When you have longitudinal data and doing growth modeling, do you ususally standardize your dependent variables (continuous variable, for example, measurement in height) first or you use the raw measurement directly? I thought Willett (1997, p217) discouraged using standardized scores in growth modeling; but after reading it more carefully, I am not sure whether he discouraged the use of standardized measurement only in the "difference score approach" or in the more general growth modeling on longitudinal data. Could you please provide some comments as well as references on this topic? Thanks!
Willett, J.B. 1997. "Measuring Change: What Individual Growth Modeling Buys You." Pp. 213-243 in Change and Development: Issues of Theory, Method, and Application, edited by E. Amsel and K.A. Renninger. Mahwah, NJ: Lawrence Erlbaum Associates.
bmuthen posted on Saturday, May 07, 2005 - 11:46 am
Using raw measurements is the way to go - you want to see how individuals (and time-specific means and variances) change over time for an outcome that has a constant metric. For a good discussion in an education context, see the Seltzer et al article "The Metric Matters". His email is firstname.lastname@example.org
John Chen posted on Saturday, May 07, 2005 - 2:15 pm
I am running a two-part model(because the dependenet variable has preponderous of zeros), using mlr estimator. Uaually it took one and a half hours to converge. I also notice there are QNs after EMs in the Tech8 showing on the computer screen. It is a problem? What shall I do so the program may take less time to converge?
bmuthen posted on Saturday, May 07, 2005 - 4:40 pm
QN is not a problem, but merely indicates that the optimization engine finds it advantageous to switch from EM to Quasi-Newton iterations.
The computations take longer when you have many dimensions of integration and a large sample - the number of dimensions is shown in the output and is for this model a function of how you specify the binary part of the model. Here, 3 is already a large number. If, for example, you have a quadratic function, you can fix the quadratic growth factor variance at zero and get only 2 dimensions of integration. You may also try getting an approximate ML solution by changing from the default of 15 integration points per dimension to say 10 or even 5. Such an approximate approach can be useful while you are exploring different models. You can then rerun
Hi Dr. Muthen, I have five measures (Y1 – Y5) of a latent construct (L) all measured on a different scale, and all assessed annually for 4 years. I wish to do latent growth curve modeling (LGCM) with the 4 time point data on (Y1 – Y5) to estimate the intercept and slope of the latent construct (L). Since the five Y's are on different scales, I should not average them without first standardizing them, and further since I intend to do LGCM I don't think its appropriate to use averages of standardized scores. Would it be okay to take averages of unstandardized Y1 – Y5 scores for each wave and then use these average scores to create intercepts and slopes for L?