[Page 23 continued]

The results of the meta-analysis are presented in three main sections:

[Page 24]

[Page 25]

Descriptive Statistics for Study Characteristics

The 37 studies coded (see References and Table 1) included 88 samples across the six dependent measures and 25,367 persons (mean per study was 634, with a range of 16 to 820). Of the total sample size across studies, 9,230 (36%) reported experiences of CSA. We did not conduct tests for homogeneity because they are required only when samples are theoretically very disparate in nature

[Page 26]

(Rosenthal, 1995). The studies ranged in publication date from 1981 to 1995, with the majority of studies being published in 1994 (65%; 24 studies). The majority of studies were retrospective-longitudinal in nature, used samples of convenience, assessed participants through standardized questionnaires, and computed uni-variate and multi-variate statistics in relation to the specific examined outcomes.

Effect Size Analysis

A variety of approaches exist for the computation of effect sizes within the d family (e.g., Cohen's d, Hedges's g). In the current research, Glass's delta was selected. Glass's effect size is the difference between population means divided by the standard deviation of the population control group (Rosenthal, 1994). According to Rosenthal (1994), this estimate is preferred when the standard deviations based on the two different conditions differ greatly from each other.

Table 1 summarizes the unweighted effect sizes computed for each study. The effect sizes for each of the 88 samples ranged from a minimum value of -.69 to a maximum value of 3.31. A positive effect size in this analysis indicates that experiences of CSA had negative outcomes, whereas a negative effect size indicates that CSA had positive consequences in relation to the examined outcomes.

Studies differ from one another in many methodological and substantive ways (Shadish & Haddock, 1994). To take these differences into account, it was necessary to use a procedure that would justify the combination of the 37 studies used in this meta-analysis. Appropriate weights by study sample size were calculated to minimize the variance. Such a weighting assumes that studies with larger samples have a smaller variance and in turn are better estimates of effect size. Shadish and Haddock termed this a quality rating, which is the only standardized weighting scale for studies to date. It was this quality-weighted version that was used to compute the weighted average effect size.

As shown in Table 2, the average unweighted effect size and average weighted effect size were calculated for all the dependent variables across studies. The most conservative estimate of the influence of CSA on the outcomes is the average weighted d value, which also revealed the smallest variability (standard deviation ranged from .03 to .07 across the dependent measures) and the narrowest confidence limits (e.g., .41 to .47 for depression). Therefore, we can be 95% confident that the prevalence rate of PTSD as an outcome due to CSA in the general population falls within the boundaries of .37 to .43. In contrast, the average unweighted d, which most closely reflected the raw data, had the greatest variability (standard deviation ranged from .30 to. 79) and largest confidence intervals (e.g., -.20 to 1.38 for sexual promiscuity).

Table 2 also presents the results of the raw counts and percentages in the form of an average weighted binomial effect size display (BESD). This display demonstrates the practical importance of any effect indexed by a correlation

[Page 27]

[Page 28]

coefficient (Rosenthal, 1994,1995). The correlation (r) refers to the difference in outcome rates between the experimental and control groups, whereby the column and row totals always sum to 100. The results indicate that there is a substantial increase in negative outcome for each identified dependent variable.

The specific point increases can be found in Table 3 [ > ??? Table 2 ???]. According to the DSM-IV (American Psychiatric Association, 1994), community studies have revealed a lifetime prevalence for PTSD ranging from 1% to 14%. To be conservative, using the upper limit of 14% and an average weighted effect size r of .20, the current meta-analysis indicates that there is a 143% (20/14 x 100) increase in risk of developing PTSD symptoms following CSA among the general population. Similarly, there is 

Thus, the magnitude of the effect of CSA on all the outcomes examined has substantial practical significance that requires immediate public attention and action. 

We computed an estimate of the Fail Safe N^*1 to assess the file drawer problem, which assumes that journals contain 5% of the studies that show Type I errors, whereas the file drawers in research laboratories are filled with 95% of studies showing non-significant results (Rosenthal, 1991 ). We found that 73 unreported PTSD studies averaging a null result would have to exist somewhere before the overall results of the current meta-analysis could be reasonably ascribed to sampling bias. Similarly, 95 unreported studies on depression, 39 studies on suicide effects, 49 on sexual promiscuity, 15 on the victim-perpetrator cycle, and 5 on poor academic performance would be required to refute the results of the present meta-analysis. Accordingly, we can accept with confidence that the results of the current meta-analysis were not a function of failure to publish non-significant results.  

*1 The following formula reported by Wolf (1986) was used to estimate the number of unpublished studies that would be necessary to include so as to reverse the present  findings and produce an effect size equal to zero (i.e., Fail Safe N): ES = effect sizes in  the present study per domain; N = number of studies in the meta-analysis; N_fs.05 = number of studies required to nullify present mean ES at the .05 level of significance: N_fs.05 = 
(sum of ES}^{2
 ___________}     - N.  
     (1.645)

Analysis of Mediating Variables 

A number of potential mediating variables were studied using uni-variate analyses. Specifically, these variables included gender, socio-economic status, type of abuse reported, age of abuse, relation of the victim to the perpetrator, and number of abusive incidents. 

Because regression analyses were not possible,

[Page 29]

Table 3:
Potential Effect of Mediating Variables on Average Unweighted d and Weighted d Outcomes

[Page 30]

a series of analyses of variance (ANOVAs) were conducted, and the results are presented in Table 3. We could not obtain results for some of the outcomes measured due to their homogeneous nature and lack of variability. A total of 36 uni-variate tests were executed one at a time, and none of the mediating variables was found to be statistically significant. These results are summarized for both the unweighted ds and weighted ds.