Predictive analysis of crime forecasting

FORECASTING THE FUTURE OF
PREDICTIVE CRIME MAPPING

by

Elizabeth R. Groff
National Institute of Justice

and
Nancy G. La Vigne
The Urban Institute

Abstract: While the use of mapping in criminal justice has increased
over the last 30 years, most applications are retrospective - that is,
they examine criminal phenomena and related factors that have al-
ready occurred. While such retrospective mapping efforts are useful,
the true promise of crime mapping lies in its ability to identify early
warning signs across time and space, and inform a proactive approach
to police problem solving and crime prevention. Recently, attempts to
develop predictive models of crime have increased, and while many of
these efforts are still in the early stages, enough new knowledge has
been built to merit a review of the range of methods employed to date.
This chapter identifies the various methods, describes what is required
to use them, and assesses how accurate they are in predicting future
crime concentrations, or "hot spots." Factors such as data require-
ments and applicability for law enforcement use will also be explored,
and the chapter will close with recommendations for further research
and a discussion of what the future might hold for crime forecasting.

INTRODUCTION
Methodological rigor in crime prevention initiatives has increased
significantly in the last two decades. This is a result of partnerships
between researchers and practitioners as well as the introduction of

Crime Prevention Studies, volume 13, pp.29-57.

Elizabeth R. Gruff and Nancy G. LaVigne

more user-friendly analytic software programs, including Geographic
Information Systems (GIS) (Crime Mapping Laboratory, Police Foun-
dation, 2000; Report of the Task Force on Crime Mapping and Data
Driven Management, 1999; Weisburd and McEwen, 1997). GIS is
often credited for providing a valuable analytic tool for the identifica-
tion and analysis of crime problems as well as the development and
assessment of crime prevention programs (Groff, 1996; Groff et al.,
2000; La Vigne and Wartell, 1998, 2000). GIS has been used to pro-
duce maps depicting crime "hot spots" as well as to conduct spatial
analyses that suggest relationships between crime and characteris-
tics of the social and physical environments in which crime concen-
trations occur (Rich, 1995; Sherman and Weisburd, 1995; Weisburd
and McEwen, 1997).
The use of GIS in law enforcement has increased significantly,
and the variety of applications for crime control and prevention is
quite broad (Dunworth et al., 1998; La Vigne and Groff, 2001; Ma-
malian and La Vigne, 1999; Crime Mapping Laboratory, Police Foun-
dation, 2000). Most applications, however, are retrospective - that is,
they examine criminal phenomena and related factors that have al-
ready occurred. While such retrospective mapping efforts are useful,
the true promise of crime mapping lies in its ability to identify early
warning signs across time and space and inform a proactive ap-
proach to police problem solving and crime prevention. Such efforts
necessitate predictive models that identify "hot spots" of crime and
disorder, as well as areas where crime is abating.
Recently, attempts to develop predictive models of crime have in-
creased,1 and while many of these efforts are still in the early stages,
enough new knowledge has been built to merit a review of the range
of methods employed to date. This chapter identifies the various
methods, describes what is required to use them, and assesses how
accurate they are in predicting future crime concentrations, or "hot
spots."2 This review covers methods ranging from simple predictions
based on the locations of past events to highly sophisticated model-
ing methods employed by researchers. Factors such as data require-
ments and applicability for law enforcement use will also be explored,
and the chapter will close with recommendations for further research
and a discussion of what the future might hold for crime forecasting.
It should be noted that it is beyond the scope of this paper to pro-
vide detailed information and accompanying algorithms for the meth-
ods reviewed herein. Rather, the intent is to describe these methods
with an emphasis on data requirements, ease of use, and applicabil-
ity to crime prevention. Each method is appropriately cited to enable
the reader to obtain more specific information and provide guidance
to those wishing to employ the methods themselves.

-30-

Forecasting the Future of Predictive Crime Mapping

WHY PREDICT CRIME?
The ability to predict the locations of future crime events can
serve as a valuable source of knowledge for law enforcement, both
from tactical and strategic perspectives. From a traditional policing
perspective, predictive mapping can inform a police department's de-
ployment efforts, helping to allocate patrols more efficiently and re-
duce response times. Despite the increased emphasis on proactive
policing, the core of police work remains that of responding to calls
for service, making effective deployment strategies paramount to a
well-functioning police department.
From a more proactive standpoint, problem-oriented policing ef-
forts may be enhanced by a more accurate scanning of areas with
crime problems, in that one can examine both distributions of past
crimes and predictions of future concentrations. As detailed below,
some of these predictive methods also provide information on "lead-
ing indicators," or explanatory variables, which can aid in the analy-
sis stage of problem-oriented policing. Such indicators offer the abil-
ity not only to predict future crimes, but to identify underlying
causes of those future hot spots. Thus, predictive mapping can assist
in the identification of crime problems and enable officers to target
intervention efforts to very narrowly defined geographic areas.
Predictive mapping holds promise for improving the identification
of areas in which to focus interventions, but it also may improve the
way those interventions are implemented. A common criticism of
crime prevention efforts is that crime is simply displaced — most of-
ten geographically — rendering the intervention ineffective. Studies
have demonstrated that displacement is not at all inevitable, and
that when it does occur 100% of crime is not displaced (Eck, 1993;
Hesseling, 1995). Nonetheless, successful crime prevention strategies
consider potential displacement possibilities and craft interventions
based upon those considerations. Thus, predictive mapping can help
law enforcement to anticipate areas of displacement, and may lead to
targeting the intervention in a broader geographic area in order to
reduce the possibility of its occurrence. Likewise, predictive mapping
may be used to enhance the potential "halo" or "diffusion of benefits"
of an intervention, whereby it has a beneficial effect beyond the
places that were targeted. Clarke and Weisburd (1994) offer ways in
which diffusion of benefits may be enhanced, including the concen-
tration of resources on highly visible or attractive targets to give the
impression to potential offenders that the intervention is more wide-
spread that it actually is. Predictive mapping can help identify those
attractive targets, thus bolstering this spillover effect.

-31-

Elizabeth R. Groff and Nancy G. La Vigne

This discussion of the potential uses of predictive mapping is by
no means exhaustive, but it is designed to illustrate the valuable
contributions that predictive models might offer, and emphasizes the
importance of this relatively new area of inquiry. Before reviewing the
various predictive methods that have been employed, it is important
to examine the theoretical underpinnings of this topic.

