I want to test direct & indirect effects of continuous observed variable (X) & mediator (MEDVAR) on an observed 3-category DV(NOM3OUT). Total N=160 & some missing data (assumed MAR) on NOM3OUT. I planned to use multiple imputed data sets (generated outside Mplus)& pool the indirect effects estimates across imputed data sets w/ TYPE=IMPUTATION.
DATA: FILE IS IMPLIST.DAT; TYPE = IMPUTATION; ANALYSIS: ALGORITHM=INTEGRATION; INTEGRATION=MONTECARLO;
MODEL: MEDVAR ON X (P1) COVAR1 COVAR2 COVAR3; NOM3OUT ON MEDVAR (P2 P3) X COVAR1 COVAR2 COVAR3; MODEL CONSTRAINT: NEW(IND1 IND2 MEDOR1 MEDOR2 INDOR1 INDOR2); IND1 = P1*P2; IND2 = P1*P3; MEDOR1 = exp(P2); MEDOR2 = exp(P3); INDOR1 = exp(P1*P2); INDOR2 = exp(P1*P3); If I understand, this would yield logit coeff's for P2 &P3. Is it feasible to estimate & test indirect effects by forming products of ordinary & logit regression coeff's? If I exponentiate the estimated indirect effects, does that yield odds ratios? If not, is there a way to estimate probit coeff's directly or could the logit coeff's be converted to probit as part of the model constraint statement (e.g., multiplying logit coeff's by 0.625)?
I looked through the paper you mentioned, but have not been able to figure out how to apply the principles and examples to our particular case. Might you have any further or more specific thoughts about our proposed model/method? Many thanks in advance for any guidance you can offer!
You can try to first do this treating NOM3OUT as a binary variable (one category vs the other two) - there are explicit formulas for that in my paper (see binary outcome, continuous mediator). Otherwise, you may need a statistical consultant to help you.