4. Concluding remarks
Our analyses using 11,096 pairs of incidences of
residential burglary (each pair consists of the same
incidents geocoded using older and newer address
information, respectively) indicate that the kernel
density estimation with a cell size of 25 25 m and a
bandwidth of 500 m may work quite well in
absorbing the poorer precision of geocoded locations
based on data from older recording system,
whereas in several areas where older recording
system resulted in very poor precision level, the
inaccuracy of incident locations may produce
artifactitious and potentially misleading patterns
in kernel density maps.
It should be noted that the present analysis is
limited to a single offense type and a single set of
ARTICLE IN PRESS
Fig. 8. Distribution of incident densities of akisu (residential burglary)(ban-level AGC): example of areas where address expressions do not
allow ban-level geocoding.
1104 Y. Harada, T. Shimada / Computers & Geosciences 32 (2006) 1096–1107
parameters in the kernel density estimation. Further
analyses are needed for a better understanding of the
impact of the precision of address geocoding on the
estimated density of crime locations in a variety of
settings. Nevertheless, our analysis here appears to
have provided a starting point and to have suggested
a direction for future studies. By accumulating
empirical evidence along this line, we will be able to
develop practical guidelines, as well as caveats, for
legitimate and meaningful use of kernel density
estimation in the analyses of spatial patterns of crime.