Profiling The Ripper

[jmenges]

Researchers and authors John Malcolm and Jon Lee Rees join Jonathan Menges in a discussion about behavioral and geographic profiling.
Does profiling have value when applied to the hunt for Jack the Ripper?



Available to stream or download now from the following link:

http://www.casebook.org/podcast/listen.html?id=236

Also in Overcast, Castbox, Apple & Google Podcasts, and wherever else FREE roundtable discussions about the Whitechapel Murders can be found.

Thanks to John Malcolm and Jon Rees for participating in this discussion...

And Thank You for listening,



JM

[Scott Nelson]

One of the most comprehensive pieces ever on profiling the Ripper in two parts: Murder Most Foul

Ripperologist 153 and 154.

[Sam Flynn]

A very interesting talk, folks. Many thanks.

PS: To Jon Rees and others for info, the "braf" in "Cwrw Braf" rhymes with "starve", not "staff". </pedantry>

[Lipsky]

Interesting.

First observations:

Concerning geography
1. Virtual geoprofiling is useful to narrow down the theatre of operations and produce a probable "lair"/home for our man.
2. Geographical aspect of the profile 'after' the murders has no other practical aspect than to confirm the perpetrator was local, not just living "there".
3. Geographical vicinity is strengthened by physical vicinity. This de facto excludes the absurd notion that these women didn't know each other.
4. Expansion of geographical correlation to prior attacks at Spring expands the pool of victims to include prior women. There is further correlation on MO
5. The killer's comfort zone is valid with prior Spring attacks.
6. The decreasing more sporadic post-MJk are also included in the geographical pattern with sufficient consistency.
7. This does not include torsos and floating half corpses whatsoever.

Concerning time (the mirror image of space in the ST 2-D system)
1. The time window of the events is narrowed down to a few months, expanded to 2 more years with each victim per year as the pattern fades out to the completion of the task. Swift attacks at spring 1888, climax at the autumn of terror, settling score with loose ends in subsequent period.
2. Time window is very narrow on all known murder attacks, successful or otherwise, regardless of extent of mutilation.
3. Narrow time window ensures (1) minimization of any physical evidence down to near-zero (2) swift entry and exit of a crime scene (3) very little witnesses and if any, hardly credible due to speed of events
4, The time aspect demonstrates that the killer is more of an agent in an operation than a sex offender/lunatic. This is precise, designated, deliberate, controlled. This is a plan carried out swiftly by a man on a mission.

Concerning psychology
1. None of the contemporary doctors attached to the case produced any serious profile or even cared to, at best they thought of the man as insane and sex-driven.
2. Debates of doctors at inquests/assisting police (when they were not disagreeing with each other) were as to whether there was always one knife, one and very specific course of mutilating actions.
3. Analysis of the perpetrator revolved mostly around his medical skills and the degree of said skills.
4. Contemporary medical approach ignored (1) the swift termination of the victim's life BEFORE any mutilation (not a sadist), (2) lack of any direct or indirect gratification (purpose was to shock spectators and terrorize a small group of them with inner workings) (3) why these locations were chosen, (4) why these DATES were chosen (mmmkay a weekend but why "this" and not "That"). A contemporary newspaper was more spot-on, two days before the MJK murder (one and a half, one might say). Interestingly, this was also declared by Dr. Winslow (1910, so it is de facto in retrospect) who identified the time cycle of the the dates as did that newspaper.
5. Dr Winslow is the first to provide a bold, if flawed, profile of the man.
6. Dr Winslow employed that profile to expand the number of victims way beyond the monolithic "five".
7. Dr Winslow's time window coincides with the mirror image of geographical correlation and purpose of the killer (to deflect the schemes of a few greedy folks at Whitechapel who employed sad drunk desperate prostitutes who were on their hands for reasons not needed to expand on in this thread).
8. Douglas profile is spot on in some truly "common sense" areas, that still get scorned (!) by some policemen who think all starts and ends in "caught in the act" or "i watched him do it" situations. Goes to prove why Dahmer and others slipped out of their hands in their faces. Profiling provides the general pattern hence the ability ot strike pre-emptively.
9. Douglas profile fails when assumes that the killer was a sex offender with a purpose for gratification.
10. Detached childhood (Dahmer), weird relationship with parents and especially mother (Bundy) is present in our man. Such background also breed professional military special-forces assassins and field agents with no strings attached. This is where the profile coincides with our man.
11. The argument of "sheer physical force" when dealing with confused, desperate, middle aged alcoholic prostitutes driven in a scheme way over their heads is overblown. A disciplined trained man with a purpose can fit the image conveniently. As could almost anyone else in Dorset street with a temper and a pint (and there was too many of them) --- but such "skills" alone do not complete a swift set of operational attacks.
12. Douglas profile and sex offender serial killers continue until arrested, usually because they lose grip on their tactics/antics and their delusional megalomania goes unchecked, causing them to fall short on their misbeliefs. Bundy thought he could seriously represent himself in court, or later on, hold off his execution by admitting to crimes, areas of further bodies and addiction to pornogrpahy (lol, good one Teddy boy). Zodiac thought his ludicrous puzzles demonstrated genius instead of pathetic confused loner with too much time on his hands. Ian Brady thought he could redecorate his house with blood in front of a disciple who could just go to bed as if nothing happened and not go to the police (!). Until their madness betrays them, and they get arrested....Or, if they die abruptly. Our man here never lost hold. The MJK butchering was not the beginning of a downfall leading to expose, arrest, or asylums. It was disciplined, as was the laying low afterwards. He never went ranting mad. We would have known. Leave them barbers, polish jews and hairdressers/wigmakers at peace (they suffer enough already).
13. Our man is deranged, in the sense that he has no moral restraint as to the fulfillment of the plan. Like the hideous Reinhard Heydrich, cold, methodical, brutal in a clinical way. But way more cautious than that pompous fascist creep.
14. The first full fledged mutilation that initiates the climax of terrorism, Tabram's murder, holds specific weight for the killer as a time instance. This is the transition from phase I (Spring attacks) to the next phase of the terrorism (since the scheme devised against him and the interests he represented seemed to go on after the first storm was tried to be covered up by those who pimped/guided these unfortunate women). These specifics of the killer, are where motive becomes distinct from the purpose. Those distinctions provide the deranged singularity of the murderer. That, and his only other public act of terrorism -- the Lusk letter.
15. Handwritting forensics from a psychiatric point of view , the direct reference to the Letters from Hell book (a jolly read, yes?) and the erased (hence failed to provide public outrage as planned) graffiti provide the canvas of this deranged personality. It is "in character", but to do so he dives in his soul. Method acting, from the heart.
16. Too much attention on that boring dilemma of medical skills (or not). Wrong angle. A disciplined solider can learn fast "some basic knowledge" which was the only level of medical "Awareness" recognized by the medical doctors studying the mutilations. More dramatic and difficult to achieve was that "special ops" skill of entrying and exiting a crime scene around policemen on a beat, plaincloth officers, vigilantes and ruffians of all sorts and trades. Simply being local doesnt cut it. A nerdy medical student wouldnt know where to turn. A sex offender would be tempted to either prolongue the pain before death or indulge in the mutilations. Our man was always restrained. He knew when, for how long, and how. And mostly, he knew WHERE. Locations were chosen beforehand, probably scouted in prior excursions to screen and decide on efficiency.
17. Antisemitism was classic tactic of decoy (special ops again?). What better confusion in an area where poor Jews resided en masse? And sadly in Europe antisemitism was already rising as a tendency both in the lower stratta and the aristocracy/capitalist elite, with the known catastrophic results some decades later. but for our man, antisemitism is a tactic, a manouver choice, like his "cannibalism". See however that in the letter he doesnt belittle women OR the police? This is an act of terrorising the locale by taunting their hero of the day, the self-proclaimed vigilantes' head, the only one with local knowledge and (some, at least) street wisdom to protect the prostitutes (and possibly to make them share the blackmailing scheme with them -- a nice way out of bancrupty, Mishter Lusk?).
18. Our man was the opposite of these ladies --- he probably abstained from alcohol. Like a good special-ops, a sober mind is a swift and efficient mind (albeit a tiny deranged, ja?). On the contrary, he would make sure his victims were neatly disoriented (another aspect of special ops). The suggestion that he "had a few" to "relax his inhibitions" before the murder is laughable. A relaxed mind gets caught. And this man had already no inhibitions. Nor needed to "transit" from a quiet phase to a loud, murderous, savage one. He was savage at all times, and quietly, viciously so. (Dr winslow failed here, lets give him due though for trying more honestly than some unsuitable folks with a penchant for memoirs, name-calling and miserable, inept asylum patients).


