Comparison of complete-case, pairwise available-case, and maximum likelihood missing data methods
Date of Award
Doctor of Philosophy (PhD)
Monte Carlo, Psychological tests, Statistics
Complete-case (CC), pairwise available-case (PW), and maximum likelihood (ML) missing data methods were compared in a Monte Carlo experiment. A multivariate normal distribution, observations missing at random, and a monotone pattern of missingness was assumed. There were three blocks of variables that form the monotone pattern of missing data. The parameters of interest were the mean vector, variance, and covariance.
The sample centroid and sample variance-covariance matrix were generated directly. An algorithm developed by Odell and Feivesan (1966) was used to generate the variance-covariance matrix. The mean vector was generated by SAS' RANNOR function.
The exogenous variables were: the number of variables within each block; correlations among the variables; initial block's sample size; and attrition percents. The average root mean square error (rmse) was the criterion used to analyze the results of the Monte Carlo experiment. A split plot factorial, SPF-2222-3, with four between factors and one within factor was conducted three times, once for each parameter.
The attrition percent was successful in discriminating among the missing data methods with the mean, and variance. The other exogenous variables did not discriminate among the missing data methods.
For the mean PW is recommended for the 20% attrition rate. ML is recommended as the first choice for the 40% attrition rate, followed by PW. For the variance, PW is recommended for both the 20% and 40% attrition rates.
There was no interaction between the missing data methods and any of the exogenous variables for the covariance.
Fraser, Christine Anne, "Comparison of complete-case, pairwise available-case, and maximum likelihood missing data methods" (1996). Psychology - Dissertations. Paper 104.