11.1 Case-Based Reasoning

In many ways, the process of criminal investigative analysis, or behavioral profiling of violent crime, truly is an amazing example of data mining and predictive analytics. At first, the process seems like magic. When confronted with a jumbled mass of clues, the investigator is able to identify a likely suspect, like a magician pulling a rabbit out of the hat. On further investigation, though, it becomes clear that are two important elements functioning in this investigative process. The first is good case management. By reviewing the case thoroughly, items that have been overlooked or lost can be identified and addressed. Similar to identifying where missing puzzle pieces belong, this process in and of itself can result in tremendous clarification in the direction of an investigation and significant progress toward closing the case. It allows the investigator to reveal the big picture and identify any readily apparent trends and patterns. In many ways, this is similar to the process of data cleaning and management. By thoroughly reviewing and organizing the case, missing details or data are identified and, if possible, addressed.

Similarly, using data mining and other automated methodologies can prompt consideration of evidence in a different light. By distilling the collected narrative information through a categorical filter, the information can be further characterized and categorized into smaller, readily identifiable groups that have value from an investigative perspective. How does this work in the field? Overkill, as discussed above, and many other factors considered during investigation can be broken down into “yes/no” or dichotomous decisions. Using this type of approach, decision trees can be developed that guide the investigative process, which ultimately can be used in automated motive determination algorithms and even strategic characterization models of likely suspect attributes. These and other data management techniques can add more structure and meaning to investigative information.

The second key piece in the behavioral analysis process is pattern recognition. While this might sound easy, any good investigator knows that it can be much more difficult than it sounds. Investigators who are particularly good at this task are those able to move beyond some of the obvious details and identify underlying themes and patterns. In addition, they will begin trying to match these elements in the current case to their internal database of previously solved cases in an effort to identify a common theme or pattern that might give them some insight. This process of comparing current evidence against historical memory to see what fits can be described as case-based reasoning. Case-based reasoning is a model of learning in which people comprehend new experiences within the context of previous ones,⁴ a process that can be modeled using expert systems or artificial intelligence.

In some ways, a good investigator can be thought of as a closet analyst. The process and approach associated with criminal investigations is one of the most analytical aspects of law enforcement. Attributes that are associated with a good patrol person or an excellent member of the SWAT team are very different from those factors that are associated with a good investigator. The comparison between the Hawaii Five-O and Columbo detective shows provides a nice parallel. McGarrett and Danno are out there in the thick of things, catching bad guys, while Detective Columbo seems to be stumbling around in a clumsy fashion. Yet the viewer can almost see Columbo’s wheels turning and watch him mentally chewing on the evidence as he chews on his cigars, trying to make the pieces fit in an effort to understand the puzzle. In addition, Columbo always was “bothered by” the pieces that did not fit, highlighting gaps in an alibi or oddities in the crime scene. In many ways, Columbo is a walking example of data mining, case-based reasoning, and anomaly detection rolled into one baggy trench coat. Again, not very elegant, but it works.

Cold case investigation can be seen as an extreme manifestation of the analytical side of the investigative process. These very special investigators and analysts are truly unique in their ability to move beyond the superficial nature of the evidence and look for the underlying form or commonalities with other solved cases in the investigator’s repertoire. While this may sound almost like a Zen-like approach to investigation, much of it comes back to excellent case management, anomaly detection, and some superb case-based reasoning.

Replicating this process through the use of computer programs or artificial intelligence algorithms certainly confers some benefits. First, a computerized database has the ability to store and quickly recover a large amount of information, with an associated capacity that extends well beyond the memory storage and retention abilities of a human. Expert systems have the ability to store, retain, and simultaneously consider information from more cases than any single investigator is likely to encounter in his or her career. In addition, the information is not subject to the same memorial decay and errors in interpretation over time that can occur with the human memory.

