This question reflects the dilemma I face in analyzing path models using AMOS with missing data.
I have a dataset with n = 116, missing rate = 6% (Missing at Random) for a full path model with about 20 parameters. My analysis is performed using SPSS (with the imputation add-on) and AMOS 20.
Given the small sample size and the many parameters to estimate, I'm using the bootstrap option in AMOS (pls let me know if this is not a good idea). However, AMOS doesn't allow bootstrap if the dataset has missing values. Which of the following options makes the most sense?
Don't do anything to the missing data. Use maximum likelihood in the AMOS. Give up the bootstrap option and the estimation of modification indices.
Use data imputation that yields one complete dataset: 1) Expectation Maximization in SPSS; or 2) Single imputation with AMOS data imputation (with regression/stochastic regression/Bayesian estimate options).
Perform multiple imputation in SPSS. But the question is how one combines the multiple datasets to one dataset without reducing the power of multiple imputation. Would it be acceptable to take average of all imputated values to create one complete dataset for the subsequent path analysis in AMOS?