Sample-level analysis.

To examine the intensity of CSA psychological effects or correlates, we first meta-analyzed the sample-level effect sizes from the 54 samples for which these could be computed (sample-level effect sizes are listed in the Appendix ). ^[*3]  

[*3] Appendixes containing other effect sizes for other analyses in the Results section (i.e., symptom-level, moderator analyses, male-female differences, family environment-CSA relations, and family environment-symptom relations) can be obtained by writing to Bruce Rind.

The resulting unbiased effect size estimate, based on 15,912 participants, was r _u= .09, with a 95% confidence interval from .08 to .11. Because this interval did not include zero, the obtained result was statistically significant (i.e., SA students were less well adjusted than controls). This difference in adjustment between SA and control students was small, however, according to Cohen's (1988) guidelines; in terms of variance accounted for, CSA accounted for less than 1% of the adjustment variance.

A chi-square test of the homogeneity of the sample-level effect sizes revealed that they were not homogeneous, chi ² (53) = 78, p < .01. In an attempt to achieve homogeneity, we examined the distribution of sample-level effect sizes to determine whether outliers existed. We defined outliers to be effect sizes that were at least 1.96 standard deviations away from the unweighted mean effect size (i.e., falling in the extreme 5% of the distribution). Three outliers were found ( r = .36 in Jackson et al., 1990 ; r = .40 in Roland et al., 1989 ; r = -.25 in Silliman, 1993 ) with z scores of 2.71, 3.16, and -3.60, respectively. The Jackson et al. study included only incest cases in the CSA group, and the Roland et al. study included a large proportion of incest cases. Moreover, Neumann et al. (1996) also found the Roland et al. result to be an outlier. Measures used in these studies from which effect sizes were computed included: the SAS, BDI, RSE, and DSFI ( Jackson et al., 1990 ); the MMPI form R ( Roland et al., 1989 ); and the LOC and TSCS ( Silliman, 1993 ). These measures were all used in other studies whose effect sizes were not outliers, implying that the outlying results were not a function of these measures. Removing these outliers resulted in homogeneity, chi ² (50) = 49.19, p > .50, based on k = 51 samples, with N = 15,635 subjects. The recalculated unbiased effect size estimate (r _u= .09) and the 95% confidence interval (.08 to .11) were unchanged after rounding. The obtained small unbiased effect size estimate implies that, in the college population, the magnitude of the relationship between CSA and adjustment is small, which contradicts the assumption that CSA is associated with intense harm in the typical case.

Symptom-level analysis.

Next we examined the magnitude of the relationship between CSA and adjustment at the symptom level. Table 3 presents the results of the 18 symptom-level meta-analyses. The table shows for each meta-analysis the number of independent samples ( k ), the total number of participants in these samples ( N ), the unbiased effect size estimate (r _u), the 95% confidence interval of r _u, and the homogeneity statistic ( H ) based on the  chi-square test.

Initial meta-analyses yielded 8 homogeneous and 10 heterogeneous results. In an attempt to achieve homogeneity with heterogeneous sets, we examined the distribution of effect sizes within each of these sets to detect outliers, as defined previously. We removed all such deviant effect sizes and then recomputed the meta-analyses. If homogeneity was achieved in a particular set, then the search for outliers stopped for that set. Otherwise, the

[Page 32]

reduced set of effect sizes was examined for new outliers, and, if found, the outliers were removed and the meta-analysis was performed again. If the set of effect sizes was still heterogeneous and no additional outliers were found, the set was considered to be heterogeneous. This procedure resulted in achieving homogeneity in 7 of the 10 initially heterogeneous sets, yielding 15 out of 18 homogeneous sets. Effect sizes remained heterogeneous only for hostility, self-esteem, and sexual adjustment. Of the 9 effect sizes removed in the 7 sets that became homogeneous, the majority came from two of the studies that contributed to the heterogeneity of effect sizes in the sample-level meta-analysis-5 from Roland et al. (1989) and 1 from Jackson et al. (1990) . These six effect sizes and one additional effect size from Bendixen et al.'s (1994) female sample were removed from the upper end of their distributions. Two effect sizes were removed from the lower end of their distribution ( Fishman, 1991 ; Fromuth & Burkhart, 1989 , Southwest sample). Measures on which removed effect sizes were based in Jackson et al.'s and Roland et al.'s studies were listed previously in the sample-level meta-analysis section; Bendixen et al. and Fishman used investigator-authored items, whereas Fromuth and Burkhart used the SCL-90-R. Many studies with nonoutlying effect sizes used investigator-authored items and the SCL-90-R, implying that the outlying results were not a function of the measures used.

In Table 3 , the original numbers (i.e., number of samples, number of participants in these samples, unbiased effect size estimate, and homogeneity statistic) associated with the heterogeneous results for the seven sets that became homogeneous are shown in parentheses, whereas the numbers associated with the reduced homogeneous sets appear directly under the column headings. Removing outliers showed itself to be productive in achieving homogeneity; further, this procedure had little effect on effect size estimates, indicating that the large majority of effect size estimates can be considered to be reliable estimates of true effect sizes in the college population. The unbiased effect size estimates for all 18 symptoms were small according to Cohen's (1988) guidelines. The effect size estimates ranged from r _u= .04 to .13. Despite these small values, all effect size estimates, except for one (locus of control), were statistically significantly greater than zero, as is indicated by their 95% confidence intervals. These findings indicate that, for all symptoms but one, CSA participants as a group were slightly less well adjusted than control participants. The small magnitude of all effect size estimates implies that CSA effects or correlates in the college population are not intense for any of the 18 meta-analyzed symptoms.

Table 3
Meta-Analysis of 18 Symptoms Associated With Child Sexual Abuse From College Samples

Note
k represents the number of effect sizes (samples);
N is the total number of participants in the k samples;
r_u is the unbiased effect size estimate (positive values indicate better adjustment for control subjects);
H is the within-group homogeneity statistic (chi square based on df = k - 1).
Cutting or trimming outliers was performed when effect sizes were heterogeneous in an attempt to reach homogeneity. Original numbers, before cuttinng ore trimming, are shown in parentheses. 95% confidence intervals are based on final (cut or trimmed) distributions.
² Cutting or trimming outliers failed to produce homogeneity; thus, only original numbers are shown.
* p < .05 in chi-square test.