UMD Theses and Dissertations

Permanent URI for this collectionhttp://hdl.handle.net/1903/3

New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a given thesis/dissertation in DRUM.

More information is available at Theses and Dissertations at University of Maryland Libraries.

Browse

Filters

2008 - 200912010 - 20141

Settings

Author: search.filters.author.Weng, Chin-FangSubject: search.filters.subject.Statistics

Search Results

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    FIRST ORDER AUTOREGRESSIVE MIXED EFFECTS ZERO INFLATED POISSON MODEL FOR LONGITUDINAL DATA - A BAYESIAN APPROACH
    (2014) Weng, Chin-Fang; Mislevy, Robert J; Measurement, Statistics and Evaluation; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    The First Order Autoregressive (AR(1)) Mixed Effects Zero Inflated Poisson (ZIP) Model was developed to analyze longitudinal zero inflated Poisson data through the Bayesian Approach. The model describes the effect of covariates via regression and time varying correlations within subject. Subjects are classified into a "perfect" state with response equal to zero and a Poisson state with response following a Poisson regression model. The probability of belonging to the perfect state or Poisson state is governed by a logistic regression model. Both models include autocorrelated random effects, and there is correlation between random effects in the logistic and Poisson regressions. Parameter estimation is investigated using simulation studies and analyses (both frequentist and Bayesian) of simpler mixed effect models. In the large sample setting we investigate the Fisher information of the model. The Fisher information matrix is then used to determine an adequate sample size for the AR(1) ZIP model. Simulation studies demonstrate the capability of Bayesian methods to estimate the parameters of the AR(1) ZIP model for longitudinal zero inflated Poisson data. However, a tremendous computation time and a huge sample size are required by the full AR(1) ZIP model. Simpler models were fitted to simulated AR(1) ZIP data to investigate whether simplifying the assumed random structure could permit accurate estimates of fixed effect parameters. However, simulations showed that the bias of two nested models, ZIP model and mixed effects ZIP model, are too large to be acceptable. The AR(1) ZIP model was fitted to data on numbers of cigarettes smoked, collected in the National Longitudinal Study of Youth. It was found that decisions on whether to smoke and on the number of cigarettes to smoke were significantly related to age, sex, race and smoking behavior by peers. The random effect variances, autocorrelation coefficients and correlation between logistic and Poisson random effect were all significant.
  • Thumbnail Image
    Item
    Fixed versus Mixed Parameterization in Logistic Regression Models: Application to Meta-Analysis
    (2008) Weng, Chin-Fang; Slud, Eric V.; Mathematical Statistics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    Three methods: fixed intercept generalized model (GLM), random intercept generalized mixed model (GLMM), and conditional logistic regression (clogit) are compared in a meta-analysis of 43 studies assessing the effect of diet on cancer incidence in rats. We also perform simulation studies to assess distributional behavior of regression estimates and tests of fit. Other simulations assess the effects of model misspecification, and increasing the sample size, either by adding additional studies or by increasing the sizes of a fixed number of studies. Estimates of fixed effects seem insensitive to increasing the sample sizes, but the deviance test of fit is biased. Conditional logistic regression avoids the possibility of bias when the number of studies is very large in a GLM analysis and also avoids effects of misspecification of the random effect distribution in a GLMM analysis, but at the cost of some information loss.