Skip to first unread message
cezre...@gmail.com
Feb 28, 2019, 5:42:07 PM2/28/19
to lavaan
Hello,
I have a latent variable change model I have fit below. I have measurements at 2 time points and I have 3 groups. When I include all 3 groups, I get the standard "Warning message: In lav_object_post_check(object) : lavaan WARNING: some estimated lv variances are negative" ". I copy the output below for the model with 3 groups. However, when I drop one of the groups (that seems to be the problem), the code runs fine. So this leads me to believe that this is NOT a model mis-specification issue but, rather, something with the data of the group I dropped. I do know the post-scores decrease some, relative to the pre-scores, for that group and I am wondering if that might be contributing to the problem. I don't see any extreme values (i.e. outliers) that might cause this issue. Any thoughts on what might be driving the error and how to deal with it? I can provide the output for running the model when I drop the one group, if that is helpful.
mod2<-'
pre.wm=~1*RSpan1+1*SSpan1 #Specifies the measurement model for pre.wm
post.wm=~1*RSpan2+equal(pre.wm=~SSpan1)*SSpan2 #Specifies the measurement model for post.wm
post.wm ~ 1*pre.wm #Fixed regression of post.wm on pre.wm
dCOG1 =~ 1*post.wm #Fixed regression of dCOG1 on post.wm
post.wm ~ 0*1 #Intercept of post.wm constrained to 0
post.wm ~~ 0*post.wm #Fixes variance of post.wm to 0
dCOG1 ~ 1 #Estimates intercept of change score
pre.wm ~ 1 #Estimates intercept of pre.wm
dCOG1 ~~ dCOG1 #Estimates variance of change score
pre.wm ~~ pre.wm #Estimates variance of pre.wm
dCOG1~pre.wm #Estimates the self-feedback parameter
RSpan1~~RSpan2 #Allows residual covariance on RSpan across T1 and T2
SSpan1~~SSpan2 #Allows residual covariance on RSpan across T1 and T2
RSpan1~~RSpan1 #Allows residual variance on RSpan at T1
SSpan1~~SSpan1 #Allows residual variance on SSpan at T1
RSpan2~~equal("RSpan1~~RSpan1")*RSpan2 #Allows residual variance on RSpan at T2
SSpan2~~equal("SSpan1~~SSpan1")*SSpan2 #Allows residual variance on SSpan at T2
RSpan1~0*1 #Constrains intercept of RSpan to 0 at T1
SSpan1~1 #Estimates the intercept of SSpan at T1
RSpan2~0*1 #Constrains the intercept of RSpan to 0 at T2
SSpan2~equal("SSpan1~1")*1 #Estimates the intercept of SSpan at T2
'
fitmod2 <- lavaan(mod2, data=all,group='Group' , estimator='mlr',fixed.x=T,missing='fiml') #fixed.x=T regards all exogenuous covariates as estimated
summary(fitmod2, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)
Warning message:
In lav_object_post_check(object) :
lavaan WARNING: some estimated lv variances are negative
> summary(fitmod2, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)
lavaan 0.6-3 ended normally after 55 iterations
Optimization method NLMINB
Number of free parameters 42
Number of equality constraints 3
Number of observations per group
1 70
2 78
5 99
Number of missing patterns per group
1 2
2 2
5 2
Estimator ML Robust
Model Fit Test Statistic 5.252 5.775
Degrees of freedom 3 3
P-value (Chi-square) 0.154 0.123
Scaling correction factor 0.909
for the Yuan-Bentler correction (Mplus variant)
Chi-square for each group:
1 5.252 5.775
2 0.000 0.000
5 0.000 0.000
Model test baseline model:
Minimum Function Test Statistic 271.474 234.809
Degrees of freedom 18 18
P-value 0.000 0.000
User model versus baseline model:
Comparative Fit Index (CFI) 0.991 0.987
Tucker-Lewis Index (TLI) 0.947 0.923
Robust Comparative Fit Index (CFI) 0.990