I mean really, was our man someone who couldn't stop flexing his weenie? Or someone who had just landed from the South Pole or whatever?
Oh wait, he exited the crime scene as a woman. Did he enter it as such or did we just discover the fastest crossdresser in history
("oh hi, here, Police Constable, hold me knif while i fix my victorian eyeshadow...smokey-eyes is "in phase" with Mitre Square, mmm? this IS City territory after all?")
Yes, my humour sucks at the wee hours.

[DrDMode]

Two comments made at the outset proved significant as the podcast wore on...one was that none of the participants are experts in the field (sadly that was all too apparent as matters 'progressed' - not that it stopped them from pontificating freely!) and the second was that the discussion had begun in a conversation in a wine bar. More than once I mused to myself that it never really deserved to go any further than a chat over a couple of drinks (sorry guys but I can't deny that impression). At best it could have filled a ten minute mini talk, each speaker having three minutes or so and then wrap it up...much of what occurs here is repetitive (one of the participants in particular but I will refrain from identifying the man in question) and talks in and around the same territory, frequently in very generalised terms. When examples are given and cases cited they are overwhelmingly negative in keeping with the tone of dismissal...a realistic attempt at objectivity would not have gone amiss. I normally listen to the podcast as I do monotonous chores, but sad to say this is the first casebook podcast that has heightened rather than alleviated the monotony. Not destined to become one of the classics alas

[John Malcolm]

Thank you for your candor, Dr. D. I can probably guess which participant you are singling out as being particularly repetitive and "frequently in very generalised terms". I do suffer from both maladies, frequently and particularly. It's OK, I can take it. Maybe I can even work on these issues, hopefully to the benefit of all (like yourself) who admirably endure the monotony. Or maybe it wasn't me you were talking about, which may indicate some degree of paranoia on my part. But instead of demonstrating your expertise in the art of condescension (which does unfortunately strike a nerve), maybe it wouldn't be too much to ask of you to add some insight to the conversation. I'm sure all of the participants would appreciate some positive input on the subject, as it is obviously something you have enough interest in to have suffered through this less-than-classic podcast. I certainly welcome your thoughts on the subject. But I find your comments about the podcast (e.g. "a realistic attempt at objectivity") to be, at very most, counterproductive.

[JeffHamm]

Had a listen to this and, as I do a bit of research on geographical profiling, just thought I would make a few comments. First, it is well worth noting that the spatial analysis of offense locations is, like any analysis, influenced by the quality of the data being inputted. Garbage in, garbage out, as they say. Second, it is entirely correct that geographical profiles are suitable to some cases, and not to others. A highly mobile offender, as mentioned, will result in a moving target, so to speak.

On the other hand, and this probably needs more emphasis, a geographical profile does not solve a case. Cases are solved by detectives and investigators, who have to take in and consider all of the sources of information. It is a fact of life that some evidence will appear contradictory, be in error, or seem anomalous even after a case is solved. DNA, for example, does not by itself solve a case, it simply provides evidence that someone's cells were present at a particular location. It would not, for example, be surprising if Barnett's DNA were found in Miller's Court given he lived there quite recently (obviously not a possible type of evidence in 1888, just using as an example).

