mike zyphur posted on Friday, February 03, 2006 - 5:31 pm
Hi Bengt/Linda, I tried to find this issue addressed elsewhere on the message board, but couldn't find it. Perhaps you would be kind enough to address it here:
In an LCA with multiple indicators with variances that are VERY different (e.g., age in months and 5-point likert-type scale responses), to what extent will first standardizing the variables allow each variable to be treated equally in the process of class formation? While the point of the LCA is to make the variables independent within each class, how strongly influenced is class formation by variables which have much larger variances than the other variables?
If it is greatly influenced, what does this mean for LCAs which mix continuous and binary indicators? While the variances of continuous variables can be easily changed by, for example, standardization, clearly this is not true for binary indicators. Does this pose problems for LCAs which have both continuous and binary indicators? Is what I'm bringing up even an issue?
Thank you for your time!
bmuthen posted on Friday, February 03, 2006 - 5:39 pm
Unlike k-means clustering, LCA is not hurt by variances being different across variables. The model allows them to be different. The key to success in LCA is when variances - and most importantly means - vary across classes for a given variable.
If the variances are very different across variables, however, you may want to scale by 10 or 100 say, to make it numerically easier to do the analysis. But don't standardize.