Robust Tucker-Lewis Index (TLI) 0.940
Loglikelihood and Information Criteria:
Loglikelihood user model (H0) -1149.830 -1149.830
Scaling correction factor 0.984
for the MLR correction
Loglikelihood unrestricted model (H1) -1147.204 -1147.204
Scaling correction factor 1.049
for the MLR correction
Number of free parameters 39 39
Akaike (AIC) 2377.661 2377.661
Bayesian (BIC) 2514.527 2514.527
Sample-size adjusted Bayesian (BIC) 2390.897 2390.897
Root Mean Square Error of Approximation:
RMSEA 0.095 0.106
90 Percent Confidence Interval 0.000 0.228 0.000 0.242
P-value RMSEA <= 0.05 0.226 0.192
Robust RMSEA 0.101
90 Percent Confidence Interval 0.000 0.225
Standardized Root Mean Square Residual:
SRMR 0.012 0.012
Parameter Estimates:
Information Observed
Observed information based on Hessian
Standard Errors Robust.huber.white
Group 1 [1]:
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
pre.wm =~
RSpan1 1.000 0.722 0.691
SSpan1 1.000 0.722 0.783
post.wm =~
RSpan2 1.000 0.770 0.713
SSpan2 (SSp1) 1.030 0.300 3.440 0.001 0.793 0.810
dCOG1 =~
post.wm 1.000 0.541 0.541
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
post.wm ~
pre.wm 1.000 0.938 0.938
dCOG1 ~
pre.wm -0.099 0.203 -0.486 0.627 -0.171 -0.171
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.RSpan1 ~~
.RSpan2 0.262 0.164 1.598 0.110 0.262 0.458
.SSpan1 ~~
.SSpan2 0.028 0.154 0.185 0.854 0.028 0.086
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 0.151 0.091 1.660 0.097 0.362 0.362
pre.wm 0.003 0.118 0.028 0.978 0.005 0.005
.RSpan1 0.000 0.000 0.000
.SSpan1 (.21.) 0.061 0.094 0.655 0.513 0.061 0.067
.RSpan2 0.000 0.000 0.000
.SSpan2 (SS1~) 0.061 0.094 0.655 0.513 0.061 0.063
Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 0.168 0.112 1.504 0.133 0.971 0.971
pre.wm 0.522 0.118 4.408 0.000 1.000 1.000
.RSpan1 (.16.) 0.571 0.191 2.985 0.003 0.571 0.523
.SSpan1 (.17.) 0.330 0.160 2.061 0.039 0.330 0.387
.RSpan2 (RS1~) 0.571 0.191 2.985 0.003 0.571 0.491
.SSpan2 (SS1~) 0.330 0.160 2.061 0.039 0.330 0.344
R-Square:
Estimate
post.wm 1.000
dCOG1 0.029
RSpan1 0.477
SSpan1 0.613
RSpan2 0.509
SSpan2 0.656
Group 2 [2]:
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
pre.wm =~
RSpan1 1.000 0.719 0.706
SSpan1 1.000 0.719 0.700
post.wm =~
RSpan2 1.000 0.493 0.522
SSpan2 1.436 0.421 3.407 0.001 0.708 0.775
dCOG1 =~
post.wm 1.000 NaN NaN
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
post.wm ~
pre.wm 1.000 1.459 1.459
dCOG1 ~
pre.wm -0.264 0.216 -1.226 0.220 NaN NaN
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.RSpan1 ~~
.RSpan2 0.039 0.116 0.333 0.739 0.039 0.067
.SSpan1 ~~
.SSpan2 -0.003 0.098 -0.028 0.978 -0.003 -0.007
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 -0.086 0.102 -0.846 0.398 NaN NaN
pre.wm 0.128 0.117 1.096 0.273 0.178 0.178
.RSpan1 0.000 0.000 0.000
.SSpan1 0.024 0.118 0.202 0.840 0.024 0.023
.RSpan2 0.000 0.000 0.000
.SSpan2 -0.020 0.145 -0.135 0.893 -0.020 -0.021
Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 -0.037 0.075 -0.495 0.621 NaN NaN
pre.wm 0.518 0.133 3.895 0.000 1.000 1.000
.RSpan1 0.520 0.158 3.289 0.001 0.520 0.501
.SSpan1 0.538 0.129 4.167 0.000 0.538 0.509
.RSpan2 0.650 0.132 4.924 0.000 0.650 0.728
.SSpan2 0.334 0.150 2.229 0.026 0.334 0.400
R-Square:
Estimate
post.wm 1.000
dCOG1 NaN
RSpan1 0.499
SSpan1 0.491
RSpan2 0.272
SSpan2 0.600
Group 3 [5]:
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
pre.wm =~
RSpan1 1.000 0.693 0.709
SSpan1 1.000 0.693 0.683
post.wm =~
RSpan2 1.000 0.502 0.576