Second, an expert system theoretically has no bias—bias that can favor a particular scenario and color the course of future evidence collection and investigative process. It is not swayed by the suspicious boyfriend, or by the shifty looks and dubious alibi given by the jilted lover. Similarly, it is not likely to even subtlely discard or overlook evidence that is contrary to a favored hypothesis or outcome. This certainly is not to suggest any lack of professionalism or ethics on the part of sworn investigators. They are some of the most honor-bound, committed professionals in law enforcement. Rather, human investigators bring their internal norms, life experiences, and feelings to every case. While this can make them incredible champions for the victims and their families, it also shaves away some of their objectivity. It is almost impossible to look at the face of a dead child and remain objective; however, even a little compromise in objectivity can be associated with a concomitant decrease in efficacy. Computers, on the other hand, do not care.

Finally, expert systems are not bothered by crime scene details; they can get beyond the “yuck” factor that occasionally can catch even the most seasoned investigator. Certain victims and certain scenes affect investigators in different ways. Those that are affected by everything that they see are limited in their ability to effectively investigate violent crimes. On the other hand, those who do not respond to anything probably have been on the job too long and should consider a change of pace in an effort to remain effective and retain their humanity. In short, even the most staid investigators are likely to be bothered by something that they encounter. This limits their objectivity and can compromise their ability to reduce the crime to common elements that can be characterized and compared to previous cases. Artificial intelligence systems truly embody the philosophy of “just the facts, ma’am” because they have no capacity for anything else (at least at the time of this writing).

Does this mean that we should discard detectives in favor of computer modeling programs and expert systems? Absolutely not! Expert systems, no matter how “smart,” lack one element critical to effective and meaningful data mining: domain expertise. Without an “expert” in the field to evaluate the nature of the evidence and information collected and to evaluate critically the value and validity of any created models, the risk for significant errors in logic and interpretation would seriously limit our ability to use these tools. Moreover, it is absolutely essential that professionals within law enforcement and public safety minimally have significant involvement in the analytical process. The best scenario would be one in which analytical and investigative personnel work together in the data mining and modeling process, perhaps with some outside help and support from statistical experts. It bears repeating that it seems to be relatively easy to teach data mining and predictive analytics to law enforcement personnel. These folks know where the data come from and how the models will be used. They intuitively know what is available for models and when. As such, they are much less likely to make critical errors that result in the creation of a model based on circular logic, which essentially requires information that is dependent upon the output as an input or indicator variable. Many investigators are natural data miners, given the intuitive nature of their work, and seem to embrace the approach when given the opportunity. On the other hand, trying to convey the internal norms, historical knowledge, and accumulated domain expertise from an active investigative career to data mining experts has proven to be extremely difficult.

In sum, data mining and predictive analytics can enhance the investigative process, particularly through many of the automated pattern recognition programs and scoring algorithms. However, the use of expert systems alone has limited value and could significantly compromise the investigative process.

11.2 Homicide

At its most basic, homicide as a crime can be categorized and divided in many ways, most frequently based on victim-perpetrator relationship and motive.⁵ Figure 11-3 depicts a possible decision tree for murder. The first branch of the tree is divided by three possible victim-perpetrator relationships: familial or domestic, acquaintance or known, and stranger. The second division in the branches includes possible motives associated with stranger murder. Finally, within the various motives, it is possible to characterize the motive even further. In this case, the drug-related homicides have been divided using the Tripartite Model of Drug-Related violence developed by Goldstein.⁶ Briefly, this model describes three types of drug-related violence: Psychopharmacological, economic compulsive, and systemic violence, which are related to the different role that drugs and/or the drug-related lifestyle may play in violent crime.

Figure 11-3 Possible decision tree for categorizing murder.

While the Goldstein model might be helpful for counting crime or determining the proper focus for needed crime prevention strategies, it frequently adds little value to the investigative process because the nature of the victim-perpetrator relationship and true motive generally are not confirmed until a suspect has been identified and apprehended.

Information collected during the investigative process can be subdivided and used to determine the nature of the crime and possible suspects. As can be seen in Figure 11-4, much of the information collected and considered in the investigative process can be reduced to simple binary choices or sets, reinforcing the fact that data preparation is one of the most important steps in the data mining process.

