Effectiveness of Pretrial Community-Based Diversion in Reducing Reoffending by Adult Intrafamilial Child Sex Offenders

Research design

A retrospective quasi-experimental research design was applied in this study. The treatment variable in the study had two levels: offenders accepted into the Program (coded as 1) and offenders who were declined entry into the Program and experienced the regular court process (coded as 0). The accepted group (n = 92) was defined on the basis of intention to treat and comprised participants who were offered a place in the Program without distinguishing treatment completers from noncompleters.

Of the 92 participants who composed the accepted group, 58% (n = 53) completed treatment within the Program and 42% (n = 39) did not complete the Program (32 breached the terms of the Program and 7 voluntarily withdrew from the Program before completion). The declined group (n = 120) were intrafamilial sex offenders who were referred to the Program but were declined treatment and attended traditional court and sentencing procedures. Participants were declined entry if they did not meet the statutory criteria specified above or were deemed unsuitable by the Program Director.

Before comparing the two groups on sex offending outcomes, propensity score matching was conducted to control for known differences between the accepted and declined groups. Groups were compared on outcomes related to their absolute risk of sexually reoffending, time to sexually reoffend, and relative risk reduction in sexual offending.

Although a sample of 212 intrafamilial sex offenders is relatively large in comparison to prior studies of this sex offender subgroup, it is small in terms of evaluating the effectiveness of a treatment program using quantitative methods. An analysis of the statistical power of the research design was conducted (Rosner, 2006). Using an effect size consistent with that found for similar treatment programs of a 6% point difference, with an alpha value of .05 and a sample of 212, the power of the design was estimated as .45. This indicated a 45% chance of finding a statistically significant difference when there is an actual difference of 6% between the treatment and control groups and is a low level of statistical power.

Participant Sample

All 214 adult child sex offenders referred by the New South Wales Courts to the Diversion Program between 1989 (the Diversion Program’s inception) and 2003 were eligible participants. A separate study drawing on the same sample compared 96 biological and 118 nonbiological fathers with respect to their demographic features, characteristics of the index victim and index offense, prior offending, and sexual, violent, and nonviolent recidivism (Titcomb, Goodman-Delahunty, & Waubert de Puiseau, 2012).

Two offenders were excluded from further analyses: one participant who died soon after his last contact with the Program and one participant whose file lacked sufficient information about variables relevant to this study. Therefore, analyses were conducted on a participant sample of 212.

Of the participants, a relatively small proportion was Indigenous (6.6%, n = 14). The remaining participants identified as Caucasian. At the time of the index offense, participants ranged in age from 23 to 68 years (M = 39.60, SD = 7.54), with a majority (44.3%, n = 94) between 35 and 43 years of age. As displayed in Table 1, participants in the accepted and declined groups were comparable on most demographic variables. Employment history was the sole variable that differed more widely in the two groups: Accepted participants had a more stable employment history than did declined participants.

OPEN IN VIEWER

Table 1:

Demographic Characteristics of Participants in Accepted and Declined Groups

		Accepted Group		Declined Group
Variable		%	n	%	n
Continuous variable (M, SD)
Age in years		39.50	7.13	39.65	7.90
Categorical variables (%, n)
Indigenous	Yes	5.40	5	7.50	9
	No	94.60	87	92.50	111
Employment*	Steady	73.30	66	52.90	63
	Sporadic	11.10	10	16.80	20
	Difficulty maintaining	8.90	8	11.80	14
	Unknown	6.70	6	18.50	22
Marital status at time of index offense	Single or never married	1.10	1	3.40	4
	Legal marriage	69.60	64	60.50	72
	Common law marriage	16.30	15	25.20	30
	Separated	8.70	8	6.20	8
	Divorce	4.30	4	3.40	4
	Widowed	—		0.80	1
	Unknown	—		0.80	1
History of violence against spouse or partner	No history	34.50	29	26.60	25
	Some history	38.10	32	30.90	29
	Extensive history	17.90	15	16.00	15
	Insufficient information	9.50	8	26.60	25

Note. Values are % and n unless otherwise noted.

*

p < .05.