What geographical profiles do is produce what they call a "jeopardy map", which is basically an attempt to indicate areas of higher and lower probabilities. The notion is that an offender's anchor points will be reflected in areas of higher probability. An offender's residence is a common anchor point, but it is not always the case. It may be a place of work, a place of worship, a favorite pub, or club, and that sort of thing. Offenders may have multiple anchor points (a stable residence and a stable work place, for example, might result in two anchor points).

The underlying idea behind the analysis is that an offender has to decide to commit an offense, whether spur of the moment or after long planning, is not as important as the fact that when in that location they decided it was ok. Anchor points are simply locations where we spend lots of our time in our daily lives. For a serial offender, that might even include the areas they go patrolling when looking for potential victims. Anyway, the more time you spend in an area, the more familiar that area naturally becomes, and the more familiar an area is, the safer it becomes.

As a result, approaches to geographical profiling like those adopted by Rossmo, or Canter, analyse the spatial pattern of the offenses based upon the idea the offender is attempting to minimize that sense of risk. Rossmo argues for a buffer zone, that locations too close to an anchor point, will be perceived as higher risk than those further away. However, this competes with the fact that the further away from our anchor points a location is the less familiar one will be with it, so increased distances results in increasing risk as well. This results in two competing forces, one that suggests offenses should be "repulsed" from anchor points, and another in the opposite direct, pulling offenses closer. The end combination is that offenses should tend to bunch up and cluster at some intermediate range. His program (called Rigel), is based upon that notion. Cantor's program (Dragnet), does not have a bufferzone (no "repulsion"). Pragmatically, both programs do equally well (which I'll get to later).

Underlying both of them is data, which comes from (as mentioned) solved cases. In short, information from a large number of different series are combined which then allows one to work out a probability distribution. For example, what percentage of offenses are within 0.1 km of an offender's home? What about between 0.1 and 0.2? 0.2 and 0.3? and so forth. With an unsolved case, one can choose a location anywhere on the map, and from that, work out the distribution of offense distances from that point, and compare what you get with the general pattern. The more similar the observed distribution is to the general pattern, the higher the probability that location is an anchor point. (It's a bit more complicated than just that, but that's the gist of it). Now, not all offenders will fit that general pattern, but that's hardly a surprise, and it is emphasized (or should be), that the analysis is a general pattern and suggests areas that have a higher probability than others of being where actual evidence may be found. It is also stressed that if evidence leads the investigation to a low probability area, the profile should be ignored. It's an informed suggestion that is better than chance, but it is not magic, and it is not evidence, and there will be offenders that are not "like the others." Investigations that get tunnel vision based upon any kind of profile have, unfortunately, not kept that in mind.

Another approach to the spatial analysis comes from the fact that any offense will involve some sort of journey, travelling to and from the offense. If you have a pattern of offenses, committed by the same offender, then they have made multiple journeys. The pattern of offenses can be analysed on the basis of trying to extract the underlying journey pattern, rather than an attempt to model risk evaluation. (This has been what I've been playing with). I won't go into the details of how this works specifically, but in the end it results in a probability map as well, again suggesting areas that have a higher probability than others of being productive.

One cannot evaluate if a profile approach is "good" or "bad" by simply looking at example cases. Showing that a profile got it wrong this time, for example, just means that something that has a less than 100% accuracy rate will sometimes make errors. And showing that something "got it right this time", could just mean that sometimes it's lucky. What one has to do is compare a profiling approach to chance level performance by examining how well it does over a number of cases. And when that is done, geographical profiling is quite a bit better than chance. This is because there are patterns to be found, and they are useful. They are not, however, magic bullets, and they often look much like one might expect (but why that is considered a bad thing is beyond me, it's just a mathematical representation of what offenders tend to do).

The jeopardy maps are not intended to zoom in on one particular house, but rather direct attention to areas worth considering. The larger an area over which an offender commits offenses, the larger the resulting potential search space is. Geographical profiling is an attempt to reduce that search space, or to prioritize areas in the search space. If you have offenses that are distributed over a few hundred square miles, you would be happen if you had a good chance of locating the offender after only having to search 10 or 20 square miles even though that's still a pretty big area.

So, while Rossmo's "hot spot" focused on a street that might have hundreds of people in doss houses, that is still a lot less than having to search the entire area. Remember, it's really about suggesting an order to which the search space should be searched to minimize the total amount you have to search. It's not about finding the needle in the haystack, but about suggesting which farm's haystacks to check first, then second, then third, etc. Unfortunately, since JtR was never identified, we can't determine if this case is, or is not, one in which the profile output actually would have done that. They don't always, this may, or may not, be one of those cases.

The marauder/commuter division, is another point that was well worth mentioning. Basically, the first thing to do is fit the smallest possible circle you can around all the offenses. That's called the crime-range. Offenders who live inside that circle are called marauders, and those who live outside it are commuters. In general, about 80% of offenders will be marauders, and this applies to arson, murder, and rape (I can't recall if burglary also finds this division). Geographical profiles are generally based upon the notion that a given series is a marauder, meaning lives within the crime range. That means right off the bat there's a 20% chance it's not going to work. And that is not insubstantial. Specifically, though, there are some caveats to that. First, the crime range is a data thing, you need offenses to define it, but it is an estimate of what I'll call the criminal "territory", which is the area where the offender is willing to commit offenses, even if they haven't yet. The fewer offenses you have in a series, the more the observed crime range will underestimate the criminal territory. So what can happen is that as a series gets longer, an offender who was originally a commuter suddenly becomes a marauder. Studies that have tried to estimate the percentage of marauders vs commuters have reported different breakdowns, but this often is a result of those that have reported lower percentages of marauders also tend to have a sample where the series length is shorter. One of the things I've been trying to work out is a way to factor in the series length to give a better estimate of the territory size. There are some people working on ways to try and determine if a series is more likely to be a commuter than a marauder, though, because if one could determine that division, then at worst one could just say "this case is not well suited to a spatial analysis", or better "develop a model that is better at profiling commuters".

