Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form

An Index of Local Sensitivity to Nonignorability and a Penalized Pseudolikelihood Method for Data with Nonignorable Nonresponse

Zhu, Fang (2008) An Index of Local Sensitivity to Nonignorability and a Penalized Pseudolikelihood Method for Data with Nonignorable Nonresponse. Doctoral Dissertation, University of Pittsburgh. (Unpublished)

Primary Text

Download (830kB) | Preview


The public health significance of this study is to provide researchers and practitioners more improved methods to analyze data with missing values as such data get prevalent in practice. When data are missing at random (MAR), the missing-data mechanism can be ignored. Otherwise, the mechanism needs to be modeled. Further sensitivity analyses are often necessary to evaluate the impact of alternative mechanism assumptions on the inferences. For data with nonignorable nonresponse, a pseudolikelihood method was developed, where specification of the mechanism is not necessary. A sensitivity analysis for this method and extensions to nonparametric and semi-parametric regression models were proposed in this thesis.An index of local sensitivity to nonignorability for the maximum likelihood method (ISNIML) for data with missing outcome values where the missing-data mechanism was modeled by a logistic regression was developed. It is used to evaluate how a small deviation from MAR affects the maximum likelihood estimate. A new index of local sensitivity to nonignorability (ISNIPL) was proposed for this pseudolikelihood method in this thesis. Compared with ISNIML, it has the advantage that functional specification of the missing-data mechanism is not required. Depending on whether or not the distribution of the covariate can be parametrically modeled, two versions of this ISNIPL were derived. Simulations suggested that ISNIPL is very close to ISNIML when the likelihood is correctly specified by the latter. But it does not require assumption on the function form of the missing-data mechanism. The analysis of a real dataset was used to highlight their differences and utility.In the second part, a penalized pseudolikelihood (PPL) method was developed for semi-parametric regression models with the following form: y = xâ + g(t) + error, where g is an unspecified function and can be estimated by a natural cubic spline, for data with nonignorable nonresponse. Two cross-validation methods were considered to find the optimal smoothing parameter. Simulations suggested that PPL with the traditional cross-validation method yields less biased estimates of the parameter of interest and the nonparametric function. This PPL method was also illustrated in analysis of a clinical dataset.


Social Networking:
Share |


Item Type: University of Pittsburgh ETD
Status: Unpublished
CreatorsEmailPitt UsernameORCID
ETD Committee:
TitleMemberEmail AddressPitt UsernameORCID
Committee ChairTang, Gonggot1@pitt.eduGOT1
Committee MemberWahed, Abdus Swahed@pitt.eduWAHED
Committee MemberRollman, Bruce LRollmanBL@upmc.eduBRR1
Committee MemberChang, Chung-Chou
Committee MemberMazumdar, Satimaz1@pitt.eduMAZ1
Date: 28 September 2008
Date Type: Completion
Defense Date: 10 June 2008
Approval Date: 28 September 2008
Submission Date: 1 August 2008
Access Restriction: No restriction; Release the ETD for access worldwide immediately.
Institution: University of Pittsburgh
Schools and Programs: School of Public Health > Biostatistics
Degree: PhD - Doctor of Philosophy
Thesis Type: Doctoral Dissertation
Refereed: Yes
Uncontrolled Keywords: nonignorability; Miss at random; pseudolikelihood
Other ID:, etd-08012008-115251
Date Deposited: 10 Nov 2011 19:55
Last Modified: 19 Dec 2016 14:36


Monthly Views for the past 3 years

Plum Analytics

Actions (login required)

View Item View Item