Depends what you mean by “types“, I guess…?
Given the variation in the types of sexual convictions that exist – adults vs. children, contact vs. non-contact, stranger vs. acquaintance, etc. – it’s no surprise there have been various attempts to classify those convicted into groups. We did it with approximately 2,000 of them.
In a recent study, we used a statistical technique called latent profile analysis to identify groups from a large dataset of psychological test scores. These were pre-program scores on 92 scales from 17 psychological measures, from a sample of 2,245 adult men who had participated in the Sex Offender Treatment Program between 2003 and 2014 across England and Wales.
As far as we were aware, you can’t simply plug 90-odd scales into a latent profile analysis and see what happens. You identify a smaller number of variables representing the breadth of the construct of interest. For us, that was dynamic risk related to sex offences – psychological factors associated with future sexual reconvictions. We needed to pare our data back a little.
Some 40 scales measured constructs we now know not to be statistically associated with sexual reoffending. The remaining 52 scales were first subjected to a statistical clustering technique called exploratory factor analysis, which classified them into four factors: (1) socio-affective and emotional management; (2) sexual preoccupation/interests (child specific); (3) sexual preoccupation/interests (non-child specific); and (4) pro-offending thinking. Quelle surprise! – the scales were chosen for the SOTP specifically to represent those domains of risk (socio-affective, cognitive, and sexual), as articulated in the Structured Assessment of Risk and Needs.
We chose six scales that (a) were most highly associated with their domain (via their “factor loading“) and (b) were a good representation of the underlying construct for the domain. ‘Why six scales for four domains, though?‘ Good question. For sexual preoccupation (non-child), we chose one scale representing rape fantasy and another representing sexual obsession (subtly different, we thought). Then for pro-offending thinking, we chose a scale related to beliefs about adult victims and another one for beliefs about child victims. These six scales were:
- Impulsivity and carelessness.
- Sexual fantasies related to child molestation.
- Sexual fantasies related to rape.
- Sexual obsession and preoccupation.
- Pro-offending beliefs related to child molestation.
- Pro-offending beliefs related to rape.
Latent profile analysis (LPA) uses a Gaussian mixture model to… (no, I’m just kidding – no complex stats here, please). It’s a statistical technique used to uncover hidden (“latent“) groupings within a larger population. Like sorting a bag of marbles, each marble sporting a different array of colors. Some strike you as disproportionately characterized by blue, grey, and white, so you group those together. Others appear redder, yellower, and… purpler? More purple. So, you group those together too. And so on.
We grouped people based on psychological test scores. But the goal is the same: to identify distinct profiles in a larger population that have similar patterns of latent characteristics, thus creating smaller, more homogenous groups that share common traits.
But this grouping could go on forever! Piles upon piles and piles of marbles! Sub-division ad infinitum! Salami-slicing ad absurdum! As it turns out, when deciding what number of profiles is correct, one cannot have one’s CAIC and eat it too! (A little stats joke for you there.) We used a variety of metrics to make sure we stopped at the optimal number of profiles – enough to represent enough variation in our data but not so many that the results can’t be generalized beyond our data.
Five profiles, our metrics told us. They were:
- Low psychological impairment (51% of the cases): Relatively lower scores on all six scales.
- Impulsive (8%): Higher impulsivity and large variation in sexual obsession.
- Distorted thinker (12%): Both relatively higher rape- and higher child-related pro-offending thinking errors.
- Rape preoccupied (8%): Relatively higher rape fantasy and sexual obsession.
- Child fantasist (20%): Relatively higher child fantasy.
But are they really smaller, more homogenous groups with common traits? Reeeeally…? To answer that, we can validate. We validated in two ways.
First, we held back a random subset representing 20% of the larger dataset to re-run our analysis. If our profiles are a true representation of groups that exist in the real world, then they should reappear in any relevant dataset we test for them in. And they did. Solution validation? Check.
Second, if our profiles are a true representation of groups that exist in the real world, we should be able to make predictions about their performance on other relevant variables. We used the low-psychological impairment profile as our baseline or reference point (i.e., we compared the other four profiles to it, given it represents an absence of character: a null profile, if you will) and checked to see if our profiles differed in criminogenically meaningful ways. And they did. Criterion validation? Check. For example:
- The impulsive group had the lowest IQ; the child-fantasists the highest.
- The impulsive group had the most general priors; the child-fantasists the fewest.
- The rape preoccupied and child fantasists had the most sexual priors.
- The “thinkers” and “fantasists” were older; the impulsive and preoccupied, younger.
And, interestingly, the other four profiles were all lower on socially desirable responding compared to the low psychological impairment profile. Even the smart folk at the British Journal of Psychiatry granted us leave to delve into a little post-hoc resulting to see if the LPI profile was truly just a shower of fibbers, falsifiers, and fabricators. Not quite the case, but some were prevaricating. After a swift Saunders correction of the test scores for those in the LPI profile, approximately one-third could subsequently be allocated to one of the other latent profiles, once socially desirable responding was controlled for.
All great, so far. But so what? What’s the point?
Well, sexual violence is a pervasive international public health concern and evidence-based methods for classifying criminals is critical to efficient use of finite (and often stretched) criminal justice resources. This includes better planning of interventions between those who require interventions focused more on self-regulation and emotion management and less on sexual interests and post-offence rationalisation (and vice versa). Or being able to more efficiently distribute pharmaceuticals aimed at reducing sexual preoccupation and arousal. While the usual caveats apply about needing more and more evidence to drill down on groups, we hope to have added a data-driven typology of men convicted of sexual offences to the literature.