Table 2 displays the proportion of each index offense-related characteristic in the accepted and declined groups. On most index offense characteristics, the groups of participants were similar; two characteristics revealed larger differences. Participants in the declined group were less varied in their range of sexual offending (committed only one type of sexual behavior; e.g., sexual touching) and were more likely to have abused their victim(s) in an unplanned and impulsive manner compared to participants in the accepted group.

OPEN IN VIEWER

Table 2:

Index Offense Characteristics of Participants in Accepted and Declined Groups

		Accepted Group		Declined Group
Variables		%	n	%	n
Continuous variables (M, SD)
Number of charges laid		5.88	5.46	5.52	5.84
Age of victim in years		8.70	3.18	8.96	3.56
Duration in years		3.13	2.89	3.31	3.58
No. of victims	Range = 1–5	1.23	0.54	1.23	0.59
Categorical variables (%, n)
Relationship to victim	Biological child	46.74	43	43.33	52
	Nonbiological child	53.26	49	56.67	68
Gender of victim	Female	89.13	82	92.50	111
	Male	4.35	4	5.00	6
	Both female and male	6.52	6	2.50	3
Range of sexual offending*	Little variation	3.50	3	13.20	12
	Limited number of acts	24.40	21	34.10	31
	Variety of behaviors	70.90	61	49.50	45
	Unknown	1.20	1	3.30	3
Sexual contact	None	—		0.80	1
	Touching	16.30	15	15.00	18
	Penetration	83.70	77	84.20	101
Premeditation*	Unplanned or impulsive	1.10	1	2.50	3
	Some planning	20.00	18	32.50	39
	Clearly planned	77.80	70	60.80	73
	Unknown	1.10	1	4.20	5
Use of weapon	None	100.00	92	97.50	117
	Blunt object	—		1.70	2
	Unknown	—		0.80	1

Note. Values are % and n unless otherwise noted.

*

p < .05.

Table 3 displays the prior offense records of the participant groups. Participants in the declined group had more extensive past offending rates than did accepted participants. Accepted participants were less likely to endorse sexual abuse myths than were their declined counterparts. Otherwise participants in both groups appeared similar in relation to prior offending characteristics.

OPEN IN VIEWER

Table 3:

Prior Offending and Tolerance of Sexual Abuse by Participants in Accepted and Declined Groups

		Accepted Group		Declined Group
Categorical variables (%, n)		%	n	%	n
Static-99 scores	Zero level of risk	69.57	64	62.50	75
	Some level of risk (1-7)	30.43	28	37.50	45
Some prior offending*		46.70	43	60.80	73
Prior sexual offending		12.00	11	10.80	13
Tolerance of sexual abuse*	No endorsement of sexual abuse myths	62.00	57	13.80	13
	Agreed with some sexual abuse myths	31.50	29	50.00	47
	Endorsed many sexual abuse myths	6.50	6	36.20	34

Note. Values are % and n unless otherwise noted.

*

p < .05.

Data Sources

Dependent Measures of Reoffending

The outcome variable was sexual reoffenses determined by an official report to the NSW Police. Reported sexual reoffenses were defined at the most inclusive level to include sexual reoffenses with adults and extrafamilial and intrafamilial victims including contact and noncontact offenses. Police reports were used in preference to arrest and conviction rates as report rates remain verifiable and were likely to capture with more accuracy the level of sexual reoffending compared to arrest and conviction rates. Although report rates offer a more accurate rate of sexual reoffending compared to arrest or conviction rates, it is acknowledged that report rates may include charges that are subsequently dropped (Gelb, 2007).

The follow-up period was operationalized as the number of days between the date of each participant’s final contact with the Program (on assessment completion, treatment completion, breach or withdrawal) and October 1, 2007. Days when offenders were in custody and unavailable to reoffend were subtracted based on information provided by the Department of Corrective Services to avoid inflating the desistance period. The follow-up period ranged from 2.8 to 18 years for the full participant pool (N = 212), with median follow-up times for the accepted group and declined groups of 4,618 days and 3,451 days, respectively. Cox regression was used to provide a test of treatment effect as it accounts for any impact of differences in follow-up time between the groups (Prentky, Lee, Knight, & Cerce, 1997).

