Issues related to output of latent growth model with covariates

77 views
Skip to first unread message

xuey qiny

unread,
Nov 19, 2020, 9:42:21 AM11/19/20
to lavaan
Hi Professors, 

I have several questions related to the output of latent growth model with covariates. 

For example, I fit a linear latent growth model with 4 time points' panel data, and here is the code I used in lavaan: 

# i means latent intercept and s means the latent slope
i=~ 1*x1 + 1*x2 + 1*x3 + 1*x4
s=~ 0*x1 + 1*x2 + 2*x3+ 3*x4 

# latent growth model with covariates 
i~cov1+cov2
s~cov1+cov2

From this latent growth model, I could get the mean of i and s, and the estimates of i and s regressing on cov1 and cov2 from the summary of the fit model. My questions are:  

(1) If I would like to get the estimated mean of i and s for the whole population with covariates, is it calculated like this?   
 Estimated Intercept Mean = 
                                                  Estimated Intercept + 
                                                  Estimated Slope (cov1)* Sample Mean (cov1) + 
                                                  Estimated Slope (cov2)* Sample Mean (cov2)
If this formula is correct,  it must be calculated by hand? or I could use some functions in lavaan? I have used Lavpredict before, but I think it could give i and s for each observation, but could not give the estimated mean of whole population. 

(2) When I calculated the Estimated intercept mean for the whole population from the question (1), how can I determine whether it deviates from some values, for example, mean slope deviating from 0? 

(3) If I could use the formula in question (1), how to set the mean value for cov1 and cov2 if they are categorical variables? 

Thank you very much. 

Regards 

Xueying 

Patrick (Malone Quantitative)

unread,
Nov 19, 2020, 10:23:56 AM11/19/20
to lav...@googlegroups.com
Xueying

If you mean-center cov1 and cov2 before analysis (e.g., with scale() ), questions 1 and 2 take care of themselves. The model-implied means of i and s become the estimated intercepts of i and s, and by default they are tested against zero in the output.

However, question 3 puts a wrench in things. There is no such thing as the mean of a categorical variable. If the variables are nominal, you could effect-code and then mean-center the effect codes.

Pat

--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lavaan/8ab350f2-dd30-4f13-a943-46758d6e33b6n%40googlegroups.com.


--
Patrick S. Malone, Ph.D., Malone Quantitative
NEW Service Models: http://malonequantitative.com

He/Him/His

Xueying Qin

unread,
Nov 19, 2020, 10:59:54 AM11/19/20
to lav...@googlegroups.com
Hi Patrick, thank you very much for the quick response, I understand now. I would like to ask a further question, my latent growth model is actually a bivariate latent growth model with cross-lagged effects between the latent residuals of the two manifest variables at different time points, so the model is more complicated than what I wrote in the above email, however, the latent growth part is just like what I just showed. So my question is, if I mean-center cov1 and cov2 before analysis, will this action have an impact on the cross-lagged parameters? (change values or even direction?)

Thank you very much again. 

Regards 

Xueying 


You received this message because you are subscribed to a topic in the Google Groups "lavaan" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lavaan/-YupG7LJUEY/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lavaan+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lavaan/CAJc%3DyOGpcP4k3i9nnPciJ%3DS-0OVsD6KG6_qvMRCvdsFLJe23fg%40mail.gmail.com.

Patrick (Malone Quantitative)

unread,
Nov 19, 2020, 1:00:51 PM11/19/20
to lav...@googlegroups.com

Xueying Qin

unread,
Nov 19, 2020, 2:55:25 PM11/19/20
to lav...@googlegroups.com
Thank you very much, Patrick. 

Xueying 

xuey qiny

unread,
Nov 20, 2020, 9:53:34 AM11/20/20
to lavaan
Hi Patrick or other Professors, 

I have run the analysis following your suggestions, and I would like to confirm with you some of the points with the results. 

My cov1 and cov2 are continuous variable and binary variable, respectively. Before fit growth model,  I set cov2 as dummy variable which has two levels, 0 and 1, and then I used scale function to mean-center both cov1 and cov2. 

