Message/Author 

Anonymous posted on Saturday, March 16, 2002  3:27 am



Dear Linda/Bengt I have 4 timepoint data. The means of the outcome variable are 1.24, 1.29, 1.34, and 1.30. I expected that a quadratic LGM might fit well. However, I obtained a negative variance of the quadratic factor. Thus, I tried equal error variance, error covariance, and/or factor covariance. I still had the problem. Finally, I decided to free slope factor loadings: 0,*, *, 1, where * means a free parameter. My question is how to interpret the effect of a covariate on the slope factor. For example, if the effect is positive, then how can I interpret the effect of the independent variable on the slope factor? It seems to me that it is not possible to interpret it because the slope is positive and then negative. Am I right? Then, for this type of free slope models, can't we evaluate the effect of a covariate on the slope factor? Thank you very much for your time. 


The relationship between the covariate and the slope growth factor can be interpreted for the time period between 0 and 1, for example, the first and fourth time points. The slope is the rate of change from the first to the fourth time points. Assuming that your variances are small enough so that the increase followed by a decrease is significant, you could also investigate the quadratic model a bit more by fixing the variance of the quadratic factor to zero (don't forget to fix the corresponding covariances to zero also). The negative quadratic variance may simply reflect that individuals are very similar with respect to the downturn at the fourth time point. 

Anonymous posted on Saturday, March 16, 2002  6:23 pm



Dear Bengt Thanks so much for your prompt answers. May I ask you two questions based on your answers? 1. When we free slope factor loadings, as you said, The slope is the rate of change from the first to the fourth time points. I still cannot understand how to interpret the effect of a covariate on the slope factor (factor laodings increase and then decrease). Suppose that we fix slope loadings as 0, 1, 2, and 3. Also suppose that the effect of a covariate on the slope factor is positive. Then, our interpretation would be that the higher on the covariate, the higher on the slope, that is the higher change rate. However,suppose that we fix slope loadings as 0, *, *, and 1. Also suppose that the effect of a covariate on the slope factor is positive. Then, how can we interpret the effect? Given that the slope is not linear, the interpretation is not quite clear. 2. Suppose that we use a quadratic model where we fix the variance of the quadratic factor to zero. Then, how can we evaluate the effect of a covariate on the quadratic factor when its variance is 0? Thank you very much again. 

bmuthen posted on Sunday, March 17, 2002  4:42 pm



When choosing time scores as 0, *, *, 1, the slope is defined as the change from time 1 to time 4. The mean of the slope is the mean difference between time 1 and time 4. So a significant influence of a covariate x on the slope says that x is significantly influencing the change from time 1 to time 4. Note, however, that technically speaking the slope, and therefore the covariate, influences the outcome also at times 2 and 3. The time 2, 3 influence is however multiplied by the estimated time scores for those two time points, which means that the significance of the influence at those points has to be judged using "the delta" method. But you need not concern yourself with that if you don't want to. Regarding the quadratic factor variance, you want to distinguish between two cases. (1) the variance is zero without covariates. (2) the residual variance is zero with covariates. Having found (1), the residual variance often becomes positive when adding covariates because of using more information  you can try that. In (2), this is ok because if the covariate has an influence it means that the slope actually does have variation (it just was not detected without the covariate). I think this is sometimes referred to as "varying" slopes rather than random slopes in the HLM lit. 


Thank you very much for Mplus. I am trying to model data collected at four time points; intake, 2 months, 4 months, and 12 months. My problem is being able to specify the terms to include in the model (constant, linear, quadratic, etc). The means for the four time points are 18.4 (intake), 12.7 (2 months), 13.6 (4 months), and 14.4 (12 months)  lower scores are better. That is, people seem to respond to the treatment well initially, then fall back to some extent. This type of data must be common in some treatment fields. Any suggestions as to what sort of curve I should try to fit? Your advice would be greatly appreciated. Thank you. Peter 


You might consider a piecewise growth model with one intercept factor and two slope factors  one for the initial drop and one for the subsequent rise. The language would be: i BY y1y4@1; s1 BY y1@0 y2@1 y3@1 y4@1; s2 BY y1@0 Y2@0 y3@1 Y3@5; You can then have different covariates for the decline and for the subsequent rise. 


Linda, you are a marvel. Thanks heaps for your helpful advice. I fitted the model as you suggested and came up with a very nonsignigicant Chisq (p = .35) and a CFI of .99. With many thanks, Peter 


I'm glad I could help. 

Anonymous posted on Monday, December 01, 2003  12:06 pm



Hi, I ran slp by y1@0, y2@1, y3* y4* (taken measures at baseline, 1 week, 3 months, and 6 months) and got the following results (Estimates S.E. and Est./S.E.) suggest that the slope is not linear (if linear, expect 0,1,12,24). Slp BY y1 0.000 0.000 0.000 y2 1.000 0.000 0.000 y3 36.916 18.477 1.998 y4 4.543 2.010 2.261 How do i determine whether to add a quadratic term or piecewise growth modeling? Could you please let me know the syntax? Also, say that you are now adding the covariates to the above model. Can I fix the slopes to what I got from the above unconditional model or do i let the program to reestimate them? Thanks a million for your help. 


You can try a quadratic model. I don't think piecewise would work here. Example 22.1B in the Mplus User's Guide shows how to set up a quadratic growth model. Basically, you just add a quadratic growth factor. The time scores for the quadratic factor are the squared linear time scores. I would not fix the slopes when adding covariates. 

