國立高雄大學圖資館 |

Language: English

Back

Marginally specified conditional models for longitudinal outcomes with possibly non-ignorable non-response.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Marginally specified conditional models for longitudinal outcomes with possibly non-ignorable non-response.
Author:	Wilkins, Kenneth Joseph.
Description:	109 p.
Notes:	Adviser: Garrett Martin Fitzmaurice.
Notes:	Source: Dissertation Abstracts International, Volume: 65-05, Section: B, page: 2176.
Contained By:	Dissertation Abstracts International65-05B.
Subject:	Biology, Biostatistics.
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3132025
ISBN:	0496792652

Marginally specified conditional models for longitudinal outcomes with possibly non-ignorable non-response.
Wilkins, Kenneth Joseph.

Marginally specified conditional models for longitudinal outcomes with possibly non-ignorable non-response.[electronic resource] - 109 p.

Adviser: Garrett Martin Fitzmaurice.

Thesis (Ph.D.)--Harvard University, 2004.

Despite efforts at minimizing the occurrence of missing values within many longitudinal studies, they remain a frequent hindrance to analyzing change in the targeted outcome over time. In particular, when the probability of non-response is thought to depend in some way on unobserved values, the non-response process is deemed non-ignorable and, to curtail potential bias, joint modeling of outcomes and variables indicating non-response is required. An unavoidable aspect of such joint modeling is the need to adopt supplemental assumptions, unverifiable from the data at hand, concerning the dependence of the probability of non-response on unobserved outcomes; one must explore questions of interest under a plausible range of such assumptions. Two overall approaches to joint modelling are selection models and pattern-mixture models, with respective formulations that result in distinct advantages and disadvantages. This work focuses on developing alternative approaches that capitalize on the advantages, by means of marginally-specified conditional models.

ISBN: 0496792652Subjects--Topical Terms:

227395
Biology, Biostatistics.

Marginally specified conditional models for longitudinal outcomes with possibly non-ignorable non-response.
LDR:03545nmm _2200289 _450 001 162721
005 20051017073516.5
008 230530s2004 eng d
020 $a 0496792652
035 $a 00149222
035 $a 162721
040 $a UnM $c UnM
100 0 $a Wilkins, Kenneth Joseph. $3 227865
245 1 0 $a Marginally specified conditional models for longitudinal outcomes with possibly non-ignorable non-response. $h [electronic resource]
300 $a 109 p.
500 $a Adviser: Garrett Martin Fitzmaurice.
500 $a Source: Dissertation Abstracts International, Volume: 65-05, Section: B, page: 2176.
502 $a Thesis (Ph.D.)--Harvard University, 2004.
520 # $a Despite efforts at minimizing the occurrence of missing values within many longitudinal studies, they remain a frequent hindrance to analyzing change in the targeted outcome over time. In particular, when the probability of non-response is thought to depend in some way on unobserved values, the non-response process is deemed non-ignorable and, to curtail potential bias, joint modeling of outcomes and variables indicating non-response is required. An unavoidable aspect of such joint modeling is the need to adopt supplemental assumptions, unverifiable from the data at hand, concerning the dependence of the probability of non-response on unobserved outcomes; one must explore questions of interest under a plausible range of such assumptions. Two overall approaches to joint modelling are selection models and pattern-mixture models, with respective formulations that result in distinct advantages and disadvantages. This work focuses on developing alternative approaches that capitalize on the advantages, by means of marginally-specified conditional models.
520 # $a Selection models holds an advantage by directly parameterizing the marginal mean responses. In contrast, mean regression parameters within a pattern-mixture model are conditional on non-response pattern, lacking the desired interpretation. Pattern-mixture models do hold advantage, however, in adopting supplemental assumptions of non-response within their formulation, with a transparent method for ensuring that the resulting model is identified. This work develops a likelihood-based model for longitudinal binary responses with possibly non-ignorable dropout that capitalizes on each of these advantages, so is denoted a 'hybrid' model. Parameterization and estimation within this framework, however, can be burdensome for certain applications. We thus propose a marginally-specified pattern-mixture model that accommodates both monotone and non-monotone missingness in binary outcomes; it directly parameterizes change in the marginal mean in terms of regression parameters while incorporating assumptions of non-response via a model for the conditional mean given missingness pattern, constrained accordingly such that the model is identified. Estimation of parameters follows by solving modified generalized estimating equations. We then extend the generalized linear model formulation of this framework for application to diverse types of response variables. Illustrative examples accompany each formulation.
590 $a School code: 0084.
650 # 0 $a Biology, Biostatistics. $3 227395
650 # 0 $a Statistics. $3 182057
690 $a 0308
690 $a 0463
710 0 # $a Harvard University. $3 212445
773 0 # $g 65-05B. $t Dissertation Abstracts International
790 $a 0084
790 1 0 $a Fitzmaurice, Garrett Martin, $e advisor
791 $a Ph.D.
792 $a 2004
856 4 0 $u http://libsw.nuk.edu.tw/login?url=http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3132025 $z http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3132025