HI Linda and/or Bengt, I am putting in a NIMH application to renew our longitudinal study of risk for anxiety disorders and depression in 600 high school students. As per an earlier post on this discussion board, we will propose to use FIML to accomodate missing data in our latent growth curve models. I am wondering, though, what implications the FIML approach to missing data has for power analysis? I have been relying on Table 2 from MacCallum, Browne & Sugawara (1996) but am a bit uncertain as to how I should think about my sample size. I am assuming the sample size values given in their Table are based on having complete data and if we anticipate having complete data from say 300 of our participants should I be looking in their column corresponding to a sample size of 300 or 500 (500 is their largest value)? Or is there reason to believe that power will lie somewhere between that associated with a n of 300 and that associated with a n of 600? Thanks in advance!
Power is strongly reduced by missing data. Rules of thumb are not possible given dependence on so many factors. It is probably better to do a Monte Carlo simulation. Look at UG example 11.2 for how to do this in the context of growth modeling with attrition. See also
Muthén, L.K. & Muthén, B.O. (2002). How to use a Monte Carlo study to decide on sample size and determine power. Structural Equation Modeling, 4, 599-620.
which is in pdf on our web site under Papers, Miscellaneous.