Can you tell me a little more about the hidden patterns you're investigating? Are you hoping to find some correlation between demographics of offenders and type of criminal act or features of the crime itself or are looking at this longitudinally for developing and shifting crime trends?

Could you tell me a little more about the hidden patterns you are investigating?  Are looking to find correlations between offender demographics and type of crime or features of the crimes or are you looking at it longitudinally for developing and shifting crime trends?

Yes, I am trying to find the associations between the crimes and the offenders. The ultimate goal is to conclude whether a particular crime has been committed by a suspect. 

Thanks

From a legal standpoint, evidence of a defendant's character or past crimes is not admissible in court to prove the person acted in accordance with the character trait (see Rule 404, Federal Rules of Evidence or Rules of Evidence in the appropriate jurisdiction).     

I don't think you can conclude whether a suspect has committed a crime.  Fundamentally, the determination of whether some committed a crime is factually determined by a jury.  Even a determination of guilt only satisfies the legal standard.  A verdict (either guilty or not guilty) fails to satisfy the moral determination of innocence.  Ultimately, your question settles on a rule of thumb I was taught by an older lawyer many years ago when I first started practicing:  "There's only 2 ways to ever really know if someone's innocent: 1. You saw them do it; or, 2. You did it.  Anything else is open to question. [NOTE:  Even a confession can be suspect.  Is the suspect protecting someone? Is he fearful for his safety? Etc.]

If we could determine guilt by an empirical model, we could do away with the jury system and allow an algorithm to make the determination.  Think about the reasons we have not chosen that route.

I don't know whether I've answered your question.  I hope I haven't discouraged you completely.  I offer these observations only as a means of challenging you to think more about your premises and thesis.

Yes. As you mentioned the final results of a computational model will always be subjected to cognitive biases and so, a computational method can never be trusted.

But the research is based on not concluding a particular person as an accused but as a suspect, so that it would help the investigators to narrow down their search. 

See the works of spesialists from Russia.

From your explanation, I am envisioning that your have data on 52,000 criminal incidents as well as data on the offenders, and are trying to determine if certain types of offenders (or those with certain characteristics) are more likely to commit (or have committed) certain crimes.  For example, perhaps after examining all these cases and the offenders you find that rapists are statistically more likely to have come from a single parent home, tend to minimally employed, but have higher than average intelligence, while armed robbers tend to have had problems in their school years, suffered physical abuse in the home, and over 10 juvenile offenses.

Are you trying to determine that given a type of crime (burglary) that there are some characteristics that are associated with committing that type of offense?  I’m not sure what cognitive biases you might encounter if you are using some very concrete measures, for example the number of prior minor offenses, the number of prior serious offenses, the number of juvenile offenses, single or two parent home, number of school disciplinary incidents, results of psych and personality evaluations, percentage of year fully employed, etc.  The completeness in which you consider other possible correlates might give rise to questions of cognitive bias if reviewers feel you missed some of the important correlates of criminogenic behavior.  A thorough review of the criminological theory literature should provide you with enough variables to be confident your correlates are clearly associated with criminal behavior.  As you are looking for hidden patterns, I’m assuming that you hope to find previous unknown correlates, or unique correlates to certain offenses.  You could take a grounded theory approach and not forward a hypothesis as to what possible hidden patterns you suspect and instead let the statistical evidence demonstrate the significant associations, rather than let claims of cognitive bias taint the objectivity of your exploration.

Maybe you could tell us a little more about the data and your approach you are using, to better inform an approach to the literature.  Obviously, guilt can not be determined from characteristics that are indicative of the likelihood of offending but if you are looking more at profiling there should be some good literature out there regarding serial rapists and killers, and drug smugglers, but I'm not sure if there is anything that examines profiling for run of the mill felonies.

Frank makes some excellent observations.  Very good alternative view of the question.

You might also want to check into some of the literature by Wright and Decker, qualitative approaches as to how robbers and burglars operate, the mentality, MOs and some of the commonalities.  This might demonstrate some patterns and characteristics particular to these types of offenders.

Wright, R. T., & Decker, S. H. (2011). Armed robbers in action: Stickups and street culture. UPNE.

Wright, Richard T., and Scott Decker. Burglars on the Job. Boston, MA: Northeastern University Press, 1994.

Hi Frank,

Thank you very much for your very good explanation.What I have is that the grave crime data of last 5 years. It is a huge pile of a data set which has not yet been subjected to any statistical analysis. My model will definitely be subjected to cognitive biases. That's why I am worrying very much about finding the best set of parameters in developing the model.

My big question is how much information do you have related to the offenders?  If you want to produce a model that will assist the investigator in narrowing down a population to a group of suspects or further narrow down a group of existing suspects you'll need to have a large amount of detailed information on the offenders/suspects.  For example in the U.S. if we only analyzed documents related to the average felony case, we would be able to determine in most instances-age, race, height, weight, eye and hair color, the existence of scars and tattoos, address (city, neighborhood) currently employed or not (maybe with a length of employment), type of employment-all of which are self reported by the offender or determined by justice officials.  Prior adult criminal history, prior juvenile criminal history (this could be both formal charges and just arrests) are often utilized in prosecution and so are also available.  Sometimes financial stability, as well as employment, maybe length of time at the residence, may be indicated in bail and indigent defense applications.  Other documents may indicate a next of kin (likely their emotionally closest relative), and maybe their custody of children, and whether they had been drinking alcohol or consumed drugs.

