Analyses of the CSA-symptom relations indicated that college students with a history of CSA were, on average, slightly less well adjusted than college students without such a history. The question arises as to whether these relations were causal in nature. That CSA usually or inevitably causes harm is a basic assumption of many mental health care workers and child abuse researchers. The self-reported effects data, however, do not support this assumption. Nevertheless, self-reports by themselves cannot be taken as firm evidence for or against the role of CSA in causing harm, because people are frequently unaware of the causes of their behavior or current states when causal relations are ambiguous or complex (cf. Nisbett & Wilson, 1977 ). Therefore, we addressed the issue of causation further by considering family environment. Research using clinical samples has consistently shown that family environment and CSA are confounded (e.g., Beitchman et al., 1991 ), which weakens the argument that CSA-symptom relations in these samples are causal. We analyzed the relationship between family environment and CSA in the college samples to determine whether they were confounded as a first step in examining whether CSA caused symptoms.
|
Family factor |
k |
N |
ru |
95% CI |
H |
Abuse and neglect |
5 |
1,098 |
.19 |
.13 to .25 |
2.36 |
Adaptability |
3 |
976 |
.13 |
.06 to .19 |
20.38* |
Conflict or pathology |
9 |
4,906 |
.14 |
.12 to .17 |
0.74 |
Family structure |
4 |
3,803 |
.09 |
.06 to .12 |
6.54 |
Support or bonding |
13 |
3,288 |
.13 |
.09 to .16 |
36.46* |
Traditionalism |
5 |
836 |
.16 |
.09 to .22 |
8.26 |
Note
k represents the number of effect sizes (samples);
N is the total number of participants in the k samples;
ru is the unbiased effect size estimate;
95% CI is the 95% confidence interval for ru;
H is the within-group homogeneity statistic (chi square).
A positive ru indicates better family adjustment or functioning in the control than sexual child abuse (CSA) group.
* p < .05 in chi-square test.[Page 39 continues]
The confounding of CSA and family environment raises the possibility that CSA may not be causally related to symptoms in the college population or may be related in a smaller way than uncontrolled analyses have indicated.
To address this issue, we examined the relationship between family environment and symptoms. All studies providing statistics assessing the relationship between these two factors were coded. For each study, effect sizes were computed for all family environment-symptom relations.
Additionally, for each study, a study-level effect size was computed; this value represents the mean effect size based on Fisher Z transformations of all family environment-symptom relations in that study. A series of symptom-level meta-analyses and a study-level meta-analysis were then performed.
Table 11 provides the results of the meta-analyses of the symptom-level and study-level effect sizes. Symptoms that had only one effect size were not meta-analyzed.
The effect sizes ranged from r = .04 to .49. All effect size estimates based on two or more effect sizes were significantly greater than zero, as indicated by their 95% confidence intervals. Five of the seven effect sizes based on single samples were significantly greater than zero.
In the majority of cases, effect size estimates were based on a small number of samples and the effect sizes used to derive these estimates were heterogeneous. This latter finding is not surprising, given the heterogeneous collection of family environment measures for any given symptom. These estimates should therefore be viewed with caution.
Nevertheless, with the exception of two
measures based on single samples, the effect sizes were generally medium in size, in
contrast to the CSA-symptom and CSA-family environment effect sizes, which were generally
small. The study-level effect size estimate was
In terms of variance accounted for, family environment outperformed CSA in explaining symptoms by a factor of 9.
These results imply that, in the college population, family environment is a more important predictor of symptoms than is CSA
(see below for a discussion of the statistical validity of comparing CSA-symptom and family environment-symptom relations).