Figure 11-4 Possible decision tree for categorizing behavioral and forensic evidence.

While startling to many, violent crime can be very homogeneous when viewed in the proper light. One feature of the behavioral analysis of violent crime is the characterization of an incident based on common elements. These include victim as well as scene characteristics, which ultimately are associated with common perpetrator characteristics. The goal is to develop a profile or model of the type of person likely to have committed the crime, based on the known characteristics of the crime. In many ways, the behavioral analysis of violent crimes is similar to algebra. Ultimately, the objective is to solve for “X,” the suspect’s identity, which is done by systematically revealing and examining the other elements of the equation.

Whether consciously or not, the death investigator goes through a series of “yes or no” questions, often as early as when the first call comes into the dispatch center. This information is used to begin characterization and categorization of the crime based on characteristics of the victim(s) and crime scene(s). These include, but are not limited to the following:

Was there more than one crime scene? Was the victim moved or the body dumped?

Was it a weapon of opportunity, or did the suspect bring the weapon to the scene?

Was the victim at high risk for violence, or did the suspect assume risk in selecting this particular victim?

The list goes on, but a considerable amount of information can be described as binary in nature, meaning that it can be answered by a simple “yes” or “no.” Other information can be divided into categories or sets (e.g., race)—information ideal for data mining and predictive analytics.

These similarities can be used to link similar crimes perpetrated by the same individual or group of individuals. They also can be used to link similar crimes perpetrated by similar individuals or groups of individuals. In the former case, it is important to link crimes in a series because, like in a puzzle, individual elements present in one incident that were missing or overlooked in other incidents can contribute to a greater understanding of all crimes in the series. In other words, the whole often is greater than a sum of the parts. It also is important to identify all of the crimes associated with a particular individual for prosecutorial purposes. Presenting a linked series can be helpful in jogging the memory of a reluctant suspect.

Associating a crime or series of crimes to known suspects associated with solved cases can be helpful in at least two ways. First, it can provide guidance as to the type of individual likely to have perpetrated a particular crime or series. In contrast to television and movie portrayals of “profilers,” the behavioral analysis of violent crime does not identify a specific individual. Rather, it associates a crime or series with perpetrator characteristics that can be used to guide the investigation. By relating a current investigation to an earlier, solved case, additional elements can be highlighted (e.g., things that you might want to look for, why they did something, and investigative strategies, particularly interviewing).

11.3 Strategic Characterization

Frequently in an actual investigation, the nature of the victim-perpetrator relationship is unknown. It is the crime scene characteristics and victim lifestyle factors that suggest a possible motive, which then is used to identify possible suspects. In fact, it is this type of strategic characterization of likely suspects that is embodied in the behavioral analysis process.

Some of our earliest work using advanced statistics for characterizing violent crime involved the development of automated motive determination models.¹¹ Again, this work arose out of some lively discussions regarding whether it is possible to accurately model violent crime using automated methods. The cases used for this study included 25 juvenile murderers incarcerated in the Commonwealth of Virginia Juvenile Correctional Centers from February 1992 to July 1996. Information pertaining to the victims, suspects, injury patterns, and the behavioral characteristics of the assault were identified and analyzed in an effort to determine whether it was possible to determine the motive using automated methods.

For the first analysis, all of the information that could possibly be obtained was put into the model. The analytical approach selected was discriminant analysis, which is a classification modeling algorithm. These tools are covered in greater detail in Chapter 7; in brief, discriminant analysis is a multivariate statistical approach that can be used to identify factors that are useful in determining group membership. Generally, one of the assumptions with the use of discriminant analysis is that the variables used to create the model are continuous. While some of the data available for this study were continuous (e.g., age), most of the relevant information was either binary or categorical. As a modeling tool, however, discriminant analysis is robust enough to accommodate a violation of this assumption. Basically, the type of error most likely to occur with a violation of this type is that a model will not emerge, even though there might be a relationship in the data.¹² The issue of errors and how they should be evaluated is addressed in Chapter 4. In a motive determination model, any error that would misdirect an investigation could potentially be catastrophic, as it would waste investigative resources. Moreover, delay in the investigative process could seriously compromise the ultimate ability to solve the case. Therefore, failure to identify a model due to the violation of the data type assumption was determined to be acceptable after thoughtful consideration of the ultimate role for any potential models.

