Dear Dr. Muthèn: I have to compute a latent variable (ABS – affective balance score) that represents the latent difference of two latent variables (PA – positive affect and NA negative affect). Then I need to regress the obtained latent score (ABS) on another latent variable (IQ). I have heard of computing latent difference score which should be NA = PA@1 and then computing NA on IQ – where NA represents now the ABS. Is that syntax correct? OR do you have other suggestions how to best compute the difference of two latent variables? Thank you
Jiawen Chen posted on Tuesday, August 26, 2014 - 4:04 pm
Dear Dr. Muthèn:
I wonder if I can model the latent difference score of two different variables because so far all I have come across is about modeling latent difference score of the same variable across different waves. I have a variable called intrinsic work values, measured by three items: "if you are to look for a full-time job, how important are the following job characteristics (5 point scale): a)job is interesting; b)I make most of the decisions; c)I feel accomplished." I have another variable called intrinsic work rewards, measured by exactly the same items, only the question was asked differently, "describe how strongly you agree with the following characteristics about your current job (5 point scale)." So if I want to study how the discrepancy between work values and rewards predicts a host of outcome variables, is it appropriate to model the latent difference score between them at the same wave? Can I also do a third-order latent growth model of the latent difference score across multiple waves? Or some other method may be more suitable for what I want to study? Thank you very much!
It seems to me that one can do some version of latent difference modeling also with different variables, but I haven't tried.
Lixin Jiang posted on Wednesday, September 03, 2014 - 9:07 pm
I have two time points measures (NSsad, SSsad), which have been counter-balanced in my study. Now I want to use gender and personality trait to predict this latent change. See below that I have used it as latent growth model. However, it seems to be misleading as it is not growth itself. How do I model this "latent difference/change score model"? Thanks.
Lixin Jiang posted on Thursday, September 04, 2014 - 4:21 pm
Thank you, Roger. The results based on your codes were more consistent with the regression results from SPSS.
Meike Slagt posted on Thursday, November 13, 2014 - 9:43 am
Dear Dr. Muthen,
I'm trying to estimate latent change scores, using ordinal data.
I can get one part of that model to work (the measurement part), I can get the other part of that model to work (the difference scores, using factor scores saved from the measurement part as input for the difference scores), but I can't get the complete model to converge. That is, I get parameter estimates, but no standard errors. I'm providing the Mplus code I used in the next post, because it doesn't fit in this post anymore.
I would really appreciate advice on this: Why is my model not converging; Is my sample size just to small for this (N=190); Am I forgetting certain parameter constraints leading my model to be unidentified?
Thank you so much in advance!
Meike Slagt posted on Thursday, November 13, 2014 - 9:47 am
!Measurement Part, factorial invariance across time
neg1 BY PreN_Boos PreN_Ver PreN_Bang (1-3); neg2 BY PostN_Boos PostN_Ver PostN_Bang (1-3);
!Difference scores !Works when I save factor scores of neg1 and neg2 and use those as input. Doesn’t work when I try to estimate measurement part and difference scores in one model.
Dif_Neg BY neg2@1; !define latent change by T2 neg2 ON neg1@1; !autoregression T2 T1 neg2@0 neg1; !var at T2=0, estimate var T1 & change [neg2@0 Dif_Neg neg1]; !mean at T2=0, estimate mean T1 & change Dif_Neg ON neg1; !intercept change association
This model results in a warning: THE MODEL ESTIMATION TERMINATED NORMALLY THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING THE FOLLOWING PARAMETER: Parameter 10, [ DIF_NEG ] THE CONDITION NUMBER IS -0.147D-16.
I need to model a difference/change in the factor scores between the baseline model T and the model at T2. Both models are based on the same set of respondents. The goal is to estimate the change and save the corresponding score to further use in the regression analysis.
I wonder whether you could comment on the following approach or, perhaps, point me at the better alternatives. My intuition was to test for the invariance in two measurement models to ensure the scales have identical metric at T and T2, save the factor scores and compute the difference. This being said, I am not sure whether this approach would be perfectly appropriate. I think I stumbled upon Linda's answer somewhere on the forum that invariance should not be tested on the same sample of respondents. I'd appreciate any suggestions.