I fit the latent growth model as this: 

# i means latent intercept and s means the latent slope
i=~ 1*x1 + 1*x2 + 1*x3 + 1*x4
s=~ 0*x1 + 1*x2 + 2*x3+ 3*x4 

# latent growth model with covariates 
i~1 + cov1 + cov2
s~1 + cov1 + cov2

I used lavaan function, summarized the fit model and obtained six estimates for "i ~ 1 ", "i ~ cov1" , "i ~ cov2", "s ~ 1 ", "s ~ cov1" , "s ~ cov2". 

So I would like to confirm that, from the previous post,  the estimates for "i ~ 1 " and  "s ~ 1 " are the implied means which represents the mean i and mean s for the whole population, is it correct? 

I am not very confident with my understanding and would like to ask for confirmation again. 

Thank you very much. 

Xueying 

Patrick (Malone Quantitative)

unread,
Nov 20, 2020, 10:16:25 AM11/20/20
to lav...@googlegroups.com
" the estimates for "i ~ 1 " and  "s ~ 1 " are the implied means which represents the mean i and mean s for the whole population,"

Correct, assuming the model is properly specified--i.e., growth in the population is linear with time.

Xueying Qin

unread,
Nov 20, 2020, 11:08:50 AM11/20/20
to lav...@googlegroups.com
Thank you for the clarification, Patrick. For the mean-centering ,shall I set scale =T as well? Currently I only write the function like this: scale (x, center=T, scale=F), and I think it is enough. Is that correct? 

Regards 

Xueying 

Patrick (Malone Quantitative)

unread,
Nov 20, 2020, 12:11:10 PM11/20/20
to lav...@googlegroups.com

Xueying Qin

unread,
Nov 20, 2020, 12:21:06 PM11/20/20
to lav...@googlegroups.com
Hi Patrick, thank you soooooooo much, now everything is clear and I have no more questions. 

Regards 

Xueying 



xuey qiny

unread,
Nov 21, 2020, 11:59:08 AM11/21/20
to lavaan
Hi Patrick, 

This is Xueying again, and I hope this is the last time I post my question for this issue. In my bivariate (x and y) latent growth model with cross-lagged effect, I actually have several x's and y's which could be fit in the model, and I compared the results of before and after mean-centering for each combination of x and y, I found most of the scenario gave the exact same estimates of cross-lagged effects for before and after mean-centering analysis, but there are still a few combinations gave different estimates. And for these different results, I checked the CFI, RMSEA, Chi square and p values, most of them seem not to bad, for example for one combination of x and y, chi2=54.591, df=24, p<0.001, CFI=0.959, RMSEA=0.049. I remember you have said that mean-centering won't change the parameter of cross-lagged effect, so how could I explain the differences? due to small sample size? My sample size is not so big, about 500+, and I also have cluster for the data. 

Regards 

Xueying 

Patrick (Malone Quantitative)

unread,
Nov 21, 2020, 12:33:34 PM11/21/20
to lav...@googlegroups.com
You're still talking about mean-centering the covariates, right? Not x and y? Because that wouldn't be called for.

If you're only mean-centering covariates, it's possible the issue is that centering works differently for multilevel models (which you didn't mention before).There's an entire literature on the implications of different centering strategies for multilevel data.

Xueying Qin

unread,
Nov 21, 2020, 12:57:18 PM11/21/20
to lav...@googlegroups.com
Yes, Patrick. I am still talking about the mean-centering. x and y, here represent the two manifest variables I used in the bivariate latent growth model, I am sorry for the confusion. 

I think I may have made the problem too complicated if I include the mean-centering analysis. In my latent growth model, I am concerned more with the estimates of cross-lagged effects, not the estimates of the latent growth part. I posted this question, mainly because I am afraid readers or reviewers would also have interests on the latent growth parameters, even I could display all the estimates to them, but they are not the whole picture of the latent growth with covariates. 

Hopefully you could understand my concern, and thank you very much for your help. 

Regards 

Xueying 

Reply all
Reply to author
Forward
0 new messages