Anonymous posted on Thursday, March 24, 2005  9:12 pm



I am running a three time point LGM with the third slope factor pattern coefficient freely estimated (0, 1, *). I have several questions related to the interpretation of the resultant parameters: 1) The mean of the slope was 1.53. I am interpreting this to represent only the change between the first two assessments  is that correct? 2) Following this, does the significance of the slope parameter only indicate significant change between the first two measurements? 3) Similarly, does significant slope variance only refer to significant variability in growth between the first two measurements? 4) Is there a way to estimate significance of change or significance of variability between the second and third measurements? 5) I have two DVs and I estimated this same model for both constructs (0,1,*). The correlation between the slope factors that is estimated via a parallel process growth model was .438. Does this only represent the relationship for the change between the first and second measurements? Thanks in advance for your assisstance. 

bmuthen posted on Thursday, March 24, 2005  11:27 pm



1. yes 2. I assume you refer to the significance of the slope mean  then, yes. 3. yes 4. yes 5. yes I guess you want an elaboration of 4 (and 5) An easy way is to do a second run with changed time scores (*, 1, 0), where you need to give a suitable negative starting value for *. A more complicated way is to use the estimates from the run you mention and recognize that the mean change is estimated by the product lambda*[s]. This product can be given a SE since the variance of the product can be computed  but that is more complicated. 5 can also be done via the changed time scores approach. 

Anonymous posted on Friday, May 20, 2005  6:20 pm



Hi, I am running a growth mixture model and have a question about the time scores. 1) Is there a theory or best practice recommendation concerning fixing the time score to the measurement occasion versus letting it be freely estimated? 2) I am interested in overall growth looking at an earlier post I am thinking I can code my growth process time measurements as: i s q wt1 @0 wt2 @.20 wt3 @.40 wt4 @.60 wt5 @.80 wt6 @1; 3) Do you see any problem with this (given this model has a quadratic term). 4) I would conclude that I can interpret the influence of background variables on the slope as that influence for the process from time one to time 6. 5) I ran both a freely estimated time score model and a fixed time score model. I am wondering why there is such difference in the R squared within class between the two models. In the freely estimated model the R squared of the observed outcomes is around .3 for all three classes. In the model where I fix the time score the R squared is around .17 or .20 within all three classes (except the first time point is .30). 6) Finally, when estimating the freely estimated time score model the graph that is produced is not usable—that is, the axis of time has odd spacing (0, 1 for the first time period—then quite uneven), not what you would like for a publication. What would your recommendation be for graphing a freely estimated time score model? Thank you so much for your assistance!! 

bmuthen posted on Friday, May 20, 2005  11:52 pm



1) With 2 growth factors you have to fix at least 2 time scores  and 0 and 1 are typically used. With 3 factors, fix 3 values. It seems like a good baseline model to start with is with fixed time scores (and I assume you have edquidistant time points here given your scoring). If need be, this can then be followed by freeing some time scores to capture accelerated or decelerated growth. But a quadratic growth curve is already quite flexible so that would be a last resort. 2)3) That's the way to do it. The quadratic factor will use the square of these time scores automatically. 4) This would be correct for linear growth (2 growth factors), because time 6 has time score 1, but you have to take into consideration that you have both a linear slope s and a quadratic slope q both of which contribute to the mean at time 6. 5) Sounds like your development deviates a bit from quadratic. Take a look at each individual's observed trajectory using the Mplus graphics component. Perhaps you need a cubic, or a piecewise model. 6) The spacing of the time axis for the plot will be that of the time scores (fixed and estimated). You can plot at other time points. But first resolve the growth function; hopefully you don't need estimated time scores in a quadratic. 

QianLi Xue posted on Wednesday, July 18, 2007  6:58 pm



Hello, I'm trying to fit a GCM for two parallel processes for continuous outcomes using the following syntax: model: i1 s1  TRAILB1TRAILB6 at TIME1TIME6; i2 s2  WALKSP1WALKSP6 at TIME1TIME6; s1 ON i2; s2 ON i1; and suppose I find out through exploratory analysis that the relationship between, for example, s1 and i2 is not linear, but can be approximated by a twopiece linear spline. Can I do this in MPLUS? Thanks! 


No, Mplus does not do splines. 

J.W. posted on Friday, February 15, 2008  7:40 pm



My quadratic growth model with 6 time points (specified by “MODEL: I S Q  Y0@0 Y1@1 Y2@2 Y3@3 Y4@4 Y5@5;” in Mplus) works very well. However, orthogonal polynomial time functions did not work. I tried the following orthogonal polynomial coefficients: 1) The integer orthogonal polynomial coefficients from statistics texts: Intercept Linear Quadratic 1 3 5 1 2 0 1 1 3 1 0 4 1 1 3 1 2 0 1 3 5 2) the general orthogonal polynomial coefficients generated from SAS: Intercept Linear Quadratic 0.4082 .5976 0.5455 0.4082 .3586 .1091 0.4082 .1195 .4364 0.4082 0.1195 .4364 0.4082 0.3586 .1091 0.4082 0.5976 0.5455 In both cases, THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. Are there any examples/studies that apply orthogonal polynomial time functions in growth model using Mplus? Your help will be appreciated! 


Please send the input, data, output, and your license number to support@statmodel.com. 

J.W. posted on Monday, February 18, 2008  4:15 pm



Linda, the problem was in data format. The Model works fine now. Thanks a lot anyway! 


I have a question about growth modeling and fitting a particularly difficult growth pattern. I have a data set in which the DV is an oscillating outcome. However, I'm interested in the prediction of specific aspects of this oscillation. In particular, I am interested in predicting the individual variation in amplitude of the "wave" and "frequency" of the wave. My question is how to specify a growth model with an oscillating or typical wave form. 


That falls into the realm of nonlinear growth models, that is, models that are nonlinear in the random effects. This is in general not possible in Mplus. You may want to search the literature for nonlinear growth models. One book is by Marie Davidian, but I don't think it covers oscillation phenomena. 


Thank you Bengt, I will look up the book by Marie Davidian and continue the search. If I find something applicable, I will post it for anyone who is interested. 


That would be good. 