These are just what are readily available in court and offense related documentation.  Very thorough readings of arrest reports, witness statements, and court transcripts can reveal much more about an offender but of course this would simply not be practical for a large number of cases.  Other sources of information that could be used to define suspects that I mentioned previously like education records, IQ, personality assessments,  (do they have a personality disorder or of a certain type of personality) home environment information, medical care and health (mental illness, prenatal exposures) might also be available from various sources but would be difficult and time consuming to obtain on a large sample in building a database.

With all that being said, I'm curious to know what kind of offender data you have in your dataset.  If it is not detailed enough what you'll end up with is a generality-e.g. almost all serious crimes are committed by males between the ages of 16 and 25 who come from poor neighborhoods, had some kind of minority status, and were frequently unemployed, as well as using intoxicants during or prior to the crime.  That's pretty much the basic picture of criminality risk factors in the U.S. (in a nutshell) as might be available from offense or court documentation so to find the hidden or unknown correlates, you need to have detailed offender information so I am curious as to what variables you have. 

I hope this helps

I guess another approach might be is to analyze it from the components of the offense if you have limited offender data and lots of offense data. Look at all the features of the offense-time of day, time of year, geographic location , amount and type of items stolen, value of items, weapon used, multiple offenders used, disguises or concealed identity, threats of violence, age of victims, sex of victims, victims injured, type of theft (deception, force, stealth) for the particular offenses.  Once accomplished you can compare it to your available offender data.  Are certain crime features associated with certain types of offenders?  Do the features of some crimes not cleared by arrest match features of crimes where an offender or suspect has been named?  For example-Does the location of the offense show a clear statistical significant connection to residence location of the offender?  If so, then for an unsolved crime this connection between offense location and offender location could be utilized to narrow a focus area in looking for the offender.  Perhaps your findings from the dataset show that industrial area burglaries are overwhelmingly committed by older, more experienced (i.e. longer criminal history) burglars but that residential burglaries are most often committed by juveniles.  This information can then be utilized to narrow a focus on the more likely offender for unsolved crimes of certain types.  In essence, what you're doing is developing the modus operandi for different types of crimes, see who fits the MOs based on available offender data and then suggest the likelihood that similar types of offenders committed the unsolved crimes.

Hope this helps as well.   

Hi Frank,

Thank you very much. Yes, my focus is to find the relationship between an MO of a crime and an MO of a particular offender. I found that the methods available in the literature seem to have a very low accuracy due to the cognitive biases. Also, there seem to be some problems in assigning some quantitative measurements to qualitative properties. I could only find a less amount of literature (related to computer science). I could find many articles related to social sciences (mainly sociology : criminology). But, in many of the sociological papers, the researches have been focused towards the description of qualitative facts. They have not much tried in developing any statistical models. If you can provide me any literature, it would help me a lot.

Regards.

Hi M.A.P.,

I'm not sure what you've all looked through in the lit but following are some citations that might be useful.  Hierarchal modeling and logistic regression prediction examples from this lit might be suitable for what you are doing.  The studies don't necessarily address your particular research approach, much of the prediction work is done on violence or recidivism using survival time but I think that you can adapt the basic methodologies. Multilevel modeling maybe appropriate when consider location and non location related components of the offenses as well as folding in offender data into your predictions. Many offense variables could be treated as dichotomous variables to construct predictor scores or matrices and rather than determining the likelihood of an individual offending or recidivating, you can relate the likelihood of a particular offense having been committed by a particular offender based on criteria established from the data set.  After constructing what appears to be a useful predictive model you can demonstrate its validity on another sample of offenses with known offenders to demonstrate its predictive ability.  I think "Prediction in Criminology" by Farrington should provide some good insight in how to construct some effective models and the other literature should provide some alternative approaches as well.

Sorry nothing more specific (or newer) but maybe these will help at least as a starting point.

Wooldredge, J., Griffin, T., & Pratt, T. (2001). Considering hierarchical models for research on inmate behavior: Predicting misconduct with multilevel data. Justice Quarterly, 18(1), 203-231.

Farrington, D. P., & Tarling, R. (Eds.). (1985). Prediction in criminology. SUNY Press.

Caplan, J. M., Kennedy, L. W., & Miller, J. (2011). Risk terrain modeling: brokering criminological theory and GIS methods for crime forecasting. Justice Quarterly, 28(2), 360-381.

A Multidimensional Contingency Table Analysis of Parole Outcome: New Methods and Old Problems in Criminological Prediction. David J. van Alstyne and Michael R. Gottfredson. doi: 10.1177/002242787801500205 Journal of Research in Crime and Delinquency July 1978 vol. 15 no. 2 172-193

Stepwise Fitting of Logit Models With Categorical Predictors in the Analysis of Parole Outcomes On the van Alstyne and Gottfredson Study. Camil Fuchs and John Flanagan. doi: 10.1177/002242788001700209 Journal of Research in Crime and Delinquency July 1980 vol. 17 no. 2 273-279

Thank you very much Frank 

Hi Dr.Xingan,

Yes, some data mining method will have to be applied. But, the major concern of this discussion was to come up with an approach to the whole problem in the sense of dimentionality reduction as the first step.