Yi-Fen Tseng posted on Wednesday, September 29, 2010 - 3:03 pm
A proper fit for a measurement portion with four correlated errors is first obtained:
LAS BY few s2cat twopw4; PI BY adult fedu fwork private; MD4 BY md1w4 md2w4 md3w4 md4w4; PD4 BY pd1w4 pd2w4 pd3w4 pd4w4; ND4 BY nd1w4 nd2w4 nd3w4 nd4w4; CD BY ap wd ad as dq; few WITH s2cat; ND4w4 WITH ND2w4; AS WITH WD; DQ WITH AD;
Then the path portion was followed: MD4 ON LAS PI; PD4 ON LAS PI; ND4 ON LAS PI; CD ON LAS PI MD4 PD4 ND4;
The message indicates "no convergence, number of iterations exceeded." Is the problem in the default number of iterations? Something else? Could you please advice? Thank you very much.
Yi-Fen Tseng posted on Wednesday, September 29, 2010 - 11:23 pm
In the modification, number of iterations is increased to 10000 and it then can converge. However, a warning message appears and the S.E. cannot be computed.
Some variables are found to be ill-scaled because the differences in variances between variables are huge. Will you recommend to rescale the variables and try again? Do you have other reminder? Thank you very much.
Yes, you should rescale the variables using the DEFINE command. Divide each variable by a constant that brings the variances between one and ten.
Yi-Fen Tseng posted on Thursday, September 30, 2010 - 10:09 pm
Tech4 shows that most of the 20 estimated covariance for abovementioned latent variables are within the absolute value of 1, except for 3 of them. The covariance between PD4 and PD4=32.583; PD4 and ND4=28.602; ND4 and ND4=48.483.
Is this a sign of ill-scaled problem or linear dependency among latent variables or related to both? Based on the tech4 information, should I build up the path model in small pieces first in order to see what latent variable brings in the problem or rescale the variables first? Thank you for your advice.
Yet to rescale the variables, SEM models for one latent variable a time with that of the main endogenous variable were run for each path. Problems mainly appear to be two types: (1) residual covariance is not positive definite involving one specific indicator; (2) one indicator measuring one factor is found to measure another factor. Should I solve these issues before rescaling the variables? Thank you so much.