The initial results were extremely promising. Using information related to recent victim drug use and suspect substance use history, it was possible to accurately categorize 85% of the cases as drug-related or not drug-related. One interesting finding regarded the direct relationship between the suspects’ use of illegal drugs, particularly marijuana, and their involvement in a drug-related homicide. Our earlier work had confirmed conventional wisdom among narcotics detectives: Most successful drug dealers do not use the drugs that they sell.¹³ In fact, substance use has been associated with an increased risk for firearms assaults among drug sellers. Whether our finding was a cause or a consequence was not entirely clear, but substance use was not a healthy choice, particularly among those involved in the sale and distribution of illegal narcotics. Therefore, this result was somewhat surprising. Further examination revealed that the suspects’ substance use generally involved marijuana, while the recent victim use included cocaine and opiates (e.g., heroin). This finding was consistent with a model of drug-related violence that proposed different subtypes of drug-related violence associated with different types of suspect, similar to a division of labor within drug distribution networks.¹⁴ In particular, the suspects who emerged in this preliminary study were very similar to the “enforcers” described in this model.

The end result of this preliminary study was support for the notion that violent crime could be modeled using advanced statistics, as well as additional knowledge regarding our understanding of drug distribution networks and the proposed division of labor. From an investigative standpoint this was somewhat helpful, in that it offered additional information pertaining to the types of individuals likely to commit drug-related violence; however, it provided little more in the way of enhanced investigative efficacy. In other words, the results of this study indicated that recent victim use and suspect use were good indicators of drug-related violence. From an investigative standpoint, this creates circular logic: To identify the suspect, it is helpful to know the motive; to determine the motive, we need information regarding the suspect’s substance use; to know the suspect’s substance use habits, we need to know who the suspect is, which is the original question. Clearly, further work was needed.

11.4 Automated Motive Determination

More recent work in this area has been confined exclusively to the information that is available early on in an investigation. Again, drug-related violence represents a good area of study for several reasons. First, drug-related violence frequently drives the overall violent crime trends in many communities plagued by serious increases in violent crime;¹⁵ the homicide rate often rises and falls as a direct consequence of the drug-related homicide rate. Therefore, addressing drug-related violence can significantly reduce the violent crime rates in these communities.

Drug-related crimes, particularly violent crimes, can be difficult to solve in a timely fashion. Witnesses may be reluctant to come forward or may be unreliable. In some cases, drug-related violence is seen as “part of doing business,” reducing sympathy for the victim and making others less likely to get involved. The more time that elapses, however, the less likely it is that a murder will be solved. If significant progress is not made in the first few days of an investigation, it becomes increasingly unlikely that the case will be solved. Finally, and perhaps most importantly, drug-related violence appears to be relatively homogeneous and amenable to modeling, which makes it a good candidate to evaluate automated motive determination scoring algorithms.

Using automated classification techniques similar to those used earlier, drug-related homicides were analyzed in an effort to develop a model that could be used to automatically determine a motive using information available early in an investigation. The relevant variables that emerged were location, location, location. While this is an overly simplistic interpretation of the results, the most salient fact to emerge from this analysis was that certain areas were associated with an increased risk for drug-related violence (Figure 11-5).

Figure 11-5 Drug-related violence models can be deployed through the use of mapping tools. These facilitate the development of proactive deployment strategies. This particular map highlights differences in victim characteristics, which might have additional value from an operational perspective.

