|Defining area at risk and its effect on loss estimation|
Defining area at risk and its effect on loss estimation
Catastrophe loss estimations combine both hazard and vulnerability data. While hazard attributes such as intensity distributions are usually represented at a spatially-explicit pixel level, vulnerability information, such as dwellings, population and insurance portfolio data, is usually only available at coarse areal units, such as postcodes. This spatial mismatch can be a source of error since the elements at risk are implicitly assumed to be uniformly distributed across the areal unit and any spatial inhomogeneity within the unit ignored. As a consequence, the entire areal unit is improperly treated as being at risk. Here, in one of a suite of related studies on the impact of different averaging schemes on catastrophe modeling, we define occupied residential area as the area at risk and seek to assess its effect in a hailstorm loss estimation model.
Figure 1. A schematic representation of hazard intensity distribution, a postcode and identified residential areas
Figure 1(a) shows a symmetric pattern of hazard intensity about the epicenter of an event. Increasing density of colour corresponds to increasing intensity – earth shaking intensity or peak ground acceleration in the case of an earthquake, or hailstone size in the case of a hailstorm. Traditional modeling practice attributes the intensity within the postcode to its centroid. (Postcode units were created for administrative purposes and their boundaries lack any physical meaning for catastrophe loss estimation.) The centroid is determined by boundaries that can be arbitrary, and the centroid for a polygon of irregular shape may not even be located within the postcode! In that case, assigning the hazard intensity to the centroid could be quite misleading.
In Figure 1(b), we see that the residential area is further away from the epicentre than the centroid of the postcode, and that a lower intensity should be assigned to the elements at risk or else losses will be overestimated.
Using Sydney as an example, residential areas at risk were identified through street buffers, and vulnerability data remodeled using an approach called dasymetric mapping that transforms data from arbitrary areal units to physical settlement areas. Occupied residential areas are assumed to be physically linked by street networks. A commercial street database – StreetWorks™, was used to derive the residential area with a buffer distance of 100 m either side of the street segment. Spatial layers of local and national parks, lakes and rivers were also used to refine boundaries. The final residential area was then converted to a pixel format with a spatial resolution of 100 m. The Sydney study area covers 8879 km2, whereas the derived residential area comprises only 22.0% of this. Loss estimation should focus on this sub-set.
In this example, separate houses were chosen as the insured elements. Numbers of separate houses are available from CDATA 2001 for postcodes and census collection districts (CCD). There are a total of 255 postcodes and 6674 CCDs contained within the study area. The number of separate houses is 919,898, and an average sum insured of AUS $200,000 was assumed. Vulnerability data were represented at the two areal unit forms and their corresponding residential forms. The postcode represents the coarsest resolution whereas the CCD-based occupied residential areas form the finest scale.
Figure 2: Distribution of loss potential from a scenario hailstorm
Hail was used as our hazard of choice to address the key theme of this article. (The April 1999 Sydney hailstorm, which caused an estimated total insured loss of AUS $1.7 billion, was the most costly natural disaster in Australia.) In this example, identical scenario hailstorms with a hypothetical loss potential given in Figure 2 were imposed on the 140 km × 160 km grid - a total of 22 400 events. The distribution of loss was assumed to follow a lognormal function, with a SW-NE oriented elliptical footprint (semi-major axis 25 km and semi-minor axis 10 km) and maximum loss percentage for buildings (claims as a percentage of the total sum insured) of 3.5%. These characteristics are broadly consistent with those of severe Sydney hailstorms such as the March 1990 and April 1999 events.
Catastrophe losses were calculated for each hailstorm over all postcode and CCD units and the values assigned to the pixel where each hailstorm was centred. We then compared the spatial distribution of loss estimates calculated using the vulnerability data in both postcode and CCD-based residential area forms. Figure 3(a) shows the spatial distribution of estimated losses using vulnerability data at the CCD-based residential areas. Depending on the location of the hailstorm, losses range from below $1 million to $600 million.
Figure 3. (a) Calculated losses using vulnerability data at CCD-based residential area form for scenario hailstorms occurring at difference locations. (b) Loss estimation differences using the vulnerability data at postcode form and CCD-based residential area form.
Figure 3(b) shows the spatial distribution of loss estimation differences on a pixel basis. These differences are relative to the loss estimated for the postcode form, and negative differences imply that the loss estimate for CCD-based residential areas was higher. Differences range between minus 161.3% and plus 99.5%. The areal proportions of the most extreme underestimation [-161.3%, -50.0%) and overestimation [50.0%, 99.5%] groups occupy almost one third of the study area! For inner city suburbs dominated by continuous residential areas, loss estimation differences are modest because the chances of mis-allocating proportions of the portfolio to non-residential areas are small. In contrast, significant overestimations occurred for storms centred within non-residential areas (e.g., national parks, region of east Picton). Assuming homogeneity at the coarse postcode level results in allocating portfolio data to these non-residential areas and incorrectly computing larger losses. Such errors are unlikely when portfolio data at the CCD-based residential area, which more realistically reflects the true location of dwellings, is used.
Loss estimation differences using the vulnerability data in other forms were also compared. While differences were often large, the main conclusion to emerge was that loss estimates using CCD and CCD-based residential area forms were relatively close. This is good news for catastrophe modelling and is attributed to the fact that the fine resolution of CCDs reflects the general distribution of residential areas quite well.
Current loss estimation practice largely uses portfolio data at a postcode level; it is logical to anticipate that better loss estimates could be achieved through a more accurate definition of the area at risk. Our empirical findings suggest that this is indeed the case, particularly for hazards that affect only a small proportion of the area under consideration, as is true for hailstorms.
If the above analysis were repeated for some other hazards such as damaging earthquakes, which may have much larger areal footprints than do hailstorms, our results may well be less sensitive to the way in which vulnerability data is input. However in this case, it is the scale at which local soil conditions, which can amplify seismic ground motion and increase damage to buildings, may change which may dictate model accuracy. This, in combination with the non-linearity of relationships between building damage and ground shaking intensity, means that fine-grained averaging is again inescapable, if we are ever to model Probable Maximum Losses with any confidence.
For more information please contact:
Keping Chen, John McAneney,
Russell Blong or Roy Leigh
FloodAUS risk ratings provide estimates of the Average Recurrence Interval of flooding at ground level (GL), 1 metre above ground level, and 2 metres above ground level.
FloodAUS is a GIS-based risk assessment methodology developed by Risk Frontiers to estimate mainstream flood risk address-by-address using:
FloodAUS risk ratings may be purchased by postcode or urban area. Alternatively, Risk Frontiers offers a consulting service to analyse company specific portfolios. Risk Frontiers can extract maximum benefit from the data to quantify the flood risk for your portfolio.
Spatial distribution of address risk ratings
Sample FloodAUS database
of the 24 urban areas
E-mail firstname.lastname@example.org for data prices and quotes for consulting services.
For further information about FloodAUS
check out the Risk Frontiers web site: