I specified a cross-lagged panel analysis in Mplus with four variables (x1 and y1 measured at T1 and x2 and y2 measured at T2) in the following way:
x2 on x1 y1; y2 on x1 y1; x1 with y1; x2 with y2;
When doing so I got a warning stating that my independent variables x1 and y1 were probably categorical but specified as continuous. My independent variables are indeed categorical but I cannot find how I can specify this in my model. Could you indicate how I can do that?
The scale of independent variables do not need to be specified. Also, means, variances, and covariances of independent variables should not be specified in the MODEL command as these parameters are not part of a regression model. When you specify them, the variables are treated as dependent variables and distributional assumptions are made about them.
rongqin posted on Thursday, July 12, 2012 - 11:41 am
I am trying to do a cross-lagged model between a continuous variable and a categorical variable. I wonder whether I should keep the within waves correlation in the cross-lagged model. In addition, i don't understand why the model with the within waves in have bigger sample size than models without. Here is my syntax: usevariables are VvdelaaV WwdelaaW vaip waip; CATEGORICAL ARE waip; MISSING ARE ALL (99999); analysis: parameterization = theta; model: WwdelaaW on VvdelaaV; XxdelaaX on WwdelaaW: waip on vaip; xaip on waip; vaip with VvdelaaV; waip with WwdelaaW; WwdelaaW on vaip; XxdelaaX on waip; waip on VvdelaaV; xaip on WwdelaaW;
Yes, should keep the within waves correlation in the cross-lagged model. See the TECH4 or RESIDUAL output for the correlations.
rongqin posted on Friday, July 13, 2012 - 11:27 am
Thanks a lot for your response!
Sorry, I am a bit confused: because in some cases you recommended that in cross-lagged path model when there is one categorical variable: "means, variance, and COVARIANCE of independent variables should not be specified in the MODEL command as these parameters are not part of a regression model. " However, for my model, you said within-wave correlation should be included (which also include the correlation between two independent variables?)
I tried the model with within-wave correlations in anyway, the mean/intercepts/thresholds for vaip (.085) was much smaller than that of waip (1.309). And the estimated covariance between vaip and VvdelaaV (.072) was much smaller than correlations between waip and WWDELAAW (.319).Could that be because I did not and could not specify independent variable vaip as categorical. Therefore, when I keep within correlations: vaip with VvdelaaV; in the model, Mplus treat vaip as continuous data? Many thanks for your advice! Yongking
You should make a distinction between exogenous variables and endogenous variables. Exogenous variables are correlated. Their correlations are, however, not model parameters. With endogenous variables, you are estimating a residual covariance not a correlation.
rongqin posted on Monday, July 16, 2012 - 12:01 pm
Dear Linda, Thanks a lot for the clarification. As I understood from you. I don't have to put the within-correlation between independent variables. I adjusted my syntax to the following: usevariables are VvdelaaV WwdelaaW vaip waip; CATEGORICAL ARE waip; MISSING ARE ALL (99999); analysis: parameterization = theta; model: WwdelaaW on VvdelaaV; XxdelaaX on WwdelaaW: waip on vaip; xaip on waip;
waip with WwdelaaW;
WwdelaaW on vaip; XxdelaaX on waip; waip on VvdelaaV; xaip on WwdelaaW; My questions are: 1)Is the syntax correct? Is it ok to use Theta estimation in this situation? 2)what if i need to control for gender? then my independent variables become dependent variables? In that case, I should include correlation between them in the cross-lagged model?
first, I apologize if my question is very basic but I don't seem to be able to find the answer on my own.
I'm testing a cross-lagged model with two time points and with one categorical (s) and one continuous (E) variable, both measured at times 2008 (8) and 2011 (11). The continuous variable is modelled as a latent variable with 3 indicators.