THE ROLE OF THEORY IN PREDICTIVE MAPPING
As described in detail below, various means of forecasting crime
events and locations exist, and not all of them can be considered
"modeling." Some methods are strictly atheoretical, relying on past
events to predict future ones. Other methods, however, are developed
by modeling the behavior of likely offenders, making it important to
review the theory underlying these efforts because theory can play an
important role in guiding the selection of independent variables, or
leading indicators.
Perhaps the most germane theories for forecasting purposes are
the rational choice perspective and routine activities theory. Both
assume that crime is purposive and that individuals are self-
determining: when people commit crime, they are seeking to benefit
themselves, and certain calculations are involved in determining
whether the criminal act will yield positive results (Clarke, 1997).
Thus, offenders are influenced by situational and environmental
features that provide desirable — or undesirable — offending oppor-
tunities. These theories are based upon the belief that criminals en-
gage in rational (if bounded) decision-making (Becker, 1968; Cornish
and Clarke, 1986), and that characteristics of the environment offer
cues to the offender that promising opportunities for crime exist
(Brantingham and Brantingham,1978, 1981; Newman, 1972; Cohen
and Felson, 1979; Harries, 1980; Wilson and Kelling, 1982).
The practical implications of these theories are that even moti-
vated criminals may nonetheless be deterred from committing crime
if they perceive a potential target to be too risky, to involve too much
effort, to yield too meager a profit, or induce too much guilt or shame
to make the venture worthwhile (Clarke, 1997; Clarke and Homel,
1997). From a predictive modeling perspective, then, these theories
have the potential to guide the selection of independent variables
with a focus on those that characterize desirable targets — and in
turn, desirable locations — of crime. Further, theory-based modeling
enables us to identify which factors influence crime target selection,
and thus inform crime prevention efforts. The models described be-
low include an assessment of whether they are supported by theory,
and the extent to which they inform prevention efforts.

-32-


THE ROLE OF GIS IN PREDICTIVE MODELING
While geographic information systems (GIS)3 are most often asso-
ciated with data aggregation and display, the technology is capable of
serving a variety of purposes. In terms of crime forecasting, GIS can
be used at the front end, as a data manipulator; in the analysis
phase as a spatial analysis engine; at the back end, for display pur-
poses; or throughout the research project. Currently, GIS is used
most frequently in the front end as a geocoder and data aggregator.
The ability to geocode4 records in a database to coordinates on the
earth's surface unlocks the potential for spatial analysis of phenom-
ena. Once these locations can be displayed, they can be aggregated
to whatever boundary is appropriate for the analysis. On the back
end, GIS is most often used as a visualization tool. The crime fore-
casts that are generated by statistical models can be displayed both
on the screen to facilitate interactive analysis and in the form of hard
copy maps, which are more portable. Both types of output can be
used to visually identify concentrations and patterns and to commu-
nicate those findings. Finally, GIS has great potential as a data
analysis tool in and of itself. The rest of this section describes how
spatial analysis and "map algebra" can be used in a GIS to develop
spatial models to predict crime.
A layer of polygon grids in a vector GIS or a raster GIS are re-
quired in order to take advantage of grid cell-based modeling.5 In
both types of GIS, the study area is divided into a series of equal-
sized cells that together form a grid. Each cell is assigned a value
based on the quantity of the variable being measured that it con-
tains. A grid is created for each attribute to be used in the model.
One advantage to raster GIS is that it is easy to represent continuous
data such as distance from another cell (e.g., distance from a major
road) or degree of concentration (e.g., density of crime). Once the in-
dividual layers are calculated, they can then be used as parameters
in a mathematical equation.
Another capability of a raster GIS is the ability to incorporate the
effects of neighboring grid cell values on a grid cell. In ArcView's
Spatial Analyst Extension, this is known as a "focal function" (ESRI,
1998). The focal function computes a new value for each cell in a grid
based on the "neighborhood of cells" defined.6 This is analogous to a
spatial lag using the queen pattern since all cells that share a side or
a corner are included up to the specified neighborhood size. These
new "smoothed" cell values can then be used in the final model.
The use of methods that involve GIS tends to require both broader
and more in-depth skill sets. While only a basic level of GIS knowl-
edge is needed to display the results of a statistical technique on a

-33-

Elizabeth R. Groffand Nancy G. La Vigne

map, geocoding requires more skill, and using extensions such as
Spatial Analyst® to build models requires even more specialized
knowledge. Encouragingly, there have been enhancements to Spatial
Analyst® that have made it easier to learn and to use (Ormsby and
A M , 1999).7
There are several major advantages to using GIS in developing a
model for forecasting crime. As mentioned before, a GIS can use data
in a spreadsheet and spatially enable it through geocoding. Once
spatially enabled, those data can then be aggregated to whatever
areal unit is most appropriate to the analysis. These functions of a
GIS are important whether or not the model is implemented in GIS.
Finally, the ability to visualize patterns in the data cannot be over-
stated and makes GIS a valuable tool for communicating the results
of an analysis. Thus, the phrase "predictive crime mapping" used
throughout this chapter is broadly applied to a variety of methods
that use GIS in any number of points throughout the forecasting
process.

REVIEW OF METHODS

Hot Spots
The most common method of "forecasting" crime in police depart-
ments is simply to assume that the hot spots of yesterday are the hot
spots of tomorrow. Crime analysts prepare maps of crimes that have
already occurred and those maps are used to deploy officers and to
identify areas in need of intervention. While surprisingly scant re-
search exists to test this assumption, the few studies we have identi-
fied suggest that the effectiveness of this approach depends upon the
time period employed. Spelman (1995) found that examining past
crimes over a one-month period is not a particularly powerful pre-
dictor — hardly better than chance, yet one year of data predicts with
90% accuracy.8 This suggests that hot spots may flare up and di-
minish over relatively short time periods, but that these flare-ups
nonetheless occur in the same places over time, creating longer-term
trends. Thus, law enforcement agencies that examine last week's
crime statistics to deploy patrols may find it more useful to identify
hot spots based on an entire year's worth of data.
One means of testing the persistence of hot spots is to analyze the
extent to which they coincide over the course of several years. Ad-
ams-Fuller's (2001) examination of hot spots of homicide in three
U.S. cities found that the vast majority of homicide hot spots per-

