Detrending data PreviousNext
Mplus Discussion > Dynamic Structural Equation Modeling >
Message/Author
 Niyantri Ravindran posted on Friday, June 07, 2019 - 4:36 pm
Hello,

I have a question about trends in data (e.g., linear, quadratic) when using RDSEM. Does the covariance structure for autocorrelations with residuals account for trends, or is there a need to remove linear/quadratic/other trends prior to analysis? Thanks in advance!
 Bengt O. Muthen posted on Saturday, June 08, 2019 - 4:00 pm
I think you want to include a trend in the modeling - so if you have Y ON X (and Y^ residual AR), you could add Y ON TIME in some form. See our 2017 Hopkins training of Short Course Topic 12 and 13 videos and handouts on our website.
 Ming-Chi posted on Saturday, June 29, 2019 - 1:57 am
Could you please tell me, where I can find the simulation syntax about this paper. I need learn.
Tihomir Asparouhov, Ellen L. Hamaker & Bengt Muth (2017) Dynamic Latent Class Analysis, Structural Equation Modeling: A Multidisciplinary Journal, 24:2, 257-269, DOI: 10.1080/10705511.2016.1253479
 Bengt O. Muthen posted on Saturday, June 29, 2019 - 12:36 pm
This is not yet available in Mplus.
 Jon Heron posted on Monday, September 16, 2019 - 8:49 am
Until I watched Bengt's cross-classified analysis presentation on Friday I had been thinking that I needed to de-trend my own data to enable me to make the necessary stationarity assumption.

I have 203 participants each with 52 weeks of depression scores.

My first step was to use MplusAutomation to run a series of AR(k) models (k = 0,1,...,8) on each participant in turn, and then record the p-value for the kth lagged effect as k was increased to determine how many lags would be needed to describe the autocorrelation.

Eyeballing the raw data for anyone who cannot be described by AR(2) shows a range of weird and wonderful curvilinear patterns.

If some people display no systematic change whilst others are linear and others even something closer to cubic do you think cross-classified still doable?


many thanks, Jon
 Tihomir Asparouhov posted on Tuesday, September 17, 2019 - 8:30 am
Yes, it is doable. Take a look at User's Guide example 9.39. Because the random coefficient for time, time^2, time^3 is subject specific you would be able to model linear, quadratic, cubic, or no trends within the same model. If a subject has no trend the random coefficients for the powers of time should be near zero.
 Jon Heron posted on Tuesday, September 17, 2019 - 8:49 am
Awesome, thanks Tihomir :-)
 Jon Heron posted on Thursday, September 19, 2019 - 9:28 am
So, in Bengt's video example on smoking urge and adding a time trend, he jumps back and forth between a cross-classified model and a two-level model and shows how a linear trend can be added to both.

Am I right in thinking that the (somewhat simpler) twolevel approach only works is there is within-wave variation in TIME_t such that it can be added as another time-varying covariate?

In my own data I merely have 52 waves so perhaps that's why I'm not getting anywhere when trying to replicate the twolevel model.


cheers, Jon
 Tihomir Asparouhov posted on Thursday, September 19, 2019 - 12:25 pm
1. I think I was too quick on my first reply. You don't need to be using cross-classified. That is needed only if you need to have a random effect that is time specific. But reading your description you are not talking about that at all. You should be using type=twolevel for that kind of modeling.

2. The two level is an RDSEM model and the cross-classified is a DSEM model, so they won't be super easy to compare. You would have to be going through equations
(65-66) and (70-72) in http://www.statmodel.com/download/DSEM.pdf
The models would be easy to compare if you were using ^ (RDSEM) instead of & (DSEM) but ^ currently is not available for cross-classified (only &). But this shouldn't concern you if you move back to two-level models where you can use ^ instead of &.

3. You can think of Time as a time varying covariate and there is no requirement on the time variable. It should be the same syntax. You can even run the RDSEM model with the autoregressive parameters fixed to 0 and that should give you exactly the two-level model.
 Jon Heron posted on Thursday, September 19, 2019 - 11:51 pm
brilliant, thanks Tihomir
 Jon Heron posted on Friday, September 20, 2019 - 3:02 am
No question, just an update:

I estimated a "mean(y)/Phi/logv" model - I guess you'd call it a 2-level model for a repeated continuous variate with random intercept, random AR(1) autocorrelation and random residual variance. Very catchy.

