Go to >

Semi-partial correlational Analyses

32

Go to >

Contrast analyses

33

Go to >

Simple correlations

34

Go to >

Moderators concerning aspects of the CSA experience

34

Semipartial correlational analysis.

To examine whether the variability in sample-level effect sizes could be accounted for by moderator variables, we performed multiple regression analyses. We focused on the sample-level rather than symptom-level effect sizes because of the substantially larger sample-level data set, which is more appropriate for multiple regression analysis. As in other meta-analyses (e.g., Oliver & Hyde, 1993 ), we performed multiple regression specifically to obtain correlations between each moderator and the effect sizes while controlling for other

[Page 33]

moderators, because of the possibility that the moderators were confounded. We focused on semipartial correlations. This moderator analysis was based on a weighted multiple regression procedure, using a weight of N - 3 for each sample, which represents the reciprocal of the variance for an effect size r , thereby producing the best linear unbiased estimate (cf. Hedges, 1994 ); this approach is consistent with the use of unbiased effect size estimates. The sample-level effect sizes were regressed on the three variables that were coded for each sample: level of contact (0 = both noncontact and contact sex, 1 = contact sex only ), level of consent (0 = willing and unwanted sex, 1 = unwanted sex only ), and gender (0 = male, 1 = female ). Examining the relationship of gender with the effect sizes was done to address the issue of gender equivalence. As discussed previously, it is widely believed that contact sex is more severe or serious than noncontact sex; therefore, it was of interest to test whether this factor would account for variability in effect sizes. Finally, it was expected that unwanted sex would be associated with larger effect sizes; hence, level of consent was examined as a moderator. Results from this analysis regarding level of consent and level of contact are likely to be conservative (i.e., their relationship with the effect sizes may be underestimated) because the first level of each variable overlaps with the second level (e.g., willing and unwanted sex overlaps with unwanted sex only). Also entered into the regression equation were two two-way interactions: Contact × Gender and Consent × Gender. The Contact × Consent and Contact × Consent × Gender interactions were not included because no male samples consisted exclusively of cases of unwanted contact sex and only one female sample consisted exclusively of unwanted contact sex. Finally, because outliers can skew correlational results, we excluded from the multiple regression analysis the three outliers identified previously in the sample-level meta-analysis. Four studies containing both men and women were also excluded, because they did not report results separately for the two genders.

The regression model was marginally significant, F (5, 41) = 2.09, p = .09. Significance tests of predictors were based on adjusting their standard errors to obtain a correct model for multiple regression involving effect sizes (see Hedges, 1994 ). Three predictors were significantly related at the .05 level to the effect sizes: consent, gender, and the Consent × Gender interaction. The other two predictors, contact and Contact × Gender, were not related. The semipartial correlations between these latter two predictors and the effect sizes were, respectively, sr (41) = .15 and -.13 (two-tailed p s > .30). A second regression model was run, eliminating the two nonsignificant predictors in the previous model. This new model was statistically significant, F (3, 43) = 3.18, p = .03; all three predictors were significantly related to the effect sizes at the .05 level. The semipartial correlations between the effect sizes and the predictors of consent, gender, and Consent × Gender were, respectively, sr (43) = .33, .38, and -.36 (all two-tailed p s < .05). These results indicate that unwanted sex and being female were each associated with poorer adjustment. These results have to be qualified, however, because of the significant Consent × Gender interaction.

Contrast analyses.