-34-


sisted over time,9 suggesting that past history may indeed be an ac-
curate predictor of future hot spots, at least in the long-term.
Adams-Fuller (2001) also attempted to understand the root
causes of hot spots by examining their socio-economic and environ-
mental characteristics. She found that most historical homicide hot
spots had public housing, were located in economically depressed
sections of cities, contained drug markets, and had major thorough-
fares running through them, providing easy access into and out of
the area. Her research clearly illustrates the ability to integrate theo-
retical explanations of crime with the search for hot spots. In fact,
hot spot methods are one-dimensional without the inclusion of con-
textual variables.
There are many ways in which researchers and analysts identify
hot spots (for a thorough review, see Jefferis, 1998). The most sim-
plistic approach is to use GIS to create graduated circles, the radii of
which reflect the number of events. While this method requires
minimal GIS skills, it also has its drawbacks, in that these circles
can often overlap, making it difficult to visually discern patterns of
concentrations. A more commonly applied hot spot method among
researchers is the use of spatial statistical software such as STAC10
or CrimeStat.11 These methods generate a set of ellipses that repre-
sent the highest concentrations of points.
In recent years, methods for visualizing hot spots have increas-
ingly relied on a raster GIS to interpolate a surface of crime based on
reported crime events. The analysis results have the look of a
"weather map," and are extremely popular for communicating crime
patterns for a jurisdiction. However, this type of hot spot identifica-
tion treats the known crime events as a sampling of the continuous
surface of all crime. In other words, it creates data points in geo-
graphic locations where crimes have not occurred, based on averages
between actual data points. As a consequence, large-scale maps (e.g.,
at the neighborhood level) often depict higher crime rates than were
reported to the police. This disconnect between reported crime and
interpolated crime has yet to be adequately resolved. An additional
caveat with any of these methods is that the output is based upon
user-defined criteria (e.g., band width and search area) and there are
no standard guidelines for what those criteria should be.
The popularity of these methods stems from their relative ease of
implementation — at least as compared to other methods discussed
in this chapter — as well as their ease of interpretation. The ease of
implementation is directly attributable to the existence of software to
automate the algorithms for identifying clusters and drawing an el-
lipse. The methodology for this was outlined almost 25 years ago by
Getis and Boots (1978), but languished until incorporated in a soft-

-35-


ware product. Since 1998, several other free software programs have
automated the creation of hot spots using ellipses.12 Thus the cur-
rent popularity of hot spots may be due to the fact they are relatively
easy to generate and understand.

Repeat Victimization
The above research indicates that temporally aggregated hot spots
may serve as accurate predictors of crime, but that relying on shorter
previous time periods for predictive purposes is less effective. The
exception to these research findings relates to "hot dots" (Pease and
Laycock, 1996) rather than hot spots: that is, the repeat victim
rather than the high-crime area. The concept of repeat victimization
is now well established in the criminological literature (for an early
review, see Farrell, 1995): those individuals or places that have been
victimized once are likely to be victimized again, and the time course
to subsequent victimization is a few short months (Anderson et al.,
1995; Farrell and Pease, 1993; Polvi et al., 1990). This research sug-
gests that past victimizations of individual addresses, places, and
businesses can be very accurate predictors of future victimizations,
even when relying on the previous month's victimization.
The crime prevention benefits of focusing on repeat victims to pre-
vent subsequent crimes is well established (for a summary of the
preventive impact of repeat victimization strategies on crime, see
Pease, 1998), and raises the question: Can repeat victimizations of
individuals and places be used to predict not just hot dots, but hot
spots? Very few researchers to date have examined the extent to
which hot spots are composed of repeat victimizations, except for
those who have focused on residential burglary. Both Bennett (1996)
and Townsley et al. (2000) found that one-third of all burglaries re-
ported in the hot spot areas under study were repeat burglaries.
While Morgan (2001) found a lower degree of repeat victimization
concentration within high-crime areas, the areas under study com-
bined multiple census districts and thus were larger than the average
hot spot. In his research, Morgan also found what he terms "near
repeats": residences close to repeat victims were likely to be victim-
ized. Finally, a recent publication by Farrell and Sousa (2001) con-
cludes that while repeat victimizations and hot spots do coincide,
some hot spots experience more repeat victimization than others and
may vary based upon crime type.
The research on repeat victimization suggests that much more
could be learned from further examination of the composition of hot
spots and the extent to which repeat victimizations could be used to
predict not just future victimizations, but future hot-spot areas. Such

-36-


studies should explore this question across different sized hot spots
as well as different types of crime problems, and should explore the
"near repeat" concept in further detail. Depending on future research
findings in this area, repeat victimization has the potential to provide
a simple method that could be employed by users of all skill levels
and does not require the use of GIS. The limitation of this approach
is that it does not tell us specifically what it is about the predicted
hot spots that makes them hot, limiting the extent to which the
method would inform prevention efforts.

Univariate Methods
There are a variety of univariate methods available to predict
crime. These methods use previous values of one variable to predict
its future value. They are attractive because of their straightforward-
ness: univariate methods require a minimum of data collection since
they involve only one variable. Additionally, they are atheoretical and
thus do not demand any thought about which variables should be
included in the analysis. These methods range from simple random
walk and naive lag 12 to more sophisticated models that incorporate
both seasonally and time trends. Among police practitioners the
most frequently used crime prediction methods are so-called "naive"
univariate ones (Gorr et al., 2002; Gorr and Olligschlaeger, 2001).
The two naive univariate methods used by police are the random
walk13 or and the naive lag 12.14 The random walk method is a good
predictor of series in which there are frequent pattern changes (e.g.,
to predict stock market behavior) because it reflects those changes
immediately. However, it is a poor predictor when the series to be
forecasted has seasonally or time trends (Gorr et al., 2002; Gorr and
Olligschlaeger, 2001). While these basic univariate methods are by
far the most straightforward methods of predicting crime, they are
also unfortunately by far the least accurate (Gorr et al., 2002; Gorr
and Olligschlaeger, 2001).
More sophisticated univariate methods are available that more ac-
curately predict crime levels by including seasonally in the model,15
accounting for time trends using exponential smoothing16 and pool-
ing data17 (Gorr et al., 2002; Gorr and Olligschlaeger, 2001). How-
ever, the addition of these steps also makes the methods more com-
plicated for the user.
While all the more sophisticated univariate methods offer im-
provements over the simpler ones, the exponential smoothing meth-
ods have two main advantages as tools for crime forecasting. First,
they offer the ability to account for changes in crime over time rather
than relying on the current period to forecast the next period. Sec-

