FROM: Grubin, D & Wingate, S.; Criminal Behaviour and Mental Health, 6, 349-359 1996 Whurr Publishers Ltd. Pages 349 - 357

DON GRUBIN AND SARAH WINGATE
University of Newcastle/Newcastle City Health Trust, Newcastle upon Tyne

Abstract

This paper identifies from the literature risk factors for reoffending in convicted sex offenders. Unless all sex offenders are to be incarcerated for indeterminate lengths of time, some means to separate higher from lower risk men is clearly necessary. The complexities of the move from identifying high-risk groups to predicting risk in specific individuals is discussed. It is argued that the focus should be shifted from the search for single, putatively predictive variables to an attempt to understand why recidivism occurs.

Introduction

When sex offenders reoffend the media interest and public anger that follow are often profound. This has been particularly evident in the United States where a number of sexual psychopath laws have been passed in various states, usually after the commission of particularly notorious sex crimes by known offenders (Grubin & Prentky, 1993). A frequently discussed example of this is Washington State's Sexual Predator Act which allows for the civil commitment of sex offenders for life if they are deemed to be sexual predators and likely to reoffend, together with community notification when sex offenders are released from prison (Wettstein, 1992; Brody and Green, 1994).

Public concern about recidivism by sex offenders, however, is in contrast with the relatively low levels of sexual reoffending that are in fact the case. The tension between public perception of sex offenders representing a high recidivism risk and the evident reality of statistics that demonstrate relatively low levels of sexual reoffending is in part influenced by a confusion between frequency and severity of reoffending, in part by the higher risk associated with some offenders, and in part by the nature of sex offending itself where any reoffence may be seen as unacceptable. Fisher and Thornton (1993), for example, have observed that there are 'a relatively large number of offenders who offend at a low rate, perhaps even just once, and a smaller number who offend at a relatively high rate' (p. 108). Amongst this latter group will be men whose reoffences are both frequent and severe. Some workers have argued that indefinite incarceration is appropriate in this small number of cases (Harris, Rice & Quinsey, 1993), but the difficulty is in accurately identifying those at most risk of serious reoffending whilst avoiding the unnecessary detention associated with false positive predictions.

In theory, a greater understanding of what predicts sex offence recidivism would be beneficial not only in deciding who needs to be locked up and for how long, but also in identifying those with particular treatment needs. But how can high-risk offenders be identified? Are actuarial approaches using static demographic variables superior to clinical ones based on changing circumstances and less quantifiable experience? Or is the prediction of sexual reoffending in specific cases little more than a gamble based on an imperfect understanding of the relevant odds?

Sex offender recidivism

Most sex offenders are not reconvicted for sex offences. Kaul (1993), for example, cites a 1960s follow-up of over 2900 Danish sex offenders (Christiansen, 1965) in which just 10% were convicted of another sex offence over a period of 12 to 24 years, although time at risk does not appear to have been considered in this study. A more recent UK report found that only 7% of a randomly selected sample of over 900 sex offenders (men with either current or past convictions for sex offences) released from prison in 1987 were reconvicted of a sex offence over the next four years (Marshall, 1994). Similarly, a meta-analysis involving 61 studies that included nearly 29 000 sex offenders followed up on average for four to five years found that sexual recidivism was 13% for child molesters and 19% for rapists, and reoffences of non-sexual violence 10% and 22% respectively (Hanson & Bussière, 1995). [Cfr. Hanson & Bussière 1998]

These studies suggest that less than one in five of a general sample of sex offenders released from prison go on to commit a further sex offence. Similar figures are not available for convicted sex offenders not sent to prison, but one would expect the recidivism rate to be even lower in this group. Such low reoffending rates are in contrast with recidivism studies in released male prisoners in general amongst whom reoffending rates are in the range of 50% over two years and 60% over four (Home Office, 1994), though in a 15 to 30 year Canadian follow-up study of non-sexual criminals released from prison the recidivism rate was over 80% (Hanson, Scott & Steffy, 1995). These findings would appear to support West's (1987) claim that the typical sex offender appears in court once and then never again, at least for further sex crime.

