2.3 Aggregated health data and the modifiable areal unit problem

A key feature of many spatial analyses involving health data is that data originally relating to individual patients – or even individual health events – that have been aggregated over geographical areas (for example the Province of Gauteng) or population sub-groups (for example counts of deaths among males aged 5-14).

The modifiable areal unit problem is a term for the effect on analysis of aggregating data on individuals into a set of areas (e.g. districts, counties, census tracts, or regions, say).

This type of data aggregation may be necessary for various reasons. Firstly, it may be that confidentiality constraints prevent the publication of individual-level information by the agency collecting the data. This is particularly likely to be the case with sensitive health-related events. There is generally a trade-off between the amount of attribute detail that can be provided and the size of the aggregation group – which is often a geographical area. The problems associated with the specific size and definition of geographical areas of this type is known as the modifiable areal unit problem. The modifiable areal unit problem comprises two related aspects, known as (a) scale and (b) aggregation effects: the scale effect is that pattern and relationships in our aggregated data vary according to the scale over which they have been aggregated. The patterns we see in small units may well be different from those we see in larger units. The ecological fallacy relates to the fact that relationships between variables that we observe at one level of aggregation may not necessarily apply if the data were observed at the individual level – or indeed at any other level of aggregation. The Aggregation effect is that, at any given scale, they vary simply according to where we place the boundaries between the zones. In a classic study of the Modifiable Areal Unit Problem (or MAUP), Openshaw (1983) showed how the aggregation effect could produce a shift in the correlation coefficient from -0.99 to +0.99 in an analysis of voting data for the state of Iowa. By reviewing evidence on the scale effect from many studies, Openshaw (1983) also showed how a relationship between two variables often appears stronger when it is investigated using larger areal units as opposed to smaller units.

Even when we have data on the location and health of individuals, it is frequently the case that we wish to relate health event data to environmental, behavioural and predisposing factors which are only available to us in aggregate form. This occurs even if event data are referenced as individual points but are assigned to polygons representing census zones or environmental values. Even more complex is the situation in which event data aggregated to one set of areal units are overlaid on a data layer for a different set of areal units and it is necessary to attempt some form of areal interpolation. In each of these cases, it is particularly important to understand the characteristics and limitations of areally aggregated data.

More recently, techniques have begun to emerge that tackle some of the problems that the Modifiable Areal Unit Problem raises. One such technique is multi-level modelling; another related technique is called geographically weighted regression (GWR). GWR in particular is a technique that is now available in ArcGIS in the spatial statistics part of the ArcToolBox

print friendly pdf

---

---