For what it's worth, the model to be estimated is hierarchical CFA, in which several constructs load on a general construct.
Analyze a model for the 2 time points jointly. Check scalar measurement invariance. Fix the factor mean(s) at zero for the first time point and let them be free to be estimated for the second time point. No need for factor scores. The factor mean(s) at the second time point represent(s) the difference from the first time point.
I am new to Mplus, so please excuse my basic question. I would like to calculate latent difference scores for two latent constructs in the following model: y = (a-b) + (c-d) with a = construct with the indicators a1,a2,a3, b with b1,b2,b3, c with c1,c2,c3, and d with d1,d2,d3.
Time points are analogous to groups in that scalar measurement invariance makes it possible to identify a factor mean difference. That was my reply. I don't know about specific references. With 2 time points there is not much else one can do.
Sara Geven posted on Monday, September 07, 2015 - 10:32 am
Dear Prof Muthen, I would like to analyze a change in a factor measured at two time points. However, I would also like to incorporate the effect of the factor score at wave 1 on the change in factor scores.
This is not possible when I use the code described above (as the path from delin1 to diff is constrained):
I am hoping to use latent difference score modeling to examine changes in two processes, however variables at each time point are theoretically and empirically binomial. Do you know of a resource for latent difference score modeling with binomial variables?
Sara Geven posted on Wednesday, September 09, 2015 - 1:19 pm
Thank you for your willingness to help. I just solved the problem. One item intercept had to be fixed for identification.
Seth Frndak posted on Friday, December 18, 2015 - 8:18 am
I am attempting to create a LDS model with 4 timepoints. The first three timepoints 1-3 are 1-year apart. The 4th timepoint is 4 years after the 3rd timepoint. So you could explain it as timepoints 1-3 & 7.
I know that one assumption of the LDS model is that the autoregressive paths are constrained to 1. If an autoregressive path is modeled between 3 and 7, is there a way to estimate the longer time gap without violating model assumptions?
Furthermore, would additional parameterization be necessary to estimate the time 7 difference score after taking the adjusted autoregressive path into account?
Seth Frndak posted on Friday, February 19, 2016 - 8:08 am
What are the implications of skewed data in a latent difference score framework?
It seems (Burt and Obradovic 2012) that skewness of the baseline scores can create spurious relationships with covariates. I have been unable to find any resources on LDS and highly skewed data. Can you point me in the right direction?
Mark Prince posted on Friday, February 26, 2016 - 9:31 am
I am attempting to run a monte carlo simulation for a bivariate latent change score model and am getting the error:
*** FATAL ERROR THE POPULATION COVARIANCE MATRIX THAT YOU GAVE AS INPUT IS NOT POSITIVE DEFINITE AS IT SHOULD BE.
My syntax works with real data (using code I got from a workshop with Jack McArdle). I am using the parameter estimates from that output in the simulation to look at power, but I get the error above when I do that.
I was wondering if it would be possible to run a latent difference score (LDS) model in MPlus with individually-varying times for each of the latent variables. Also, the time points are not necessarily consecutive since an individual can have data at up to four time points (covering a span of roughly four years) and in that case, I would use only the earliest and latest time points for the LDS. I'm using one construct at two different time points for the LDS with covariates and a distal outcome. If you could point me to suitable references/MPlus code for a LDS model with individually-varying times, I would really appreciate it. Thanks for your help.
I am trying to calculate the latent difference between informants, on the same variable, which difference I then want to predict. I tried two different ways. First the way Linda wrote on April 23, 2012:
In this case the mean of the Diff does not appear on "Model Results" but only on "Sample Statistics for Estimated Factor Scores".
Second, I tried: [f1] (a); [f2] (b); diff BY; diff BY f1@1f2@1; [diff] (c); Model Constraint: c = a-b;
In the latter case, the Mean of the Diff appears in the model results.
Although the estimate of the mean Diff is almost identical in both methods (they differ by 0,001 unstandardized), the SE's are totally different (0,246 and 0,027 respectively). According to the first method, mean Diff is not significant.
Furthermore, comparing "a" and "b" by means of Wald test also results in significant difference.
How do you explain this difference in methods? Since I would like to predict the Diff, how should I proceed?