The (important part of) syntax looks like this:
Variable: categorical is s8 s11;
Model: E8 BY e18 e28 e38; E11 BY e111 e211 e311; E11 ON E8 s8; s11 ON E8 s8;
s8 WITH E8; s11 WITH E11;
I'm not sure if I should mark s8 as categorical or not. If I don't, the MPlus output says "WARNING: VARIABLE S8 MAY BE DICHOTOMOUS BUT DECLARED AS CONTINUOUS." But I suppose this doesn't matter because s8 is an independent variable and thus its scale shouldn't matter?
However, my actual problem is that I get different results depending on whether I mark s8 as categorical or not. Specifically, in the above model, the path 's11 ON E8' is significant if s8 is marked as categorical, but non-significant if it is not.
Furthermore, if s8 is marked as categorical, MPlus doesn't compute standardized parameter estimates for the regression paths. Do you have any suggestions?
The variable s8 should not be on the CATEGORICAL list. This list is for dependent variables only. If you are using the default WLSMV, you should related s8 and e8 using the ON option not the WITH option, for example,
Thank you very much for your help! The model works well now and the results make sense.
ellen posted on Sunday, October 20, 2013 - 9:53 pm
Hi, I am running a cross-lagged model based on a 2-wave longitudinal data set. Below is my measurement model.
Variable: Names are a b c d e f g h i j k L; Categorical Are a-f; Missing All (-11); Model:
XTime1 by a (1); XTime1 by b (2); XTime1 by c (3);
XTime2 by d (1); XTime2 by e (2); XTime2 by f (3);
YTime1 by g (4); YTime1 by h (5); YTime1 by i (6);
YTime2 by j (4); YTime2 by k (5); YTime2 by L (6);
a with d; b with e; c with f;
g with j; h with k; i with L;
In the measurement model output, it indicates that there was NOT a significant bivariate correlation between the two latent constructs of XTime1 and YTime2 (r=.012, p =.735).
However, when I performed the cross-lagged structural model by indicating that,
XTime2 ON XTime1 YTime1; YTime2 ON YTime1 XTime1; XTime1 with YTime1; XTime2 with YTime2;
Somehow the output showed that XTime1 significantly and negatively predicted YTime2 (B= -.08, p = .006), even though the correlation between XTime1 and YTime2 was NOT significant (and slightly positive) in the measurement model (r=.012, p =.735).
You are comparing a covariance to a partial regression coefficient.
Kaigang Li posted on Tuesday, November 19, 2013 - 8:32 am
Hello Drs. Muthen,
The syntax below is to fit a cross-lagged stability model with categorical variables W1-W3. Given that W1 is independent variable, so I only claimed W2 and W3 as categorical.
CATEGORICAL are W2 W3 ; Model: W3 on W2; W2 on W1;
S3 on S2; S2 on S1;
W1 with S1;
Based on your answer on the very top of this page "means, variances, and covariances of independent variables should not be specified in the MODEL command" it seems that "W1 with S1;" should not be specified.
Kaigang Li posted on Tuesday, November 19, 2013 - 10:18 am
Thanks Dr. Muthen,
Then I am little confused.
When answering the question posted at the line "Sointu Leikas posted on Tuesday, August 13, 2013 - 2:24 am" above, Linda suggests using "s8 ON e8;" instead of "s8 WITH E8;" S8 and e8 are independent variables.
So, no matter which statement is used, it means that the correlation between s8 and e8 needs to be specified. However, the outputs are different between including and not including that statement.
Would you please clarify this as I think I may have misunderstood some part?
Lina Homman posted on Tuesday, February 25, 2014 - 5:49 am
I am running a cross lagged panel analysis with 3 time points and two measures. I am just wondering whether I should include residual correlations between measures at time point 2 and 3? I find that if I do not do this I get significant cross lagged paths while if I do include them they are not significant.
