How do I decide what the right starting values are?

When using the raw data option in Mx, you need to provide starting values for each free element in your model. The default starting value is zero, which usually lies too far away from the real values and may lead to optimization problems in Mx. Matrices containing estimates of means, or effects of covariates can be started at the value that is calculated in a program such as SPSS, SAS or Excel.

As a rule of thumb for choosing starting values for matrices containing standard deviations in a variance component model, one calculates the observed variance ('var') in a program such as SPSS, SAS or Excel, divides this value by the number of variance components ('n') in the vc model, and then takes the square root: starting value sd = SQRT (var/n).

For example, the observed variance of HEIGHT is 90. We want to fit a variance decomposition model including three (A, C, and E) sources of variance (90/3 = 30). Since we use a Cholesky decomposition A, C, and E are modeled as X*X', Y*Y', and Z*Z' respectively. So we need the square root of 30 (5.5). So starting values for matrices X, Y, and Z are 5.5.