| Summary: | 博士 === 國立臺灣大學 === 公共衛生學系 === 85 === In the applied regression analysis, data on some components of
the regressors may be incomplete for parts of
study subjects due to missingness or measurement error.
Discarding the incomplete observations may lead to
severe bias and efficiency loss for the estimation of the
parameters. Two new classes of estimating functions
which incorporate both the complete and incomplete data to
estimate the regression coefficients are proposed in this
dissertation: one is a weighted estimating function which
combines complete and incomplete data with a weight
accounting for heteroscedastic variations from the two
sources of data, and the other is a simultaneous
estimating function which estimates all the relevant parameters
simultaneously. Both the two proposed estimating functions
yield consistent estimates for regression coefficients
without the need to specify a correct model for the
incomplete covariates, and do not suffer the ''curse of
dimensionality'' encountered in the existing
nonparametric methods. Simulation study reveal that
the efficiency property of the proposed estimators is
satisfactory. In particular, the proposed methodologies are
computationally convenient and may be widely practicable in
daily data analysis.
|
|---|