SSpan2 1.494 0.619 2.414 0.016 0.750 0.699
dCOG1 =~
post.wm 1.000 0.521 0.521
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
post.wm ~
pre.wm 1.000 1.380 1.380
dCOG1 ~
pre.wm -0.309 0.196 -1.576 0.115 -0.818 -0.818
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.RSpan1 ~~
.RSpan2 0.023 0.165 0.136 0.892 0.023 0.046
.SSpan1 ~~
.SSpan2 0.142 0.167 0.851 0.394 0.142 0.249
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 -0.149 0.103 -1.448 0.148 -0.567 -0.567
pre.wm -0.067 0.098 -0.686 0.493 -0.097 -0.097
.RSpan1 0.000 0.000 0.000
.SSpan1 -0.092 0.102 -0.902 0.367 -0.092 -0.090
.RSpan2 0.000 0.000 0.000
.SSpan2 0.095 0.201 0.473 0.636 0.095 0.088
Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 0.023 0.075 0.304 0.761 0.331 0.331
pre.wm 0.480 0.110 4.384 0.000 1.000 1.000
.RSpan1 0.475 0.139 3.414 0.001 0.475 0.497
.SSpan1 0.550 0.095 5.762 0.000 0.550 0.534
.RSpan2 0.507 0.159 3.197 0.001 0.507 0.668
.SSpan2 0.590 0.284 2.078 0.038 0.590 0.512
R-Square:
Estimate
post.wm 1.000
dCOG1 0.669
RSpan1 0.503
SSpan1 0.466
RSpan2 0.332
SSpan2 0.488
Terrence Jorgensen
Feb 28, 2019, 10:42:36 PM2/28/19
to lavaan
when I drop one of the groups (that seems to be the problem), the code runs fine.
The code seems to run fine even in the example you posted. There don't seem to be any equality constraints across groups, so I don't see why results should differ when you remove a group. I suspect it has something to do with the fact that only the first group has any df, which is probably due to your use of the equal()* operator instead of simply assigning labels to parameters. When you have multiple groups, I'm not sure the equal()* operator applies that constraint across all groups. Better to label your parameters.
Terrence D. Jorgensen
Assistant Professor, Methods and Statistics
Research Institute for Child Development and Education, the University of Amsterdam
cezre...@gmail.com
Mar 1, 2019, 2:23:00 PM3/1/19
to lavaan
Thanks for your response, Terrence. I have updated the code as you suggested; but the problem still remains. Group 2 still has a negative variance for dCog1 (and the warning goes away when I drop Group 2 from the analysis). I am not sure what you mean by only the 1st group has any df. Any other thoughts? Is it critical that I don't have a negative variance before proceeding to use the model results? Everything else is fine...
mod2<-'
pre.wm=~1*RSpan1+1*SSpan1 #Specifies the measurement model for pre.wm
post.wm=~1*RSpan2+1*SSpan2 #Specifies the measurement model for post.wm
post.wm ~ 1*pre.wm #Fixed regression of post.wm on pre.wm
dCOG1 =~ 1*post.wm #Fixed regression of dCOG1 on post.wm
post.wm ~ 0*1 #Intercept of post.wm constrained to 0
post.wm ~~ 0*post.wm #Fixes variance of post.wm to 0
dCOG1 ~ 1 #Estimates intercept of change score
pre.wm ~ 1 #Estimates intercept of pre.wm
dCOG1 ~~ dCOG1 #Estimates variance of change score
pre.wm ~~ pre.wm #Estimates variance of pre.wm
dCOG1~pre.wm #Estimates the self-feedback parameter
RSpan1~~RSpan2 #Allows residual covariance on RSpan across T1 and T2
SSpan1~~SSpan2 #Allows residual covariance on RSpan across T1 and T2
RSpan1~~a*RSpan1 #Allows residual variance on RSpan at T1
SSpan1~~b*SSpan1 #Allows residual variance on SSpan at T1
RSpan2~~a*RSpan2 #Allows residual variance on RSpan at T2
SSpan2~~b*SSpan2 #Allows residual variance on SSpan at T2
RSpan1~0*1 #Constrains intercept of RSpan to 0 at T1
SSpan1~c*1 #Estimates the intercept of SSpan at T1
