Is multiple imputation recommended to handle attrition? I have PRE-POST data from a psychotherapy group intervention, including measures on anxiety, depression, QoL, coping, self-esteem, as well as demographic variables. I´m performing a t-test to examine changes from PRE-POST and a path analysis to examine the effect of coping on anxiety, depression, and QoL at T2 while controlling for the T1 measures.
The data is showing an attrition rate of about 20% (from N=55 at T1 to N=45at T2).
My questions:
1-Is it recommended to use multiple imputation to handle attrition in this case?
2-Could anyone please point me out to a good article/resource related to this situation? 
3-Does multiple imputation work at the item level? I am missing entire responses to instruments but also responses to specific items within instruments.
 A: A good introduction to multiple imputation is http://www.stefvanbuuren.nl/mi/MI.html; you might particularly want to look at the 2002 Schaefer and Graham paper linked from that page, which talks about your issue of losses of individual items versus loss of entire instruments.
The main assumption you require is that the missing data are "missing at random," in the technical sense explained in those references. Essentially, it's OK if the probability of missingness depends on the values of observed data, but it's not "missing at random" if the probability of missingness depends on the values of the missing data.
So applying that principle to your case, you are probably safe in imputing the individual items. You do, however, have to think about whether, say, those who dropped out of your study tended to be those for whom your interventions did not work and became more anxious or depressed.
Also, consider your audience. Some clinicians are highly reluctant to believe results of multiple imputation, clinging to the incorrect idea that it's just "making up data." If your data are "missing completely at random," in the sense that the probability of missingness doesn't depend on either the missing data or the observed data, then restricting to those who completed both instruments just costs you a bit of power, which might be worth it in practice for convincing that type of audience of your result.
My initial comment, in retrospect, was misleading and I will delete it. You should certainly investigate whether dropping out is related to pre-intervention variable values, but not for the reason I suggested.
