Teorías del Poder.

2.1.4.1 Psychometric properties.

The interrater reliability of the traditional needs/risk assessment instruments and gender informed variables were examined on a random selection of 15 cases collected and coded independently by a second rater. Interrater reliability was evaluated by calculating the intraclass correlation coefficient (ICC). Two-way, mixed, single rating ICCs were explored, with absolute values reported for individual items, while consistency values were reported for subtotal and total scores, consistent with best practices (McGraw & Wong, 1996; Nichols, 1998; Shrout & Fleiss, 1979). The ICC not only examines the correlation among the ratings, but also considers potentially different anchor points among raters as it is a measure of agreement, not association

(Bartko, 1991; McGraw & Wong, 1996). In general, reliability coefficients around .70 or higher are considered to be acceptable, while values of .80 are typically regarded as moderate to high levels of reliability (Murphy & Davidshofer, 2005).

Internal consistency was examined using Cronbach’s alpha. Note, as Green and Lissitz, and Muliak (1977) clarify internal consistency implies interrelatedness of items, but not necessarily a unidimensional construct. Cortina (1993) suggests caution when using alpha alone to determine if a scale reflect a unidimensional construct, especially when a scale has a larger number of items, as higher alpha values would be expected even with lower item intercorrelations and factor analytic techniques should be used to clarify the dimensions of a scale. This may be especially relevant to the current examination, as by design these scales (and the gender informed variables) measure multiple items in a number of domains or subconstructs thought to be related to the construct of risk for reoffence which is acknowledged to be heterogeneous.

2.1.4.2 Predictive validity.

First, correlations were conducted on each of the total scores of the actuarial risk instruments with general recidivism, violent recidivism, revocations, and incidents of institutional misconduct, to examine the strength and direction of these relationships. Correlations of the actuarial risk instruments and gender-informed variables and severity of recidivism were also examined.

Cohen (1992) also suggests that correlations in the .10 range represent small effect sizes, in the .30 range medium effects sizes and in the .50 range large effect sizes. However, Meyer and colleagues (2001) remind researchers should be “satisfied” with correlations at the .10 to .19 range, “pleased” with correlations in the .20 to .39 range, and “rejoice” at values that are higher given the difficulty of consistently achieving uncorrected univariate correlations greater than .30 in real world settings.

Bonferroni adjustments or other methods of adjusting statistical significance levels (i.e., p-values) to address the family wise error rate were not employed in the current study. Rothman (1990) cautioned against making adjustments for multiple comparisons as he argued that data being examined are not random numbers, but represent actual observations governed by natural laws and suggests that researchers should be less reluctant to explore leads that may be incorrect to avoid missing potentially important findings. He highlighted that there is no formula

that can replace the critical evaluation of results observed. Further, Perneger (1998, p. 1236) suggested that “Bonferroni adjustments are, at best, unnecessary and, at worst, deleterious to sound statistical inference.” This is because such adjustments create new methodological problems (e.g., increases the risk of a type II error; which may be especially problematic in new areas of study), and the Bonferroni method implies that all null hypotheses are true simultaneously (Pernerger, 1998) which is not the focus of the current investigation. That is, the goal of the current study is not to say whether all variables are related to all outcomes or not, which a Bonferroni correction would allow, but instead to examine each variable in its own right. Instead, Pernerger (1998) suggested describing statistical methods and rationales as well as possible interpretations to help the reader make reasonable conclusions without Bonferroni adjustments. There are a growing number of researchers who have supported this position (Cabin & Mitchell, 2000; Moran, 2003; O’Keefe, 2003). In the current study, the examination of multiple definitions of related variables also provides some form of within study replication for the findings.

Further, confidence intervals were also provided where possible as they are a good alternative to statistical significance corrections (Garamszegi, 2006). Thus, these correlational relationships were also illustrated with ROC analyses. These analyses produce Areas under the Curve (AUCs) which, for the purpose of this study, represent the probability that a randomly selected recidivist will have a higher risk score than a randomly selected nonrecidivist (Hanley & McNeil, 1982). Moreover, AUCs are not affected by base rates or selection ratios, which is especially important when predicting violent recidivism (Mossman, 1994; Quinsey, Harris, Rice, & Cormier, 2006) given that only 11% of this sample recidivated violently. AUC values at .50 represent chance prediction, while those at 1 represent perfect prediction (Mossman, 1994; Quinsey, Harris, Rice, & Cormier, 2006). Rice and Harris (1995) proposed that AUCs of 0.64 to 0.70 represent a moderate effect size and AUCs of 0.71 and higher represent a large effect size. 2.1.4.3 Predictive validity of risk levels.

Survival analyses allow the examination of time to an event, as well as the rate of this event relative to a comparison group (Luke & Homan, 1998). In this study, survival analyses allowed the cumulative percentage of offenders who did not recidivate (that is, those who have “survived”) to be tracked during the available follow up period. Further, by grouping offenders based on their risk ratings their relative performance could be compared. While there are a

number of nonparametric statistics that can be used to ensure that the groups (in this case actuarial risk levels) actually differ in their survival probabilities, with a small sample size or longer follow up time Luke and Homan (1998) suggest interpreting the generalized Wilcoxon or Tarone-Ware tests. For, these tests earlier observations are weighted more because there are more cases remaining in the risk pool and the tails of these curves can less reliable.

