Four women in the control condition who believed they had consumed alcohol and indicated that they felt intoxicated were deleted from further analyses, thus reducing the sample size to 230. There were no placebo manipulation failures. Participants in the low alcohol condition had a mean BAL of .034% (SD = .008) immediately before reading the stimulus story and a mean BAL of .033% (SD = .007) after completing dependent measures. Participants in the high alcohol condition had a mean BAL of .062% (SD = .007) just before reading the stimulus story and a mean BAL of .079% (SD = .012) after completing the dependent measures. This demonstrates that participants were on the ascending limb or at peak BAL while completing the dependent measures.
Participants rated their level of perceived intoxication on a scale of 0 (not at all intoxicated) to 6 (extremely intoxicated) before the experimental story and at the end of the dependent measures. Prior to reading the story, women in all four beverage conditions significantly differed from one another in perceived intoxication, F(3,226) = 260.94, p < .001 (Mhigh alc = 4.19, SDhigh alc = 1.04; Mlow alc = 3.57, SDlow alc = 0.89; Mplacebo = 2.38, SDplacebo = 1.54; Mcontrol = 0.05, SDcontrol = 0.34). Likewise, after completing the dependent measures, women in all four beverage conditions still significantly differed from one another in their perceived intoxication, F(3,226) = 121.64, p < .001 (Mhigh alc = 4.15, SD
high alc = 1.40; Mlow alc = 3.30, SDlow alc = 1.43; Mplacebo = 2.78, SDplacebo = 1.61; Mcontrol = 0.23, SDcontrol = 0.94).
The current study operationalized alcohol consumption by utilizing simple coding procedures for regression models. The first code (expectancy effect) contrasted control participants with placebo participants. To test the pharmacological effect of alcohol, the second code contrasted control participants with alcohol participants (both .04 and .08 participants combined). Preliminary analyses showed that the effects of the low and high dose groups on primary appraisals did not differ from each other for any partner risk group; thus, they were collapsed for all analyses.
Data Analytic Plan
To test the model represented in , we employed a multi-groups modeling approach using Mplus 3.0 (Muthen & Muthen, 2004
). Multi-groups modeling tests whether the factor structure underpinning the relationships described in were similar across all partner risk conditions (Rigdon et al., 1998
). Therefore, within a single Mplus model, a path analysis was run for each partner risk condition. Mplus then determines model fit using information derived from all three models. Maximum likelihood estimation was selected because it is robust to violations of normality (Chou & Bentler, 1995
). Model fit was assessed with several absolute and incremental fit indices, including chi-square, root mean square error of approximation (RMSEA), and comparative fit index (CFI; Bentler, 1990
; Bentler & Bonett, 1980
; Bollen, 1989
; Browne & Cudeck, 1993
). Although a nonsignificant χ2
demonstrates that the model fits well, it is dependent on sample size and significant values are often accepted if other indicators of fit are good. RMSEA values less than .08 and CFI values over .90 indicate good fit (Browne & Cudeck, 1993
; Hoyle, 1995
All significant correlations among the measured variables were in the expected directions (see ). Across all three partner risk conditions, primary appraisals were positively correlated with secondary appraisals, which were negatively correlated with assertive condom request and positively correlated with likelihood of engaging in unprotected sex. Finally, assertive condom request and likelihood of engaging in unprotected sex were negatively correlated at each level of partner risk. Furthermore, across all three partner risk conditions, general intention to have unprotected sex was negatively related to assertive condom request and positively related to likelihood of unprotected sex. In the low and high risk conditions, general intention was also positively related to secondary appraisals.
Means, Standard Deviations, and Correlations among Predictor Variables
Multiple Regression Analyses
Given the number of contrast codes and possible interaction terms that existed across the models and resulting concerns about power, multiple regression analyses were used to simplify later path analytic models. A series of hierarchical multiple regression analyses were run to examine the significance of the general predisposition by alcohol contrast interaction on primary appraisals within each partner risk group (Cohen, Cohen, West, & Aiken, 2003
). The first set of analyses examined the two-way interaction between general predisposition and the pharmacological effect contrast code (control vs. alcohol participants). The first step included general predisposition and the pharmacological contrast code; the second step included the multiplicative interaction term. Across all three partner risk conditions, the interaction term was not a significant predictor of primary appraisals, unknown: B = .04, t(77) = 0.14, p = ns; low: B = −.25, t(75) = −1.57, p = ns; high: B = .16, t(75) = 0.84, p = ns. The second set of analyses examined the two-way interaction between general predisposition and the expectancy effect contrast code (control vs. placebo participants). The first step included general predisposition and the expectancy contrast code; the second step included the multiplicative interaction term. Across all three partner risk conditions, the interaction term was not a significant predictor of primary appraisals, unknown: B = .42, t(77) = 1.29, p = ns; low: B = .03, t(75) = 0.13, p = ns; high: B = .21, t(75) = 0.77, p = ns. Because the two-way interactions were not significant in any of the above analyses, the interaction term was dropped from the path analytic models; therefore, only main effects were examined.