Maja Wiest posted on Wednesday, July 29, 2009  1:30 pm



Dear Linda & Bengt, it is possible to implement an exponential function of growth in a GMM like it is usually done to model learning curves that level off after a certain time? I like to model adaptation trajectories after a critical life event and I am mostly interested in then and where people start to level off in ratings of life satisfaction. Thank you very much. 


The relationship between the outcome and the growth factors must be linear. However, a nonlinear growth shape can be specified using the values of the time scores. See the Topic 3 course handout. 

anonymous posted on Friday, October 02, 2009  8:09 pm



Drs. Muthén and Muthén, Thank you very much for this site and for the courses you offer. My question concerns modeling nonlinear growth using time scores and is related to previous posts. I am using age adjusted standard scores measured at four time points. Based on the spacing of the time points, the appropriate way to set up the time scores would be i s 0 1 2 2.5. However, examining the means over time, i s.25 0 1 2 may be better and the fit is improved. I have not had success in fitting a piecewise model with the four time points. Is it incorrect for the time scores to be set up to reflect nonlinearity without using a specific function? 


If you select time scores to fit the data better, I am not sure how you would interpret your results. Time scores should reflect the distance between measurements. If you fit a linear model, what do the modification indices show? 


Also, you may want to reconsider using ageadjusted standard scores for growth modeling. See an article by Mike Selzer called The Metric Matters which may help explain why. I'm not sure where it is published but you can probably find it by searching Jstor and other venues under Selzer. 


I assume that with time scores of: 0 1 3 5 a significant test of slope mean and variance between 0 and 1 also applies to later intervals 1 and 3 or 3 and 5. Is that correct? I assume that because the one unit change is nested within the later intervals. However, when one has time scores like: 0 1 2.71 4.42 my assumption mentioned above would not hold, because the one unit change is not nested within the later intervals (e.g., from 2.71 to 4.42)!? 


The significance of the slope mean (say alpha) refers to the modelestimated mean change in the outcome from 0 to 1 which is alpha. The change from 0 to 3 is 3*alpha, so the significance of alpha is relevant here too. There would be an issue if the time scores were estimated parameters, but that doesn't seem to be the case here. 

Hugo Hesser posted on Thursday, January 28, 2010  10:18 am



Thank you for Mplus! I’m trying to model data from an RCT with a short and longterm followup, collected at four time points: pre intervention, 2 month, 6 month and 42 month. We are interested in comparing the effects of two treatments and overall improvement for the entire sample over time. Sample means and individual slopes seem to suggest that individuals decrease rapidly in symptoms between pre and two month (intervention phase), and continue to decrease between 2 and 6 month but at a slower rate (shortterm followup), and then slightly increase between 6 and 42 month (long term followup). This development over time seems to be similar in both treatments (ie., linear piecewise) Is it possible to run a piecewise growth model with only four time points? And more importantly, do you believe that this might be the right way to go, as there seem to be three unique phases (i.e., initial large drop, continued improvement, and a slight deterioration at the final time point) in our data. If not, could you please guide us in another direction. We have tried a quadratic model but it fits very poor. We would greatly appreciate any advice that you can give us regarding a shape that might fit our data. 


Four time points is not enough for a piecewise model without some restrictions. You could consider the following model or some variation of it: MODEL: i s1  y11@0 y12@1 y13@1 y14@1; i s2 q  y11@0 y12@0 y13@1 y14@2; s1@0; q@0; 

Hugo Hesser posted on Saturday, January 30, 2010  9:56 am



Dear Linda, Thank very much for your useful suggestion of a model, which resulted in a very nice fit (although maybe over fitted). This is definitely in the right direction we want to go. Unfortunately, we ran into some negative residual variance at the last time point. This is probably due to that the variable at that time point is not normally distributed (evidence of a floor effect, many people are doing good). Hopefully we can figure this out on our own (estimator = MLR). But if you have some ideas we would love to hear them. We have some additional questions though: 1. Just to be sure, our time scores for our piecewise model were:1 piece = 0111, 2 piece = 0017, to specify unequal distance between measurement occasion in the 2nd piece (0, 2, 6, 42 month). Is this correct? 2. We would like to center at the last time point (42 month), to be able to interpret the effect of a timeinvariant covariate on endpoint. Is this doable in our model. If so, how would you set it up? We are so glad that you are willing to share you extensive knowledge. With many thanks, Hugo 


To obtain the time scores with centering at the last time point, try: 1 0 0 0 7 7 6 0 Be sure you get the same chisquare. 


I have a very basic question: What is the difference between the piecewise approach and freeing slope loading factors to account for an initial decline and then an increase? For example, I have 6 times of data equally spaced by 1 year. From time 1 to time 2, there is a decrease; then from times 2 through 6, there is an increase. If I free the slope loading factors would I do the following: i s  cp1 cp2 cp3@0 cp4@1 cp5@2 cp6@3; If I do the piecewise approach, would it look like the following: i BY cp1cp6@1; s1 BY cp1@0 cp2@1 cp3@1 cp4@1 cp5@1 cp6@1; s2 BY cp1@0 cp2@0 cp3@1 cp4@2 cp5@3 cp6@4; 


The two models cannot really be compared. They have different parameters and model the data differently. I would probably use two sequential models where each piece has an intercept growth factor. You cannot identify a linear model for two measurement occasions so for that part you would need to fix the variance of the slope growth factor to zero. 


Thanks, Linda. Is my code correct? When I use the piecewise code listed above, I get the error THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 15. THE CONDITION NUMBER IS 0.352D16. Parameter 15 is S1 


Please send the full output and your license number to support@statmodel.com. 


