibility analyzes in the healthcare sector and for various other purposes (e.g. local supply in food retail) (Augustin et al. 2015; Wieland 2011, 2015a). Aside from the aspect of acquisition costs compared to commercially sold trail networks or routing software, the freely available and usable crowdsourced data (i.e. collected collaboratively by a large number of volunteers) of this project is also of high quality and completeness: the constantly expanding OSM trail networks are, as was found in a study six years ago, quite comparable to the data from commercial providers (Neis et al. 2012: 5ff.). OSM therefore represents a valid and freely available source of “big spatial data” that has great potential for accessibility analyzes in regional health services research.
In the case of the above-mentioned study on small-scale health care in southern Lower Saxony, OSM data was also used, with accessibility being modeled in the form of the shortest car journey time to the nearest facility. The accessibility from the health facilities was shown in cartographic form and the accessibility from the districts was prepared in a table (Wieland/Dittrich 2016: 17f.). Table 2 shows the accessibility of general practitioners and psychotherapy practices as well as pharmacies based on the 420 sub-areas and broken down into four accessibility levels (car travel time up to 5 minutes, 5 to 10 minutes, 10 to 15 minutes and 15 to 20 minutes). Map 2 shows the car accessibility of the pharmacies based on the locations themselves in the form of the same four accessibility levels.
This shows that in the entire study area, each of the three health facilities mentioned is within a car park.
Travel time of a maximum of 20 minutes can be reached. Table 2 also shows the number of inhabitants at each travel time level: For example, the nearest general practitioner from 255 sub-areas with a population of 407,061 residents can be reached within 5 minutes by car. From a further 149 sub-areas with a total of 64,366 inhabitants, access to the nearest general practitioner's practice is five to ten minutes' drive away. Travel times of more than ten minutes by car can only be found in 16 areas with a total of 6,737 inhabitants. Depending on the spatial concentration or spread of the facilities, the distribution across the accessibility zones for pharmacies and psychotherapists is different, although in both cases at least the majority of the population in the study area can reach the nearest facility of the respective type within five minutes by car.
Although this transport-related accessibility appears to be a better alternative to pure density values, this one-dimensional concept of small-scale accessibility has several theoretical-conceptual problems that significantly reduce its informative value: On the one hand, the accessibility model always only takes into account the nearest location and ignores every other destination , which is only a few meters or seconds further away from the point of departure of the inquirer. Accessibility is therefore operationalized as the accessibility of the closest provider, which implies that the physical proximity of any facility of the respective type (e.g. family doctor's practice) is the only relevant care criterion. This is also reflected in the fact that not individual doctors are taken into account as destinations, but the (point) locations of their practices: In the case of group practices or practice groups, these combination locations are only counted once in the one-dimensional accessibility analysis.
In summary, it can be seen that the spatial reference level has a decisive influence on the analysis results and that the significance with regard to spatial care structures in the healthcare system can be improved by additionally taking transport accessibility into account. Both methods have specific weaknesses: While density values ignore transport accessibility and are based on (actually constructed) administrative boundaries, at least the one-dimensional accessibility analysis reduces doctor care to a simple proximity relationship without taking aspects such as selection into account; In addition, objective travel times can be subjectively perceived very differently depending on the respective purpose. It is precisely at these points that model-based procedures come into play, which were primarily developed in the context of retail and location research; These spatial interaction models can be used both for the multidimensional operationalization of accessibility and for modeling spatial demand patterns, which will be presented in the second part of the article. <<