MISSING DATA IN LINEAR MODELS WITH CONSTANT INTERCLASS CORRELATION AND RANDOM BLOCK DESIGNS WITH BINARY DATA

BARRY KURT MOSER, Purdue University

Abstract

This paper covers two distinct topics: Missing data in linear models with constant interclass correlation and random complete block designs with binary data. The missing data question is addressed in Chapter 1 with an example problem solved in Chapter 2. Two models are examined in Chapter 3 for the random block design problem with example problems run in Chapter 4. An analysis procedure for linear models with constant interclass correlation and missing data is developed in Chapter 1. Estimates of the missing data and the unknown parameters are given under mild side conditions. The estimators of the missing values have the same expectation as the missing data. Those for the unknown parameters are maximum likelihood estimators. A six step procedure for calculating the estimates is given in Chapter 2. The computer program used to run the calculations is listed in Appendix 1. Two analysis procedures for random block designs with binary data are described in Chapter 3. In the first procedure maximum likelihood estimators (M.L.E.) of the unknown parameters are derived for a logit random block effect model. Likelihood ratio tests and the asymptotic distribution of the M.L.E.s are derived. In the second procedure ordinary least squares estimators (O.L.S.E.) of the unknown parameters are developed for a linear form of the logit random block effect model. These O.L.S.E. are consistent, unbiased, and equal to the generalized least squares estimators. The numerical calculations for these two procedures are explained in Chapter 4 with two example problems. The computer program which performs the maximum likelihood estimation is listed in Appendix 2.

Degree

Ph.D.

Subject Area

Statistics

COinS