I want to fit latent growth models for some continuous observed variables measured at 3 equally spaced time points. Because the observation times were equally spaced, I began my analyses using fixed time scores of 0, 1 and 2. However, plots, descriptives, model fit output, etc. indicated that some of the observed data were nonlinear. Also, some of the observed data have a pattern of growth followed by a marked reversal of growth (e.g., increasing means from time 0 to 1, followed by decreasing means from time 1 to 2). For the data indicating nonlinear growth, I can use free, logarithmic or exponential time scores, but can I also manually fix selfdetermined nonlinear time scores (e.g., 0, 1, 1.3) to improve fit? For the data having a reversal in growth, can I use a negative time score (e.g., 0, 1, .75) to improve fit? Finally, when evaluating the chisquare test of model fit, higher pvalues are desired, correct? Thanks! 


I would free the last time score. 


Dr. Muthen, Thank you for your prompt reply. I free the last time scores, but on some models I get a message stating "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED. FACTOR SCORES WILL NOT BE COMPUTED DUE TO NONCONVERGENCE OR NONIDENTIFIED MODEL". can you tell me why this happens? Also, when comparing model fit between free time scores, fixed time scores, logarithmic time scores, is there a fit statistic that is preferred for use in this regard? Thanks! 


Please send the full output and your license number to support@statmodel.com. 


Dr. Muthen, I am current trying to develop unconditional models for three health behaviours using linear and quadratic LCM. I have run both linear and quadratic LCM's for all three health behaviours that I am evaluating. I noticed that the RMSEA demonstrates slightly better fit for each quadratic model over its linear counterpart; difference of 0.004 for physical activity, difference of 0.006 for alcohol, and difference of 0.02 for smoking. Are these differences substantial enough to warrent using a quadratic LCM over a linear LCM? Is there another kind of test to determine the shape of the curve rather than just comparing RMSEA's between linear and quadratic models? I also noticed that when comparing the linear and quadratic LCM, the slope means change substantially in magnitude and sign (ie: positive to negative). Is this typical when comparing linear and quadratic LCM? Thanks 


If the mean of the quadratic growth factor is not significant, I would not include the quadratic growth factor. For difference testing of nested models, you should use chisquare or the loglikelihood. 


Dear Linda/Bengt, I am trying to identify the best fitting growth function for offending (measured with the use of a cumulative scale) measured repeatedly over the course of 6 years (five measures were done on consecutive years and the last one was made 3 years after the fifth measure). The observed means of offending are 17.46; 15.06; 21.23; 16.79; 20.53; and 49.02. At first, I tried to free slope loadings: 0, 1, *, 3, 4, *, but I got the warning about “no convergence.” Next, I tried to fit piecewise regression with four different slopes, but I got the warning concerning “nonpositive latent variable covariance matrix.” When I fixed variance to zero, the warning disappeared, but I would rather not do that because I want to estimate the effect of a covariate on the slope(s). Could you please advise me on what sort of curve I should try to fit? Your advice would be greatly appreciated. Thank you. Arina. 


This sequence of means seem difficult to fit a curve to. I would investigate why the means jump up and down so much. Perhaps that is a function of using a cumulative scale. 


Dear Dr Muthen If I wanted to test between using a linear or quadratic growth factor I should be using difference testing of nested models using chisquare or the loglikelihood. How would I do this? Is there an example somewhere? Thanks Wayne 


It's a little tricky to do this testing with chisquare because the more restricted model has not only the quadratic growth factor mean but also its variance fixed at zero  a variance fixed at zero is on the border of the admissible parameter space so chisquare is not formally valid. You can certainly test if a fixed, i.e. zero variance, growth factor is needed by simply looking at the significance of the quadratic growth factor mean. 


Dear Dr Muthen Could I use a loglikelihood for difference testing or does that have the same problems as the chisquare? If the quadratic growth factor mean is significant does that mean that I should choose and quadratic growth curve over a linear one? If the quadratic growth factor mean is not significant, does that mean I should use a linear model? Sorry for all the questions, but I am not familiar with difference testing. Thank you Wayne 


Loglikelihood difference testing has the same problems as chisquare difference testing. Yes. Yes. 


Hello I know that "q" refers to a quadratic growth curve model. What is the syntax for a cubic growth curve model? Thank you Wayne 


Growth factors do not need to have a specific name. They are named as any other variable. See pages 601608 of the Version 6 User's Guide for a description of how to specify various types of growth models. 


Hi I recently ran linear and quadratic growth curve models. Both models fit the data fairly well. Is there an alternative to the linear and quadratic models that I could test to see if it fits the data better? Would allowing some of the factor loadings to be free be one alternative in which the data may fit better? Thanks Wayne 


You should look at the mean of the quadratic growth factor. If it is not significant, you should use the linear model. 

Rod Bond posted on Thursday, November 25, 2010  10:23 am



Hi, I am modeling change in quality of life at 4 times – 0, 3, 6, 12 months. Sample means indicate a significant increase from T1 to T2 but thereafter no mean change. I have tried: 1. Freely estimating time scores for T3 and T4. The slope variance is negative and the residual variance at T1 is large. If s@0, fit is worse and still large T1 residual. If residuals constrained to be equal, fit better and the slope variance significant. 2. Fitting a quadratic function. Gives poor fit and a negative q variance. If q@0, fit poor and large T1 residual. 3. Transforming the time scale. Best is a root transformation where time is raised to the power of 0.1. This provides a good fit, although T1 residual now smaller than at other 3 times. Constraining residuals to be equal does not significantly worsen the fit and the slope variance is now significant. 4. Piecewise model. Time for s1 are 0 1 1 1, and for s2 are 0 0 1 3. s1@0 for identification. The model fits well, but gives large T1 residual. My inclination is to go with the root transformation of time because it fits the data well and gives me a significant slope variance that can be related to covariates. But do you agree? Also, I am unclear about whether to constrain the residual variances to be equal. In effect, should lack of fit be pushed into slope factor variance? Rod 


It's a common situation. Alt. 3 might be good in terms of fitting well, but I am concerned that 13 all make if harder to interpret covariate effects. Alt 4 is better that way, although you say it has a T1 problem. What about Alt 5, using sequential processes  the first one for T1 and T2 and the second one for T3 and T4. Adding a second intercept makes the fitting easier. True, you have to give up a slope variance for each piece, but you gain the advantage of having covariates be able to influence 4 different growth factors, each with a clear interpretation. 

Hemant Kher posted on Thursday, November 25, 2010  8:32 pm



Dear Professors, I have a question regarding my exponential growth model that involves 6 waves of data. I am able to determine the exponential rate (alpha), but not its variance. In other words the model implies that all worhers learn at the same rate. The code I have used for my model is given below: model: i BY y1y6@1; s BY y1* (L0) y2 (L1) y3 (L2) y4 (L3) y5 (L4) y6 (L5); y1y6 (1); [i*30 s*20]; i*1 s*1; i with s*0; [y1y6@0]; rate BY y1@0; rate@0; rate with i@0 s@0; model constraint: New (rate); L0 = 1(exp((0)*(1*rate))); L1 = 1(exp((1)*(1*rate))); L2 = 1(exp((2)*(1*rate))); L3 = 1(exp((3)*(1*rate))); L4 = 1(exp((4)*(1*rate))); L5 = 1(exp((5)*(1*rate))); Is it possible to estimate the variance for the exponential parameter "rate" in this model? 

Rod Bond posted on Friday, November 26, 2010  1:04 pm



Bengt Thanks for your advice. I know this has come up before in this forum, but I am confused as to how you can relate a slope with the variance fixed at 0 to a covariate. I can see that the model runs but what does it mean if, say, I find a significant effect of gender on the s1 slope? Rod 


It simply implies that the slope mean is different for males and females. 

Rod Bond posted on Friday, November 26, 2010  2:29 pm



If the slope mean differs according to values on a covariate, then the slope has a variance? Would the interpretation of the gender effect be different if, with more time points, the slope variance was free? I'm still confused! 


If you regress s1 on gender you have 2 potential sources of variance in s1, gender and the residual in the regression. I am saying that if the residual variance is 0 your model says that s1 has different means for males/females. And yes, s1 has a variance due to gender. But the residual variance is 0  so 100% Rsquare. Note that when s1 is regressed on a covariate, the Mplus statement s1; refers to its residual variance. Or, maybe you are referring to an uncondional analysis (no covariates) finding that s1 has no variance. In that case, it is still of interest to regress s1 on gender for 2 reasons. Adding the covariate changes the model and an s1 variance may be detected  due to both covariate and residual. 

Rod Bond posted on Friday, November 26, 2010  4:51 pm



Bengt Thanks. I now understand! Rod 

Ting Dai posted on Wednesday, March 21, 2012  12:30 am



Dear Drs. Muthen, I have 5 waves of data, and descriptives show it might be a cubic growth pattern, however, the complex PSI matrix had a nonpositive definite. My questions are: 1) Should I try to fit a reduced cubic model? 2) In Mplus, how can I fit a reduced cubic model that reduces the slope and the quadratic term? Is it by setting s@0; [s@0]; q@0; [q@0]; , or should I only set the factor variances to be zero? 3) Would a linear model with intercept and slope but freeing the Times 24 slope paths be a better option than a cubic or a reduced cubic? 4) Any sources I could cite to support the use of a model described in 3)? Thanks very much! 


12. You would fix only the variances to zero not the means. 3. I would use a cubic model over a model with free time scores. I would use a free time score model when one value is not in line with the others. I would not free several time scores. 4. No. 

Ting posted on Wednesday, March 21, 2012  2:37 pm



Thanks, Dr. Muthen, for a quick response. Just so I am sure I get what you said: A cubic or reduced cubic model is better than a linear model with slope paths set as 0,*,*,*,1? And you wouldn't worry that 5 time points are too few to fit a cubic model? Thanks, Ting 


Yes, I think if cubic fits it is better than a model with free time scores. With so many fixed parameters, five time points may be okay. Try it out and see. 


Dear Drs. Muthen, I am fairly new to Mplus and I am trying to run a growth mixture model using data with four time points. I have used the following syntax to run the model: i s q YDEser@0 DE3ser@1.5 DE4ser@3 DE5ser@4.5; i s q@0; I conducted a series of analyses and determined based on the BIC and on LMR that a fourclass model fits the data best. Unfortunately, I have not been able to replicate the best loglikelihood, regardless of the starting values I use. I am not sure what starting values I should use, but I have used up to 2000 random starts. I am afraid that the solution I have is only a local one. (1) How can I over come this problem? Is my model mispecified? Thanks in advance! 


I also have a set of 5 timeinvariant covariates that I would like to include in the model. I ran the model with the four classes and included the covariates even though I wasn’t sure whether the 4class model described above was a good one. After running the model, I found that the classifications/proportions for each class changed, which I attributed to the covariates because some have missing data. However, after conducting a search on this discussion board, I found that this may be due to a misspecification in my 4class model and not to missing data. (2) What can I do to address this issue? Thank you, I appreciated any help you can give me! 


You don't say how many final starts you give. It should be about 1/4 of the initial starts, for example, 2000 500. You must replicate the best loglikelihood before you add covariates. When classes change when covariates are added, it points to the need to direct effects from the covariates to the outcomes. See the following paper on the website for further information: Muthén, B. (2004). Latent variable analysis: Growth mixture modeling and related techniques for longitudinal data. 


Related to the post on july, 07, 2010 to july, 16, 2010... In my growth model the quadratic mean is not significant (the quadratic variance, however, is significant). Based on the assumption that a quadratic model is only needed when the quadratic growth factor mean is significant I wanted to report the necessity of a quadratic model very briefly (and elegant!?) as a chisquare test with 1df. More specifically, I transformed the ttest of the quadratic mean to a chisquare test by squaring the zvalue of the quadraticmean test and stated that including a quadratic factor does not provide a better model fit as compared to the linear model. Is that correct? Do I have to take the quadratic growth factor variance into account in any way? Btw., BIC of the quadratic model was worse as compared to the linear model. 