There is, of course, the issue of whether or not the profiles are better than a trained and experienced investigator? That's a bit up in the air, really. The thing is, that humans are really good pattern recognizers, and that's basically what a geographical profile is, a pattern recognition analysis. In some ways, the fact that the outputs look unsurprising to someone who has studied and investigated lots of cases is a good thing, it should look sensible. It's just a way to mathematically calculate that pattern.

Now, despite my pointing out that specific cases are not a good basis for evaluating whether or not geographcal profiles are useful, and that's not my intent here, the podcast did mention Dennis Radar, and in a way that suggested his case was unlikely to result in a profile that would be helpful because of his tendancy to search and stalk. Below is the output from the routines I've been working on. The red squares are the offense locations, the bright blue square is Dennis Radar's home location. His home falls in "zone one" (the yellow and magenta areas combined). In his case we see zone 1 (which reduces the search space to 2.5% of the total) is split between the north (Park City) and south areas. His wife worked at the VA Hospital (zone 4), and Radar used to drive her to work when he was unemployed (just before the start of his crimes) and he later was employed by ADT (I think just slightly to the west of the southern zone 1 in zone 7 or 8, but I've never been able to confirm the address). Anything in zone 20 or less is better than what a random (chance) search would produce on average (chance of course could range from zone 1 to zone 40, with zone 1 and 40 only happening 2.5% of the time each; basically, randomly placing zones should end up with 2.5% of offenders in each of the zones) In other words, the output is picking up on a couple of anchor points.

Also note, though, it's not about laser pinpointing an offender, it's just a way of suggesting how to search, and if it results, more often than not, in reducing that search time that is a good thing. But it should never, ever, be used to exclude viable leads or potential suspects. Just like the white van that's mentioned, it suggests a vehicle description that might be worth checking out, but that doesn't mean suspects in other vehcles should be excluded on that basis alone. The sighting was "of note", not "definitive" because nobody saw someone in the white van actually shoot anyone, or even have a gun. An error in how that information was used was the problem, not the information itself. That's the biggest thing to remember, they are information, they are better than chance, but they do not replace actual evidence or actual investigative work.

- Jeff

[Michael W Richards]

I think that the geographical implications revealed by a murder site mapping are that the targeted area is identified as being much smaller than in most serial type cases, and in a far more concentrated population than is usual. I think that tells us one thing at least, if not where the most probable nesting site would be...that he had to be on and off the streets quickly. Which really would confirm what most people realize anyway, he was a local man. In my opinion, these were local men, but that's another matter.

[Trevor Marriott]

Had a listen to this and, as I do a bit of research on geographical profiling, just thought I would make a few comments. First, it is well worth noting that the spatial analysis of offense locations is, like any analysis, influenced by the quality of the data being inputted. Garbage in, garbage out, as they say. Second, it is entirely correct that geographical profiles are suitable to some cases, and not to others. A highly mobile offender, as mentioned, will result in a moving target, so to speak.

On the other hand, and this probably needs more emphasis, a geographical profile does not solve a case. Cases are solved by detectives and investigators, who have to take in and consider all of the sources of information. It is a fact of life that some evidence will appear contradictory, be in error, or seem anomalous even after a case is solved. DNA, for example, does not by itself solve a case, it simply provides evidence that someone's cells were present at a particular location. It would not, for example, be surprising if Barnett's DNA were found in Miller's Court given he lived there quite recently (obviously not a possible type of evidence in 1888, just using as an example).

What geographical profiles do is produce what they call a "jeopardy map", which is basically an attempt to indicate areas of higher and lower probabilities. The notion is that an offender's anchor points will be reflected in areas of higher probability. An offender's residence is a common anchor point, but it is not always the case. It may be a place of work, a place of worship, a favorite pub, or club, and that sort of thing. Offenders may have multiple anchor points (a stable residence and a stable work place, for example, might result in two anchor points).

The underlying idea behind the analysis is that an offender has to decide to commit an offense, whether spur of the moment or after long planning, is not as important as the fact that when in that location they decided it was ok. Anchor points are simply locations where we spend lots of our time in our daily lives. For a serial offender, that might even include the areas they go patrolling when looking for potential victims. Anyway, the more time you spend in an area, the more familiar that area naturally becomes, and the more familiar an area is, the safer it becomes.

As a result, approaches to geographical profiling like those adopted by Rossmo, or Canter, analyse the spatial pattern of the offenses based upon the idea the offender is attempting to minimize that sense of risk. Rossmo argues for a buffer zone, that locations too close to an anchor point, will be perceived as higher risk than those further away. However, this competes with the fact that the further away from our anchor points a location is the less familiar one will be with it, so increased distances results in increasing risk as well. This results in two competing forces, one that suggests offenses should be "repulsed" from anchor points, and another in the opposite direct, pulling offenses closer. The end combination is that offenses should tend to bunch up and cluster at some intermediate range. His program (called Rigel), is based upon that notion. Cantor's program (Dragnet), does not have a bufferzone (no "repulsion"). Pragmatically, both programs do equally well (which I'll get to later).

Underlying both of them is data, which comes from (as mentioned) solved cases. In short, information from a large number of different series are combined which then allows one to work out a probability distribution. For example, what percentage of offenses are within 0.1 km of an offender's home? What about between 0.1 and 0.2? 0.2 and 0.3? and so forth. With an unsolved case, one can choose a location anywhere on the map, and from that, work out the distribution of offense distances from that point, and compare what you get with the general pattern. The more similar the observed distribution is to the general pattern, the higher the probability that location is an anchor point. (It's a bit more complicated than just that, but that's the gist of it). Now, not all offenders will fit that general pattern, but that's hardly a surprise, and it is emphasized (or should be), that the analysis is a general pattern and suggests areas that have a higher probability than others of being where actual evidence may be found. It is also stressed that if evidence leads the investigation to a low probability area, the profile should be ignored. It's an informed suggestion that is better than chance, but it is not magic, and it is not evidence, and there will be offenders that are not "like the others." Investigations that get tunnel vision based upon any kind of profile have, unfortunately, not kept that in mind.