Even when one looks at studies of apparently higher risk offenders, reoffence rates remain relatively low. For example, in a sample of over 300 sex offenders who had committed more serious offences (i.e. those who had received sentences of at least four years) released from English prisons in 1980, it was found that just 15% of rapists and about a third of child molesters were reconvicted for a sex offence over the next 10 years (Thornton & Travers, 1991). Similarly, a large-scale research project in California designed to evaluate a sex offender treatment programme for prisoners aimed at more serious offenders has had difficulty in demonstrating any impact of treatment because of the low base rate of reoffending in the non-treatment controls: only 14% of 289 untreated child molesters and 14% of 80 untreated rapists followed for an average of about four years had been rearrested (as opposed to reconvicted) for new sex offences (Marques, Day, Nelson & West, 1994; Marques personal communication).

There are, of course, well-known difficulties in interpreting reoffending data, and it might be argued that the reoffences of sex offenders are less visible than those of other types of offender. First, although sexual reconviction rates are known to be proportional to actual offending rates (Hindelang, 1974), sex offences are generally under-reported, and most measures of sex offence recidivism will be an under-representation of true reoffence rates. Marshall and Barbaree (1988), for example, found that unofficial records and contacts uncovered 2.4 failures, and 2.7 victims, for every one documented by official statistics. Victim surveys imply even more unreliability in the official data, suggesting that up to 80% of sex offences may go unreported (Mayhew, Elliot & Dowds, 1989).

Undetected reoffences, however, are a problem for all recidivism studies. In terms of sex offenders, even studies with intense follow-up that do not depend on convictions do not reveal massive numbers of reoffences (Marshall & Barbaree, 1988; Marques et al., 1994). If anything, the numbers of undetected offences by non-sex offenders are probably greater in relation to crimes with higher base rates such as theft and general violence, particular given the fact that once an individual has been convicted of a sex crime he becomes known to the system..

Another potential problem in interpreting recidivism studies is the length of follow-up. Soothill and Gibbens (1978), for example, followed 174 men convicted of sex offences against girls under 13 for 24 years. They found that 11% of the sample had been reconvicted of a sex or violence offence over a five-year at-risk period, but by the end of the study the figure had increased to 18%, with no decrease in the level of severity. Other studies carried out by this group in relation to rapists and incest offenders (Soothill, Jack & Gibbens, 1976; Soothill, 1980; Gibbens, Soothill & Way, 1981 ) suggest that these other types of sex offenders are also at risk of reoffending for many years. Similarly, a 15-year minimum follow-up of 197 child molesters released from Canadian prisons between 1958 and 1974 found that over this period 42% were reconvicted for sexual or violent crimes, but about a quarter of these reconvictions occurred between 10 and 31 years after prison release (Hanson, Steffy & Gauthier, 1993). The authors do not distinguish between sexual and violent offences in this study, but in a more recently published analysis by the same group involving 191 of the child molesters, 35% were reconvicted of a sexual offence in the same follow-up period (Hanson, Scott & Steffy, 1995).

Thus, whilst the base rate of the behaviour underlying sex offending may not be high, it is persistent. It is important, however, to look more closely at those individuals whose reoffending is delayed for many years. In the Soothill and Gibbens (1978) study, all the reoffending by men with three or more previous convictions took place within five years, whilst those who did not reoffend for 10 years or more had just one or two prior offences. This suggests that long term follow-up is less necessary amongst some groups of sex offenders than others. It appears that shorter term follow-up is probably sufficient for more serious offenders, at least if a history of three or more previous convictions is taken as an indication of this (Thornton & Travers, 1991).

Whilst one must be cautious when looking at recidivism data, those who downplay them are at risk of overstating their case. Official statistics and follow-up studies in fact do appear to give a reasonable though conservative estimate of the extent of reoffending by sex offenders. What they make clear is that although sex offenders on the whole are not at great risk of reoffending, there is a need to identify the minority who are.