The results were distilled and deployed through maps, which allowed for proactive, risk-based deployment specifically targeting drug-related violence. One interesting finding that emerged with the maps was that one particular area was associated with victims who were more likely to be employed. Further analysis of the location revealed a drug market frequented by users from the surrounding localities, which was consistent with employed victims. While this initially appeared to be nothing more than an interesting factoid, the operational implications quickly became apparent. Drug violence associated with employed victims was consistent with victims who were buyers. One way to reduce the violence in these areas was to reduce the number of potential victims. Demand reduction approaches offered the perfect solution to this finding, while different approaches could be used in other markets associated with different types of drug-related violence.

The opportunity to more fully characterize crime and tailor specific crime reduction approaches based on a thoughtful analysis of the problem is one powerful benefit of using data mining and predictive analytics in the development of operational strategies. Just as we cannot please all of the people all of the time, there is no universal “crime reduction” strategy that will work for every situation all of the time. While this seems entirely logical, it is very difficult to even begin to match an appropriate operational plan to a particular issue if the nature of the problem has not been characterized and defined. For those who would say, “Utilize, don’t analyze,” this approach offers a solution for evaluating operational approaches in a meaningful way—something that increasingly has become a requirement as resources are increasingly limited.

When the time of day, day of the week, and victims’ employment status were included in the analysis, the accuracy of the model increased concomitantly. Again, this model included only that information available early in an investigation; however, when the new information was added to the model, it became too opaque to have any value in the deployment process. Therefore, these results were deployed through a web-based package, which provides for 24/7 analytical capacity.¹⁶ As most homicide detectives can attest, people generally do not kill each other when it is convenient for us, particularly when drugs are involved. Most analytical teams, however, work normal business hours. By deploying algorithms through these web applications, operational personnel have the benefit of analytical help when they most need it: early in an investigation, when the case is progressing most rapidly.

11.5 Drug-Related Violence

Like homicide in general, drug-related violence can be categorized and divided further, based on the specific motive for the violence. When asking why this is relevant, it is important to remember that the more an investigator knows about what happened and why, the higher the likelihood that a suspect will be developed, effectively investigated, and prosecuted.

In his tripartite model of drug-related violence, Dr. Paul Goldstein¹⁷ has proposed three categories of drug-related violence. A fourth type of drug-related violence also should be considered that has been described as the “outlaw lifestyle.”¹⁸ Briefly, this type of drug-related violence is related to the fact that drug dealers, like police officers, generally are “on the job 24/7.” What this means is that to function effectively in an extremely predatory environment, the individual needs to always be “on.” A drug dealer involved in a violent drug market is going to have difficulty maintaining his safety when he is selling if he is perceived as being weak in other domains of his life.

A second benefit of analyzing violent crime in this fashion is that the outcomes can be used to guide the development of enforcement strategies. For example, a finding that in a certain area associated with open-air drug markets, most of the victims of drug-related homicides were employed is consistent with drug users being killed. A possible enforcement strategy in this situation could include aggressive demand reduction techniques in an effort to keep potential victims out of harm’s way. On the other hand, victim characteristics consistent with dealer rivalries or gang disputes over markets or territories would require a different approach.

11.6 Aggravated Assault

Many aggravated assaults can be viewed as incomplete or poorly planned homicides. Similarly, the line that separates a lethal from a nonlethal assault often reflects the quality of medical care or timeliness of response, rather than some specific intent on the part of the suspect. With that in mind, homicides and aggravated assaults can be viewed more as a continuum, rather than as two separate and distinct entities. When we explore this a little further, it makes sense based on what we know of violent crime. For example, the number of innocent victims associated with drive-by shootings supports the fact that drive-bys frequently do not go according to plan. Vehicle movement, unreliable weapons, and random bystanders all contribute to the variability associated with this type of violence. Similarly, it is not unusual for there to be several nonlethal assaults that precede a domestic homicide. Therefore, many of the same approaches to the analysis of murder can be applied to nonlethal assaults. One important difference from a data mining perspective, however, is that there usually are more nonlethal assaults than homicides. While this is a good thing for a community, it also is a very good thing for the analyst, as there generally are more incidents available for analysis and modeling.