I would report that the quadratic growth factor mean is not significantly different from zero using the ztest and that the linear model has a lower BIC. I don't think more needs to be done. 


Thank you. Recently I've heard or read (may be it was in this forum but I can't remember) that the Qvariance (besides Qmean) also can be used as a criterion to accept or reject a quadratic growth model (so, following this, you need a nonsignificant Qmean AND a nonsignificant Qvariance to reject a quadratic growth model). What do you or others think about this argument (or does it depend on the specific research question whether Qvariance is of interest to decide about the growth function?). 


I would not focus on the variance when the mean is zero. I think this may be over interpreting the details of the variance structure. 


Dear. Linda I studied and analyzed using LGM. My data was non linear and the model was be with free time scores(T1@0 T2* T3* T4* T5@1). How can be the numerical formula of growth factor loading and estimates of the slope mean expressed ? I can find an formula of either linear growth modeling or piecewise growth modeling but an nonlinear with time free scores. 


You would use the formulas in the Topic3 course handout on the website. Substitute the estimated time scores for the asterisks. 


I¡¯d like to thank you for your help. In the formula, yti = ¥ç0i + ¥ç1ixt + ¥åti I can't understand how 'xt(estimated time scores)' is determined? What's formular of the estimated time scores of nonlinear model or spline model? 


It would come from the value that is estimated in the analysis. 


I am most grateful to you for such a prompt reply. I know that the value is estimated in the analysis. But I would like to know how can the value estimated ..and what is the formula for estimating the value? I tried to search for the formula in Technical Appendices. But I can't find how can the 'x parameters' estimated? 


It is estimated as a factor loading in a CFA. 

Meg posted on Friday, May 18, 2012  10:33 pm



Hi Linda, I am running a LGM with 4 waves of data. The data is not equally spaced (1 year, 5 years, and 7 years). I have set the slope factor loadings to 0 * * 1. The estimates for factors 2 and 3 are .014 and .271. I'm not sure how to interpret the negative estimate for the second loading. 


Please send your output and license number to support@statmodel.com. 


Dear Drs.Muthen, I am conducting a cohortsequential latent growth curve analysis with 3 time points. Is it acceptable to model a curvilinear model (with quadratic term) with 3 just time points ? I ask because a naive analysis suggests the growth to be non linear and my understanding was that one would need a minimum of 4 time points to examine curvilinear growth. Any suggestions will be very helpful Thanks, Charu Mathur 


You need at least four time points to identify a quadratic growth model. 


Thanks a lot, Dr.Muthen. Will this requirement of a minimum of 4 data points hold true in the following scenario as well? we are linking 8 cohorts from 118 years, each cohort being measured every 5 years, for 3 times each. Thanks, Charu Mathur 


If you are using 1 to 18 as time, then you have 18 timepoints so you are fine. 


Thanks Much,Dr.Muthen. I have a couple of followup questions regarding the cohortsequential latent growth analysis, I am conducting. We have 8 cohorts, and each cohort has 3 waves of data with 5 year interval. EDA shows a quadratic growth I examined a CSLGC with Q term with BMI as an outcome. The variances were negative and also, there were several warning messages therefore, I constrained the variances to zero and the model converged. Did this happen because the growth factors had zero variance and I was estimating a random effects model? Notably, Q term was significant in both random and fixed effects models. Subsequently, I ran a dual domain or parallel process model to examine simultaneous growth between BMI and screen time. A univariate analysis showed BMI to have Q growth with fixed effects, and screen time exhibited linear growth with random effects. However, the dual domain model did not converge. I changed the screen time growth factor variances to zero, lowered convergence criteria etc but nothing helped. Interestingly, when I modeled growth as linear for both BMI and screen time with random effects, the dual domain model converged. Do you think the problem is over identification or is it something else? Will appreciate your help, Charu Mathur 


I would need to see the outputs to understand your situation. Please send them and your license number to support@statmodel.com. 


Dear Linda and Bengt, I have been working on the nonlinear growth models (4 time points) and I would like to have some help interpreting the results. Best fit model is the model i s  y1@0 y24@1 y3* y4* with free loading time scores(also the theory supports this conclusion). 1. Slope in the first psycholocial functioning is .29 and the estimated time scores are .91 (T2) and .93 (T3) 2. Slope of the second functioning is .13 with estimated time scores .09 (T2) .18 (T3). I need help to intepret these results regarding the free loads. I know the slope represents the growth between time points 0 and 1. But what does the estimated time scores indicate? Thank you very much. 


Their deviations from the linear time scores 2 and 3 tell you about their deviation from linearity. 


Thank you very much. Have a great weekend. 


I recall that there was a previous message discussing Mplus' inability to estimate BLUPs. In a quick search of the discussion board, I was unable to find the message. When random intercepts and/or coefficients are used, might there be a way of labeling a plot of the randomly determined curves by relevant observation? This strikes me as an important capabilty. In fact, because it may be difficult to test whether the use of random coefficients meets the necessary assumptions, such labeling increases the interest in using such random coefficients. Also, I haven't been able to analyze all the complications associated with obtaining BLUPs in the SEM setting. However, I believe these would be quite important, when feasible. In longititudinal regression analysis, one of BLUPs interesting properties is significantly attenuating the multicollinearity that is inevitably present when numerous categorical dummy varaibles are used, say, to analyze the effect of, individual schools on some dependent variable. So, obtaining BLUPs (and the associated test statistics), when they make sense in SEM, would also be a very nice addition to Mplus. 