Another approach to the spatial analysis comes from the fact that any offense will involve some sort of journey, travelling to and from the offense. If you have a pattern of offenses, committed by the same offender, then they have made multiple journeys. The pattern of offenses can be analysed on the basis of trying to extract the underlying journey pattern, rather than an attempt to model risk evaluation. (This has been what I've been playing with). I won't go into the details of how this works specifically, but in the end it results in a probability map as well, again suggesting areas that have a higher probability than others of being productive.

One cannot evaluate if a profile approach is "good" or "bad" by simply looking at example cases. Showing that a profile got it wrong this time, for example, just means that something that has a less than 100% accuracy rate will sometimes make errors. And showing that something "got it right this time", could just mean that sometimes it's lucky. What one has to do is compare a profiling approach to chance level performance by examining how well it does over a number of cases. And when that is done, geographical profiling is quite a bit better than chance. This is because there are patterns to be found, and they are useful. They are not, however, magic bullets, and they often look much like one might expect (but why that is considered a bad thing is beyond me, it's just a mathematical representation of what offenders tend to do).

The jeopardy maps are not intended to zoom in on one particular house, but rather direct attention to areas worth considering. The larger an area over which an offender commits offenses, the larger the resulting potential search space is. Geographical profiling is an attempt to reduce that search space, or to prioritize areas in the search space. If you have offenses that are distributed over a few hundred square miles, you would be happen if you had a good chance of locating the offender after only having to search 10 or 20 square miles even though that's still a pretty big area.

So, while Rossmo's "hot spot" focused on a street that might have hundreds of people in doss houses, that is still a lot less than having to search the entire area. Remember, it's really about suggesting an order to which the search space should be searched to minimize the total amount you have to search. It's not about finding the needle in the haystack, but about suggesting which farm's haystacks to check first, then second, then third, etc. Unfortunately, since JtR was never identified, we can't determine if this case is, or is not, one in which the profile output actually would have done that. They don't always, this may, or may not, be one of those cases.

The marauder/commuter division, is another point that was well worth mentioning. Basically, the first thing to do is fit the smallest possible circle you can around all the offenses. That's called the crime-range. Offenders who live inside that circle are called marauders, and those who live outside it are commuters. In general, about 80% of offenders will be marauders, and this applies to arson, murder, and rape (I can't recall if burglary also finds this division). Geographical profiles are generally based upon the notion that a given series is a marauder, meaning lives within the crime range. That means right off the bat there's a 20% chance it's not going to work. And that is not insubstantial. Specifically, though, there are some caveats to that. First, the crime range is a data thing, you need offenses to define it, but it is an estimate of what I'll call the criminal "territory", which is the area where the offender is willing to commit offenses, even if they haven't yet. The fewer offenses you have in a series, the more the observed crime range will underestimate the criminal territory. So what can happen is that as a series gets longer, an offender who was originally a commuter suddenly becomes a marauder. Studies that have tried to estimate the percentage of marauders vs commuters have reported different breakdowns, but this often is a result of those that have reported lower percentages of marauders also tend to have a sample where the series length is shorter. One of the things I've been trying to work out is a way to factor in the series length to give a better estimate of the territory size. There are some people working on ways to try and determine if a series is more likely to be a commuter than a marauder, though, because if one could determine that division, then at worst one could just say "this case is not well suited to a spatial analysis", or better "develop a model that is better at profiling commuters".

There is, of course, the issue of whether or not the profiles are better than a trained and experienced investigator? That's a bit up in the air, really. The thing is, that humans are really good pattern recognizers, and that's basically what a geographical profile is, a pattern recognition analysis. In some ways, the fact that the outputs look unsurprising to someone who has studied and investigated lots of cases is a good thing, it should look sensible. It's just a way to mathematically calculate that pattern.

Now, despite my pointing out that specific cases are not a good basis for evaluating whether or not geographcal profiles are useful, and that's not my intent here, the podcast did mention Dennis Radar, and in a way that suggested his case was unlikely to result in a profile that would be helpful because of his tendancy to search and stalk. Below is the output from the routines I've been working on. The red squares are the offense locations, the bright blue square is Dennis Radar's home location. His home falls in "zone one" (the yellow and magenta areas combined). In his case we see zone 1 (which reduces the search space to 2.5% of the total) is split between the north (Park City) and south areas. His wife worked at the VA Hospital (zone 4), and Radar used to drive her to work when he was unemployed (just before the start of his crimes) and he later was employed by ADT (I think just slightly to the west of the southern zone 1 in zone 7 or 8, but I've never been able to confirm the address). Anything in zone 20 or less is better than what a random (chance) search would produce on average (chance of course could range from zone 1 to zone 40, with zone 1 and 40 only happening 2.5% of the time each; basically, randomly placing zones should end up with 2.5% of offenders in each of the zones) In other words, the output is picking up on a couple of anchor points.

Also note, though, it's not about laser pinpointing an offender, it's just a way of suggesting how to search, and if it results, more often than not, in reducing that search time that is a good thing. But it should never, ever, be used to exclude viable leads or potential suspects. Just like the white van that's mentioned, it suggests a vehicle description that might be worth checking out, but that doesn't mean suspects in other vehcles should be excluded on that basis alone. The sighting was "of note", not "definitive" because nobody saw someone in the white van actually shoot anyone, or even have a gun. An error in how that information was used was the problem, not the information itself. That's the biggest thing to remember, they are information, they are better than chance, but they do not replace actual evidence or actual investigative work.

- Jeff

You cannot rely on this type of profiling. If the killer did not reside in Whitechapel but simply came into the area to find his victims and then after killing them left the area, then all of this is worthless.

https://www.trevormarriott.co.uk

[Michael W Richards]

You cannot rely on this type of profiling. If the killer did not reside in Whitechapel but simply came into the area to find his victims and then after killing them left the area, then all of this is worthless.

www.trevormarriott.co.uk

The fact that the killer needed to be off the streets quickly, for the reasons I indicated, suggests that he had accommodations locally.

