Anonymous posted on Monday, July 19, 2004 - 1:08 pm
I am having trouble with the estimation of a four-wave simplex model using categoric variables. I'm sorry to ask such a general question, but do I have the code right? When I run the same model treating the data as continuous, it works fine. I'm trying to sort out whether I have a problem with my data or my code. When I run the code below, I get the following error:
THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 12.
DATA: FILE IS ma1.dat; VARIABLE: NAMES ARE x1-x4; CATEGORICAL ARE x1-x4; MISSING ARE .; ANALYSIS: TYPE = H1 MISSING; PARAMETERIZATION=THETA; ESTIMATOR=WLSMV; MODEL: [x1$1 x2$1 x3$1 x4$1] (2); [x1$2 x2$2 x3$2 x4$2] (3); [x1$3 x2$3 x3$3 x4$3] (4); [x1$4 x2$4 x3$4 x4$4] (5); t1 BY x1; t2 BY x2; t3 BY x3; t4 BY x4; x1 (1); x2 (1); x3 (1); x4 (1);
You have four factors with single indicators. Just get rid of these and do your regressions using the x variables. With a single group analysis, the residual variances of the categorical variables are fixed to zero. They can only be part of a multiple group or growth model.
Anonymous posted on Tuesday, July 20, 2004 - 11:23 am
Actually, in my case, it is the partitioning of the variance that I am most interested in. I should have mentioned that x1-x4 are the same indicators measured longitudinally. I think that what is missing from my code is a scaling of the true-score variance. Adding the following line seems to yield sensible results:
Indeed, this is not a growth model, but a four wave autoregressive model. That is, I have a single item measured longitudinally in a panel study. I would like to estimate a model like Wiley & Wiley (1970) using categoric variables.
I have estimated such models using the THETA parameterization and constraining the thresholds equal across occasions of measurement.
I would then like to divide my sample into groups based on birth cohort and estimate the model again in a multiple group framework. This approach often fails. In attempting to estimate these models, I have constrained the thresholds to be equal both across occasions of measurement and across groups. Is this later constraint necessary? What I am really interested in is knowing if the ratios of true score variance to total variance (reliability) varies between birth cohorts. I think that the varying observed distributions between groups is the source of my difficulty.
In the multiple group framework, I attempt to estimate two models, 1) a fully-constrained model with all parameters equal across groups (H0), and 2) a model where the groups are estimated independently, save for the equal threshold constraint within and between groups (H1). It is this latter model that often fails. This failure also seems to be related to the number of response categories, such that dichotomies fail to yield a proper solution more often than do items with a larger number of observed levels.
I think the only way to understand what you are experiencing is to see the two ouptputs -- constrained and non-constrained. Can you send them along with your license number to firstname.lastname@example.org. There are many issues that could be coming into play.