Dear all,
I am running an SEM in R. However, the model does not fit with reporting 'lavaan WARNING: some estimated lv variances are negative'. Any suggestion or solution?
I guess the problem might be the correlation between two variables (i.e. Land, Off). What do you think? Here is the output. Thank you.
> SEM #fitting SEM model > fit fit lavaan 0.6-5 ended normally after 77 iterations Estimator ML Optimization method NLMINB Number of free parameters 14 Number of observations 242 Model Test User Model: Test statistic 8.352 Degrees of freedom 7 P-value (Chi-square) 0.303
> summary(fit,standardized=TRUE) lavaan 0.6-5 ended normally after 77 iterations Estimator ML Optimization method NLMINB Number of free parameters 14 Number of observations 242 Model Test User Model: Test statistic 8.352 Degrees of freedom 7 P-value (Chi-square) 0.303 Parameter Estimates: Information Expected Information saturated (h1) model Structured Standard errors Standard Latent Variables: Estimate Std.Err z-value P(>|z|) Std.lv Std.all Land =~ L12 1.000 0.245 0.744 L11 -0.229 0.095 -2.398 0.016 -0.056 -0.252 Off =~ O11 1.000 NaN NaN O12 0.660 0.225 2.930 0.003 NaN NaN O13 0.647 0.223 2.897 0.004 NaN NaN Regressions: Estimate Std.Err z-value P(>|z|) Std.lv Std.all Y1 ~ Land -0.020 0.050 -0.396 0.692 -0.005 -0.032 Off -0.074 0.090 -0.820 0.412 NaN NaN Covariances: Estimate Std.Err z-value P(>|z|) Std.lv Std.all Land ~~ Off 0.020 0.004 4.637 0.000 0.929 0.929 Variances: Estimate Std.Err z-value P(>|z|) Std.lv Std.all .L12 0.049 0.024 1.996 0.046 0.049 0.447 .L11 0.047 0.004 10.547 0.000 0.047 0.937 .O11 0.050 0.006 7.880 0.000 0.050 1.174 .O12 0.041 0.004 9.503 0.000 0.041 1.085 .O13 0.044 0.005 9.666 0.000 0.044 1.076 .Y1 0.022 0.002 10.923 0.000 0.022 0.998 Land 0.060 0.026 2.354 0.019 1.000 1.000 Off -0.007 0.004 -1.990 0.047 NaN NaN Warning messages: 1: In sqrt(ETA2) : NaNs produced 2: In sqrt(ETA2) : NaNs produced 3: In sqrt(ETA2) : NaNs produced
Hello Qirui,
In my experience, the most likely cause for negative error variance estimates (so-called Heywood case) is a data set that has too much collinearity. You should have a look at collinearity statistics for the variables under investigation. There are other possible causes, one being when correlations/covariances were estimated using pairwise deletion (and there are considerable instances of missing values) such that the system of values is not coherent.
These (collinearity statistics) can usually be obtained by treating all the variables as IVs/predictors in a regression subprogram (be sure to add in some other variable as the DV!). This should help identify "troublesome" variables.
Cristian's suggestion of an alternate estimation algorithm can help you to get an answer, should you decide to forge ahead with your data as is.
Good luck with your work.
Reference:
Heywood, H. B. (1931). On finite sequences of real numbers. Proceedings of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character 134(824), 486–501. doi:10.1098/rspa.1931.0209
You can change the ML estimator to WLSMV, each factor must be put separately and then perform correlation of the factors if it is an oblique model.
Hello Qirui,
In my experience, the most likely cause for negative error variance estimates (so-called Heywood case) is a data set that has too much collinearity. You should have a look at collinearity statistics for the variables under investigation. There are other possible causes, one being when correlations/covariances were estimated using pairwise deletion (and there are considerable instances of missing values) such that the system of values is not coherent.
These (collinearity statistics) can usually be obtained by treating all the variables as IVs/predictors in a regression subprogram (be sure to add in some other variable as the DV!). This should help identify "troublesome" variables.
Cristian's suggestion of an alternate estimation algorithm can help you to get an answer, should you decide to forge ahead with your data as is.
Good luck with your work.
Reference:
Heywood, H. B. (1931). On finite sequences of real numbers. Proceedings of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character 134(824), 486–501. doi:10.1098/rspa.1931.0209
- Badges
- Science method