Dear Drs. Muthen, I am trying to conduct a bivariate latent difference score analysis, and I have two variables estimated at only two time points. I tried to adjust a syntax I found that refer to several time points, but I think I am doing something wrong. I get the message that "THE DEGREES OF FREEDOM FOR THIS MODEL ARE NEGATIVE. THE MODEL IS NOT IDENTIFIED". I attach the syntax. I would appreciate your help in understanding my mistake: (chy and chx are the latent change variables).
I am trying to estimate a random slope for the effect of a latent difference score on the slope of a multilevel latent growth model. However, it seems that thiis not possible as I get fatal error messages. Here's my syntax for the within part. What am I doing wrong?
Maybe I should be more precise. The error message I received was: THE ESTIMATED WITHIN COVARIANCE MATRIX COULD NOT BE INVERTED. COMPUTATION COULD NOT BE COMPLETED IN ITERATION 1. CHANGE YOUR MODEL AND/OR STARTING VALUES.
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ERROR IN THE COMPUTATION. CHANGE YOUR MODEL AND/OR STARTING VALUES.
I am assuming this has to do with the fact that I am trying to estimate a random slope for an effect of a latent variable (CHperprb) on another latent variable (slope). So I would like to know whether this is in theory possible. If it is, then it may have to do with my particular data and I am willing to send it to support.
I am somehow surprise that the Chi_Square is zero (which makes sense since the model is just identified), but the between SRMR is .oo4 (changing the estimator from MLR to WLSMV resolves the problem, but i can't explain why). Is something wrong with the syntax? Many thanks in advance!
I am trying to do a latent change score model with a predictor variable: 'Bin_INSCHSUSP'. The data is complex (weighted with PSU and Strata). However, I get an warning message that says the PSI is not positive definite. Problem involving 'delta'. Please help.
Analysis: Type = COMPLEX ; MODEL:
!Latent Construct--time point one
BIGmath1 BY X1MTHID X1MTHEFF (1) X1MTHINT (2);
TOTmath1 BY BIGmath1*(3) X1TXMTH(4) zHiMath09(5);
!Latent Construct--Time Point 2
BIGmath2 BY X2MTHID X2MTHEFF(1) X2MTHINT(2);
TOTmath2 BY BIGmath2*(3) X2TXMTH(4) zHiMath12(5);
X2TXMTH WITH X1TXMTH; zHiMath12 WITH zHiMath09; X2MTHID WITH X1MTHID; X2MTHEFF WITH X1MTHEFF; X2MTHINT WITH X1MTHINT;
Rachel Jang posted on Thursday, October 04, 2018 - 10:11 am
Dear Dr. Muthèn:
I ran LDS using the code below. However, I keep getting the error message (THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE) and obtaining bad model fit. Is there anything I can do to remove the error message and improve model fit? Thank you so much for your help in advance!
WARNING:THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE.THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES. CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE DX2.
Rachel Jang posted on Thursday, October 04, 2018 - 10:12 am
Please send your output to Support along with your license number. We ask that postings be limited to one window.
Daniel Lee posted on Wednesday, December 26, 2018 - 8:32 am
Hi Dr. Muthen,
Are there any Mplus-related manuscripts or videos that cover the inclusion of distal outcomes in latent difference score models?
I would like to estimate a model where the latent difference scores predict a dependent variable. Is the syntax as simple as just including one additional line of syntax indicating the following (deta is the latent difference scores):
Daniel Lee posted on Friday, May 17, 2019 - 9:11 am
Hi Dr. Muthen, I have a hopefully quick question about a a latent change score model with covariates.
If the effect of a covariate (x) on the latent change variable is positive, but the mean of the latent change variable is negative, does that mean that the covariate is contributing to a larger decrease in the change variable from y(t-1) to y(t)?
I am conducting a latent change score model and have a question about the use of transformations across waves. Out of my four weeks of data, only one week's data had serious skew, which led me to perform a square root transformation. My question: is it acceptable to use the original (untransformed) data for 3 of the weeks and the transformed data for the one week that had a skew? I was not sure if all of the weeks' data needed to undergo a square root transformation so that the range was consistent or if the transformation would not affect the analysis of change scores. Thank you!