Centering approach in DSEM PreviousNext
Mplus Discussion > Dynamic Structural Equation Modeling >
Message/Author
 Ruixue Zhaoyang posted on Monday, November 05, 2018 - 1:09 pm
Hi,

I have two questions regarding the centering approach used in DSEM for the lagged variable and the level-1 covariates respectively. (1) From my understanding, the latent centering approach is used for the variable with the autoregressive parameter by default. So in the following syntax, we use the raw (uncentered) data for y. Is this correct (in Mplus V8.0)? (2) Is latent centering approach also available for level-1 covariate x, which has variability on both within-person and between-person levels? If so, how do we specify it? Or do we have to use the group-mean center approach for x as follow?

variable:
lagged=y(1);
within=x;

define: center x (groupmean);

model:
%within%
s1 | y on y&1;
s2 | y on x;
 Tihomir Asparouhov posted on Monday, November 05, 2018 - 1:57 pm
The Y variable is latent centered (since 8.0). The covariate X can be latent centered (since 8.1) using script like this

variable:
lagged=y(1);

model:
%within%
s1 | y on y&1;
s2 | y on x;
%between%
y; x;

so you basically have to take out the observed group mean centering statement and the within= specification. No other changes are needed.
 Ruixue Zhaoyang posted on Monday, November 05, 2018 - 2:49 pm
Thank you very much for your explanation!

I'm using Mplus 8.0 for DSEM analysis. Does it mean that in 8.0, the observed group-mean centering approach needs to be used for covariate X to get accurate estimates for within-person and between-person effects?

Thank you!
 Tihomir Asparouhov posted on Monday, November 05, 2018 - 3:31 pm
You can use the observed centering approach if the cluster sizes is more than 100, but generally you should try to move to 8.2, not just because of the latent centering but also because you might want to use the RDSEM method instead and include auto-correlation for the covariate as well. See the first three papers
http://statmodel.com/TimeSeries.shtml
 Annette Brose posted on Wednesday, April 10, 2019 - 1:30 pm
Hi,

I am currently familiarizing myself with DSEM and have a question on the meaning of the intercept when using the latent covariate approach to centering.

The model is simple: the dependent variable y (occasions nested in persons) is regressed on the time varying predictor variable x.
The Mplus code is

VARIABLE:
NAMES = id y x;
USEVARIABLES = y x ;
WITHIN = ;
BETWEEN = ;
CLUSTER = id;
ANALYSIS:
TYPE = TWOLEVEL RANDOM;
ESTIMATOR = BAYES;
PROCESSORS = 2;
BITERATIONS = (2000);
MODEL:
%WITHIN%
s | y on x;
%BETWEEN%
y on x;
x;
[x];
y;
[y];
s;
[s];
y with s;

The estimated intercept for y is very different from the intercept of a similar model in which I use the observed group mean centering approach as described in Asparouhov & Muthen (2018). Instead, the intercept seems to be reflect a value for x = 0. The estimated slope, instead, is quite similar to the slope of the group mean centering approach. That is, the latent mean centering approach seems to center the Level-1 predictor x around the latent means of x (similar to what happens in group mean centering), but the Level-2 part of x does not seem to be centered. Is my observation correct or am I doing something wrong here / misunderstand something? What exactly is happening here?

Happy to hearing from you. Thank you.
 Tihomir Asparouhov posted on Thursday, April 11, 2019 - 2:32 pm
Correct. The Level-2 part of x is not centered. Centering on Level-2 is not as important as centering on Level-1. Centering on Level-2 is nothing more than subtracting a constant from a variable, i.e., it is a simple model reparameterization / change of scale. You can add
DEFINE: CENTER x (GRANDMEAN);
if you want it centered.
 Annette Brose posted on Monday, April 15, 2019 - 10:36 pm
Thank you!
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a public posting area. Enter your username and password if you have an account. Otherwise, enter your full name as your username and leave the password blank. Your e-mail address is optional.
Password:
E-mail:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: