Panel Data Econometrics PreviousNext
Mplus Discussion > Multilevel Data/Complex Sample >
 Sanjoy posted on Thursday, March 02, 2006 - 10:00 pm
Dear Professors

Following is the situation we have to handle ... could you tell me please whether can we estimate the model with MPlus 4 (my version is Base + Mixture)

1. We have 386 observations (Y= continuous dependent variable, X1= time dependent covariates, X2=constant covariates) over 22 Time periods.

2. For Some observations the dependent variable (Y) is interval censored.

3. The Panel is "unbalanced" in a sense that for some of the observations the relevant time period is less than 22

4. We want to estimate "fixed effect" and "random effect" both

Thanks and regards

PS: Thanks to the MPlus team for coming up with the version 4.0. Wish all of them the best in their research pursuit.
 bmuthen posted on Friday, March 03, 2006 - 3:02 am
That looks quite doable, except I don't know what interval censored y implies.
 Sanjoy posted on Friday, March 03, 2006 - 6:55 pm
Thank you Professor ... this is what I have meant by censoring

Y's in our model represent repayment by a firm (i) ... we have data recorded from 1980 (t=0) to 2002 (t = 22)

1. There are some firms for whom who have repayment information made at t= 0 as well as t=22 ... doubly censored.... (we do not know when did they start and when could they finish)

2. Some firms are on repayment at t=0 and end their repayment at say t=15 ...left censored...(we do not know when did they start but know when did they end)

3. Some firms start their repayment at say t=10 and still keep repaying at t=22 ... right censored ...(we know when did they start but do not know when could they finish)

 bmuthen posted on Saturday, March 04, 2006 - 10:30 pm
Does this relate to discrete-time survival modeling? See the Muthen-Masyn (2005) JEBS paper on our web site (Recent Papers).
 S.Arunachalam posted on Monday, September 08, 2014 - 6:12 am
Respected Prof. Muthen. In my current project I have a panel data (variables collected over time) as follows:
X variable named NP is for three time periods (Np_t1, NP_t2, NP_t3);
Y variables named lnP is for four time periods (lnP_t1,lnP_t2,lnP_t3,lnP_t4)
Two Moderator Z1, Z2 (interaction variables) are time-invariant
As the focus is not in growth pattern of X or Y variable I am not using growth curve modeling. So I need to go for multilevel modeling of this data, in long format? However if I turn to long format, there is a challenge in the 'data structure' as X is only for three periods, Y is for four periods. So I kindly request your advice on how to model the equation: Y is function of (X, Z1, Z2, X*z1, X*z2) and to guide which format of the two is best; either using:
1.) wide format: but no need of growth pattern just account for the clustering of records within time like in multilevel
2.) OR using long format

Please advice. Thanking you very much in advance.
 Bengt O. Muthen posted on Monday, September 08, 2014 - 10:57 pm
I think you should take the wide approach. This way you can also model how the Y variables are related across time and see if the Xs have lagged effects on the Y's.
 S.Arunachalam posted on Tuesday, September 09, 2014 - 5:40 am
Resp. Prof. Muthen. Thank you for your guidance. If I am using wide format then should I write individual regression for every time point? i.e.
y4 on x3 (b1);
y3 on x2 (b1);
y2 on x1 (b1);
y2 on y1;

In that case should I create interaction terms also for every time point? Would interaction regressions on yi be set to be equal across time?

This problem is arising because I can't tell Mplus to cluster in wide format (similar to using multilevel in long format).
 Bengt O. Muthen posted on Tuesday, September 09, 2014 - 9:41 pm
When you do 2-level (long format) modeling you correlate different time points for the same individual, so that's what you mimic in the wide approach. The x's are correlated by default (not part of the model parameters) and you can correlate the y's in several ways, one being having residual correlations (y2-y4 WITH y2-y4). And add interactions at every time point if needed.

Then ask for Modindices and see what you've misspecified.
 S.Arunachalam posted on Wednesday, September 10, 2014 - 2:47 am
Thank you very much for your guidance Prof. Muthen. Now I am able to better understand the parallels between long format and wide format for longitudinal or panel data.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message