-37-

Elizabeth R. Groffand Nancy G. LaVigne

ond, they are the most accurate type of method when the goal is to
forecast "small to medium-level" changes in crime (Gorr and Ollig-
schlaeger, 2001)
While the univariate techniques outlined above share many char-
acteristics, they vary in the sophistication of both the software em-
ployed and the skills required to use them (Table 1). All of the tech-
niques use area-level data and the results can be displayed easily in
a desktop GIS. Both the random walk and the naive lag 12 are very
easy to compute and can be calculated within standard spreadsheet
packages. The classical decomposition model and the exponential
smoothing models are more sophisticated and are most easily calcu-
lated using a standard desktop statistical package (e.g. SPSS® or
SAS®). The more sophisticated models also require an analyst with
more advanced statistical training. Consequently, the investment in
personnel and equipment is higher for the more sophisticated models
than the simpler ones.
In an effort to shed some light on the comparative accuracy of
these forecasting methods and others, Gorr et al. (2001) employed a
rolling-horizon experimental design to test 10 different combinations
of data-driven methods and univariate models to account for season-
ality and time trends (Gorr et al., 2002; Gorr and Olligschlaeger,
2001). Of the techniques tested, the Holt exponential smoothing with
pooled seasonally was the most accurate, and the simple exponen-
tial smoothing model with pooled seasonality was second best. These
results clearly demonstrate that using citywide, pooled measures of
seasonality offers more accuracy regardless of the exponential
smoothing method used. Their findings that the random walk and
Naive Lag 12 methods were the least accurate at forecasting crime is
of immense importance because these statistics are widely used in
the field (Gorr et al., 2002; Gorr and Olligschlaeger, 2001).
The good news from Gorr and Olligschlaeger's (2001) study is that
the more sophisticated univariate methods predict as well as the far
more complicated multivariate methods for cases with small to me-
dium changes in crime levels. Specifically, they note that while the
pooled exponential smoothing model is not typically used in police
agencies, the model is relatively simple to implement and, if used to
forecast in areas with an average of 30 or more events per month, it
offers good forecast accuracy. Given that simple exponential
smoothing methods have already been recommended in one of the
most frequently cited crime analysis books (Gottlieb et al., 1998), the
real challenge may be in encouraging widespread adoption of these
methods by analysts. One strategy for achieving this goal would be to
make the methods easier for crime analysts to implement.

-38-

Predictive analysis of crime forecasting


Leading Indicators
"Leading indicator" multivariate methods focus on using current
and past values of independent variables to predict the future value
of the dependent crime variable (Gorr and Olligschlaeger, 2001). The
"leading indicators" term in the title of the model refers to specific
characteristics of areas or neighboring areas (e.g., shots fired, calls
for service, disorderly conduct offenses, etc.) for which their rise or
fall in current and previous months can be used to predict future
values of the dependent crime variable.
There are three issues that must be addressed when specifying a
"leading indicator" model related to crime (Gorr and Olligschlaeger,
2001). First, leading indicator methods require the identification of
leading indicator variables before the model can be used. Identifica-
tion of the appropriate leading indicator variables requires a thor-
ough review of the literature, and grounds this method in theory. De-
veloping theory-based leading indicators is a time consuming task
that is critical to the success of the model.18
Second, because crime forecasts are typically done for short time
periods and across smaller areas, often there are not enough events
to develop robust model parameters. Thus, it is important to pick an
areal unit that is large enough to provide adequate numbers of ob-
servations. In general, the greater the volume of crime the more reli-
able the forecasts will be, and the smaller the volume of crime the
more variable the data will be and the more unreliable the forecasts.
Gorr and Olligschlaeger (2001) determined that a grid with 4,000-foot
cells was the smallest grid cell that would still provide reliable fore-
casts.
The third issue when specifying this type of model concerns the
development of leading indicator measures at the same geographic
and time scales as the dependent variable (e.g., grid cell data that
vary by month). Geographic information systems (GIS) have made the
spatial issue much less problematic than in the past, since a GIS
enables automated aggregation of points to customized areas. Thus,
both crime types and leading indicator variables can be aggregated to
the same areal units easily. Furthermore, Gorr and Olligschlaeger
found that selected Part 2 crimes and CAD calls are leading indica-
tors of Part 1 dependent variable crimes. Hence police are among the
few organizations fortunate enough to generate their own leading in-
dicators.
One advantage of this technique is the ability to use spatial
econometric methods that enable the inclusion of spatial and tempo-
ral lags in the model. Spatial lags allow the explicit modeling of the

-44-


effects of the values in neighboring cells on the value of the subject
cell.19 Temporal lags allow the modeling of the effects during previous
time periods on the study time period.
Over all, leading indicators show great promise since they are the
only method that has the ability to "see pattern changes coming"
(Gorr, 2001 and Gorr and Olligschlaeger, 2001). Thus, they are very
good at predicting large changes in crime levels, while extrapolative
methods are better for small to medium changes. Gorr (2001) rec-
ommends that police agencies routinely use both extrapolative meth-
ods, such as the Holt two parameter exponential smoothing method
with pooled seasonality and a leading indicator model. He recom-
mends that if the leading indicator model forecasts a large change,
one should use it because the technique forecasts correctly about
half the time. If it does not forecast a large change, then one should
use the extrapolative method's results.
While promising, these methods require significant expertise on
the part of the end user. The analyst must be familiar with multivari-
ate statistical methods and have access to statistical or spreadsheet
software programs to calculate the statistics. A geographic informa-
tion system is necessary to aggregate the chosen leading indicator
measures to the areal unit used in the study, whether it is a grid cell
or some other area such as a police beat.

Point Process Model
A new method being employed by Brown (2001) and his colleagues
is based on the theory of point patterns and multivariate density es-
timation, and can best be described as a point process model (Brown,
2001). The modeling is akin to neural networks in that there is
training involved, and past data are used to predict future events. In
essence, this approach glues multivariate models together and uses
notions from kriging and density estimation (Brown, 2001).
Brown et al. (2000) developed this predictive model based upon
the preferences of offenders, or what they term "event initiators": past
behavior illustrating the preferences of offenders is used to model
both when and where future crimes will occur: "...we do not regard
the past crime intensity at a site as a direct factor to influence how
soon criminals are going to strike again. However, this past behavior
does tell us about the preferences of site selectors and we directly
model those preferences..." (Brown et al., 2000:4). The output of the
model is a probability surface indicating likely areas of future crimes.
Brown and colleagues compared their point process model's pre-
dictive powers to that of using past hot spots to predict future ones,
testing both models by plotting commercial and residential burgla-

-45-


ries. For all but one comparison, the point process model statistically
outperforms the comparison model at the 90% confidence level.
Mean percentile scores are calculated to demonstrate how well the
model performs, essentially indicating more about how the model
performs over all than about its ability to forecast at a particular lo-
cation. The model narrows down at-risk areas to 15-25% of all po-
tential areas, thus enabling law enforcement to better target re-
sources. The smaller the percent, the better the model, because po-
lice can more narrowly focus resources. According to Brown (2001),
the model is limited from producing smaller percentages because ex-
planatory data were generated from census block groups, which do
not provide enough variation across space.
While this method shows promise, in its current form it is not
possible for others to replicate and apply it because it was custom
programmed for a specific research purpose. Furthermore, the ap-
proach requires both high-level programming skills as well as knowl-
edge of kriging and density estimation. Nonetheless, this method has
distinct advantages over others in that it is informed by theory (ra-
tional choice) and identifies which variables have explanatory power.