Table 11
Meta-Analyses of Symptoms as a Function of Family Environment Factors
Symptoms |
k |
N |
ru |
95% CI |
H |
Alcohol |
1 |
383 |
.04 |
-.06 to .14 |
-- |
Anxiety |
3 |
788 |
.34 |
.28 to .40 |
19.80 |
Depression |
5 |
1,279 |
.38 |
.33 to .43 |
22.28* |
Dissociation |
1 |
251 |
.49 |
.39 to .58 |
-- |
Eating disorders |
4 |
822 |
.21 |
.15 to .28 |
10.05* |
Hostility |
1 |
383 |
.15 |
.05 to .25 |
-- |
Interpersonal sensivity |
2 |
634 |
.32 |
.24 to .38 |
20.25 |
Locus of control |
1 |
383 |
.07 |
-.03 to .17 |
-- |
Obsessive - compulsive |
2 |
634 |
.27 |
.20 to .34 |
4.02* |
Paranoia |
1 |
383 |
.16 |
.06 to .26 |
-- |
Phobia |
1 |
383 |
.18 |
.08 to .28 |
-- |
Psychotic symptoms |
1 |
383 |
.22 |
.12 to .31 |
-- |
Self-esteem |
5 |
1,345 |
.26 |
.20 to .30 |
37.13* |
Sexual adjustment |
2 |
337 |
.23 |
.13 to .33 |
0.24 |
Social adjustment |
3 |
653 |
.41 |
.35 to .47 |
20.50* |
Somatization |
2 |
634 |
.22 |
.15 to .29 |
12.59 |
Suicide |
2 |
634 |
.26 |
.18 to .33 |
1.41 |
Wide adjustment |
4 |
992 |
.31 |
.25 to .37 |
12.95* |
Study level |
13 |
2,846 |
.29 |
.26 to .33 |
62.56* |
Note
k represents the number of effect sizes (samples);
N is the total number of participants in the k samples;
ru is the unbiased effect size estimate (positive values indicate greater degrees of symptoms are associated with poorer family functioning);
95% CI is the 95% confidence interval for ru;
H is the within-group homogeneity statistic (chi square).
-- dashes indicate H was not computed because only one sample was involved.
Meta-analyses were performed when k > 1.
Study-level effect sizes are mean effect sizes, based on Fisher Z transformations, of all symptom-family environment relations in a given study.
* p < .05 in chi-square test.
Results of the three sets of analyses just presented
(i.e., meta-analyses of the relationships between CSA and symptoms, CSA and family environment, and family environment and symptoms)are consistent with the possibility that the small but statistically significant CSA-symptom associations found in the studies reviewed may have been spurious.
This possibility is suggested by the logic of semipartial correlational analysis, or equivalently, hierarchical regression analysis ( Keppel & Zedeck, 1989 ). These analyses are useful for determining whether a significant relationship between two variables remains significant after controlling for extraneous factors.
The necessary conditions for a significant relationship to be reduced to nonsignificance are as follows:
The analyses presented above demonstrate that the first three of these conditions were generally satisfied.
Further, the finding that the mean correlation between CSA and symptoms ( r = .09) was somewhat smaller than that between CSA and family environment ( r = .13), which in turn was substantially smaller than that between family environment and symptoms ( r = .29), suggests that many significant CSA-symptom relations might be reduced to nonsignificance with statistical control. To address this possibility directly, we coded all studies that employed statistical control (see Table 12 ).
Table 12
Study |
Type of control |
Significant results |
|||
N |
Before |
After |
% reduction |
||
Brubaker, 1999 |
Separated categories |
1 |
1 |
0 |
100 |
Cole, 1988 |
Hierarch. Regression |
5 |
3 |
0 |
100 |
Collings, 1995 |
ANCOVA |
10 |
8 |
6 |
25 |
Fromuth & Burk, 1989, mw |
Hierarch. Regression |
13 |
6 |
6 |
0 |
Fromuth & Burk, 1989, se |
Hierarch. Regression |
13 |
0 |
0 |
- |
Fromuth, 1986 |
Hierarch. Regression |
13 |
4 |
1 |
75 |
Gidycz et al., 1995 |
Path analysis |
3 |
0 |
0 |
- |
Greenwald, 1994 |
Hierarch. Regression |
1 |
0 |
0 |
- |
Harter et al., 1988 |
Path analysis |
2 |
1 |
0 |
100 |
Higgins & McCabe, 1994 |
Hierarch. Regression |
2 |
2 |
0 |
100 |
Lam, 1995 |
Multiple regression |
3 |
0 |
0 |
- |
Long, 1993 |
Multiple regression |
2 |
1 |
0 |
100 |
Pallotta, 1992 |
ANCOVA |
13 |
6 |
0 |
100 |
Yama et al., 1992 |
ANCOVA |
2 |
2 |
1 |
50 |
Totals |
83 |
34 |
14 |
59a |
Note. N indicates the number of symptom measures whose relation to child sexual abuse (CSA) status was examined (or was intended to be by the study authors) by using statistical control. "
Before" indicates the number of relations significant before applying statistical control; " After" indicates the number of significant relations after applying statistical control. " Reduction" indicates the percent of significant relations that became nonsignificant after statistical control.
-- Dashes indicate that persentage reduction was not computed because all results were initially nonsignificant;
ANCOVA = analysis of covariance;
mw = Midwest; se = Southeast.
a Based on the percent of total significant relations that became nonsignificant after control. The unweighted percent reduction was 83%.Page 39 continued]
Coding involved recording for each study the type of statistical control used, the number of symptoms whose relationships with CSA were controlled for, the number
[Pasge 40]
of significant CSA-symptom relations before statistical control, and the number of significant CSA-symptom relations after statistical control. [*4]
[*4] It would have been preferable to code and examine effect sizes before and after statistical control, rather than the number of (non)significant relations. Because of inadequate reporting of the statistics that resulted from statistical control, this procedure could not be used.
Table 12 displays the
results of this coding. In the last column the percentage of reduction from before to
after statistical control is provided.
Statistical control was used in 13 studies with 14
samples-in some cases control was not used because nonsignificant correlations between
symptoms and family environment obviated this procedure, although the researchers had
planned to use statistical control; these samples are included in this analysis.
In all cases but one (i.e., Brubaker, 1991 ), statistical control involved using statistical procedures such as hierarchical regression or analysis of covariance (ANCOVA). Brubaker (1991) imposed control by separating her participants into mutually exclusive categories
(i.e., no abuse, CSA only, psychological abuse only, physical abuse only, followed by combinations of these abuse types).This deconfounding procedure has been used recently by other researchers examining noncollege samples, who have shown that when CSA is isolated, its negative correlates tend to shrink considerably or disappear (e.g., Eckenrode, Laird, & Doris, 1993 ; Ney et al., 1994 ).
Of 83 CSA-symptom relations, 34 (41%) were significant before statistical control. Only 14 (17%) remained significant after statistical control.
It is important to note that, within any given study, multiple CSA-symptom relations were not independent, because they were based on the same sample. It may therefore be more appropriate to use only one result per study (e.g., percentage of reduction) to evaluate the effects of statistical control.
Using this approach, the overall reduction from statistical control was 83% (as opposed to the 59% reduction using dependent relations).
One additional study, not shown in the table and not included in the above analysis, also used statistical control ( Wisniewski, 1990 ). This study was based on 3,187 female college students drawn from 32 colleges and universities that were fairly representative of all institutions of higher learning in the United States.
Unlike the other studies using statistical control,
which held extraneous factors constant for all participants (with or without CSA) in a
single analysis, Wisniewski conducted four separate analyses using path analysis, one for
each separate group of participants (i.e., no CSA, nonincest CSA, incest CSA, and
nonincest CSA with adult revictimization).
For all CSA participants, she constructed a CSA
severity score that reflected the degree of felt victimization from and negative reactions
to the CSA. Results of her analyses revealed that CSA did not contribute to current
adjustment for nonincest or incest CSA participants and contributed to only a small degree
( b weight = .02) in the case of incest with adult
revictimization subjects. Wisniewski found that other factors, particularly family
violence, best explained current adjustment.
Results from studies using statistical control supplement the analyses of the intercorrelations among CSA, symptoms, and family environment. They provide direct evidence that the majority of significant CSA-symptom relations examined in the college samples may have been spurious.
These results imply that significant CSA-symptom relations in studies based on
[Page 41]
college samples cannot be assumed to represent effects of CSA. Although the results of the analyses of statistical control, as well as analyses of the CSA-symptom-family environment relations, do not prove that CSA-symptom relations are spurious in the college population, they specifically do not support the assumption that a basic property of CSA is that it causes psychological injury.
Statistical validity..
In comparing CSA-symptom and family environment-symptom relations, as well as statistically controlling for family environment when assessing CSA-symptom relations, several statistical issues may relate to the validity of these analyses.
It is possible that the CSA-symptom association may be underestimated relative to the family environment-symptom association.
First, often unstandardized measures of CSA may have less reliability than measures of family environment. Lower reliabilities translate into attenuated correlations ( Glass & Hopkins, 1996 ; Hunter & Schmidt, 1994 ).
