Mike Todd posted on Monday, March 12, 2007 - 12:27 pm
My colleague and I are attempting to examine mediation (X-->M-->Y) in a multilevel context where X and M are dichotomous, Y is continuous, and all 3 variables are measured 252 times for each of 55 individuals. Our interest is in testing purely within-person associations (effect of symptoms (X) at day t on mood (Y) at day t+1 as mediated by sleep disturbance at night t).
While there are certainly straightforward multilevel regression methods for estimating and testing indirect effects in all-continuous variable models, I've yet to find an analogous approach for models where either M or Y is categorical. So, it has been recommended that we use Mplus.
The major question is, can such an analysis reasonably be conducted in Mplus. Whether we choose the typical multivariate repeated measures approach (ala latent curve modeling) or the multilevel (TYPE=TWOLEVEL) approach, it seem that we would run into problems given the very large number of variables (~750) and the relatively small number of cases (55).
You can do this in Mplus as a 2-level model for 3 variables (X, M, Y) with repeated measures over time as level 1 and individuals as level 2. You do this using ML estimation with logit regression for the binary dependent variables. I can send you a paper by MacKinnon et al which is under review and that endorses the Mplus product approach to indirect effect estimation with a binary outcome. Although the Mplus Model Indirect features are not available for the numerical integration algorithm that is neeed, you can use Model Constraint to produce the product of slopes that is of interest.
Mike Todd posted on Tuesday, March 20, 2007 - 10:28 am
Thanks so much, Bengt. Please do send the MacKinnon paper. I would like to see what Dave has come up with on this.
Mike Todd posted on Thursday, October 11, 2007 - 4:20 pm
I've started working with these data again.
I've successfully set up and run simple 3-variable multilevel path models of the form X->M->Y (including a direct path from X to Y), where X and M are binary and Y is continuous and all variables are at the within-person level (i.e., Level 1).
I have 4 questions now:
1) In your opinion, would it make sense to obtain bootstrapped estimates and standard errors for the product of the X->M and M->Y paths to test the indirect effect of X on Y?
2) If yes to question 1, is there a way to set this up in Mplus 4.21 where TYPE=TWOLEVEL and ALGORITHM=INTEGRATION?
3) If yes to question 2, is there a relevant example online and/or in the manual?
4) Will version 5 allow for the MODEL: INDIRECT specification under the INTEGRATION algorithm?
Dear Dr. Muthen, I am attempting to examine multiple mediation analyses with repeated measurements (N = 119). The predictor X and the two mediators M1 and M2 were assessed at Time 1, Time 2, and Time 3. The criteria Y was assessed at Time 4. All variables are continuous. My intention is to test multiple mediation 'controlled' for time. Should it be conducted within a multilevel model or should I examine the mediation with time as a control variable? How could it be implemented in MPlus?
Use a single-level analysis with data in wide format.
Sabrina Krys posted on Wednesday, February 15, 2017 - 1:29 am
Thank you Dr. Muthen,
but how can I implement a multiple mediation using intercepts and slopes in mplus? Which specifications have to be done? Do you have any article or chapter where a mediation with data in wide format is conducted? I hope my understanding is not fundamental wrong.
Sabrina Krys posted on Thursday, February 16, 2017 - 12:15 am
I thought you were referring to growth analysis. If not, what is meant by 'single level analysis using data in wide format'? My question is, whether I can analyze mediation analyses with repeated measures (dependent samples) or not. Should I analyze each mediation individually or is it possible to integrate them in one model? My approach was to put my data into long format, build a group variable (=time), and control for it. But the problem is that the samples are dependent.
For other interested readers: I found another paper that might be interesting within the context of mediation with repeated measures. Itīs Preacher (2015) - Advances in Mediation Analysis. Also could be of interest: Vancouver & Carlson (2015) - All Things in Moderation, Including Tests of Mediation.
Sorry, but I have another question: The following input uses the mentioned data in long format with a group (=Time) variable (t1, t2, t3). Is it possible to put the three samples together and control for time?
M1 ON x group; M2 ON x group; Y ON x m1 m2 group; m1 with m2;
I think Mplus won't complain but I am not sure you get at the indirect effect correctly this way. There is a big literature on longitudinal mediation - also outside the SEM references you give. See, e.g., this contribution from the causal inference area:
VanderWeele (2011) in J of Consulting and Clinical Psych.