Artificial Neural Networks
One of the earliest efforts to do predictive crime mapping was that
of Olligschlaeger (1997), who employed a "feed-forward network with
backpropogation" to predict areas where future drug markets will
emerge. Best known to laypeople as artificial intelligence, the type of
neural network model employed by Olligschlaeger is capable of
learning extremely complex space-time patterns (Olligschlaeger,
1997). According to Olligschlaeger, "The goal is to map the input
units to a desired output similar to the way in which the dependent
variable is a function of the independent variables in a regression
analysis. The difference is that regression analysis uses linear direct
mapping whereas multi-layer feed-forward networks use non-linear
direct mapping" (Olligschlaeger, 1997:325). In essence, the network
is trained by feeding it past data and adjusting the weights assigned
to the input units; when the network is processed, the error signal is
fed back, or "backpropogated" through the network to adjust the
weights until, ultimately, the error signals are minimized (Ollig-
schlaeger, 1997). GIS was employed in conjunction with this model
in order to process spatial and temporal data, including data aggre-
gation and determination of spatial and temporal lags. This was ac-
complished by overlaying a grid and summing the data points that
fall into each cell, as well as employing contiguity measures.

-46-


The network was used to predict drug markets. Inputs included
35 months of weapon-related, robbery, and assault-related calls-for-
service; the relative proportions of residential and commercial prop-
erties in each cell; and a seasonally index. The model's performance
was tested against Ordinary Least Squares (OLS) regression analysis
and the random walk method. A comparison of R-squareds indicated
that the network's forecasting abilities were superior (Olligschlaeger,
1997). In more recent work, Olligschlaeger and Gorr (2001) found
that neural networks outperform multiple regression leading indica-
tor models when the set of leading indicators is rich and numerous.
While this method holds promise, there are no statistical tests of
significance associated with it, precluding the ability to determine
which inputs (independent variables) are providing predictive power.
The model is somewhat atheoretical and the method requires a high
degree of expertise, making it difficult to replicate.

Polygon Grid/Raster GIS Methods
As stated earlier, GIS can be used throughout a research project
to incorporate spatial relationships in the crime forecast. In Groff and
La Vigne (2001), we used a combination of polygon grid cells and
raster-based GIS to generate an opportunity surface for residential
burglary. We identified a set of variables based on the theories of
routine activities, rational choice and environmental criminology, and
used GIS to operationalize those variables.
We were very interested in modeling the effects of the values of
surrounding properties on a particular property. For example, we
wanted to model the effect of having a substandard housing unit or
vacant unit nearby. In order to incorporate the effect of surrounding
grid cell's values, each layer was recalculated using a focal neighbor-
hood function within the GIS.20 Map algebra was used to combine the
new grid cell values in each layer to produce an overall risk index
surface for residential burglary. Reported burglaries were then plot-
ted on top of this new opportunity surface to determine how well the
model predicted. The percentage of cells that were accurately pre-
dicted was used to empirically compare the actual burglaries with the
opportunity surface. Two categories of burglaries were examined: any
burglary event and repeat burglaries.
Interestingly, the model predicted repeat burglaries better than lo-
cations with both single and repeat burglaries. This finding leads to
questions about the strength of the predictive power of repeats: could
the repeat-address locations serve as a proxy for all of these other
variables used to create the opportunity surface? If so, it would sup-
port the use of past repeat victimization addresses to predict future

-47-


hot spots. There is a definite need for more research to test the valid-
ity of this line of reasoning. However, even if repeat burglaries are
good predictors of future hot spots we still do not know why they are
more likely to be future hotspots.
The method outlined above can be used with any point data or
with polygon data. However, the amount of data available definitely
determines the level of detail that can be incorporated in the model.
Thus, in relation to methods discussed above, this approach has the
potential to be data-intensive if used at the micro level. Most data
sets that are routinely collected do not have the level of detail re-
quired to do this type of modeling. For instance, our analysis re-
quired specific characteristics of houses/properties and their sur-
rounding areas that may be difficult to obtain (e.g., housing quality,
incivilities, lighting level, etc.). The model can also be used with less
specific data sets and at a macro level of analysis, but its results are
likely to be less accurate.
In addition to having GIS skills, an analyst employing grid-based
methods must understand how to build models. The level and
breadth of statistical knowledge required will depend upon the so-
phistication of the model. In the case of the example above, the
model was designed for use by law enforcement agencies, so it was
purposely kept straightforward and simple. However, as the field
evolves, the demand for more complex and more accurate models will
continue to grow. This growth is evident in the demand for the inte-
gration of spatial statistics and GIS. If achieved, this integration
would enable an analyst to combine statistical rigor with effective
modeling of context to produce more accurate crime forecasts.

CONCLUSIONS
As mentioned at the beginning of this chapter, our purpose was to
review the current methods employed for predictive crime mapping,
from basic approaches currently used by crime analysts, to sophisti-
cated models developed by researchers. Our method was to assess
each forecasting approach on the basis of accuracy, data require-
ments, hardware and software requirements, and ease of use. What
we learned is that in many respects, this review is premature. The
more sophisticated approaches described in this chapter are still very
much in the development stages and can best be considered "alpha
versions" that have yet to be tested by the end users. Our review also
suffers from the fact that there are few published, refereed works on
these methods. Furthermore, the variability of the methods them-
selves precludes a head-on comparison of the accuracy of one versus
another. Nonetheless, this is an appropriate point in which to take
-48-


stock of both current practice and future development, as this review
may inform mid-course corrections.
Perhaps the most important finding from this review is that, while
technology has improved our ability to create, maintain and ma-
nipulate data, there is still much work to be done before we can ef-
fectively forecast crime trends. GIS has enabled the creation of geo-
graphic data (both crime and crime-related) and the integration of
data from a variety of sources. However, the most frequently used
methods for evaluating the data are the same ones that have been in
use for about 30 years. In fact, the current state of knowledge seems
to indicate that at least two of the existing and relatively straightfor-
ward methods (those based on exponential smoothing) are as effec-
tive — if not more so — than more advanced ones.
This is not to say that further examination of new methods should
be abandoned. Recent innovations have yet to gain widespread ac-
ceptance and we will not be sure of their accuracy until further re-
search is conducted. Specifically, in order to better evaluate the
methods that have been created, the field needs more head-to-head
comparisons of some of the newer, more sophisticated methods ver-
sus the more traditional, univariate ones. These comparisons must
also consider the question of what qualifies as an "accurate" predic-
tion, both in relation to the scale of the predicted area and the qual-
ity of the prediction itself. Is a 500-foot grid cell necessary, or will a
4,000-foot area (such as the one used by Gorr et al.) suffice? How
accurate does the spatial prediction have to be in order to inspire
confidence by police officers? If we can identify that there will be an
increase in crime, how accurate does the prediction really need to be
for effective intervention? Finally, how small does the unit of analysis
need to be in order to be practically useful? These are important
questions for further investigation; until they are answered, individ-
ual researchers and analysts will continue to experiment with their
own methods — possibly reinventing the wheel — rather than learn-
ing from each other.
One troubling aspect of many of the methods reviewed here is the
lack of guiding theory, which not only can help develop better models
but also helps us interpret a model's performance so that prevention
efforts are improved. The accuracy of the multivariate methods de-
pends upon the appropriateness of the variables included in the
model. Since the identification of appropriate variables is grounded in
theory, one without the other will only have limited utility. Many of
these methods are also very complicated, requiring a high level of
specialized statistical and modeling expertise as well as a large vol-
ume of data that are not often easily available on a micro level. Ulti-
mately, our goal is to find relatively a simple method that is both ac-

