ABSTRACT
In many cases of multivariate data analysis, the values of the mean vector and the variance–covariance matrix or the correlation matrix are calculated first. Various methods of multivariate analysis are actually implemented based on these values. Broadly speaking, there are four methods of processing the missing values, as described below (Little and Rubin, 1987).
Exclude all incomplete observations including missing values. The sample size will be smaller than originally planned, however. Create a data matrix consisting of only complete observations and execute the ordinal estimation procedures.
Exclude only the missing parts and estimate the mean vector based on the arithmetic mean of the observed values that exist with respect to each variable. For the variance–covariance matrix and the correlation matrix, perform estimation by using all the observed values ofeach pair of variables, if any.
10Estimate the missing values to create a pseudocomplete data matrix while retaining the original sample size and execute normal estimation procedures.
Assuming an appropriate statistical model, estimate the maximum likelihood of necessary parameters based on the marginal likelihood with respect to all observed values that have been acquired.