Predicting sex offence recidivism

It is often said that the best predictor of future behaviour is past behaviour. Various studies have demonstrated that recidivism rates in first offenders range from 10% to 21% compared with rates of 33% to 71% in those with past convictions (Marshall, Jones, Ward, Johnston & Barbaree, 1991). Marshall's follow-up study (1994) of nearly 13 000 offenders of all types released from prison in England and Wales in 1987, for example, found that men with a history of sex offences were more likely to be convicted of a sex offence over the next four years than those with no history of sexual offending: 7% of released prisoners with such a history were responsible for 31% of the subsequent sex offence convictions, a highly significant finding. The practical importance of this finding in terms of actual numbers, however, is less impressive than the statistical significance would suggest. Of the 926 men with a history of sex offending, just 61 (7%) were actually convicted of another sex offence, and though only 1% of the remaining sample were convicted of a sex offence, in real terms the latter 136 men represented twice as many future sex offenders.

Thus single variables, even those as important as a history of sex offending, are of limited use in predicting sex offence recidivism. Nevertheless, supporters of actuarial approaches to the prediction of recidivism would argue that the problem is one of collecting enough of the right variables to plug into an algorithm that will generate a reliable desk-top prediction of risk of sexual reoffending. For example, Quinsey and Walker (1992, P. 246) claim that 'enough work has been completed to establish a general consensus within the research community about the classes of variables that are valid predictors of recidivism'.

What might these variables be? Most studies focus on static demographic and historical factors such as childhood abuse, parental instability, marital status, relationship to victim, and criminal convictions. For example, Hanson, Steffy and Gauthier (1993) found that increased recidivism amongst child molesters was associated with never having been married and having previous sex offences, whilst in rapists Thornton and Travers (1991) reported that reoffending related to a combination of a current or past conviction for violence, four or more previous convictions of any sort, and age under 30 at the time of their sentences. Sugarman, Dumughn, Saad, Hinder and Bluglass (1994), in a study of exhibitionists, found that future contact sex offending could be predicted by childhood conduct disorder, early first conviction, criminal history, personality disorder and relationship difficulties.

Other groups have added more dynamic variables to the equation, in particular sexual arousal patterns. Rice, Harris and Quinsey (1990) found that both sexual and violent recidivism in rapists were well predicted by their phallometric (PPG) response to non-sexual violence in addition to their score on the Hare Psychopathy Checklist. They claim, in fact, that these two variables alone were as successful in predicting recidivism as the remainder of all the many demographic, psychiatric, criminal and offence variables used in the study.

These same workers in a later much more detailed actuarial study argued that a risk score based on the interaction of nine variables produced an estimate of the probability of reoffending by serious sex offenders that, for those with high scores, approached one (Quinsey, Rice & Harris, 1995). Not surprisingly, the variables accounting for most of the variance were the number of previous sex offence convictions and the number of previous prison sentences; the Hare psychopathy rating and deviant sexual arousal measured on PPG evaluation contributed only marginally to the variance, which is perhaps somewhat puzzling given their earlier claims regarding these latter two variables.

Statistically significant improvement in prediction does not necessarily translate into clinical relevance. In the Quinsey et al. study (1995), their algorithm was applied to a population of nearly 180 sex offenders treated in a maximum security hospital. As the risk score increased, so too did reoffending, and overall the authors claim that their method resulted in 72% correct decisions in relation to 'violent failure' and 77% correct decisions in relation to sexual reconvictions, both of which represented a relative improvement over chance of just over 40%. The vast majority of the sample, however, had risk scores giving a probability of reoffending that was less than 40%, which, even if a significant improvement over chance, is not particularly helpful to those who must make decisions about release. Thus, although the 50% success rate of tossing a coin to decide who was going to reoffend would not be as good as their method, it would not be all that much worse. It is only for those with the highest scores that the probability of reoffending becomes clinically meaningful at around 85%, but this accounted for just six men, 3% of their sample.

This study is perhaps the apotheosis of the actuarial approach, and as such it highlights well the fundamental defects inherent in it. First, it is wholly empirically driven and as such gives no reason to believe that findings from one population can be generalised to another. Second is the implication that human beings are entirely a function of their histories: static historical data do not change, which means that regardless of treatment or maturity the risk of reoffending should remain unaltered. But both these objections are practical ones that can be overcome. The most crucial difficulty arises from the fact that actuarial prediction is about groups, and unless we are talking about a high-frequency behaviour it can tell us little about individuals.