RSpan2~0*1 #Constrains the intercept of RSpan to 0 at T2
SSpan2~c*1 #Estimates the intercept of SSpan at T2
'
fitmod2 <- lavaan(mod2, data=all,group='Group' , estimator='mlr',fixed.x=T,missing='fiml') #fixed.x=T regards all exogenuous covariates as estimated
summary(fitmod2, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)
Warning message:
In lav_object_post_check(object) :
lavaan WARNING: some estimated lv variances are negative
> summary(fitmod2, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)
lavaan 0.6-3 ended normally after 34 iterations
Optimization method NLMINB
Number of free parameters 39
Number of equality constraints 3
Number of observations per group
1 70
2 78
5 99
Number of missing patterns per group
1 2
2 2
5 2
Estimator ML Robust
Model Fit Test Statistic 7.960 7.526
Degrees of freedom 6 6
P-value (Chi-square) 0.241 0.275
Scaling correction factor 1.058
for the Yuan-Bentler correction (Mplus variant)
Chi-square for each group:
1 5.267 4.980
2 1.443 1.364
5 1.250 1.181
Model test baseline model:
Minimum Function Test Statistic 271.474 234.809
Degrees of freedom 18 18
P-value 0.000 0.000
User model versus baseline model:
Comparative Fit Index (CFI) 0.992 0.993
Tucker-Lewis Index (TLI) 0.977 0.979
Robust Comparative Fit Index (CFI) 0.994
Robust Tucker-Lewis Index (TLI) 0.981
Loglikelihood and Information Criteria:
Loglikelihood user model (H0) -1151.184 -1151.184
Scaling correction factor 0.967
for the MLR correction
Loglikelihood unrestricted model (H1) -1147.204 -1147.204
Scaling correction factor 1.049
for the MLR correction
Number of free parameters 36 36
Akaike (AIC) 2374.369 2374.369
Bayesian (BIC) 2500.707 2500.707
Sample-size adjusted Bayesian (BIC) 2386.587 2386.587
Root Mean Square Error of Approximation:
RMSEA 0.063 0.056
90 Percent Confidence Interval 0.000 0.166 0.000 0.158
P-value RMSEA <= 0.05 0.361 0.400
Robust RMSEA 0.057
90 Percent Confidence Interval 0.000 0.166
Standardized Root Mean Square Residual:
SRMR 0.043 0.043
Parameter Estimates:
Information Observed
Observed information based on Hessian
Standard Errors Robust.huber.white
Group 1 [1]:
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
pre.wm =~
RSpan1 1.000 0.720 0.694
SSpan1 1.000 0.720 0.777
post.wm =~
RSpan2 1.000 0.782 0.723
SSpan2 1.000 0.782 0.801
dCOG1 =~
post.wm 1.000 0.539 0.539
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
post.wm ~
pre.wm 1.000 0.921 0.921
dCOG1 ~
pre.wm -0.082 0.120 -0.679 0.497 -0.140 -0.140
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.RSpan1 ~~
.RSpan2 0.252 0.116 2.164 0.030 0.252 0.449
.SSpan1 ~~
.SSpan2 0.038 0.101 0.379 0.705 0.038 0.112
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 0.153 0.082 1.877 0.061 0.364 0.364
pre.wm 0.003 0.117 0.023 0.982 0.004 0.004
.RSpan1 0.000 0.000 0.000
.SSpan1 (c) 0.063 0.092 0.680 0.496 0.063 0.068
.RSpan2 0.000 0.000 0.000
.SSpan2 (c) 0.063 0.092 0.680 0.496 0.063 0.064
Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 0.174 0.088 1.971 0.049 0.981 0.981
pre.wm 0.519 0.109 4.775 0.000 1.000 1.000
.RSpan1 (a) 0.560 0.135 4.157 0.000 0.560 0.519
.SSpan1 (b) 0.341 0.093 3.676 0.000 0.341 0.397
.RSpan2 (a) 0.560 0.135 4.157 0.000 0.560 0.478
.SSpan2 (b) 0.341 0.093 3.676 0.000 0.341 0.358
R-Square:
Estimate
post.wm 1.000
dCOG1 0.019
RSpan1 0.481
SSpan1 0.603
RSpan2 0.522
SSpan2 0.642
Group 2 [2]:
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
pre.wm =~
RSpan1 1.000 0.716 0.723
SSpan1 1.000 0.716 0.677
post.wm =~
RSpan2 1.000 0.607 0.620
