IBM® SPSS® Amos™ 28

In fitting a structural equation model, you have to impose constraints on the model so as to fix the unit of measurement of each unobserved variable. If you are planning to use the Bootstrap method, you should fix the scales of the unobserved variables by placing appropriate constraints on the regression weights, and not by constraining the variances of the unobserved variables. This method for fixing units of measurement is necessary for the following reason: If the scales of measurement of the unobserved variables are fixed by constraining their variances, the criterion of minimizing the discrepancy function will determine some of the regression weights only up to a sign change. That is, given one set of parameter estimates, it will be possible to change the signs of some of the regression weights without affecting the fit of the model. This is actually an example of nonidentifiability and also an example of multiple local minima, but it is a benign example unless you are bootstrapping. In bootstrapping, if the signs of some regression weights are arbitrary, their estimates will tend to 'jump around' from one bootstrap replication to another, and the reported bootstrap standard errors will be artificially inflated as a result.