To investigate the Consent × Gender interaction, effect sizes for each of the different levels of consent and gender were meta-analyzed separately, and then contrast analyses were performed comparing the unbiased effect size estimates between the different levels of each moderator. Next, effect sizes within each of the four Consent × Gender combinations were meta-analyzed separately, and then contrast analyses between unbiased effect size estimates in appropriate combinations were performed. This procedure follows the model of a main effects and then simple effects analysis in an analysis of variance (ANOVA). The contrast analyses were based on the formula presented by Rosenthal (1984) and used weighted Fisher Z transformations of the effect sizes. Within each of the two sets of Fisher Z s being compared in a given contrast analysis, the weight of a Fisher Z was its degrees of freedom (i.e., N - 3) divided by the sum of degrees of freedom for all Fisher Z s in that set. Weights in the first set were positive, whereas those in the second were negative. This weighting method resulted in a statistic (i.e., normal deviate z ) that is equivalent to Hedges's (1994) between-groups heterogeneity statistic (i.e., Q _BET, distributed as chi ²) for testing differences between two sets of effect sizes, in that the square of z is equal to the chi ²  value.

Table 4 presents the results of the four meta-analyses across the different levels of gender and consent. Effect sizes were homogeneous in all four groups and unbiased effect size estimates were all significantly greater than zero, as is indicated by the 95% confidence intervals that did not contain zero. The contrast between the female (r _u= .10) and male (r _u= .07) unbiased effect size estimates, based on 14,578 participants, was nonsignificant, z = 1.42, p > .10, two-tailed. The contrast between the unwanted sex (r _u= .10) and all levels of consent (r _u= .10) unbiased effect size estimates was also nonsignificant, z = .03, p > .10. These nonsignificant main effects are attributable to the Consent × Gender interaction, which is described next.

Table 4
Meta-Analyses of Sample-Level Effect Sizes Assessing CSA-ADjustment Relations
in College Students for Each Level of Gender and Consent

These results help clarify the significant Consent × Gender interaction found in the multiple regression analysis. Adjustment was associated with level of consent for men, but not for women. Noteworthy is the finding that SA men in the all-levels-of-consent group were unique in terms of not differing from their controls in adjustment. Because all levels of consent corresponds to social and legal definitions of CSA, these results imply that, in the college population, the association between CSA and adjustment problems is not equivalent for men and women. If the definition of CSA is restricted to unwanted sex only, however, then these results imply a gender equivalence between men and women in the association between CSA and adjustment problems.

Simple correlations.

In a further attempt to explain variability in sample-level effect sizes, we examined the association between several additional factors and the sample-level effect sizes (the three outliers were not included in these analyses). Associations were computed using weighted correlational analyses (weights were N - 3 for each sample). We coded all studies for method of assessment (e.g., face-to-face interview vs. questionnaire), type of institution (e.g., public vs. private), sampling strategy (e.g., a convenience sample of introductory psychology students vs. a broader sample of students obtained by random or pseudorandom sampling), mean age of students at time of assessment, the maximum age for a "child" in the study's definition of CSA, and whether the study was published. No method variance in assessment emerged because all studies were based on questionnaires. Similarly, type of institution did not show itself to be useful for correlational analysis because nearly all studies were conducted at state universities. For sampling strategy, we categorized studies into two groups: ones that used convenience samples of students (usually psychology or sociology) and ones that used wider samples that included students in nonsocial science courses or that were based on random or pseudorandom sampling of all students at the school. Of the 38 studies for which sampling strategy could be coded, 25 were of the first type and 13 were of the second. Sampling strategy was not related to effect sizes, r (36) = .16, p > .30, two-tailed. Regarding age of students, if CSA has early effects that diminish over time, or if it has delayed effects that emerge only as students get older, then a significant correlation between mean age of students in the sample and effect sizes would be expected (the range of mean ages in the samples went from 18.0 to 26.6 with an overall mean age of 20.8). The correlation, however, was nonsignificant, r (36) = .01, p > .90, two-tailed. Similarly, maximum age of "child" in the study's definition of CSA was not related to the effect sizes, r (44) = -.05, p > .70, two-tailed. The relationship between whether a study was published and the sample-level effect sizes was marginally significant, r (49) = .25, p = .08, two-tailed. The 27 samples with published results had a slightly larger unbiased effect size estimate (r _u= .11) than that of the 24 samples whose results were unpublished (r _u= .08).

Moderators concerning aspects of the CSA experience.

