Detecting Child-Sex Associations; why are they in the too diffcult box?

Yesterday, I was speaking at a conference addressing unconscious bias and vulnerability. I was in discussion with a number of people working in child protection, and in helping survivors of child abuse. I was reminded of some work I had done back in 2009 when I developed a version of my implicit testing platform to detect child-sex associations. Why has everyone gone quiet? The idea that we could develop a metric capable of detecting potential paedophiles was attractive. There had been a number of studies in clinical/forensic settings suggesting that implicit testing could distinguish between convicted paedophiles and other offenders. Most of those studies had some flaw in terms of sample size or the way the test had been configured (e.g. too general in the survivor groupings or the sexual constructs). And so I began to develop anew metric.

Consent agreed photographs were obtained for this purpose of children in the four target groups: Boys and Girls, appearing Under and Over 8 years of age.  A samples was taken in 2009 ( N=72 to 86 per test ) using paid staffing agency subjects from diverse backgrounds and of varying ages ( all over the age of 18 years) and gender groups who were employed to be tested on other versions of Implicitly?.  All of the subjects completed the four Child-Sex tests on the basis that testing was anonymous and that they were simply providing baseline data.  Results suggest that the large majority of adult respondents, as predicted, had negative child-sex associations (between 82.1% and 95.9% depending on the test used). Overall 28 of the 316 respondents produced positive scores. The positive child sex associations seen were all quite weak and under 200ms ( all less than one standard deviation and all but 5 of the 28 positive scores were less than half a standard deviation) and focused mainly on boys, and predominantly boys over 8 years of age. The strongest negative associations tended to be seen towards boys and girls under 8 years of age.  So far so good.

The next obvious step was to take three new samples to ensure that the test could distinguish between offender and non-offenders, but also one important new group; survivors of child sex abuse. Without the final group there was risk that people could be identified as potential paedophiles for having stronger child sex association, when their child sex associations may have been developed as a consequence of them being the victim of abuse.

And here the trail grows cold. Negotiations to access sex offenders and survivors had so many ethical and legal issues, and I have to say occasions of risk adverse managers, that I put the idea into the too difficult box. It made bias testing around ethnicity, gender, disability and sexual orientation look like a walk in the park. I returned to ploughing my social bias furrow.

But I have one final throw of the dice; that such an instrument could be used as part of the supervision of those working in child sex abuse, as a confidential and self directed test. Taken periodically, to ensure that exposure toi such work might invoke child sex associations in staff.

You can stop reading now, but if this is your thing, here is the 2009 write up:

Early trials of a psychometric test for child sex predilections to screen job applicants seeking work with children

P Jones, Shire Professional

Introduction

Paedophiles are motivated to hide their thoughts, feelings about their sexual beliefs and attraction toward children. Prospective employers with a statutory duty to protect children in their care are forced to depend upon debarring (DBS) checks of the sexual predilections of job applicants. DBS checks are carried out based upon previous convictions before a court, cautions administered and any reports or intelligence on an individual which although insufficient to warrant a conviction, were enough to arouse concerns.  The checks use only past behaviour as a predictor of future behaviour which is insufficient with populations whose offences have yet to be uncovered or are yet to begin. In dealing with offenders, the clinicians, police and social service are forced to rely on hard external evidence or self-reports (e.g. interviews or questionnaires). Given the strong motivation of paedophilic offenders to hide both their desire to offend and their sexual attraction toward children such methods are unlikely to be fruitful.

 The problems of a reluctance or inability to self disclose is well known in other areas where the predilection is hidden, for example being racist, sexist or homophobic. Psychologists have developed tests that aim to measure beliefs through indirect or implicit metrics. The Implicit Association Test (IAT; Greenwald, McGhee, & ; Schwartz, 1998) purports to measure socially unattractive associations that test takes have sought to hide (Greenwald et al., 1998; Swanson, Rudman, & ; Greenwald, 2001). The IAT is a series of simple sorting tasks where the speed and accuracy with which respondents sort words and pictures is used as a metric for the additional processing required to overcome an association between a group (the concept) and an idea (construct). It has been postulated (e.g. Gray et al, 2005) that by using Adults and Children as the Concept, along with Sexual and Non-sexual words as the Construct it may be possible to measure the strength of the child-sex association. The critical evidence for such an assertion is criterion validity as this data may relate test scores to the predilection.