11.7 Sexual Assault

Stranger rapists can be some the most disturbing predatory criminals a detective will encounter. Even the hint of a serial stranger rapist can create a climate of fear in a community. Several years ago, during a casual conversation, Dr. Paul Ferrara, the director of forensic scientists in Virginia, noted that a surprising number of the DNA “cold hits” for predatory sex crimes had come from criminals without prior histories of sex offending. At the time, Virginia was noteworthy for having created a very successful offender DNA database with broad inclusion criteria. While some states were confining their samples to convicted sex or violent criminals, the Commonwealth of Virginia obtained DNA samples from all convicted felons.

Dr. Ferrara noted that several of the offenders identified by cold hits were associated with prior property crimes, particularly burglaries. In some ways, this finding was not surprising, in light of what we know about sexual predators and some violent criminals. In reviewing violent crimes, it is not unusual to find a pattern of escalation that includes crimes that do not appear violent initially. For example, Timothy Spencer, the “Southside Strangler” and first person convicted with DNA evidence, had a history of burglaries in northern Virginia that preceded the homicides that he committed later. Similar cases have revealed a pattern of burglaries or trespassing that preceded escalation into more serious patterns of offending.

An initial review of the data revealed that approximately 40% of the cold hits were associated with offenders who had no documented history of either sex crimes or violent offending. Perhaps more importantly, had the database been confined to these patterns of offending, approximately 40% of the criminals might not have been caught and thus would have been allowed to continue to prey on their communities.¹⁹

Exploring this finding offered at least two benefits. First, the ability to characterize and identify patterns of offending that indicate an increased likelihood for escalation offers the promise of early detection, enhanced investigative efficacy, and increased community safety. This also creates increased opportunities for early detection and intervention for sex offenders, who have a pattern of offending that is noteworthy for its high recidivism rate and resistance to treatment. Second, an increased understanding of sex offenders, how they escalate, and how they prey on the community gives behavioral scientists an opportunity to better understand this particularly challenging form of criminal behavior. This greater understanding offers the promise for early intervention and the concomitant increase in public safety.

The initial study involved reviewing large correctional databases. Using discriminant analysis, models were created to determine which factors were predictive of subsequent stranger rapes. Not surprisingly, prior offense history reliably emerged as the most predictive variable. What was a shock, however, was that a prior property crime actually was a better predictor for a stranger rapist than a prior sex offense. It is important to note that discriminant analysis ideally is used with continuous variables, while offense history frequently is confined largely to categorical variables. Because discriminant analysis is such a robust statistical test,²⁰ however, it was permissible to violate this assumption. The most likely error to occur when this assumption is violated would be a failure to identify a model, which was acceptable given the nature of the question.

Subsequent manual review of the paper files associated with these offenders revealed some differences in the nature of the property crimes they perpetrated. In several cases, these offenders appeared to specifically target occupied dwellings. When someone was home at the time of the break-in, it frequently was a female alone, or a female with small children. In addition, these offenders frequently took items of little or no value, if they removed anything at all from the residence. This behavior is inconsistent with a purely economic motive for the crime.

In many ways, these crimes differed qualitatively from traditional burglaries. Reduced to its simplest form, a burglary is an economic crime in which the offender tries to maximize the yield while managing the risk of being caught. The individuals associated with a subsequent stranger rape distinguished themselves from “normal” burglars in that they incurred a greater degree of risk in preferentially targeting occupied dwellings, and generally had little to show for their efforts other than a few items that could be viewed as souvenirs. In other words, their crimes were abnormal.

The discovery and confirmation processes associated with data mining expanded our understanding of stranger rapists. This new understanding was used to generate a brute force anomaly detection system for identifying crime patterns and trends determined to be at increased risk for escalation. Perhaps more importantly, it also resulted in the development of a general principle regarding “normal” crime. Since the original study, this concept has been applied successfully to other nonviolent crimes that deviate in some way from “normal.” In other words, anything that suggests some type of secondary gain beyond or instead of the economic motive generally associated with the crime is cause for concern because it frequently indicates potential for escalation or more serious patterns of offending. Similarly, any preferential behavior on the part of the offenders that increases their risk of detection or apprehension also is cause for concern.

