2 50 CHAPTER 2 d on sexual recidivism was 0.71, 95% confidence interval (CI) [0.63, 0.79], n = 5,699, based on 22 unique samples. The Q indicated that the variability across studies was not significant (Q = 25.5, p = .23). The I2 (17.7%) indicated that the amount of variability in effect sizes among studies that could not be attributed to chance was small. 2.4.1.1.1 Moderators The Qbetween analyses for the fixed-effect model revealed significant effects for two categorical moderators. Instruments developed within the study sample produced larger effect sizes than replications on new samples (Cohen’s d respectively 0.94 and 0.67; Qbetween = 4.2, df = 1, p = .04). And instruments scored on combined locations produced larger effect sizes than instruments used either in an institution sample or in a community sample (Cohen’s d response 0.97, 0.61, 0.65; Qbetween = 6.4, df = 1, p = .04). The fixed-effect meta-regressions for the continuous moderators showed that older studies and studies including shorter follow-up periods produced larger effect sizes than more recently published studies and those with longer follow-up periods (respectively, b1 = -.02757, SE = .00964, Z = -2.859, p = .004, k = 23; b1 = -.00161, SE = .00072, Z = -2.224, p = .026, k = 21). No significant effect was found for sample size (b1 = .00002, SE = .00024, Z = .068, p = .946, k = 21). 2.4.1.2 Predictive properties for violent (including sexual) recidivism In the overall analysis for violent (including sexual) recidivism (14 unique samples), the fixed-effect weights of the individual studies (the inverse of the variance) varied between 2.21 and 500.00, with a median value of 38.52. The studies producing the largest weights, 250.00 and 500.00, had sample sizes of respectively 3,402 and 3,360 (recidivism rates were 7.8% and 6.9% respectively). These weights were more than twice as large as the next largest study weight (90.91) and more than 113 and 226 times larger than the smallest study weight, respectively. To reduce the influence of these studies, the variances were artificially increased so that their weights were at most twice as large as the next largest study weight. After adjusting the weights, we found significant variability among study effect sizes (Q = 42.7, p < .01; I2 = 70.0). Cohen’s d for the random-effects model was 0.43, 95% CI [0.29, 0.57], N = 10,368, based on 14 unique samples. 2.4.1.2.1 Moderators The Qbetween analyses for the fixed-effect model revealed no significant categorical moderators. The fixed-effect meta-regression for the continuous moderators showed that studies with larger samples produced smaller effect sizes (b1 = -.00011, SE = .00002, Z = -5.937, p < .001, k < 14) and that studies including longer follow-up periods produced larger effect sizes (b1 = .00243, SE = .00065, Z = 3.742, p < .001, k = 12).
RkJQdWJsaXNoZXIy MjY0ODMw