To perform survival analyses, the time women were released in the community was compared. For recidivists this was defined as time to first reoffence (for general or violent reoffence, respectively) from the latter of first release date or the end of the assessment period. This was compared to follow up time for nonrecidivists, which also began with the later of release or post-assessment period and ended with the date that all recidivism data were collected. Offenders were compared grouping them based on their risk scores as outlined in each instrument’s manual or by truncating the data into near equal thirds, as was done with the VRS, to create low, medium and high risk groups. Specifically, for the SIR-R, scores above and including 6 were classified as low risk, 5 to 1 as low to moderate risk, 0 to -4 as moderate risk, -5 to -9 as moderate to high risk, and -10 and lower as high risk. For the LSI-R, scores from 0 to 13 were low risk, 14 to 23 were low/moderate risk, 24 to 33 were moderate risk, 34 to 40 were medium/high risk and scores of 41 and above were high risk. For the LS/CMI, there were no scores in the very low risk level, scores from 6 to 10 were low risk, 11 to 19 were medium risk, 20 to 29 were high risk, and over 30 were very high risk. Finally, for the VRS low risk was scores up to and including 12, medium risk was 13 to 36, and high risk was 37 and up.

2.1.4.4. Incremental predictive validity.

To explore the relationship between recidivism (general and violent recidivism, as well as revocation and institutional misconduct) and the gender informed variables (childcare responsibility, presence of spouse or common-law, presence substance abuse, substance abuse intensity, illegal financial support, history of social assistance, self harm/suicide, child physical abuse, child sexual abuse, child emotional abuse, adult physical abuse, adult sexual abuse, and adult emotional abuse) were correlated. As was done for each of the risk assessment instrument, ROC analyses were also conducted for each gender informed variable with each outcome variable.

Next, a subset of the gender-informed variables was combined to look at the gender- informed composites that may help clarify some of the inconsistencies found in past research.

Namely, variables looking at presence of spouse or common law and number of dependents were combined to explore whether being a single parent is predictive of recidivism (general, violent, institutional misconduct) for female offenders in addition to whether marital status or childcare responsibilities alone were predictive. This was explored using a series of correlations.

Secondly, after examining the past victimization variables and substance abuse alone, each relevant type of victimization was combined with substance abuse to see whether the combination of victimization and substance abuse was predictive of recidivism as suggested by McClellan, and colleagues (1997). This was again explored using a series of correlations.

Next a series of backward elimination stepwise logistic regression analyses was used to determine whether the gender-informed variables and their composites predicted outcome over and above the risk assessment instruments alone. Specifically, a risk assessment instrument was entered into its own block, followed by the gender-informed variables or composites in a second block in stepwise form to determine if they increased the effect size of the model over and above that of the instrument alone. Moreover, for this study, the group of gender informed variables and the group of composites were explored initially in separate logistic regression analyses to avoid increasing the error variance by including a number of highly related variables in the same analysis (recall, the composites were derived from the gender informed variables). Once individual gender informed variables and composites were identified in separate stepwise logistic regression analyses as contributing significantly to the prediction of outcome over and above the risk assessment instrument being examined, only these gender informed variables and composites were combined in one analysis to derive a final solution representing the best combined prediction. Note, when the presence of substance abuse, abuse variables and their related composites were shown to significantly add to the predictive utility of their individual models, the presence of substance abuse with not included in the final combined model to avoid violating the independence of errors assumption for linear regression. This series of analyses was conducted separately for each risk assessment instrument

Although stepwise methods are criticized for being atheoretical, overly affected by random sampling variation and at risk of over or under-fitting the model, they are defensible for exploratory model building when previous research upon which to base a hypothesis is lacking (Field, 2009). Further, backward elimination, as compared to forward elimination, reduces the risk of eliminating a predictor involved in a suppressor effect, and thus reduces the risk of a type

two error (Field, 2009). The likelihood ratio statistic was used to examine the steps in the model to evaluate improvements in fit, as the Wald statistic can be unreliable and produced increase risk of a type 2 error also. The Wald statistic was used to examine the influence of an added variable or composite in predicting outcome within a given model. The Exp (β), or exponentiated regression coefficient, is an odds ratio statistic reflecting the degree of change in risk of outcome based on a change in that predictor variable. When Exp (β) is greater than 1 this represents a positive relationship with outcome, while a value less than one represents a negative relationship. For example, an Exp (β) value of 2 suggests that for each one point increase in the value of the predictor variable (or the presence of that variable if it is dichotomous), the risk of the targeted outcome doubles. Similarly, with an Exp (β) value of .5 the risk of outcome is cut in half with a one point increase in the predictor variable. It is especially important that the confidence interval for the Exp (β) does not cross 1 or the interpretation of the variable (i.e., the direction of the relationship) becomes unreliable (Field, 2009). See Hosmer and Lemeshow (1989) for a detailed review of logistic regression.

3.1 Results