I am currently running a cross-lagged model that looks (in a simplified version) like this:
GENERAL MODEL: StigmaTime1 BY ST1 ST2 ST3;
StigmaTime2 BY ST4 ST5 ST6;
StigmaTime1 on Intervention SupportTime1 SideEffectsTime1;
StigmaTime2 on StigmaTime1 Intervention SupportTime1 SideEffectsTime1 SupportTime2 SideEffectsTime2;
STABILITY PATHS: SupportTime2 on SupportTime1; SideEffectsTime2 on SideEffectTime1;
CORRELATION BETWEEN INTERVENTION AND SUPPORT: Intervention WITH SupportTime1; Intervention WITH SupportTime2;
with Intervention, SupporTime1, SupportTime2, SideEffects being categorical (dichotomous) variables.
Now I have two question about this model: 1. I understand that I cannot use the WITH-statement to link the intervention and support variables. Is the ON-statement then my only option?
2. For the stability paths, I now use the ON-statement. Does this mean that I have to specify SupportTime2 and SideEffectTime2 as categorical variables (CATEGORICAL ARE) because they are now dependent variables in my model?
I am having some problems with a simple cross-lag panel model. I have 2 observed variables, measured at 3 time points. One variable is a continuous, single item variable. The other variable is also continuous, but due to the skewness I’ve used the censoring (below) option. As a result I am using the wlsmv estimator.
When I estimate the stability paths between the three continuous, single-item variables I find that one of the paths is not significant (VariableA.wave2 ON VariableA.wave1), which seems odd, considering they are only measured 6-months apart, so it seems unlikely they are not significantly related to one another. When I run the correlation between the two variables in STATA, it is close to .50, however in Mplus the correlation matrix is reporting a .01.
The sample statistics appear to be reporting accurate information. Otherwise the model fit is fine, and I can’t identify anything wrong with the syntax:
MODEL: s0Opp_02b ON cov1 cov2 cov3 cov4 cov5 cov6; s0sro on cov1 cov2 cov3 cov4 cov5 cov6;
!stability paths s1Opp_02b on s0Opp_02b; s2Opp_02b on s1Opp_02b; s2sro ON s1sro;
!cross lags s1Opp_02b on s0sro; s2Opp_02b on s1sro; s1sro ON s0Opp_02b; s2sro ON s1Opp_02b;
!within-time correlated errors s0Opp_02b with s0sro; s1Opp_02b with s1sro; s2Opp_02b with s2sro;
I'm running a Cross-Lagged Analysis and would like to test whether the cross-lagged path coefficients are significanly different from one another. Could you please tell me, whether I (1) should standardize the coefficients first (via the "model constraint" command and "new" command) before conducting the test, and whether I (2) should use the "model constraint" command to compare this constrained model with an unconstrained model (by a Chi-Square test), or better use the "model test" command?
After I have calculated the standardized coefficients, should I compare them by (a) the "model constraint" command to compare this constrained model with an unconstrained model (by a Chi-Square test), or (b) use the "model test" command?
No on both a) and b). Instead, use Model Constraint and Model Test in tandem. See page 773 of the V8 UG. You express the standardized coefficients in Model Constraint and you test their difference in Model Test.
I'm hoping to clarify this discussion regarding whether the correlation between exogenous variables should be included in the model specification for a panel analysis (in this case, Cog_00 and Out_00)... I understand that exogenous variables have variances and correlations (not residual variances and covariances)..When I remove their correlation from the model, variances are not estimated for these two variables..
out_16 on out_08; out_08 on out_04; out_04 on out_00;
Cog_16 on Cog_08; Cog_08 on Cog_04; Cog_04 on Cog_00;
out_16 on Cog_08; out_08 on Cog_04; out_04 on Cog_00;
Cog_16 on out_08; Cog_08 on out_04; Cog_04 on out_00;
!Cog_00 with out_00;
Cog_04 with out_04; Cog_08 with out_08; Cog_16 with out_16;
You should not correlate exogenous variables such as those two. They are not part of the model - just like in regular regression. At the same time they are not assumed uncorrelated. You would only correlate them if you want them to be part of the model for FIML missing data handling when they have substantial missing data (see our book).
For your 2-group analysis you get chi-square contribution to fit for each group, but not CFI contribution.