This model reported problems with AR coeffs for ~25% of the cases which I put down to some participants exhibiting time-trends in their data.

I fitted a range of polynomial regressions to each participant in turn and concluded that the most complex trend could be explained by a cubic hence I de-trended *everyone* using a cubic.

Working with the detrended data (residuals plus observed means) and fitting the same mean(y)/Phi/logv model leads to no reported problems with AR coefficients, a mean Phi which has halved and a mean logv reduced by 75%. this feels like a good result as data artifacts were inflating those quantities.

I then think - I wonder if I can achieve the same outcome by working with the raw data and adding linear/quad/cubic terms as 3 additional random effects. This model shows substantial variation in these random effects but modest means (consistent with the population mean behaviour being pretty well behaved - something also borne out when I fit the simple x-classified model Bengt describes as "Quick and dirty, but not so dirty").

thanks for all the pointers!
 Bengt O. Muthen posted on Friday, September 20, 2019 - 6:33 am
Great. Write it up and send for posting.
 S. Goudarzi posted on Saturday, July 25, 2020 - 1:50 pm
I am trying to fit a multivariate model with cross lags and auto-regressions --equation 7 on McNeish & Hamaker (2019) primer on two-level dynamic structural equation models for ILD in Mplus.

I would like to modify the code to conduct a RDSEM as I need to detrend my data.
However, I cannot seem to get the code to work.
 S. Goudarzi posted on Saturday, July 25, 2020 - 1:55 pm
Here is my modified code:
DATA: FILE = DSEM_E035_fourwave.dat;
VARIABLE: NAMES ARE E035 fact4 COUNTRY WAVE;
USEVARIABLES E035 fact4 WAVE;

Missing are all (9999);
Lagged = E035(1) fact4(1);
Within=WAVE;
cluster=COUNTRY;

ANALYSIS: TYPE = TWOLEVEL RANDOM;
estimator=bayes;
biterations=(2000);
MODEL:

%WITHIN%
PHI1 | E035^ on E035^1;
PHI2 | fact4^ ON fact4^1;
PHI3 |fact4^ ON E035^1;
PHI4 |E035^ ON fact4^1;
TREND1|E035 on WAVE;! E035 is regressed on time to capture the trend over time, the
!slope is latent;

TREND2|fact4 on WAVE;! E035 is regressed on time to capture the trend over time, the
!slope is latent;
logV1|E035;
!the Within-Level residual variance of E035 is latent, !sigma^2_1i;
logV2|fact4;
!the Within-Level residual variance of neoliberalism is latent, !sigma^2_2i;

%BETWEEN%
[E035];
[fact4];
gamma_10;
[PHI1];
[PHI2];
[PHI3];
[PHI4];
Trend1];!mean of the linear trend slope, gamma_60;
[Trend2];!mean of the linear trend slope, gamma_70;
[logV1];
[logV2];


E035;
fact4;
PHI1;
PHI2;
PHI3;
PHI4;
logV1;
logV2;
trend1;
trend2;
 Bengt O. Muthen posted on Sunday, July 26, 2020 - 2:48 pm
For this model, more than say 10 time points are needed for most subjects. If this doesn't help, we need to see your data and output - send to Support along with your license number. Also, we ask that postings be limited to one window - if longer messages are needed, use Support.
 S. Goudarzi posted on Monday, July 27, 2020 - 10:43 am
Thanks much,
I wonder how many time points you need for the equation 7 (Multilevel VAR(1) ) from McNeish & Hamaker (2019).
 Bengt O. Muthen posted on Monday, July 27, 2020 - 10:48 am
See the paper on our website under Time Series analysis (DESM):

Schultzberg, M. & Muthén, B. (2018). Number of subjects and time points needed for multilevel time series analysis: A simulation study of dynamic structural equation modeling. Structural Equation Modeling: A Multidisciplinary Journal, 25:4, 495-515, DOI:10.1080/10705511.2017.1392862. (Supplementary material).
 S.Goudarzi posted on Monday, July 27, 2020 - 11:15 am
awesome, thanks much
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: