Date of Award
Master of Arts in Psychology
Wayne F. Velicer
The nature of time series designs often leads to missing data, due to its requirement of taking a large number of repeated observations on a single experimental unit. There are many methods available for estimating missing observations; some have been borrowed or adapted from univariate methods, while others have been developed specifically for the time series problem. This study compares the effects of four different methods of data estimation on time series analysis. These methods include: (1) deletion of the missing data points; (2) substitution of the mean of the series; (3) substitution of the mean of the observations adjacent to each missing data point; and (4) the use of a maximum likelihood algorithm to estimate the missing data points
Time series data were simulated for 50 different combinations of autocorrelation, slope, and proportion of data missing. Original series were 100 data points in length. Methods of data point estimation were compared in terms of the resulting time series analysis estimates of level, error variance, degree of autocorrelation, and slope in the series.
Major findings include: (1) the maximum likelihood approach is consistently accurate under all conditions tested; (2) the mean of the series is the least accurate approach overall; and (3) using the mean of the adjacent observations also has significant limitations. Results also indicate that conditions of severe negative autocorrelation in the time series lead to worse estimates of error variance, when using deletion or mean of the adjacent observations. Finally, series that have a non-zero slope result in less accurate parameter estimation.
Colby, Suzanne M., "Handling Missing Data in Time Series Analysis" (1992). Open Access Master's Theses. Paper 1571.