If you have a reason to expect a certain set of match-ups, then simply look at those select few. For instance, suppose you were using the SCL-90 as an outcome measure. You might have some theoretical rationale for expecting that empathic interventions like reflection of feelings would have an effect on "feeling that others don't understand you or are unsympathetic," but that relaxation training would have more of an effect on "feeling tense or keyed up."

The thing to be wary about is the "buckshot" approach of simply correlating every type of intervention with every item on your outcome measures. If your rating system captures 20 different therapist behaviors and you have 90 outcome items, that would be 1,800 correlations - of which at least 50-100 will be statistically significant even if there's nothing "really" there. Of course, sometimes we don't know what to expect ahead of time. If you must correlate everything with everything, then I'd recommend having a cross-validation sample: anything that doesn't hold up in both groups is probably junk.

Another option is to set your alpha level higher - probably a good idea, but in my opinion the widely used Bonferroni correction goes way too far. When you have a lot of analyses, it demands impossibly large effect sizes.

I used the MCMI and wrote a program that would take the answers and identify the specific behavior from the  list of self-perpetuating behaviors that Milon listed in his book that the patients responses indicated.. I read the description to the patient, which was usually validated as something he or she was doing. Then I targeted the behavior to achieve  change. I did it for both the adult and children s versions of the MCMI. It worked quite well with adolescents.