-49-


curate and can tell us something about why future hot spots are
likely to emerge in certain locations, so as to inform prevention ef-
forts.
Once effective crime forecasting methods are identified, perhaps
the most important challenge will be educating practitioners so they
can employ them. There is a significant leap that must take place
before crime analysts begin to use even the simplest of these meth-
ods, and this leap may well be achieved through the automation of
forecasting techniques into a user-friendly software program. A par-
allel can be drawn with the diffusion of statistical techniques and
GIS. With the advent of cheap, easy to-use software, the use of sta-
tistical software (e.g., SPSS®) increased. The same scenario occurred
with GIS software. Recent history suggests that the adoption of so-
phisticated crime forecasting software would follow the same trajec-
tory if automated tools were available. However, the development of a
new set of analytical skills will still be necessary in order to use these
new methods.
There are several conclusions to be drawn from this review of
crime forecasting methods. First, the more complicated methods are
not always better predictors. More research is needed that evaluates
the relative performance of methods. Second, many questions sur-
rounding the choices made in sophisticated models must be empiri-
cally answered before the models will accurately and consistently
perform (e.g., size of grid cell size and spatial lag). Third, additional
research is needed to identify the input variables in the multivariate
models. Choice of variables is critical to the success of the model and
must be informed by theory. Finally, the connection between the
output of models and how they translate into practice is extremely
important. In fact, perhaps the most important measure of a crime
forecasting technique may be whether it aids in crime control and
prevention.

•
Note: Points of view are those of the authors and do not necessarily rep-
resent the views of the U.S. Department of Justice or the National Insti-
tute of Justice, or the Urban Institute.
Address correspondence to: Nancy G. La Vigne, The Urban Institute,
Justice Policy Center, 2100 M Street, NW, Washington, DC 20037. E-
mail: <NLaVigne@ui.urban.org>.

-50-


Acknowledgments: We would like to thank Professor Wilpen Gorr for his
thoughtful and constructive comments on an earlier version of this
manuscript. Any errors, misrepresentations or inaccuracies are the fault
of the authors.

REFERENCES
Adams-Fuller, T. (2001). "Historical Homicide Hot Spots: The Case of
Three Cities." Doctoral dissertation, Howard University, Washington,
DC.
Anderson, D., S. Chenery and K. Pease (1995). Biting Back: Tackling Re-
peat Burglary and Car Crime. (Home Office Crime Prevention Unit
Series, No. 58.) London, UK: Home Office.
Anselin, L. (1999). Spatial Data Analysis with SpaceStat™ and ArcView®.
Ann Arbor, MI: SpaceStat (http://www.spacestat.com).
Cohen, D. Cook, W.L. Gorr and G. Tita (2000). "Spatial Analysis of
Crime." In: D. Duffee (ed.), Measurement and Analysis of Crime and
Justice, Criminal Justice 2000, vol. 4. Washington, DC: Office of Jus-
tice Programs, National Institute of Justice.
Antenucci, J., K. Brown, P. Croswell and M. Kevany, with H. Archer
(1991). Geographic Information Systems: A Guide to the Technology.
New York, NY: Van Nostrand Reinhold.
Bailey, T.C. and A.C. Gatrell (1995). Interactive Spatial Data Analysis.
Essex, UK: Addison Wesley Longman Ltd.
Becker, G.S. (1968). "Crime and Punishment: An Economic Approach."
Journal of Political Economy 76:169-217.
Bennett, T. (1996). "Identifying, Explaining, and Targeting Burglary Hot
Spots." European Journal of Criminal Policy and Research 3:113-123.
Brantingham, P.J. and P.L. Brantingham (1981). "Notes on the Geometry
of Crime." In: P.J. Brantingham and P.L. Brantingham (eds.), Envi-
ronmental Criminology. Prospect Heights, IL: Waveland Press.
and P.L. Brantingham (1978). "A Theoretical Model of Crime Site
Selection." In: Krohn and Akers (eds.), Crime Law and Sanctions.
Beverly Hills, CA: Sage Publications.
Brown, D.E. (2001). Conversation with Professor Donald E. Brown,
Chair, Systems and Information Engineering Department, Univer-
sity of Virginia. August 3, 2001.
H. Liu and Y. Xue (2000). "Mining Preferences from Spatial-
Temporal Data." Paper presented at First SIAM International Con-
ference on Data Mining, Chicago, IL, April.

-51-


Brown, R.G. (1963). Smoothing Forecasting and Prediction of Discrete
Time Series. Englewood Cliffs, NJ: Prentice Hall.
Clarke, K.C. (1997). Getting Started with Geographic Information Systems.
Upper Saddle River, NJ: Prentice-Hall.
Clarke, R.V. [1992] (1997). Situational Crime Prevention: Successful Case
Studies (2 nd ed.). Albany, NY: Harrow and Heston.
and R. Homel (1997). "A Revised Classification of Situational Crime
Prevention Techniques." In: S.P. Lab (ed.), Crime Prevention at a
Crossroad. Cincinnati, OH: Anderson.
and D. Weisburd (1994). "Diffusion of Crime Control Benefits: Ob-
servations on the Reverse of Displacement." In: R.V. Clarke (ed.),
Crime Prevention Studies (vol. 2). Monsey, NY: Criminal Justice
Press.
Cohen, L.E. and M. Felson (1979). "Social Change and Crime Rate
Trends: A Routine Activities Approach." American Sociological Re-
view 44:588-608.
Cornish, D.B. and R.V. Clarke (1986). The Reasoning Criminal: Rational
Choice Perspectives on Offending. New York, NY: Springer.
Crime Mapping Laboratory, Police Foundation (2000). "Integrating Com-
munity Policing and Computer Mapping: Assessing Issues and
Needs Among COPS Office Grantees." Final Report. Washington,
DC: Department of Justice, Office of Community Oriented Policing
Services (http://www.usdoj.gov/cops/pdf/toolbox/computer_map-
ping/e032k0046.pdf).
Dunworth, T., G. Cordner, J. Greene, T. Bynum, S. Decker, T. Rich, S.
Ward and V. Webb (1998). "Police Department Information Systems
Technology Enhancement Project: ISTEP." Washington, DC: De-
partment of Justice, Office of Community Oriented Policing Services
(http://www.usdoj.gov/cops/cp_resources/pubs_ppse/istep.htm).
Eck, J.E. (1993). "The Threat of Crime Displacement." Criminal Justice
Abstracts 25:527-46.
ESRI (1998). Working with Spatial Analyst: Three Day Course Notebook
with Exercises and Training Data. Redlands, CA: ESRI.
Farrell, G. (1995). "Predicting and Preventing Revictimisation." In: M.
Tonry. and D.P. Farrington (eds.), Building a Safer Society. (Crime
and Justice, vol. 19.) Chicago, IL: University of Chicago Press.
and K. Pease (2001). Repeat Victimization. (Crime Prevention Stud-
ies, vol. 12.) Monsey, NY: Criminal Justice Press.
and W. Sousa (2001). "Repeat Vicitimization and Hot Spots: The
Overlap and Its Implications for Crime Control and Problem-
Oriented Policing." In: G. Farrell and K. Pease (eds.), Repeat Victimi-
zation. (Crime Prevention Studies, vol. 12.) Monsey, NY: Criminal
Justice Press.

