I've run a model (1-1-1) with a single mediator and would like to extend this for a second mediator in serial. I've basically extended Model H (http://www.statmodel.com/download/Preacher.pdf) from Preacher's website. Is this appropriate? I just defined an additional pathway and indirect effect.
To be clear, this is using diary data over 7 days. So that there are 7 timepoints for each variable nested within individuals. The variables on level 2 are preceded with a "PM" and are the "person-mean". The variables on level 1 are preceded by a "WP" and are "person-mean centered".
ANALYSIS: TYPE IS TWOLEVEL;
MODEL: %WITHIN% y WP_x WP_m1 WP_m2; WP_m1 ON WP_x (aw); WP_m2 ON WP_m1 (bw); y ON WP_m2 (cw); y ON WP_x; [x@0] [m1@0] [m2@0] %BETWEEN% y PM_x PM_m1 PM_m2; PM_m1 ON PM_x (ab); PM_m2 ON PM_m1 (bb); y ON PM_m2 (cb); y ON PM_x; [y PM_x PM_m1 PM_m2]; MODEL CONSTRAINT: NEW(indb indw); indb=ab*bb*cb;; indw=aw*bw*cw;
There is nothing wrong with the syntax, but I might suggest regressing WP_M2 on WP_x and regressing y on WP_m1 (and corresponding regressions in the %BETWEEN% model). It is typical to include all these regressions in serial mediation models. Also, if your data are from a daily diary study you might have autocorrelated residuals within person, which this model does not accommodate.
Nick posted on Wednesday, March 23, 2016 - 12:42 pm
Thank you very much for the help!
The data are from a daily diary study. What would you suggest? Is there a more appropriate model that would take auto-correlated residuals into consideration? Could I use the current model and simply use lagged variables for M1, M2, and Y as covariates? I've seen this done in similar situations. Any advice or suggestions are greatly appreciated. Thank you in advance!
That might work for the 'within' model, assuming equally spaced occasions and a constant degree of autocorrelation. An alternative is to run the whole thing as a single-level model using wide format multivariate data, avoiding MSEM altogether. This is the same trick that justifies running latent growth curve models as single-level SEM.