Decomposing a mixed model custom ANCOVA

Here is my scenario. The first paragraph tells a bit about the methods we used to give you a frame of reference. The second tells about the first analysis and the third about the second. I can provide more details if anyone thinks it could help.

I recently collected data for a study about how personality influences face perception as a function participant and target sex. I had 20 male face pairs that were altered either communicate high and low levels of a personality trait and 20 female face pairs that did the same thing. Participants indicated which face they preferred from the pair, which were coded as 0 (low) and 1 (high) to create a ratio score with higher numbers being a stronger preference for high levels of that trait. This was for both sexes, which created a repeated measure of male extravert preference versus female extravert preference. Then, participants completed a perceived vulnerability to disease measure (PVD) and indicated their sex.

To analyze these data, we used a custom mixed model ANCOVA in SPSS with repeated factors over the extravert preference for both sexes (i.e., male preference for high levels of personality vs. female) and sex as a between subjects variable. Since this was a custom ANCOVA, we analyzed the personality measure (PVD) as a covariate but used it as a moderating variable. We ended up with a significant 3-way interaction with these variables and sought to decompose it. The first way we decomposed it was by splitting the file by sex and individually correlating the personality variable with preference for both target sexes. That is, we ran four separate correlations (this was done without analyzing the 2-way interaction yet) and found the personality variable negatively correlated with preferences for opposite sex faces. That is, men with higher levels of PVD preferred women with low levels of the trait and women with high PVD similarly preferred men with low levels of the trait. The rub here is that we did not yet analyze the possible preceding 2-way interaction before these correlations, which means that we may not have been justified to look at the individual correlations at this moment.

We went back and split the file by the between factor (i.e., sex) and two ANCOVAs with target face sex as the repeated measure and found only one marginally significant interaction for women. The same effects listed above held true for women, albeit marginally, and the effects were not there for men when we decomposed it that way. This would obviously be the more conventional way to analyze the 3-way interaction but the first way we analyzed it still had stronger findings even if the method we used to analyze them was less conventional.

It seems like we have an effect but we're just not sure what would be the best way to approach it after decomposing the interaction. Is there some kind of precedence of justification to analyze the data we did in the first attempt (see the three-way and individually correlate both levels of the repeated measures with the continuous moderator for both levels of the between)? Or is it best just to go with the second way we analyzed our results? Alternatively, does anyone know of other ways to decompose a mixed model ANCOVA? I was looking through different ways we could do it through syntax but I was just unsure of the normal conventions.