The hypothesized path model was tested such that paths not represented by lines in were not estimated. This model was tested across all three partner risk conditions but did not fit the data well, χ2 (42, 230) = 126.50, p < .05; RMSEA = .162; CFI = .72. Because this model did not fit well, the correlations presented in and the modification indices from the model above were used to guide the addition of three paths in an iterative fashion. The first, a path from secondary appraisals to likelihood of unprotected sex is consistent with the CMM in that likelihood of unprotected sex is an outcome that can be directly affected by secondary appraisals. Two other paths, from general intention to secondary appraisals in the low and high partner risk conditions and from general intention to assertive condom request in all three partner risk conditions, were added to allow for the possibility that general intention to have unprotected sex may be directly related to secondary appraisals and condom request intention. Each addition slightly increased model fit so that the final revised model fit the data well, χ2 (33, 230) = 38.83, p = ns; RMSEA = .048; CFI = .98. This final revised model served as the comparison model for the invariance testing (see ).
Final comparison model across all partner risk groups.
Across all partner risk conditions, all hypothesized paths related to the CMM (i.e., relationships among primary appraisals, secondary appraisals, assertive condom request, and likelihood of unprotected sex) were significant and in the expected directions (see ). Appraising the situation as having higher sexual potential was positively related to endorsement of secondary appraisals. The more women believed that the situation had sexual potential, the more they endorsed impelling cognitions related to having sex. These impelling cognitions were negatively related to assertive condom request intention and positively related to unprotected sex intention. Finally, assertive condom request and unprotected sex intention were negatively related.
In addition to the paths related to the CMM, several other paths were consistent across all three partner risk conditions. First, there was a significant negative relationship between general intention to have unprotected sex and assertive condom request and a significant positive relationship with secondary appraisals. Additionally, general intention to have unprotected sex was positively related to primary appraisal for sexual potential. Regarding the alcohol effects, both the pharmacological contrast code and the expectancy contrast code were positively related to primary appraisals across all three risk conditions.
To determine if the paths estimated across the risk conditions in this comparison model were invariant, a fully constrained model was compared to the final revised model presented above. The fully constrained model fit the data well, χ2 (53, 230) = 64.82, p = ns; RMSEA = .054; CFI = .96 and was not significantly different from the comparison model, Δ χ2 (20, 230) = 25.99, p = ns. The lack of differences between the fully constrained model and the comparison model suggest that there are no differences between the partner risk groups (see ). Overall, the model accounts for 41.4%, 42.8%, and 46.8% of the variance in likelihood of unprotected sex in the unknown sex risk, low sex risk, and high sex risk groups respectively.
Figure 3 Final model across all three risk groups illustrating the effects of alcohol and general intention to have unprotected sex on the Cognitive Mediation Model. Standardized loadings from the unknown risk model are presented. (Note that even though the path (more ...)
To verify this, each path was successively held invariant and compared to the model fit of the comparison model presented above. Every path was found to be invariant across the risk conditions except one: the path from the expectancy contrast code to primary appraisals, χ2 (35, 230) = 46.81, p = ns; RMSEA = .066; CFI = .96; Δ χ2 (2, 230) = 7.98, p < .05. To further investigate this difference in models, the path from the expectancy contrast code to primary appraisals was held invariant across two groups at a time. The first model tested the invariance of this path across the unknown and low sex risk conditions. Overall, the model fit well and was not significantly different from the comparison model, χ2 (34, 230) = 38.86, p = ns; RMSEA = .043; CFI = .98; Δ χ2 (1, 230) = 0.03, p = ns. The second model tested the invariance of this path across the unknown and high sex risk groups, χ2 (34, 230) = 44.89, p = ns; RMSEA = .065; CFI = .96; Δ χ2 (1, 230) = 6.06, p < .05. The final model tested the invariance of this path across the low and high sex risk groups, χ2 (34, 230) = 45.16, p = ns; RMSEA = .065; CFI = .96; Δ χ2 (1, 230) = 6.33, p < .05. Overall, the results indicate that the path from the expectancy contrast to primary appraisals in the high sex risk group was significantly different from the same path in both the unknown risk group and the low risk group. This suggests that the influence of beverage condition differed across partner risk conditions. In the unknown and low partner risk conditions, those in the placebo and alcohol conditions rated the situation as having more sexual potential than those in the control condition. In the high risk condition, however, there was only a significant pharmacological effect such that those who consumed alcohol believed the situation had more sexual potential than those in the control condition.