Gray et al (2005) used a child-sex IAT as the basis for discriminating between offenders and a control group. This study showed that paedophiles have an association between children and sex, whereas non-paedophilic offenders had an association between adults and sex. They considered that the IAT 'could identify a core cognitive abnormality that may underpin some paedophilic deviant sexual behaviour'.  Gannon, Rose and Williams (2009) conducted a similar study to Gray et al, using an IAT with female sex offenders. Contrary to Gray et al's study Gannon et al found that convicted female child sex offenders did not differ from non-sexual offending females on the Implicit Association Test. In fact, female child sex offenders appeared more likely to associate adults with sexual concepts rather than they did children.

 However, the metric and method used in both of these studies may have been flawed. Firstly both the Gray et al (2005) and the Gannon et al (2009) studies used relatively small samples ( N= 78 and 34). Of these participants only 18 in the Gray et al study and 17 in the Gannon and Rose study had committed paedophilic offences against children. The remainder had committed offences against children but these were not seen as paedophilia offences. Secondly, the version of the IAT used in both studies (Grenwald et al, 1998) has been the subject of much technical criticism ( see Blanton and Jaccard, 2006). Finally, both studies used only a generic word stimulus. The Adult concept was represented with words such as Mortgage, Office, Work, Mature, Beard and Marriage. The Child concept was represented by words such as Kid, Toys, Sweets, Infant, Playground and School. The construct was represented by sexual, and non-sexual (neutral) words.  This assumes that all sex offenders in the samples would respond to the word stimuli in the same way as representing their victim group. Therefore words such as infant, toy and playground may be associated very differently in the minds of different offender groups depending upon whether they were attracted to younger or older children.  Often child sex offenders have target groups of children towards whom they are sexually attracted. Offenders may be attracted to boys or girls (or both) and to different age groups. A key age in this regard is around 8 years of age.  

Method

 Using the Implicitly? IAT platform four new IATs were developed using gender and age specific images to represent the concepts rather than words. Implicitly? used a 5 stage IAT with two in-built practice trials and a full pre-test practice which raises the test-retest reliability to .83, compared to the .50 to .65 ( See Lane et al., 2007) often seen with the Greenwald et al test.  Respondents are presented with randomly selected words and pictures to sort into the correct concepts and constructs. The test presents them with two sorting rules, one where Child images and Sexual words are sorted together, and one where they are sorted oppositely. The time difference taken between the two conditions are used as a measure of the strength of association. Research shows that respondents are very rarely able to fake the results ( Steffens, 2004 ) and the Implicitly? platform has additional checks looking for faking profiles.

Consent agreed photographs were obtained for this purpose of children in the four target groups: Boys and Girls, appearing Under and Over 8 years of age. Rather than using adults as the opposite construct to the 'Child' concept the tests used images of objects which the researchers hoped would not be associated with sex by respondents. They still used a sexual and a neutral Construct just as Gray et al., and Gannon and Rose had done. This use of both neutral images (concepts) and neutral non-sexual words (construct) was designed to remove any contaminating effect from those who might find both adults and children sexually attractive which would produce no effect on the IAT. As the topic under study was the child-sex association this neutral stance looked only for a test score suggesting a child sex association, not an adult-sex association. In addition as the full extent of the sexual orientation of the subjects is unknown, using neutral objects rather than adult pictures removed the issues around which adult gender group a respondent may associate with sex. 

Sample

A samples was taken in 2009 ( N=72 to 86 per test ) using paid staffing agency subjects from diverse backgrounds and of varying ages ( all over the age of 18 years) and gender groups who were employed to be tested on other versions of Implicitly?.  All of the subjects completed the four Child-Sex tests on the basis that testing was anonymous and that they were simply providing baseline data.

A child-sex association would be represented by more rapid allocation of child pictures with sexual words producing a positive score on the test. If no association were evident, one might expect respondents to allocate child images with sexual words equally quickly as they do with neutral images. A negative child sex association, such as one might hope for and expect with most adults, would result because they less readily allocate Child images with sexual words than they do neutral words and this would produce a negative score.

Analysis

In the Gray et al (2005) and Gannon and Rose (2009) studies the raw response times and test errors were transformed to produce a D score via the scoring technique recommended by Greenwald, Nosek, and Banaji (2003). This D score uses the IAT practice trials, includes a penalty for incorrect trials, and expresses the IAT effect in terms of the variance of the latency measures. In this study D transformations were not used. Blanton and Jaccard (2006) had criticised such transformations as being about getting better measurement per se, and not in an attempt to better measure the construct in a more robust fashion. They had argued that by transforming the pure time measure (raw score) but then continuing to assume that the D score was a direct measure of the construct , the D scoring had contributed to make the metric arbitrary. Thus raw scores were used in this analysis and only a descriptive interpretation is given.