Many of these issues have been discussed in relation to general reoffending in mentally disordered patients (Gunn, 1993; Chiswick, 1995). The following example illustrates this last point. Imagine that an actuarial technique is developed that can predict with 90% accuracy who will reoffend and who will not, a figure much greater of course than anything currently available. Given 4000 sex offenders in prison with a recidivism rate of 10%, then 400 of these men will reoffend as opposed to 3600 who will not. With an actuarial tool of 90% accuracy, 360 of the 400 reoffenders would be identified successfully (and just 40 missed). We would also, however, incorrectly decide that 360 (i.e. 10%) of the 3600 non-offenders will reoffend. Thus, we would identify a group of 720 men predicted to reoffend of whom just 360 (50%) will actually do so. What this means is that we would be making good decisions about who will not reoffend (in the above example, only about 1% of those let out would do so), and we would identify well a high-risk group for continued detention, but for individuals within the high-risk group the odds of any one individual reoffending are still 50/50, that is, similar to tossing a coin.

Why do predictors of sexual recidivism predict?

Sex offenders are not simply bundles of variables. Characteristics that may be important to actuarials have little inherent meaning as they indicate associations but do not in themselves imply causation. They become useful in understanding recidivism only when their possible meaning in particular individuals is clarified.

In a review of violent recidivism, for example, Harris, Rice and Quinsey (1993) suggest that youth, marital status, psychopathy, criminal history, failure of prior conditional release, alcohol use and antisocial conduct in childhood have been demonstrated consistently to be good predictor variables. But what does this mean other than that young, single, impulsive, violent men, particularly when they drink, tend to behave in impulsive and violent ways? Similarly, in relation to men who offend against children, Rice, Quinsey and Harris's (1991) finding that reoffenders were more likely to have committed more sex offences, have more past prison sentences, suffer from personality disorders, have never married, and be sexually deviant on phallometric assessment tells us little more than that strongly paedophilic men who are prepared to act on their sexual preferences will continue to seek out children for sex.

In effect, variables like those described provide circular explanations for offending behaviour, and as such are of limited use in deciding whether an individual is likely to reoffend. For example, the reason why never-having been married is a good predictor of reoffending against children could simply be that these men have little interest in forming relationships with women; even if they did subsequently marry, their risk of reoffending is unlikely to decrease (indeed, one would be suspicious of such a man who suddenly announced his engagement, particularly if his girlfriend had children). The sample of men who have never married will include those with extreme deviance who cannot form such relationships, as well as those who may have a deficit in relation to their capacity for empathy, either of which could contribute causally to an individual's offending.

It may be, however, that being married does have a protective effect. Having a legitimate sexual partner could mean that men are less likely to find themselves in situations where their risk of offending is increased as they no longer need to hunt for partners. Alternatively, the emotional support of a spouse may decrease loneliness or feelings of inadequacy or rejection. Marshall (1989), and McKibben, Proulx and Lusignan (1994) describe the role of these negative emotions in the chain leading to deviant sexual behaviours.

Thus, the variable 'never having married' provides clinical utility only in so far as it tells us something about the personality and functioning of the man who displays it. Similarly, being in employment has been found to be a protective factor (Marques et al., 1994). But is this causal in nature, or is unemployment simply an indication that something else is not right?

The danger of ignoring the meaning of variables and concentrating instead only on their effects in particular studies is well illustrated in relation to the putative importance of the age of the offender, where different studies produce contradictory results. Young age in general tends to be a predictor of all types of crime. There are of course a variety of reasons for this, ranging from peer group pressure or incomplete socialisation to the fact that younger offenders may not have learnt to avoid detection of their crimes as competently as their older counterparts. These factors may also be at work in sexual offenders, explaining the findings of Thornton and Travers (1991) and Harris, Rice and Quinsey (1993), described above in relation to violent sex offenders. In contrast, Grunfeld and Noreik (1986) suggest that the effect of age on recidivism in sexual offenders is diametrically opposed to that expressed in most of the literature, with youth carrying a favourable prognosis. They argued that sexual offences committed by older men were the result of more deeply rooted tendencies or fixations that were unlikely to change over time. Clearly, post hoc explanations need to be treated with caution as plausible interpretations can be generated to account for any contradictory theory or finding.