[JeffHamm]

You cannot rely on this type of profiling. If the killer did not reside in Whitechapel but simply came into the area to find his victims and then after killing them left the area, then all of this is worthless.

www.trevormarriott.co.uk

Trevor is correct, and I hope that's clear from what I posted as well. Geographical profiling is not a claim about "solving the case", there are times, and not just a few, when the indicated hot spot does not correspond with a given specific offender/series. They are accurate more often than not, particularly for marauder type offenders. There is some research that is trying to improve the detection of commuter patterns because if you can weed those out right at the start, you end up with better information (in the sense that if it suggests this is not a commuter pattern you have more reason to prioritize the search along this suggested order; and if it is likely to be a commuter, it suggests you are better off searching further a field, though that might be a bit less impressive).

Again, what I cannot emphasize enough, the goal of these is to provide a suggested search pattern of the space that has to be searched. Most offenders do not live all that far from the area in which they commit their offenses. Of course, if the offenses are scattered over hundreds of km, along a major highway, you're dealing with a mobile offender. The thing is, the analysis will still likely high light that travel path, although looking at pins on a map will reveal the same thing. Again, it's not about "doing magic", it's about mathematically describing the patterns. We are good at seeing patterns, sometimes, though an objective calculation can bring our attention to a pattern that we miss because we spot a different one.

But these analyses are not about identifying a suspect per se. They are more about suggesting an efficient search, and more often than not, if you follow the suggested search you will locate an offender sooner than if you just search the space randomly.

However, you cannot view this as "oh, ok, let's ignore everything else", as relying on them to that extent is misunderstanding what they do. And faulting them based upon a misunderstanding is an error. You can't blame a lemon for not being an apple. If real evidence leads an investigation to a suspect who falls in a low probability zone, you ignore the profile and follow the lead. Real evidence trumps a probability map. These are useful because, without a lead to follow, they do result in finding real leads sooner in the majority of cases.

Another way they can be used is to prioritize a long list of "suspects" (or "persons of interest"). If you have someone who has known anchor points that fall in high probability zones, and another who does not, with nothing else to base a decision on, you would check out the former first (ideally one checks them all out, but the reality is that there are a limited amount of resources in terms of money, time, and so forth, so if you don't have something evidential that indicates which one you should prioritize, then you base the decision on which one fits the probable pattern). And, it does reduce things over the course of many investigations. It will, however, in some cases, get it wrong. That's how probabilities work. If the output suggests person A is 60% likely, and person B is 40% likely, then 60% of the time you get on the right track on your first search, and 40% of the time you delay that. If you do it by chance, then 50% of the time you're on the right track, and 50% of the time you're delayed. That change from 60/40 to 50/50 has cost implications over time, so it's worth doing, but it is also worth remembering that these are not "solutions", they are "probabilistic statements". A lot of the time we focus on individual cases, and you cannot evaluate their utility by doing that.

Most research suggests that offenders are commuters about 20% of the time, so 20% of the time, this isn't going to help (which is what Trevor is describing). While 80% of the time it will help, a 20% error is not small. It's a complicated problem, and the suggested search pattern has a lot of cautions that go with them. But when an investigation stalls, and leads dry up, they can help to reinvigorate things. They might suggest re-looking at someone who dropped off the radar pre-maturely (which happens), and so forth.

As for the JtR case? It's a small area. On the other hand, travel was generally by foot, so people couldn't travel the distances they can now. These analyses use the spatial distances internal to the offense locations, and base many of the calculations in a way that is relative to those distances. It is possible, though, that the calculations would benefit from having a large set of crime series where offenders were also on foot, as that could very well change some of the important calculations. Of course, it may also be that it makes no difference, and the same calculations would emerge. That's an empirical question, and we all may have a "common sense" idea what the answer would be - but I bet each of us have a different common sense answer. That's the problem with common sense, it's rarely common, and often, it's not even sensible once we actually do the research and find out how things do work. People are complicated, decisions are complicated, but they are based upon detectable regularities. That is what this sort of analysis is trying to extract, the underlying regularities. If the analysis correctly captures those regularities, they produce generalized answers, which will be broad and fuzzy, and "roughly correct", at least most of the time. But they will not, cannot, get it right every time. Whether or not they apply to JtR is hard to say, and will only be answerable if the case is solved.

Oh, and by the way, shifting the Kelly murder "a few streets east", type thing, makes almost no difference to the output. The analysis is fairly robust against such things. However, Stride's inclusion (which there are lots of reasons to debate), does shift things noticeably. That, however, doesn't mean anything with regards to whether or not she should be included. Linkage, that is the decision about which offenses should be included, must be performed first - this is about the analysis of the locations of the offenses deemed to be linked. Omit a case you shouldn't, and that can be bad; include a case you shouldn't, and that can be bad; can be bad because some locations contain more information than others. Stride's location has a great influence on what routes the offender travells, while Kelly's location, being fairly near Chapman's, has less information in it (it's partially redundant with Chapman's location). Add in Tabram, for example, and again, that's doesn't change thigns a great deal because that location isn't all that informative beyond Chapman, Kelly, and Eddowes, etc. There will be some shifts and slight changes, but nothing dramatic.

Anyway, just want to make it clear, these should be viewed as suggestions, not solutions. And like any good suggestion, they should be considered but not always accepted, and certainly not accepted without something more tangible that backs them up.

For example, here's the David Berkowitz series (Son of Sam). Berkowitz was a commuter, as Trevor described. He falls in zone 44, and that means you have to search all of the maruder space (zones 1-40), and then start searching the commuter zones (41-80). Now, he's found very early in that commuter search, so it is still better than a chance search for a commuter, but really, if all one did was take a profile and start searching, it would take years to get there. It was real evidence (a traffic ticket that identified his car I believe), that led them to check him out. If they had a profile like this and dismissed following up that lead because of it, that would be a misuse of the information the profile provides. (Actually, my arguement would be that, he's located high in the commuter zones, so the profile would fit him; in fact, he lives right along one of the estimated travel routes - the red and green lines are estimated paths of travel, the blue line is something else, which I won't get into).