Data Analysis

Statistical procedures to compute the propensity score analysis were performed using STATA IC11 Statistical Software with an add-on procedure called psmatch2 that enables propensity score matching, common support graphing, and covariate balance testing. All statistics were generated using two-tail tests at 95% confidence with alpha set at .05, unless otherwise stated. Pretest differences on categorical characteristics were calculated using Fisher’s exact χ² test; differences on continuous characteristics were calculated using the Mann–Whitney test.

Procedure: Propensity Score Analysis

Propensity score analysis is a popular technique to estimate causal treatment effects when comparing groups in an observational study while controlling for selection bias (Caliendo & Kopeinig, 2005). Based on each participant’s observed characteristics (e.g., pretreatment demographic and clinical characteristics), statistical analysis is used to calculate the probability of (or propensity for) each participant to receive treatment (Baser, 2006). This procedure combines each participant’s observed differences in characteristics into a single “propensity score.” Next, individual participants in the accepted group are matched to individual participants in the declined group based on nearest matches of propensity scores (Rudner & Peyton, 2006). After accepted and declined participants have been matched and covariate balance established, the outcome variables can be compared—in this case sexual reoffending rates and period of desistance from reoffending (Baser, 2006). Any differences in these outcome measures will be attributable to the Program and not to any known and observed differences that existed between the participant groups before the study began. For a full description of the rationale and methods underlying the use of a propensity score analysis, see Caliendo and Kopeinig (2005), Guo and Fraser (2010), and Joffe and Rosenbaum (1999).

Prior studies were reviewed to determine predictors of acceptance into treatment and factors contributing to sexual recidivism. The selected predictors were cross-checked against the variables available in the Program Database. Sample sizes limit the number of variables that can be included in logistic models (Weitzen, Lapane, Toledano, Hume, & Mor, 2004). Given a sample size of 212 participants, it was not possible to include all potentially relevant variables in the matching procedure. As such, only those variables that were identified as the most important in the literature and that did not suffer from excessive missing data were considered for the logistic regression model used in the propensity score matching analysis.

Using a combination of statistical significance and cross-validation approaches in an iterative process (as specified by Caliendo & Kopeinig, 2005), a logistic regression model was formed to generate a propensity score for each participant for selection into the Program. Unlike a typical regression analysis where the goal is often to build the best predictive model, in propensity score analysis the building of the logistic model is guided primarily by the utility of the model to create samples balanced across all the relevant covariates. As such, the variables ultimately used in the logistic regression model as displayed in Table 4 are those that created the best covariate balance across all available variables between the accepted and declined groups. Several variables were recoded for use in the logistic model (participants’ age at referral to the Program, number of victims, number of index charges, employment history) either because some cell sizes were too small or lacked a distribution approximating normality or because the logistic model performed better. Interactions among variables were not analyzed in this study.

OPEN IN VIEWER

Table 4:

Logistic Regression Model to Estimate Propensity Scores

Variable	Categories	Coeff.	SE	p > z	95% Confidence Interval
Risk of sexual reoffending (Static-99 score)	Low or high risk	−0.01	0.36	0.99	−0.72	0.71
Indigenous	Indigenous	−0.11	0.65	0.87	−1.38	1.17
Employment history	Poor employment	0.26	0.56	0.64	−0.84	1.37
	Steady employment	1.10	0.49	0.03	0.14	2.07
Age of participant	Younger than 36 years	−0.19	0.41	0.65	−0.99	0.61
	41 years and older	−0.49	0.38	0.20	−1.24	0.26
Number of index charges	6 or more	−0.05	0.55	0.93	−1.13	1.04
	2 to 5	0.28	0.59	0.63	−0.87	1.44
	1
Number of index victims	1 to 5	−0.40	0.46	0.38	−1.30	0.50
Any offending history	Yes	−0.59	0.35	0.09	−1.26	0.09
Any sexual history	Yes	0.29	0.53	0.58	−0.75	1.33
Female victims	Yes	−1.15	0.60	0.06	−2.33	0.05
Self-identified causes of sexual abuse:
Do not know why abuse occurred	Yes	−1.09	0.35	0.00	−1.78	−0.40
Fulfill emotional need (love or loneliness)	Yes	0.60	0.33	0.07	−0.052	1.26
Revenge	Yes	0.48	0.64	0.45	−0.77	1.74
Rejection	Yes	−0.47	0.76	0.54	−1.96	1.02
Intoxication	Yes	0.42	0.39	0.28	−0.34	1.19
Constant		0.58	0.94	0.54	−1.27	2.43