Mihyun Park posted on Friday, August 24, 2012  3:06 pm



Dear Linda and Bengt, I am running a sequential latent growth model analysis with multiple group analysis across gender. The syntax is following; Model: i1 s1  su11@0 su22* su33* su44* su55@1; i2 s2  ax11@1 ax22* ax33* ax44* ax55@0; ax11 with su11; ax22 with su22; ax33 with su33; ax44 with su44; ax55 with su55; i1 with s1; i2 with s2; s2 on i1 s1; i2 on i1 s1; It was worked ... but there are only the means estimate of i1 and s1 without i2 and s2 in the result output. Can I know the mean estimate of i2 and s2? 


s2 and is are dependent variables so intercepts rather than means are estimated. See TECH4 in the OUTPUT command to see the model estimated means. 

Mihyun Park posted on Friday, August 24, 2012  4:09 pm



Dear Linda and Bengt, Thank you for your time and prompt response. Your answer really help me out. I already tried to do TECH 4. The TECH 4 output have ESTIMATED MEANS FOR THE LATENT VARIABLES without S.E. I would like to analysis whether the s2 difference of gender is significantly different or not. If it's different across sex... what is the means and S.E. estimate of boys and girls. And then... Is it impossible? 


If gender is your only covariate and if male is zero and female is one, the regression of s2 on gender will tell you if the means are different. 

Mihyun Park posted on Friday, August 24, 2012  5:36 pm



I am so grateful for their help. (Gender is only covariate) You mean... Model: i1 s1  su11@0 su22* su33* su44* su55@1; i2 s2  ax11@1 ax22* ax33* ax44* ax55@0; ax11 with su11; ax22 with su22; ax33 with su33; ax44 with su44; ax55 with su55; i1 with s1; i2 with s2; s2 on i1 s1; i2 on i1 s1; s2 on gender; !Is that right? I'm trying to do that. But I still can't find the means and S.E. estimate of s2 for boys and girls. I only can know whether the sex difference of s2 is significantly different or not. And then...Multiple group analysis with sequential latent growth model having only gender covariate is useful method for comparing the difference of path across gender but it's not valid method for finding the difference of latent variables(e.g.slope). Is it correct? 


Your model has more than one covariate for s2. You have s2 ON i1 s1 gender; Your model does not estimate a mean for s2. It estimates an intercept for s2. It is a conditional model. One way to test if the means are different is to define the means for boys and girls in MODEL CONSTRAINT and test if these are different. 


Good Morning Linda and Bengt, I just finished watch your Topic 3 seminar on latent growth modeling. I learned a great deal, so thank you. I am working on trying to model alcohol use across 5 waves of data. As this is an adolescentemerging adulthood sample, the trends are clearly nonlinear. I have been running quadratic and unspecified models. My main problem however, is that after identifying my functional form, I want to move to a multivariate model, as well as be able to use the slopes as predictors of distal harm outcomes. This makes my decision about quadratic/unspecified models much more difficult as both have limits in terms of interpretability. I have several questions about unspecified models: 1. Based on your mplus discussions, it seems as though you recommend only freeing one time score, rather than 3 for example (T1@0 T2* T3* T4* T5@1)? Why is this? 2. I understand that the slope value indicates a 1unit linear change, so it is not actually capturing the nonlinear trend – therefore, does it make sense to use the slope as a predictor? 3. Once the time scores are estimated, could you rerun the model with those specific time scores? Would the slope value than represent the nonlinear trend over time? 4. Can you conduct multivariate LGM’s using unspecified models? And if so, do all your processes need to be on the same time metric? Thank you 


When you have free time scores, you should not predict from the slope growth factor. You can predict from the intercept growth factor changing the centering point. 1. You can free more than one time score. 2. No. 3. No. 4. Yes. No. 


Dear Dr. Muthén, In your post of March 24, you mention "product lambda". How can I compute this, or where can I find this in the output? Thank you very much! 


I can't find that post. Can you copy the entire post so I can see it? Lambda is the matrix of factor loadings. 


Dear Linda, This was the post. Kind regards, Caroline bmuthen posted on Thursday, March 24, 2005  5:27 pm 1. yes 2. I assume you refer to the significance of the slope mean  then, yes. 3. yes 4. yes 5. yes I guess you want an elaboration of 4 (and 5) An easy way is to do a second run with changed time scores (*, 1, 0), where you need to give a suitable negative starting value for *. A more complicated way is to use the estimates from the run you mention and recognize that the mean change is estimated by the product lambda*[s]. This product can be given a SE since the variance of the product can be computed  but that is more complicated. 5 can also be done via the changed time scores approach. 


Dear Dr.Muthén, Here you find the post I was referring to: bmuthen posted on Thursday, March 24, 2005  5:27 pm 1. yes 2. I assume you refer to the significance of the slope mean  then, yes. 3. yes 4. yes 5. yes I guess you want an elaboration of 4 (and 5) An easy way is to do a second run with changed time scores (*, 1, 0), where you need to give a suitable negative starting value for *. A more complicated way is to use the estimates from the run you mention and recognize that the mean change is estimated by the product lambda*[s]. This product can be given a SE since the variance of the product can be computed  but that is more complicated. 5 can also be done via the changed time scores approach. " 


My message mentioned "the product lambda*[s]". I was referring to the product of the time score at a certain time point (lambda) multiplied by the mean of the slope growth factor ([s]). If you look at the Topic 3 handout, you see how estimated means for the outcomes are computed using such a product. 

C. Lechner posted on Tuesday, June 04, 2013  5:46 pm



Dear Linda and Bengt, I have a growth model with 4 time points. A quadratic model fits well but, as is typical for quadratic models, I find it somewhat hard to interpret the parameters, and even more so the effects of a set of covariates on the growth factors. I now fit a logarithmic growth model with transformed time scores (0 0.69 1.10 1.39) that has almost identical fit. However, I wonder whether that really facilitates interpretation. For instance, what would the effect of a covariate of b=1 imply? Does it mean its effect depends on the time point? Moreover, how can I compute estimated values for each time point in this case (i.e., what exactly is the model equation)? In an empirical application, I have also seen an alternative logarithmic transformation of the time score, which is ln(t)/ln(tmax), resulting in the following vector of timescores: (0 0.50 0.79 1) Does this affect interpretation in any way? Thanks in advance! 