- Jeff

[JeffHamm]

One last little bit on this topic, for those who might be interested.

Ok, the basic problem with trying to assess whether or not geographical profiling is useful information (by which I mean the jeopardy maps are not just random blobs, but are better than random guessing) is that you can't do that by simply presenting single cases, which is what I've been doing above really. Those are just examples I've used to illustrate various points.

But those examples could just be a selection from random outputs. Do something randomly enough, and you will, by chance, be able to troll through those until you find something that looks like what you want to show.

Somehow, one needs to compare how these do relative to chance, or random searches. To make sense of what follows I just have to explain a little bit about what's been done. Over the years I've put together a collection of 32 different series of crimes, ranging from serial arson, rapes, and murders. All of which have been solved, and I've been able to find maps, of various quality, that have shown the crime locations, and the offender's residence. In some cases I've also been able to track down information concerning other potential anchor points (places of work, for example). For this collection there has to be at least 4 offenses, and I have to have identified at least one potential anchor point that would classify the offender as a marauder, since those are the cases that this sort of profiling is currently claiming to be based upon. Again, this is a very important point because there are a number of offenders that are commuters (although that term is often used with respect to the offender's residence, but really, the anchor point could just as easily be a pub they go to regularly, or their place of work, and if that location is inside the crime range, the offender should really be considered a marauder).

Anyway, we can then define the crime zone based upon the smallest circle that surrounds all the offense locations. That area generally underestimates an offender's territory (the area in which they would potentially consider committing an offense; basically, as a series continues the crime range will tend to expand as we get a more and more accurate estimate of the area in which they are willing to offend. And of course, that area may expand, or contract, over time, but I'm trying to keep the complexity at a minimum here and not end up down various rabbit holes that do emerge, and are, of course, valid concepts to consider). So, to account for this, I expand the crime range by increasing the radius of that smallest circle by 25%. The area of that circle is the "search area", and it gets divided up into 40 "zones", each comprising 2.5% of the search area (so all zones are of equal size). Now, if we just randomly picked locations in that circle until we found the pixel that contains the offender's residence, and did that over and over and over again, sometimes we might get it right on the very first guess, and sometimes it would be the very last pixel we chose (neither of those will happen very often of course). On average, though, we will find the offender after searching 50% of that search space (which we know will contain the offender's anchor point because I'm only looking at marauders, and they have to be inside this circle). Another way to put this is that if you search 50% of the search space you should find the offender 50% of the time.

Now, the resulting jeopardy map outputs are not constrained to that search space circle. But they produce a different probability value for each pixel on the whole map. So, I can just grab the highest probabilities until I have an area = 2.5% of the search space, and that is zone 1. The next set becomes zone 2, and so on. So zone 1-40 comprise an area of the map equal in area to the crime range circle I've defined. Zones 21-40 mean we are now searching an area greater than the random search would have to explore (I'm exploring if these zones might be worth considering as "commuter zones", and while Berkowitz does offer some hope it might be worth looking at that further, I'm not really convinced it will be that simple, but I don't have enough cases to explore that idea - it's just a hypothesis and needs to be put to a proper test).

Anyway, if the profiling routines, therefore, locate the anchor point anywhere inside of zone 1-20, then the profiling routine has done better than a random search would be expected to do, and I call that a "pass". If the routine locates the offender in zone 21 or more, I call that a fail (failed to do better than a random search). Failures are further classified as "misdirections" if the offender is located in zones 41-80 because that means you're now having to search more space than a random search would ever require, and as an "abject failure" if they are located beyond zone 80 (the areas where the profile more or less is saying "nothing to see here"). Similarly, passes are also classified as a "bulls eye" if the offender is located in zone 1 (as Radar is above).

In short, for the 31 cases I have, neither Rigel (Rosso's routines) nor the ones I'm working on produced any misdirections or abject failures. Both "passed" (better than chance) on 29 of the 31 cases, and both had 2 cases where the offender was found in a zone >20 (but less than 40). Both also had 8 bulls eyes (so 25.8% of the cases require searching less than 2.5% of the search space). Half of the cases are located after searching 9% and 6.5% of the search space (Rossmo's and mine, respectively), and 75% of the cases are located after searching 27.4% and 23.1%, respectively; remember random searching would detect 50% of the cases after searching 50% of the search space). While that might look like I'm trying to claim my routines are better, they're not, both approaches do equally well. What you can see below is a performance graph that summarizes performance. The diagonal dashed line is chance performance. To the extent the geographical profiles do better than random searching they will fall above that line. And as you can see, both Rigel (Rossmo's routines in purple) and (forgive me: Dr. Watson in red, a name chosen to emphasize that this is an assistant and is not the detective who solves the case) are well above that, and show similar rates of success. No, they don't always get it right, but they are far better than chance and so they do provide information worth considering. I'm still working on trying to improve the performance output, and hoping to obtain a larger sample of better quality information to work from as well. And, at some point, I want to see if I can work in a "marauder/commuter" separation side of things (since when a series is not solved yet, we can't do the ideal as I have and just analyse marauders - throw in a bunch of commuters and these performance values will go way down, but by how much remains to be seen).

Anyway, sorry for the long posts, but I think it's important for people to understand that the hype around this sort of thing is just that, hype, that might make for interesting TV or movie scenes, but those are hardly representative of reality. The reality is, these are just a way of analyzing the spatial patterns from a series of crimes. The analysis does provide information that can be helpful when used for what it is, a suggestion of locations that appear likely to be areas the offender has some sort of anchor point (a residence, a place of work, etc). With the huge amount of information that is generated in an investigation, it can be helpful to have a visual representation of that pattern, and one that indicates areas that have a better than chance possibility of being worthy of interest. But it is not magic, and it's not fool proof, and it should never be seen as something that over-rides far more telling information that is directly related to the investigation.