The effect of age and its contribution to sexual offending and reoffending will only begin to be understood when it is treated as a richer concept than simply one to be plugged into recidivism equations. For instance, Richardson, Graham, Bhate and Kelly (1995) found that most adolescent sex offenders displayed difficulty in containing their behaviours across several aspects of their lives, demonstrated for example by truancy or aggression, and it is this, rather than an individual's age per se, that needs to be addressed. In order to predict future sexual dangerousness, it clearly will be necessary first to disentangle sexual psychopathology from other antisocial behaviours, and then to gain a better understanding of the natural history of each.

The most important predictor of sex offender recidivism is of course a history of past offending, both of a sexual and a general nature. This is often presented as a profound insight, but in reality this variable does little more than distinguish a group of men who have demonstrated that they are prepared to engage in particular behaviours. Similarly, those with more past offences have demonstrated their willingness to carry on with their behaviours in spite of the risk of apprehension, whereas in those who have only one or two such convictions this willingness will be present in only some. But other factors may also be at work in relation to past offences. A history of previous offending may increase the likelihood of future arrest as the individual will be known to the police. Also, for some older sex offenders with previous prison sentences, life inside prison may have become preferred to life outside prison.

In the large meta-analysis carried out by Hanson and Bussière (1995), the predictive value of 69 variables was tested. About one-third of these were significantly related to recidivism, and although the largest single predictor was the relatively dynamic variable of sexual preference for children as measured by phallometric methods, no variable was of sufficient strength to warrant its use in isolation. [Cfr. Hanson & Bussière 1996 & Hanson & Harris 1998]

Unfortunately, meta-analysis is not particularly good at demonstrating multivariate effects, which require methodologies of a more complex type than one usually finds in follow-up studies. One exception to this is the work of Malamuth and his colleagues (Malamuth, 1986; Malamuth, Sockloskie, Koss & Tanaka, 1991), who have reported that sexual violence in university students followed up over 10 years is dependent on their sexual arousal in response to aggression, the extent to which dominance is a sexual motive for them, their hostility towards women, their attitudes to the acceptance of violence against women, their psychoticism and their prior sexual experience. Their analysis, making use of path analytic and similar statistical techniques, demonstrates well the richness that can emerge when one moves away from looking at single, static variables.

The Malamuth work also demonstrates that more dynamic variables than those related to life history can provide useful insights both in terms of understanding the causes of sexual offending in particular individuals, and in identifying treatment targets. Factors such as affect, levels of anger, general social skills, empathy and self-esteem may be crucial in specific cases, but they are not easily quantifiable for the purposes of recidivism studies. In their meta-analysis, however, Hanson and Bussière (1995) found that dynamic variables were notable in their absence from much of the research; in their large review of recidivism research, Furby, Weinrott & Blackshaw (1989) also noted a dearth of studies looking at dynamic as opposed to static risk factors. This is partly because of difficulties in assessment, but also because by their very nature they change over time, making quantification difficult.

Whilst their changeable nature makes dynamic variables problematic from an actuarial point of view, however, it is this very characteristic that makes them useful for predicting reoffending by individuals in a clinical context. As Hanson and Bussière (1995) point out, clinicians, particularly those involved in relapse prevention, have identified a number of factors such as anger, self-esteem and low victim empathy that may be important precursors to reoffending near to the time when reoffending takes place (Pithers, Kashima, Cummings, Beal & Buell, 1988). Because it is state rather than trait that appears to be the most relevant in terms of reoffending, factors that tend to be invisible to the actuaries may be crucially important to clinicians in determining when intervention is necessary.

Conclusion

There is a general consensus in the literature about those biographical variables that predict recidivism, but there is much less discussion about why these variables have the effects they do. Overall, identification of those at low risk and those at exceptionally high risk of reoffending is good, although prediction of the future offending behaviour of individuals in between is much less satisfactory. For actuarial studies to have clinical relevance, they must also be shown to be able to contribute to risk assessment in specific cases, and to help formulate treatment needs. This will require movement away from a blind reliance on variables for their own sake to a fuller understanding of how these variables relate to the phenomenon of sexual offending.

Acknowledgements

SW is supported by a grant from the Mental Health Foundation.

References