In England, the West Midlands police have conducted very successful work using predictive analytics to characterize and apprehend sexual predators. Through the use of self-organizing maps or clustering algorithms, they were able to identify clusters or groups of crimes that were similar. These similarities were based on a variety of relevant dimensions, which included method of approach, verbal themes, precautions taken to prevent detection, and victim characteristics.²¹

This work is particularly encouraging because many of the category clusters identified in this study match classifications described previously in the United States.²² This similarity suggests commonalities in some patterns of offending that might transcend national and perhaps even cultural boundaries. Therefore, unlike some business, health care, and educational applications of data mining and predictive analytics, work in the public safety and intelligence arena promises to transcend national boundaries. This further increases the number of potential end users for predictive algorithms, while enhancing the opportunities for increased resource exploitation of criminal justice data resources and predictive algorithms.

11.8 Victimology

In some ways, victimology, or the study of the relationship between certain individual attributes and the risk for violent crime, can be seen as a logical extension of risk assessment. Examining victim characteristics can be used in two ways. First, it frequently has value from an investigative standpoint because identification of victim lifestyle issues or risk factors often can suggest a possible motive and, by extension, a likely suspect.

On the other hand, some insight regarding potential risk factors associated with violent crime opens the door to meaningful, specifically targeted prevention strategies. Many victims of violence suggest that their injury was the result of “bad luck” or that they were merely in the “wrong place at the wrong time.” If that was the case, there would be very little that we could do to reduce violent crime, other than identify the perpetrators and get them off the streets as quickly and for as long as possible in an effort to reduce crime. Crime prevention would rely almost exclusively on investigative efficacy.

We have come to realize that certain victim attributes can increase or even be related directly to an individual’s risk for violent victimization. For example, prostitutes are at increased risk for sexual assault and other violent crimes due to their involvement in that pattern of criminal activity. Similarly, substance users are at increased risk for violence related to their involvement with illegal drug markets. These frequently are referred to as “lifestyle” factors, meaning that some aspect of the victim’s lifestyle increased his or her risk for violence.

This is a very important point, because if violent victimization truly is related only to “bad luck,” we are limited in our ability to prevent violent crime. But if we can identify particular behaviors or activities that are linked to an increased risk for violent crime, then we have an opportunity to try and change those behaviors, which might concomitantly reduce the risk for violent crime. By reducing the risk associated with some of these individuals, the amount of lead flying through the air also will be reduced, which will reduce the “bad luck” associated with those individuals who truly were in the wrong place at the wrong time. By drilling down into the data, it is possible to identify certain groups that are at greater risk for violence than others.

11.9 Moving from Investigation to Prevention

Improving investigative efficacy is very important, but what if violent crime could be characterized, anticipated, and even predicted? If this were possible, then we would have an opportunity to engage in proactive strategies that would prevent crime before it happened. In death investigations, the act already has been committed, but what if it was possible to anticipate who was next? Minimally, the high rate of subsequent assaults and murders documented in victims of violent crimes²⁴ identifies them as a group at extraordinarily high risk for violent assault and murder. It is unknown whether this increase is associated with the idea that each aggravated assault really represents an incomplete or poorly planned homicide, or because the same lifestyle factors that resulted in the first assault are still present. By identifying who is at risk, where, and why, traditional enforcement strategies can be matched and targeted to specific patterns of risk. Through the use of predictive analytics to create a rule set for drug-related homicides, for example, it could be determined that victims killed near a particular drug market were more likely to be employed. Additional information may indicate that this particular market was frequented by younger, relatively affluent, recreational drug users from neighboring communities. One possible approach in this situation would be a demand reduction strategy such as “reversals,” where police officers play the role of drug dealers in an effort to identify and arrest users, ultimately keeping the potential victims away from a dangerous activity or market. A different approach might be warranted for drug-related homicides involving individuals associated with other illegal drug markets. Identification, characterization, and modeling of victim risk factors represent another use for data mining and predictive analytics in our efforts to reduce violent crime.