Ok, I'll come down off my soap box now.

[Darryl Kenyon]

One last little bit on this topic, for those who might be interested.

Ok, the basic problem with trying to assess whether or not geographical profiling is useful information (by which I mean the jeopardy maps are not just random blobs, but are better than random guessing) is that you can't do that by simply presenting single cases, which is what I've been doing above really. Those are just examples I've used to illustrate various points.

But those examples could just be a selection from random outputs. Do something randomly enough, and you will, by chance, be able to troll through those until you find something that looks like what you want to show.

Somehow, one needs to compare how these do relative to chance, or random searches. To make sense of what follows I just have to explain a little bit about what's been done. Over the years I've put together a collection of 32 different series of crimes, ranging from serial arson, rapes, and murders. All of which have been solved, and I've been able to find maps, of various quality, that have shown the crime locations, and the offender's residence. In some cases I've also been able to track down information concerning other potential anchor points (places of work, for example). For this collection there has to be at least 4 offenses, and I have to have identified at least one potential anchor point that would classify the offender as a marauder, since those are the cases that this sort of profiling is currently claiming to be based upon. Again, this is a very important point because there are a number of offenders that are commuters (although that term is often used with respect to the offender's residence, but really, the anchor point could just as easily be a pub they go to regularly, or their place of work, and if that location is inside the crime range, the offender should really be considered a marauder).

Anyway, we can then define the crime zone based upon the smallest circle that surrounds all the offense locations. That area generally underestimates an offender's territory (the area in which they would potentially consider committing an offense; basically, as a series continues the crime range will tend to expand as we get a more and more accurate estimate of the area in which they are willing to offend. And of course, that area may expand, or contract, over time, but I'm trying to keep the complexity at a minimum here and not end up down various rabbit holes that do emerge, and are, of course, valid concepts to consider). So, to account for this, I expand the crime range by increasing the radius of that smallest circle by 25%. The area of that circle is the "search area", and it gets divided up into 40 "zones", each comprising 2.5% of the search area (so all zones are of equal size). Now, if we just randomly picked locations in that circle until we found the pixel that contains the offender's residence, and did that over and over and over again, sometimes we might get it right on the very first guess, and sometimes it would be the very last pixel we chose (neither of those will happen very often of course). On average, though, we will find the offender after searching 50% of that search space (which we know will contain the offender's anchor point because I'm only looking at marauders, and they have to be inside this circle). Another way to put this is that if you search 50% of the search space you should find the offender 50% of the time.

Now, the resulting jeopardy map outputs are not constrained to that search space circle. But they produce a different probability value for each pixel on the whole map. So, I can just grab the highest probabilities until I have an area = 2.5% of the search space, and that is zone 1. The next set becomes zone 2, and so on. So zone 1-40 comprise an area of the map equal in area to the crime range circle I've defined. Zones 21-40 mean we are now searching an area greater than the random search would have to explore (I'm exploring if these zones might be worth considering as "commuter zones", and while Berkowitz does offer some hope it might be worth looking at that further, I'm not really convinced it will be that simple, but I don't have enough cases to explore that idea - it's just a hypothesis and needs to be put to a proper test).

Anyway, if the profiling routines, therefore, locate the anchor point anywhere inside of zone 1-20, then the profiling routine has done better than a random search would be expected to do, and I call that a "pass". If the routine locates the offender in zone 21 or more, I call that a fail (failed to do better than a random search). Failures are further classified as "misdirections" if the offender is located in zones 41-80 because that means you're now having to search more space than a random search would ever require, and as an "abject failure" if they are located beyond zone 80 (the areas where the profile more or less is saying "nothing to see here"). Similarly, passes are also classified as a "bulls eye" if the offender is located in zone 1 (as Radar is above).

In short, for the 31 cases I have, neither Rigel (Rosso's routines) nor the ones I'm working on produced any misdirections or abject failures. Both "passed" (better than chance) on 29 of the 31 cases, and both had 2 cases where the offender was found in a zone >20 (but less than 40). Both also had 8 bulls eyes (so 25.8% of the cases require searching less than 2.5% of the search space). Half of the cases are located after searching 9% and 6.5% of the search space (Rossmo's and mine, respectively), and 75% of the cases are located after searching 27.4% and 23.1%, respectively; remember random searching would detect 50% of the cases after searching 50% of the search space). While that might look like I'm trying to claim my routines are better, they're not, both approaches do equally well. What you can see below is a performance graph that summarizes performance. The diagonal dashed line is chance performance. To the extent the geographical profiles do better than random searching they will fall above that line. And as you can see, both Rigel (Rossmo's routines in purple) and (forgive me: Dr. Watson in red, a name chosen to emphasize that this is an assistant and is not the detective who solves the case) are well above that, and show similar rates of success. No, they don't always get it right, but they are far better than chance and so they do provide information worth considering. I'm still working on trying to improve the performance output, and hoping to obtain a larger sample of better quality information to work from as well. And, at some point, I want to see if I can work in a "marauder/commuter" separation side of things (since when a series is not solved yet, we can't do the ideal as I have and just analyse marauders - throw in a bunch of commuters and these performance values will go way down, but by how much remains to be seen).

Anyway, sorry for the long posts, but I think it's important for people to understand that the hype around this sort of thing is just that, hype, that might make for interesting TV or movie scenes, but those are hardly representative of reality. The reality is, these are just a way of analyzing the spatial patterns from a series of crimes. The analysis does provide information that can be helpful when used for what it is, a suggestion of locations that appear likely to be areas the offender has some sort of anchor point (a residence, a place of work, etc). With the huge amount of information that is generated in an investigation, it can be helpful to have a visual representation of that pattern, and one that indicates areas that have a better than chance possibility of being worthy of interest. But it is not magic, and it's not fool proof, and it should never be seen as something that over-rides far more telling information that is directly related to the investigation.

Ok, I'll come down off my soap box now.

Thank You Jeff, very interesting material