|Institution:||The Ohio State University|
|Keywords:||Psychology; Pattern-Mixture Model; Longitudinal; Dropout; Missing Data; NMAR; Nonignorable Missingness|
|Full text PDF:||http://rave.ohiolink.edu/etdc/view?acc_num=osu150039066582153|
Missing data commonly arise during longitudinalmeasurements. Dropout is a particulartroublesome type ofmissingness because inference after the dropout occasioniseffectively precluded at the level of the individual withoutsubstantial assumptions.If missingness, such as dropout, is relatedto the unobserved outcome variables, thenparameter estimatesderived from models which ignore the missingness will be biased.Forexample, a treatment effect may appear less substantial ifpoor-performingsubjects tend to withdraw from the study. In ageneral sense, missing data leadto scenarios in which the empiricaldistribution of observed data is lacking nominalcoverage in someareas. Little (1993) proposed a general pattern-mixture modelapproachin which the moments of the full data distribution wereestimated as a finitemixture across the various missing-datapatterns. These models and their extensionsare flexible and may beestimated using wildly available mixed-modeling software insomespecial cases. The purpose of this work is to review the relevantmissing-dataliterature and to examine the viability ofrandom-coeffcient pattern-mixture modelsas an option for analystsseeking unbiased inference for longitudinal data subjecttopernicious dropout.Advisors/Committee Members: Cudeck, Robert (Advisor).