Hi, someone mentioned that there maybe complications with using proportion scores in mediation/moderation anaylsis (per Rose and Holmbeck). Can anyone speak to this, or point me to an appropriate literature resource?
Mediator and Moderator Effects in Developmental and Behavioral Pediatric Research. Journal of Developmental & Behavioral Pediatrics. 25(1):58-67, February 2004. ROSE, BRIGID M. M.A.; HOLMBECK, GRAYSON N. Ph.D.; COAKLEY, RACHAEL MILLSTEIN M.A.; FRANKS, ELIZABETH A. M.A.
Also see for relevent example in:
Journal of Clinical Child & Adolescent Psychology 2005, Vol. 34, No. 1, Pages 172-181 Links Among Attachment-Related Cognitions and Adolescent Depressive Symptoms. Shirk, Gudmundsen, Burwell
Page 60 of the article. It is moderated mediation but I wondered if proportion scores of one of the variables is possible? The Shirk article details the basic model I plan on using, detailed on page 173. Thanks!
Taking Figure 3 on page 60 as an example, gender is the moderator of the mediation. Instead of a dichotomous moderator, it is possible in Mplus to work with a continuous variable, which is a reasonable way to treat a proportion score (or you can use the logit transform of the proportion). Since the moderator interacts with a dependent variable in one of the relations (namely the mediator Child independence in its influence on Medical Adherence) the XWITH option should be used. XWITH is intended for use with either 2 latent variables or 1 latent and 1 observed. This means that you would specify a factor underlying the mediator, where the factor is perfectly measured by its single indicator (Child independence).
Okay. Actually, in my model, I have one predictor (maternal cognitions) and two mediators (cog attributions and social cognitions) and one DV (depression score)...the predictor is a proportion score, and my moderator is maternal BDI (probably to be used as dichotomous by median split). Is it okay to have the predictor be a proportion score? This is very helpful, thanks.
Exogenous predictors are treated as continuous variables in regression models. Only the scale of endogenous dependent variables is taken into account by the statistical model. In your model, the proportion variable needs no special attention.
Because the moderator is dichotomous, you could consider a multiple group analysis with two groups.