-52-


and K. Pease (1993). Once Bitten, Twice Bitten: Repeat Victimization
and its Implications for Crime Prevention. (Home Office Crime Pre-
vention Unit Series No. 46.) London, UK: Home Office.
Getis A and B. Boots (1978). Models of Spatial Processes. Cambridge, UK:
Cambridge University Press.
Gorr, W.L. (2000). "Approaches to Crime Predictive Modeling." Paper pre-
sented at annual Crime Mapping Research Conference, San Diego,
December.
(2001). Conversations with Professor Wilpen Gorr, H. John Heinz HI
School of Public Policy, Carnegie Mellon University. August 16 and
17, 2001.
A. Olligschlaeger and Y. Thompson (forthcoming in 2002). "Assess-
ment of Crime Forecasting Accuracy for Deployment of Police." In-
ternational Journal of Forecasting.
and A. Olligschlaeger (2001). Crime Hot Spot Forecasting: Modeling
and Comparative Evaluation. (Draft Final Report.) Washington, DC:
Office of Justice Programs, National Institute of Justice.
Gottlieb, S., S. Arenberg and R. Singh (1998). Crime Analysis: From First
Report to Final Arrest Montclair, CA: Alpha Publishing.
Groff, E. (1996). "Information Systems for Community Policing: A Micro-
Analysis of Crime." Proceedings of the Urban and Regional Informa-
tion Systems Association. Park Ridge, IL.
and N.G. La Vigne (2001). "Mapping an Opportunity Surface of Resi-
dential Burglary." Journal of Research in Crime and Delinquency
38(3):257-278.
J.K. Fleury and D. Stoe (2000). "Strategic Approaches to Community
Safety Initiative (SACSI): Enhancing the Analytic Capacity of a Local
Problem-Solving Effort." Washington, DC: National Institute of Jus-
tice (http://www.ojp.usdoj.gov/nij/sacsi/index.html).
Harries, K.D. (1980). Crime and the Environment. Springfield, IL: Charles
C. Thomas.
Hesseling, R.B.P. (1995). "Displacement: A Review of the Empirical Lit-
erature." In: R.V. Clarke (ed.), Crime Prevention Studies, vol. 3. Mon-
sey, NY: Criminal Justice Press.
Jefferis, E. (1998). "A Multi-method Exploration of Crime Hot Spots."
Presented at the Academy of Criminal Justice Science's Annual
Conference (http://www.ojp.usdoj.gov/cmrc/pubs/hotspot/intro.
pdf).
La Vigne, N.G. and E. Groff (2001). "The Evolution of Crime Mapping in
the United States: From the Descriptive to the Analytic." In: A.
Hirschfield and K. Bowers (eds.), Mapping and Analysing Crime
Data: Lessons from Research and Practice. London, UK: Taylor and
Francis.

-53-


and J. Wartell (2000). Crime Mapping Case Studies: Successes in the
Field (vol. 2). Washington, DC: Police Executive Research Forum.
and J. Wartell (1998). Crime Mapping Case Studies: Successes in the
Field. Washington, DC: Police Executive Research Forum.
Makridakis, S., S.C. Wheelright and V.E. McGee (1983). Forecasting:
Methods and Applications (2 nd ed.) Chichester, UK: Wiley.
Mamalian, C.A. and N.G. La Vigne (1999). "The Use of Computerized
Crime Mapping by Law Enforcement: Survey Results." (National In-
stitute of Justice Research Preview.) Washington, DC: National In-
stitute of Justice.
Mathsoft, Data Analysis Division, (1999). S-Plus Guide to Statistics vol. 1.
Seattle, WA: Mathsoft.
Morgan, F. (2001). "Repeat Burglary in a Perth Suburb: Indicator of
Short-term or Long-term Risk?" In: G. Farrell and K. Pease (eds.),
Repeat Victimization. (Crime Prevention Studies, vol. 12.) Monsey,
NY: Criminal Justice Press.
Newman, O. (1972). Defensible Space: Crime Prevention Through Urban
Design. New York, NY: Macmillan.
Olligschlaeger, A.M. (1997). "Artificial Neural Networks and Crime Map-
ping." In: D. Weisburd and T. McEwen (eds.), Crime Mapping and
Crime Prevention. Monsey, NY: Criminal Justice Press.
Ormsby, T. and J. Alvi (1999). Extending ArcView GIS. Redlands, CA:
ESRI.
Pease, K. (1998). Repeat Victimisation: Taking Stock. (Home Office Crime
Prevention and Detection Series, Paper No. 90.) London, UK: Home
Office.
and G. Laycock (1996). Reducing the Heat on Hot Victims. Washing-
ton, DC: Bureau of Justice Statistics.
Polvi. N., T. Looman, C. Humphries and K. Pease (1990). "Repeat Break
and Enter Victimisation: Time Course and Crime Prevention Op-
portunity." Journal of Police Science and Administration 17:8-11.
Rengert, G. F., S. Chakravorty, K. Henderson and T. Bole (2000). Evolu-
tion of Drug Markets: an Analysis of Susceptibility, Accessibility, Op-
portunity, and Police Action. Draft Final Report. Washington, DC: Of-
fice of Justice Programs, National Institute of Justice.
Report of the Task Force on Crime Mapping and Data Driven Manage-
ment (1999). Mapping Out Crime: Providing 21st Century Tools for
Safe Communities. Washington, DC: U.S. Department of Justice and
National Partnership for Reinventing Government (http://
www.npr.gov/library/papers/bkgrd/crimemap/content.html).
Rich, T. (1995). The Use of Computerized Mapping in Crime Control and
Prevention Programs. (National Institute of Justice Research in

