Partial Least Square and Bootstrapping: The Impact of Missing Data
A common problem in model fitting is missing data. a number of techniques have been developed for dealing with this problem. Most use indirect approaches in which missing values are replaced by a form of imputation. These techniques can also be applied when the bootstrap method is used to estimate the standard error of some estimator. Although booststrapping is used extensively, limited research has been conducted into the possible effects of missing values being present. To address this, we undertake a Monte Carlo simulation study of the impact of incomplete data on bootstrap parameter estimation for the Partial Least Square (PLS) component based approach to structural equation modelling (SEM). A published Customer Satisfaction dataset is used to provide distribution free, empirical data, and simulated incomplete data of varying sample sizes with various levels of missing data are taken from this. Five competing missing data methods are then applied to these data and the PLS results are compared. Overall it was concluded that Multiple Imputation was the best missing data technique for bootstrap estimation of path coefficients. Moreover, it was found that 500 bootstrap samples are sufficient to provide accurate estimates for a range (5%-15%) of missing completely at random (MCAR) data.
partial least squares structural equation modelling missing data Bootstrapping
Jastini Mohd Jamil James Wallace Reza Abdi
School of Management, Bradford University, Bradford BD9 4JL, UK
国际会议
The 6th International Conference on Partial Least Squares and Related Methods(第六届偏最小二乘及相关方法国际会议)
北京
英文
189-193
2009-09-04（万方平台首次上网日期，不代表论文的发表时间）