Effects of missing data on statistical power to detect change in family-based preventive intervention research
Savla, Jyoti S
MetadataShow full item record
Loss of statistical power in family-based preventive intervention research often results from missing data. This dissertation examines interactions between type and amount of missing data, sample size and statistical power to detect group differences in longitudinal change. Study 1, examines the statistical power of a 5-wave two-group (Treatment and Control) growth curve model with a small (.2) effect size as a function of missing data (5% to 95% missing), type of missingness (Missing Completely at Random, MCAR, or Missing at Random, MAR), and sample sizes (N=100 to N=1000). Results indicate that for moderate samples and amounts of missing data loss in power is proportionally higher. Interactions between sample size, statistical power and missing data are further considered. Study 2 considers the effects of measurement and design decisions to improve and moderate the effects of missing data on statistical power in the two-group growth curve model as a function of varying amounts missing data and sample sizes. When reliability of indicators is high, the effect of missing data was found to diminish. Furthermore, inclusion of an auxiliary variable suggested that a covariate is particularly beneficial in increasing statistical power at smaller sample sizes and with models with high percentage of missing data. Lastly, drop-out patterns were evaluated as a function of type of missing data (MCAR vs. MAR) with 50% missing data at different sample sizes. Results indicate that the pattern in which participants dropped out of a study affected power even for identical amounts of missing data. Methods to increase the statistical power in the presence of varying amount of missing data and sample size are recommended, with suggestions for future research.