C. Lechner posted on Tuesday, June 04, 2013  7:49 pm



Ok, sorry – I think I already solved part of the problem. Means of intercept and slope were estimated at zero, for whatever reason. When using the short language (i s  … ) , I get values that make more sense. So, whatever transformation of the timescores I use, I guess the slope still represents the linear change in outcome Y for a oneunit change in time, right? In other words, in the second transformation above the slope would indicate linear change between t1 and t4. Am I correct? 


The mean of the slope growth factor is the change from time score zero to time score one. 

C. Lechner posted on Thursday, June 13, 2013  12:03 pm



That's what I thought, thanks! Allow me one followup question: I want to test how a timeinvariant moderator affects intercept and slope, using a multigroup framework. With freely estimated time scores (e.g., 0 * * 1), the default is that the estimated timescores are held equal across groups. However, if one assumes that the shape of the curves may differ across groups, one could estimate different time scores in each group, resulting in differentially shaped group trajectories. So, for example: model: i s  y1@0 y2* y3* y4@1; model group2: i s  y1@0 y2* y3* y4@1; [i s]; What I was now wondering is whether in such a model one could still compare the means of the latent growth variables across groups. I guess it would work fine with the intercept, but what about the slope? Given that it represents change from time score 0 to time score 1 in each group, I would assume I can compare its means across groups, even though the estimated time scores in between 0 and 1 would be allowed to differ across groups (e.g., 0 0.5 0.7 1 vs. 0 0.33 0.66 1). Is that correct? 


You need to have the same growth model in all groups to compare growth factor means across groups. The same growth model represents measurement invariance which is required to test means across groups. 

EunJee Lee posted on Friday, August 15, 2014  1:52 am



Hi, I have a few questions of LGM linear spline. 1. Is it okay to constrain residual variance because of negative covariance warning in unconditional model? I got good model fit after constrained residual variance but I'm not sure that my data is okay to fit on lgm model if it needs constrain at the first step. 2. When I run linear spine with four time points, I gave time scores like 0 * * 1. However, mplus output showed time scores at Time 2 and 3 over 1 like 0 1.04 1.08 1. Why this happened and is it okay? How I can interprete or deal with? 3. If slope variance is not significant, is it not meaninful to go with growth mixture because no heterogeneous in trajectory is found in lgm? 4. If spline and quadratic model shows similar good model fit, which one is recommended to choose? I know I cannot do Chi different test to compare two models so I wonder how I can judge the better one. Thank you. 


1. Equal residual variances for the outcomes if often a good approximation. 2. That is a complex model. You could try a quadratic instead. 3. Yes, because you can still have variance in the intercept growth factor. 4. Seems like a quadratic is simpler. 

EunJee Lee posted on Saturday, August 16, 2014  5:31 am



Thank you very much. I have two more questions based on your comments. As follow your recommendation on question number 2, I tried quadratic but it showed not very good model fit like below. ChiSquare Test of Model Fit Value 48.339 Degrees of Freedom 7 PValue 0.0000 CFI = .918; TLI = .930; RMSEA = 0.113 The means of Time 1 to Time 4 are: T1: 62.283; T2: 59.674; T3:59.630; T4:59.565. So I thought linear is fitted as the shape of trajectory. However the model fit of linear was similar with that of quadratic so I tried linear spline and got excellent model fit but strange factor loading. How can I deal with factor loading problem? Or could you suggest another option please? And I have another question. When I compare the models, linear and linear spline, I usually use chi different test. If I give only spline model modification like variance constraint, should I give same modification on linear too for model comparison, or is it okay to compare with each model's best fit with different modification options? Thank you for your help. 


You may want to take a look at the Mplusrelated paper on our website under Papers, Growth Modeling: Kevin J. Grimm , Joel S. Steele , Nilam Ram & John R. Nesselroade (2013) Exploratory latent growth models in the structural equation modeling framework. Structural Equation Modeling: A Multidisciplinary Journal, 20:4, 568591, DOI: 10.1080/10705511.2013.824775 

EunJee Lee posted on Sunday, August 17, 2014  12:14 am



Thank you for your all kindness. 


Hello, In my data, I have participants selfreport their positive affect and negative affect each hour they are awake. Since positive affect shows a diurnal pattern (increases after waking, peaks in late afteroon, and then decreases), most people have modeled this change using a model that includes both a sine and cosine wave (so in SAS NLMIXED, the equation is: c0 + c1*sin((2*_pi*t/p)) + c2*cos((2*_pi*t/p)), where t = time and p = period (period is 24 hours in this type of analysis)). I have only seen this done in an MLM framework using the HLM software or SAS NLMIXED, and I'm wondering if a nonlinear growth model that models change in terms of a sine and cosine function can be done in an SEM framework, and if so, can this be done in MPLUS? Thank you! 

Jon Heron posted on Thursday, August 06, 2015  7:28 am



I think you can fit an SEMbased latent growth model for anything you can describe a mean function for. This might be a good start (more for the references than the content itself) Psychol Methods. 2015 Mar;20(1):84101. doi: 10.1037/met0000028. Meaningful aspects of change as novel random coefficients: a general method for reparameterizing longitudinal models. Preacher KJ, Hancock GR. http://www.ncbi.nlm.nih.gov/pubmed/25822207 


See also the 2 Grimm et al articles on our web site under Papers, Growth Modeling. They show SAS and Mplus approaches to nonlinear growth modeling. 

Back to top 