Second, CSA is usually measured as a dichotomous variable (i.e., present or absent), whose distribution tends to be skewed with a strong mode in the absent category. Low base rates for a category of interest (e.g., CSA) can attenuate correlations ( Glass & Hopkins, 1996 ; Rosenthal & Rosnow, 1991 ).
Further, the artificial dichotomization of an independent variable (e.g., CSA) can also attenuate correlations ( Glass & Hopkins, 1996 ; Hunter & Schmidt, 1994 ).
Regarding the first point, although most studies on CSA have not assessed the reliability of their measures of CSA, several have, all of which were based on college samples.
These results point to acceptable reliabilities for measures of CSA, which are comparable to reliabilities for family environment measures-for example, 8-week test-retest reliabilities on the Family Environment Scale have ranged from .68 to .86 ( Cole, 1988 ).
Furthermore, the reliability results from the first three of the studies just discussed are especially relevant, because their measures of CSA were modified versions of Finkelhor's (1979) measure; about half of the studies in the current review used modifications of Finkelhor's measure. Thus, support for acceptable reliability extends to a sizable portion of the studies under review.
The second issue concerns attenuating effects from low base rates. The more the
split between CSA and control participants deviates from 50-50, the greater the
attenuation in the CSA-symptom association will tend to be (cf. Rosenthal &
Rosnow,
1991 ).
This attenuation is quite small for a 27-73 split (e.g., female CSA), but it is
somewhat larger for a 14-86 split (e.g., male CSA).
However, the attenuation is small in
absolute magnitude for small effect sizes.
For the small CSA-symptom effect size estimates
obtained in the current review, adjusted effect size estimates based on a 50-50 split
increase at most by .03 (based on formulas provided by Rosenthal & Rosnow, 1991 ),
indicating that adjusted effect size estimates are still small in magnitude and are
considerably smaller than the family environment-symptom effect size estimate of
From an empirical point of view, it is noteworthy that, in the current review, base rates were not positively related to effect size estimates, r (48) = -.04, p > .70, two-tailed, contrary to expectations that they would be.
Finally, the relevance of artificial dichotomization to the CSA variable is weakened by the fact that CSA has generally been conceptualized as a categorical rather than continuous variable (i.e., one experiences CSA or one does not).
Nevertheless, despite this common conceptualization of CSA, several researchers have attempted to construct continuous measures of CSA and have used these measures to compare CSA with family environment in terms of their relative contribution to adjustment variance (e.g., Cole, 1988 ; Wisniewski, 1990 ).
Wisniewski's severity score of CSA discussed previously is one example. For nonincestuous SA students who were not revictimized as adults, a path analysis revealed that family violence was related to current levels of emotional distress ( b = .13), whereas CSA was not ( b = -.02). Likewise, for incestuous CSA, family violence ( b = .27) was related to emotional distress, but CSA was not ( b = -.01).
Cole constructed a severity index for CSA (composed of factors such as degree of invasiveness), which can also be viewed as a continuous measure of CSA. She found that CSA did not explain adjustment variance above and beyond that explained by various family environment factors in a hierarchical regression.
It is important to note that a continuous measure for physical abuse, constructed similarly to the severity index for CSA, was entered along with CSA in the last step of the analysis; this family environment factor, but not CSA, accounted for additional adjustment variance.
Results from these studies in which CSA was constructed to be continuous are consistent with results from studies in which CSA was treated dichotomously in terms of pointing to family environment, rather than CSA, as a significant contributor to current adjustment.
In sum, CSA-symptom relations could be underestimated relative to family environment-symptom relations because of
The foregoing discussion suggests that reliability is not problematic and that attenuation due to low base rates is of very low magnitude because effect size estimates were small to begin with.
In a similar vein, attenuation due to dichotomization, if artificial, would also be of very low magnitude because of the small effect size estimates that were obtained (cf. Glass & Hopkins, 1996 ).
Empirically, low base rates were not associated with lower effect size estimates, and CSA was relatively unimportant compared with family environment when CSA was treated as a continuous variable.
These considerations support the validity of comparing CSA-symptom and family environment-symptom relations and of assessing CSA-symptom relations when controlling for family environment. Nevertheless, precise, as opposed to relative, estimates of the contributions of CSA and family environment to adjustment may be somewhat problematic because of the possibility of low magnitude attenuations of CSA-symptom relations.