SSpan2 1.000 0.607 0.681
dCOG1 =~
post.wm 1.000 NaN NaN
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
post.wm ~
pre.wm 1.000 1.178 1.178
dCOG1 ~
pre.wm -0.062 0.155 -0.403 0.687 NaN NaN
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.RSpan1 ~~
.RSpan2 -0.049 0.099 -0.492 0.623 -0.049 -0.093
.SSpan1 ~~
.SSpan2 0.067 0.083 0.810 0.418 0.067 0.132
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 -0.111 0.118 -0.935 0.350 NaN NaN
pre.wm 0.126 0.116 1.088 0.277 0.177 0.177
.RSpan1 0.000 0.000 0.000
.SSpan1 0.028 0.118 0.235 0.814 0.028 0.026
.RSpan2 0.000 0.000 0.000
.SSpan2 -0.016 0.115 -0.141 0.888 -0.016 -0.018
Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 -0.081 0.100 -0.812 0.417 NaN NaN
pre.wm 0.512 0.130 3.936 0.000 1.000 1.000
.RSpan1 0.467 0.139 3.367 0.001 0.467 0.477
.SSpan1 0.604 0.121 4.977 0.000 0.604 0.541
.RSpan2 0.590 0.136 4.339 0.000 0.590 0.615
.SSpan2 0.426 0.086 4.963 0.000 0.426 0.536
R-Square:
Estimate
post.wm 1.000
dCOG1 NaN
RSpan1 0.523
SSpan1 0.459
RSpan2 0.385
SSpan2 0.464
Group 3 [5]:
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
pre.wm =~
RSpan1 1.000 0.682 0.710
SSpan1 1.000 0.682 0.665
post.wm =~
RSpan2 1.000 0.624 0.695
SSpan2 1.000 0.624 0.595
dCOG1 =~
post.wm 1.000 0.336 0.336
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
post.wm ~
pre.wm 1.000 1.094 1.094
dCOG1 ~
pre.wm -0.129 0.231 -0.560 0.576 -0.422 -0.422
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.RSpan1 ~~
.RSpan2 -0.043 0.119 -0.359 0.720 -0.043 -0.098
.SSpan1 ~~
.SSpan2 0.215 0.116 1.856 0.063 0.215 0.334
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 -0.136 0.113 -1.204 0.228 -0.649 -0.649
pre.wm -0.067 0.098 -0.686 0.493 -0.099 -0.099
.RSpan1 0.000 0.000 0.000
.SSpan1 -0.092 0.102 -0.902 0.367 -0.092 -0.089
.RSpan2 0.000 0.000 0.000
.SSpan2 -0.003 0.132 -0.019 0.984 -0.003 -0.002
Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.post.wm 0.000 0.000 0.000
.dCOG1 0.036 0.115 0.313 0.754 0.822 0.822
pre.wm 0.466 0.110 4.237 0.000 1.000 1.000
.RSpan1 0.458 0.127 3.615 0.000 0.458 0.496
.SSpan1 0.586 0.106 5.531 0.000 0.586 0.557
.RSpan2 0.417 0.120 3.464 0.001 0.417 0.517
.SSpan2 0.711 0.137 5.194 0.000 0.711 0.646
R-Square:
Estimate
post.wm 1.000
dCOG1 0.178
RSpan1 0.504
SSpan1 0.443
RSpan2 0.483
SSpan2 0.354
Terrence Jorgensen
Mar 6, 2019, 2:26:15 PM3/6/19
to lavaan
I have updated the code as you suggested
No, you are still only providing one label, instead of a vector of labels. For instance, this equality constraint only applies to the first group (as you can verify by paying attention to the labels in your summary() output for group 1 vs. 2 and 3).
RSpan1~~a*RSpan1 #Allows residual variance on RSpan at T1
RSpan2~~a*RSpan2 #Allows residual variance on RSpan at T2 You have 3 groups, so use a vector, as shown in the tutorial page I linked to before
RSpan1~~c(a, a, a)*RSpan1 #Allows residual variance on RSpan at T1
RSpan2~~c(a, a, a)*RSpan2 #Allows residual variance on RSpan at T2
Perhaps your problem will not occur when you fit the model you mean to.
I am not sure what you mean by only the 1st group has any df.
In your original post, the chi-squared contribution was zero in groups 2 and 3, which means those models fit perfectly (which typically only occurs when you have df=0)
Is it critical that I don't have a negative variance before proceeding to use the model results?
No, not if the CI includes positive values
summary(fit, ci = TRUE)
Reply all
Reply to author
Forward
0 new messages