-54-


Brief.) Washington, DC: Office of Justice Programs, National Insti-
tute of Justice.
Sherman, L.W. and D. Weisburd (1995). "General Deterrent Effects of
Police Patrol in Crime 'Hot-Spots': A Randomized Controlled Trial."
Justice Quarterly 12:625-648.
- P.R. Gartin and M.E. Buerger (1989). "Hot Spots of Predatory Crime:
Routine Activities and the Criminology of Place." Criminology 27:27-
55.
Spelman, W. (1995). "Criminal Careers of Public Places." In: D. Weisburd
and J. Eck (eds.), Crime and Place. (Crime Prevention Studies, vol.
4.) Monsey, NY: Criminal Justice Press.
Townsley, M., R. Homel and J. Chaseling (2000). "Repeat Burglary Vic-
timisation: Spatial and Temporal Patterns." The Australian and New
Zealand Journal of Criminology 33(l):37-63.
Weisburd, D. and T. McEwen (1997). "Introduction." In: D. Weisburd and
T. McEwen (eds.), Crime Mapping and Crime Prevention. (Crime Pre-
vention Studies, vol. 8.) Monsey, NY: Criminal Justice Press.
Wilson, J.Q. and G.L. Kelling (1982). "Broken Windows: The Police and
Neighborhood Safety." The Atlantic Monthly 249(3):29-38.

NOTES

1. The National Institute of Justice has funded a number of grants to
develop predictive models, drawing on spatial regression analysis, envi-
ronmental modeling, neural network analysis and other methods, and
having the capability of being displayed within a GIS (www.ojp.
usdoj.gov/cmrc).
2. This paper is limited to predictions of geographic concentrations of
crimes; assessments of "geographic profiling" and other predictive meth-
ods to identify the location and time of future serial crimes are beyond
the scope of this paper.
3. A GIS consists of hardware, software, and peripherals to create, store,
analyze and output geographic data.
4. The term geocode refers to the process of assigning coordinates on the
earth's surface to an address or some other location identifier (e.g. zip
code, census tract etc.).
5. The terms raster and vector refer to the two main types of geographic
information system data models. In a vector GIS all features are repre-

-55-

Elizabeth R. Groffand Nancy G. LaVigne

sented as points, lines or polygons. While a raster GIS is grid cell based.
For more information on basic GIS concepts please see K.C. Clarke
(1997) and Antenucci et al. (1991).
6. The neighborhood size can be defined by distance from the target cell
(e.g., 200 feet) or by specifying a number of cells to use (e.g., 10 cells).
7. Spatial Analyst is a raster GIS extension to ArcView. Version 2.0 of the
software includes a visual model builder that allows models to be created
and saved so that parameters can be changed and the model run again
automatically (i.e., in a "batch" mode). This is a vast improvement over
previous versions, for which each layer had to be created in raster form
and recreated every time the analyst changed a parameter of the model.
8. It should be noted that this assessment was based upon an examina-
tion of hot spots at specific types of locations — high schools, public
housing projects, subway stations, and parks and playgrounds — rather
than all hot spots distributed throughout the study area.
9. Persistence was defined as the intersection of five or more hot spot
ellipses (based on annual data) over a ten-year period.
10. The software began as a DOS-based program, and was under the
auspices of the Illinois Criminal Justice Information Authority (ICJIA).
The Spatial and Temporal Analysis of Crime (STAC) is available from the
ICJIA's website (http:// www.icjia.org/ public/index. cfm?metaSection=
Data&metaPage=StacFacts).
11. CrimeStat is a suite of spatial statistical software tools that is avail-
able from NIJ's Crime Mapping Research Center at http://www.ojp.
usdoj.gov/cmrc/tools/welcome.html#crimestat.

12. Three free software programs that contain a hot spot routine using
ellipses are ReCAP (http://www.sys.virginia.edu/research/crime.html),
RCAGIS (http://www.databasefiles.com/_dbf/0000007e.htm), and Crime
Stat (see footnote 11).
13. The random walk method simply uses the current time period to pre-
dict the next time period. One example of this method is when the num-
ber of Burglaries in May is used to predict the number of Burglaries in
June.
14. In the Naive Lag 12 model the value for the same time period in the
previous year is used to predict the value for the current year. Using the
above example, Burglaries in June of 2000 would be used to predict Bur-
glaries in June of 2001.

-56-


15. Classical decomposition is used to calculate a seasonal index for
each month in relation to other months in the series using multiplicative
adjustments to the trend model.
16. Exponential smoothing methods are typically used to forecast short-
term changes in a series and to try and balance sensitivity to structural
changes against accuracy of forecasts. Simple exponential smoothing
utilizes past values of a series and averages them with exponentially de-
creasing weights. This method takes into account both the trends and
variability in a data series. Recent values in data series with a high de-
gree of variability receive less weight than recent values in series that
show a definitive trend. This method "smoothes" out variability in the
data series and makes it easier to distinguish actual trends in the data.
Another method is Holt's Two-Parameter Linear Exponential Smoothing
method (Gorr et al., 2002; Gorr and Olligschlaeger, 2001), which inte-
grates a smoothed time trend model with the generation of two smooth-
ing constants (one for level and one for trend) to create a more accurate
forecast of crime. The smoothing constants for both techniques are gen-
erated using a grid cell search.
17. Pooling involves combining the data for areas and then computing a
common index across all those areas. Pooling provides more homogenous
data that, in turn, improve the accuracy of predictions. In Gorr et al.'s
(2002) tests they identified the best model for predicting crime as the one
that used classical decomposition to calculate pooled seasonal indices by
crime type for the whole city (Gorr et al., 2002; Gorr and Olligschlaeger,
2001). Gorr (2001) notes that one potential enhancement to the pooling
method he used in his previous research would be to pool the crime data
by type of land use (e.g., several categories of residential vs. commercial,
etc.). This method has both enough homogeneity in the data and a large
enough sample size to get more accurate predictions than those achieved
when the pooling was by crime type across the city.
18. Gorr and Olligschlaeger (2001) identified 25 measures of drug calls,
17 measure of property crime and 27 measures of violent crime that they
used as leading indicators in their work (Figures 4-6).

19. For more information about incorporating spatial and temporal lag
techniques into research see Anselin, 1999; Bailey and Gatrell, 1995 and
Mathsoft, Data Analysis Division, 1999.
20. For instance, we accounted for the influence of nuisance violations in
neighboring grid cells by having the values in those cells be used in the
calculation of the target cell.

-57-

Predictive analysis of crime forecasting

More Related Content

What's hot (19)

Similar to Predictive analysis of crime forecasting (20)

More from Frank Smilda (20)

Predictive analysis of crime forecasting