Studies were inconsistent in providing statistics on aspects of the CSA experience (e.g., force, penetration) that might affect adjustment among SA participants. We examined all studies to search for such moderators and found five types that were reported in at least two studies: force, penetration, duration, frequency, and incest. Additionally, several studies examined moderators that were composite measures that combined two or more of the moderators just listed. Some researchers provided correlations between a moderator and self-reported reactions or effects; other researchers provided correlations between a moderator and symptoms among SA participants. We meta-analyzed separately the moderator-reaction-effect and moderator-symptom relations for the different types of moderators when results for both types of relations were available (we considered individually the results from the studies examining composite moderators). In the case of moderator-symptom relations, if a study provided correlations between a given moderator and more than one symptom, then all of these correlations were averaged using Fisher Z transformations to create a single moderator-symptom relation for that study. Some studies reported only beta weights; these values were used as effect size estimates. A number of studies reported only that the relation was nonsignificant or that it was significant; in these cases, following recommended procedures by meta-analysts (e.g., Rosenthal, 1984 ), we set the effect size to zero in the former case and to the appropriate value corresponding to p = .05, two-tailed, in the

[Page 35]

second case. Because most effect size assignments were of the former type, some of the unbiased effect size estimates are likely to be underestimates of the moderator-symptom relations.

Table 6 provides summaries of the meta-analyses of the moderator-outcome relations. As shown in the table, only 3 of the 10 moderator-outcome relations reached statistical significance. The presence of force was associated with more negative reactions and self-reported effects; the magnitude of this relation was medium, r _u= .35. Incest (i.e., close familial CSA) was associated with both symptoms, r _u= .09, and negative reactions-self-reported effects, r _u= .13; the magnitudes of these relations were small. Notably, force was unrelated to symptoms, and penetration was unrelated to either outcome. Frequency (i.e., number of CSA episodes) and duration (i.e., length of CSA involvement) were also not related to outcome.

The table also displays recalculated unbiased effect size estimates (shown in parentheses next to original estimates) in cases where one or more effect sizes were estimated. These new effect size estimates were computed using only the known effect size values. The statistical significance of these recalculated values changed in only one case. Symptoms associated with penetration became statistically significant (95% confidence interval = .02 to .30). This result, however, should be viewed with caution, because it is based on the removal of more than half the effect sizes for this outcome, all of which were nonsignificant.

Table 6
Meta-Abalyses of Relations Between Aspects of the Child Sexual Abuse  Experience and
Outcome in  Sexually Abused College Students

Note.
k represents the number of effect sizes (samples);
Est. is the number of effect sizes that had to be estimated because statistics were neot provided or were inadequate;
N is the total number of participants in the k samples;
r_u  is the unbiased effect size estimate (positive values indicate worse reactions or poorer adjustment for participants who experienced greater degrees of the moderator);
values in parentheses after some r_us represent unbiased effect size estimates based on only known (i.e. nonestimated) rs;
95% CI is the 95% confidence interval for r_u based on both known and estimated rs;
H is the withwin-group homogeneity statstic (chi square based on df = k - 1).
^a Estimated effect sizes set at r = 0.
^b Estimated effect sizes based on p = .05, two tailed.
* p < .05.

Five studies examined composite measure-symptom relations. In one, a composite measure of paternal incest, force, and penetration was associated with poorer adjustment ( Edwards & Alexander, 1992 ). Composite measure-symptom relations in the other four studies, however, were nonsignificant. In these studies, the composite measures consisted of incest, frequency, force, and genital contact ( Greenwald, 1994 ); type of CSA and frequency ( Smolak, Levine, & Sullins, 1990 ); extent of physical contact and invasiveness of the sex ( Mandoki & Burkhart, 1989 ); factors such as invasiveness, duration, and frequency ( Cole, 1988 ). The inconsistency in results and in composition of the composite measures makes it difficult to draw conclusions concerning the composite measure-symptoms relations. Future research is required to address this issue by systematically documenting which combinations of moderators are reliably associated with symptoms.