A correction to the scoring was applied to account for the test direction (which construct was first paired with which concept) after Jones (2009) to deal with cognitive inertia.

Results were grouped based upon the difference in mean time taken to respond.

Analysis showed that the errors made in each test stage were independent of the overall test score or the time taken for each stage of the test.

Results

 Results from the 2009 sample suggest that the large majority of adult respondents, as predicted, had negative child-sex associations (between 82.1% and 95.9% depending on the test used). Overall 28 of the 316 respondents produced positive scores. The positive child sex associations seen were all quite weak and under 200ms ( all less than one standard deviation and all but 5 of the 28 positive scores were less than half a standard deviation) and focused mainly on boys, and predominantly boys over 8 years of age. The strongest negative associations tended to be seen towards boys and girls under 8 years of age.

Discussion around the use of IAT to detect child-sex associations will no doubt be controversial, but it is not suggested that IAT be the only method by which those seeking employment or access to children can be screened. Results should be taken in conjunction with other data (interviews, references, CRB checks) and used to direct these other tools, for example by exploring the scores with the candidate.

 IAT may prove a very useful tool in screening job applicants seeking work with children, especially where there may already be concerns. IAT might also be made target specific, for example using images of specific children in child custody or child abuse investigations. It may also have a place in forensic and clinical settings in establishing the true beliefs and intentions of those seeking custodial discharge having offended against children. However, as yet the child-sex versions of Implicitly? are not criterion referenced. Gray et al (2005) warn that a child sex association may be developed not only because an adult may hold thoughts and feelings about their sexual beliefs and attraction toward children, but because they have been a witness to or victim of abuse. Until the criterion evidence is available to show that scores are related to predilection, the 'hit' vs 'false alarm' rate established, and the metric to metric established (also Blanton and Jaccard, 2006) IAT remains a research tool and unsuitable for use in making high risk decisions. Accessing such criterion sample is no easy task. The funding for organisations such as the Lucy Faithful Foundation who work with people who have offended or who feel they may have child-sex associations is restricted and under constant pressure. and access to offenders in custody is very difficult One lower risk application which may already be possible may be in helping those working with abused children or their perpetrators to identify when their work begins to affect their own sexual associations in a coaching or health screening role. As an ipsative benchmarking measure IAT may afford early notice of developing associations.

References

 Blanton, H., & ; Jaccard, J. (2006). Arbitrary metrics in psychology. American Psychologist., 61, 27-41.

 Gannon, T. A., Rose, M. R., & ; Williams, S. E. (2009). Do female child molesters holdimplicit associations between children and sex? A preliminary investigation. Journal ofSexual Aggression, 15, 55-61.

Gray N.S., Brown A.A., MacCulloch M.J., Smith, J. & ; Snowden, R.J. (2005) An Implicit Test of the Associations Between Children and Sex in Pedophiles. Journal of Abnormal Psychology 2005, Vol. 114, No. 2, 304–308.

 Greenwald, A. G., McGhee, J. L., & ; Schwartz, J. L. (1998). Measuringindividual differences in implicit cognition: The implicit association test. Journal of Personality and Social Psychology, 74, 1464–1480.

Greenwald, A. G., Nosek, B. A., & ; Banaji, M. R. (2003). Understandingand using the Implicit Association Test: I. An improved scoring algorithm. Journal of Personality and Social Psychology, 85, 197–216.

Jones P.C. (2009) Order Effects In The Use Of The Implicit Association Test Presented at the British Psychological DOP Conference Blackpool January 2009.

Lane K.A., Banaji M.R., Nosek B. A., Greenwald A.G. (2007) Understanding and using the Implicit Association Test. In B Wittenbrink and N Schwarz (Eds) Implicit Measures of Attitude (pp. 59-102) New York. The Guilford Press.

Steffens M.C. (2004) Is the Implicit Association Test Immune to Faking? Experimental Psychology 2004; Vol. 51(3): 165D179.

Swanson, J. E., Rudman, L. A., & ; Greenwald, A. G. (2001). Using the Implicit Association Test to investigate attitude-behaviour consistency for stigmatised behaviour. Cognition and Emotion, 15, 207–230.