Kernel matching (with a bandwidth of .06) was used to match participant cases in the accepted group to cases in the declined group. To assess how well the matching procedure balanced the distribution of all known relevant covariates in the accepted and declined groups, the researchers considered descriptive statistics, t test comparisons, and Rosenbaum and Rubin’s (RR) standardized bias estimate (Rosenbaum & Rubin, 1985). There are, however, no clear criteria in the literature to determine when balance is achieved, as common t tests can be misleading because of the impact of sample size and variability on significance. Since standardized bias estimates take into consideration variability as well as difference, a standardized bias estimate less than 5% is regarded as sufficient to assume covariate balance (Caliendo & Kopeinig, 2005).

After matching, sexual recidivism rates in the two groups were compared. The marginal effect was calculated using probit analysis which increased the interpretability of the estimates given the binary outcome measures in this research (Hill, 2008). Cox proportional hazards regression (with failure rate) were performed to evaluate differences between the matched groups in the time it took for participants to reoffend sexually within the follow-up period. Cox regression was used rather than Kaplan–Meier survival analyses because of the need to include analysis weights from the propensity score matching procedure in the model. Consistent with the Campbell Protocol recommendation (Villettaz et al., 2006), we also provided estimates of the relative improvements in recidivism rates of the accepted versus declined groups matched samples (relative risk reduction or RRR). The RRR is the difference in event rates (in this case, sexual recidivism rates) between two groups, expressed as a proportion of the event rate in the untreated (declined) group (Cook & Sackett, 1995).

Propensity score analysis

Overall, the logistic regression model (Table 4) was predictive of treatment, and pseudo-R² revealed that the equation explained 33.2% of variation in the choice (LRχ² = 33.20, p < .01). The logistic regression model fitted the data well (Hosmer–Lemeshow test: p = .77), with reasonable discrimination (c statistic = .72). An excellent discrimination of the propensity score model is not ideal as such a result could lead to little or no overlap of the estimated propensity score between the groups, making it impossible to conduct matching (Weitzen et al., 2004). The Kernel matching procedure left four participant cases in the accepted group that could not be matched to another participant in the declined group because of dissimilar propensity scores. These cases were discarded from the analysis, leaving a sample of 208 cases: A total of 88 in the accepted group were matched to 120 in the declined group.

Figure 1 displays the percentage point differences for each variable between the accepted and declined groups before and after matching, illustrating how prematching differences were reduced in absolute terms. Provided in the appendix are RR standardized bias estimates and tests of differences before and after matching for each variable. After matching, no differences between the accepted and declined groups were statistically significant. In terms of RR estimates, all variables except four fell below the 5% threshold suggested as sufficient by Caliendo and Kopeinig (2005). Of the four variables above 5%, all differences were relatively small (Rosenbaum & Rubin, 1985), with none above 6.8%. The RR estimates for these variables were as follows: binary Static-99 = 6.8%, poor employment history = 6.0%, revenge motivated offense = 5.6%, and intoxication at the time of offending = 6.5%. Examination of the direction of these potential biases after matching, for all four variables, revealed higher bias estimates after matching of the presence of these risk factors among more participants in the accepted than the declined group, as shown in Figure 1. Thus, this analysis suggested that in terms of known variables, if there was any remaining selection bias, the likely result of this bias would predict more reoffending in the accepted than the declined group.

OPEN IN VIEWER

Figure 1:

Covariate Balance Achieved Before and After Kernel Matching

Occurrence of Sexual Reoffending

Table 5 displays a comparison of mean and standard deviation scores for sexual reoffending by participants before and after matching. After matching, results yielded a percentage point difference of 6.0 during the follow-up period, as participants who were accepted into the Diversion Program sexually reoffended at a lower rate (6.8%, n = 7.1) than participants who did not receive treatment and who experienced the regular court procedures and sanctions (12.8%, n = 12.9). The estimated effect size of treatment, expressed as an odds ratio, was 0.52 (95% CI = 0.18, 1.5). In other words, treatment reduced the risk of reoffending by 50%. When the incidence of outcomes is rare in the study population, generally interpreted to mean less than 10%, as is the case with sexual reoffending rates, the odds ratio is considered a good estimate of the risk ratio. Probit analysis showed that the magnitude of the reduction in sexual reoffending did not achieve statistical significance (SE = 0.05, p = .22, CI = −0.15, 0.03).

OPEN IN VIEWER

Table 5:

Observed Rates of Sexual Reoffending 1989–2007 by Group Before and After Matching

	Accepted Group		Declined Group
	%	n	%	n	Percentage Point Difference	95% CI	SE	p	Odds ratio	95% CI
Before matching	7.6	6.0	13.3	16.0	−5.7	−13.9, 2.4	0.04	.18	0.48	0.18, 1.3
After matching	6.8	7.1	12.8	12.9	−6.0	−14.6, 3.5	0.05	.22	0.52	0.18, 1.5

Period of Desistance from Sexual Reoffending

Cox proportional hazards regression analysis with failure rate (see Figure 2) showed that after matching, at a median time of 3,468 days (9.5 years), the failure rate for reported sexual offenses in the accepted group was approximately 7% compared to an approximate failure rate of 12% in the declined group. The difference in failure rates between the two groups was not statistically significant: HR 95% CI = 0.60 (0.21, 1.69), p = .334. There was no evidence that the proportional hazards assumption was violated: χ²(1, N = 208) = 1.58, p = .21.

OPEN IN VIEWER

Figure 2:

Cox Proportional Hazards Analysis of Sexual Recidivism in Accepted and Declined Groups Before and After Matching (1989–2007)

Relative Risk Reduction

Computation of the RRR yielded an outcome of 47% (12.8 – 6.8/12.8 × 100) indicating that the Program is estimated to reduce sexual recidivism in adult male intrafamilial child sex offenders by 47%.

Reduction in Sexual Reoffending

A comparison of the proportion of participants who sexually reoffended before and after matching (a 6% point difference) indicated that participants who were accepted into the Program sexually reoffended less than participants who did not receive treatment and who experienced the regular court procedures during the follow-up period. This result is consistent with previous research outcomes documented in meta-analyses that evaluated treatment effects in more heterogeneous groups of sex offenders (Hanson et al., 2002). Although the Probit analysis showed the magnitude of the reduction in sexual reoffending did not achieve statistical significance, the overall effect size for treatment was large. The lack of statistical significance when evaluating a single sex offender treatment program is unsurprising (Craig et al., 2003; Kingston et al., 2008) given the small number of reoffenders and the low rate of sexual reoffending in this study (7% in the accepted group, 13% in the declined group), producing low statistical power to detect the effects of treatment. The RRR of 47% implied that in a given population, if 100 sex offenders are expected to reoffend, after diversion to Cedar Cottage, the number likely to sexually recidivate will be reduced to 47, or just fewer than one half.

In conformity with the principle of intent to treat (Hollis & Campbell, 1999), participants in the accepted group included offenders who failed to complete treatment at the Program. Of the 92 participants composing the accepted group (before propensity score analysis was applied), approximately three fifths (58%, n = 53) completed treatment and two fifths (42%, n = 39) withdrew or were breached from the Program before completion. Since participants who drop out of treatment programs often recidivate at higher rates than do treatment completers (Hanson et al., 2002), the reduction in sexual recidivism rates observed in the current study is likely to underestimate the effectiveness of the treatment at the Program.

A more fine-grained examination of the seven participants in the accepted group who sexually reoffended within the follow-up period confirmed that the Program produced positive treatment effects. Of the seven recidivists, four had completed treatment, whereas three breached or withdrew before completion. Three of the four completers who sexually recidivated were referred to the Program in the first few years of its operation before legislative and regulatory amendments were implemented in April 1993. One reoffense recorded in that period was a charge of willful and obscene exposure 8 years after treatment completion against an offender with paranoid schizophrenia who was observed walking naked in his yard a few hours after he learned of the death of his victim. No conviction ensued. After 1993, more extensive assessment procedures were adopted by the Program. The report of an offense in the case of the participant who completed treatment after the changes in 1993 was by an adult coworker for peeping and prying when she recognized his face outside her window. This conduct involved no sexual contact and did not involve a minor, and no charge, arrest, or conviction ensued. In all, four of the seven sexual reoffenses committed by offenders referred to the Program were noncontact offenses; three were penetrative sexual reoffenses, two by offenders who completed treatment in the first few years of the Program operation, one by a noncompleter. The overall recidivism data showed that in the 14 years between 1993 and the end of the follow-up period in 2007, no official police reports or convictions of any sexual offense against a minor were made regarding any of the 34 participants who completed treatment in the Program.

As a whole, this study demonstrated the usefulness of propensity score analysis to overcome selection bias and produce empirically sound findings on the effectiveness of sex offender treatment programs where randomized trials are infeasible. These results demonstrated that the Cedar Cottage Diversion Program reduced sexual reoffending in adult male intrafamilial child sex offenders and that community-based programs to rehabilitate low-risk intrafamilial child sex offenders can be effective.

Limitations of the Study

As with all studies, certain limitations in this research must be acknowledged. Foremost among these is a limitation in the available information about offenders because of the retrospective nature of this study. Reoffending rates in this research were derived from records maintained by authorities within the state of New South Wales. Therefore, any reoffending on the part of participants that may have occurred outside of New South Wales was not recorded or included. This omission may have led to an underestimation of recidivism. Propensity score matching cannot control for unobserved selection bias (Guo & Fraser, 2010).

Other caveats include the fact that this research was conducted using a participant group exposed to only one particular diversion program. Although it is a strength of this study that all offenders referred to the Program in the period 1989–2003 were included and monitored for up to 18 years, whether these findings may generalize to other diversionary programs or other treatment program facilities is unknown. The delivery and efficacy of the Program and the precise causes of the reduction in sexual reoffending observed at the Program remain unknown. As time passed, the proportion of offenders referred to the Program who were accepted for treatment increased. For example, in the period 2004–2008, of the 29 referrals, 8 were declined and 20 or 69% were accepted (D. Tolliday, personal communication, July 7, 2008). With the passage of time, the efficacy of the Program appeared to improve. For example, recidivism data for all 138 offenders referred to Cedar Cottage in the period 1993–2004, of whom 64 were accepted for treatment, revealed no official police reports or convictions of any sexual offense against a minor by any of the 34 offenders who completed the Program. The turnover of staff and legislative changes indicated that the quality of the implementation of the Program varied in the years since its inception in 1989. These factors need to be investigated in future research.

Future Research

As discussed, the areas of limitation have given rise to the recommendation of areas of future research. Future research should also be implemented to address factors that fall outside the scope of this study that go toward highlighting the success of the Diversion Program. For example, at the Program, in addition to the rehabilitation of sex offenders, the family members of each offender, including the child victim(s), receive counseling to come to terms with their situation (Goodman-Delahunty, 2009). The design and implementation of the Diversion Program also spares child victims the burden of testifying in legal proceedings and increases the disclosure of further offenses that would otherwise have remained undetected (Pratley & Goodman-Delahunty, 2011). A preliminary estimate of the cost per offender of attending the Program was no different from the cost of traditional incarceration (Goodman-Delahunty, 2009). These potentially positive effects should be analyzed in future research to add to the evaluation of the effectiveness of the Program.

Offender motivation can be an important predictor of both treatment volunteerism and treatment entry among incarcerated adult sex offenders who victimize children (Jones, Pelissier, & Klein-Saffran, 2006). Reductions in sexual recidivism in participants who enter any sex offender treatment program may reflect levels of motivation rather than the effectiveness of the treatment program (Campbell, 2004). Further research is under way to assess the influence of dynamic risk factors in this sample using the Violence Risk Scale-Sexual Offender Version (Wong & Olver, 2010).