Longitudinal data analysis is a major component of public health care assessment. It is important to know how treatments compare over time, how diseases occurr and recurr, and how environmental or other exposures influence to a disease processes over time. Investigations of such topics involve the statistical analysis of time-to-event data in various areas of health care. Long term dental assessment of dental restorations have typically employed statistical analyses that assume independence of the restorations within the patient. Dental data naturally occur in the form of clusters. The patient is a cluster of correlated dental units (teeth) to be evaluated. Statistical analysis of the dental units without acknowledgement of within-cluster correlation can underestimate standard errors, which can erroneously inflate the significance level of between-cluster predictors in a model. The purpose of this thesis is to 1) review the statistical literature on the analysis of dental implant data, 2) create a suitable longitudinal data file of dental implant failure, 3) describe the data management and statistical methods used, 4) compare alternative statistical models to analyze clustered survival data, and 5) show how these models can be used to identify some patient-level and implant site-level predictors of implant failure. We consider logistic regression, discrete survival, generalized estimating equations and the Cox model with and without frailty, and examine the associations between implant failure and patient race, implant type, and oral location of implant. Models that ignore the clustering consistently overestimate the significance of patient race.

The Pennsylvania/MidAtlantic AIDS Education and Training Center (PA/MA AETC) is among the largest professional training providers for HIV treatment education in the world. The program seeks to improve HIV-related and primary care for underserved populations by strengthening the capacity of clinician and other providers to understand and treat these populations. Training is organized into three basic levels: didactic(Level I), skills building(Level II), clinical hands on training(Level III), a recipient-driven clinical consultation (Level IV) and technical assistance (Level V). During the 2002-2003 grant cycle(study year 1), the AETC provided a total of 847 structured training events (Levels I through III), as well as 429 clinical consultations and 628 responses to requests for technical assistance. For the grant cycle 2003-2004(study year 2), the AETC provided a total of 877 structured training events (Levels I through III), as well as 912 clinical consultations and 842 responses to requests for technical assistance. Data collected by the PA/MA AETC was obtained separately for each year and data analysis was performed using Minitab and Access. Pie and vertical bar charts were created in SigmaPlot to summarize activities over the two years. The evaluation demonstrates the nature and extent of AETC training and the ways it may contribute to enhancing the knowledge and skills of the participants. The two core evaluation questions for 2002-2004 are the following 1) do the programs reach the primary care provider audiences with a focus on Ryan White, community/migrant health centers (CMHCs), minority providers, and those serving medically-underserved and the poor and 2) do the regional programs address key content areas that address Ryan White CARE Act (the largest source of federal funding for people living with HIV/AIDS in the United States) requirements. A review of the data shows that a large percentage of AETC training participants are minorities or treat minorities and are serving a heavy minority client/patient caseload. More than half of the participants are physicians and nurses. Employees of CARE Act-funded agencies form a large proportion of the total trainees enabling them to acquire the latest HIV treatment knowledge and skills. The AETC strengthens their HIV knowledge and skills by providing training in advanced clinical management topics as well as topics relevant to understanding and working with the special populations served by many of the trainees. Collectively, these training activities contribute to enhancing both the quality and the continuity of HIV-related care provided to underserved and vulnerable populations across the Pennsylvania/MidAtlantic region by clinicians and other health care providers.

The current Mortality and Population Data System (MPDS) database contains the cause of death data and population data from 1950 to 2001, and it was designed to provide data for public health related studies. The cause of death data in the MPDS are provided by the National Center for Health Statistics and are updated annually as new cause of death data from NCHS are released. Since the actual annual population data is not available, the intercensal population figures have to be estimated based on the census population data. The population figures used in the MPDS were estimated by using year-based linear interpolations and extrapolations at the county level, while the census bureau used the cohort-component method to estimate the population. The purpose of this thesis is to compare these two population estimates using two approaches, 1) determine if there are any important differences between them, and 2) evaluate the effect of the difference on the calculation of the mortality rates. The results showed that at national, state and county level, sex was a factor that contributes to the difference between the two populations, while other factors such as year, race, and age group did not affect the difference greatly. The difference between the two population estimates mainly comes from the difference between the female groups of the two populations. The effect of the difference on the calculation of the standardized mortality ratios (SMRs) was analyzed by using data from an occupational cohort study. The results from the analysis of the occupational cohort data showed that the significance of the SMRs for each cause of death was not different when using different rates from the two population estimates. The 95% confidence intervals for the SMRs for the major categories of cause of death overlap. The SMRs calculated with new and old population estimates as reference populations were not significantly different.

This dissertation addresses regression models with missing covariate data. These methods are shown to be significant to public health research since they enable researchers to use a wider spectrum of data. Unbiased estimating equations are the focus of this dissertation, predominantly semiparametric methods utilized to solve for regression parameters in the presence of missing covariate data. The first aim of this dissertation is to evaluate the properties of an efficient score, an inverse probability weighted estimating equation approach, for logistic regression in a two-phase design. Simulation studies showed that the efficient score is more efficient than two other pseudo-likelihood methods when the correlation between the missing covariate and its surrogate is high. The second aim of this dissertation is to develop a methodology for left truncated covariate data with a binary outcome. To address this problem, we proposed two methods, a likelihood-based approach and an estimating equation approach, to estimate the coefficients and their standard errors for a regression model with a left truncated covariate. The estimating equation technique is close to completion, and once solved should be the most efficient method. The likelihood-based method is compared to standard methods of filling in the truncated values with the lower threshold value or using only the nontruncated values. Simulation studies demonstrated that the likelihood-based method has the best variance correction and moderate bias correction. The application of this method is illustrated in a sepsis study conducted at the University of Pittsburgh.

For survival data with nonproportional hazards, the weighted log-rank tests with a proper weighting function are expected to be more sensitive than the simple log-rank statistics for comparing survival data with random effects. A series of simulations were carried out to investigate how much better the weighted log-rank test performs under these situations. The nonproportional hazards data were generated by changing the hazard ratios and piecewise exponential functions. Our Monte Carlo simulation study shows the test with a newly developed weight function has an overall better sensitivity (statistical power) than the simple log-rank test and Harrington-Fleming's weighted log-rank test in detecting the difference between two survival distributions when populations become more homogeneous as time progresses (early difference). For the datasets with middle difference, the test with the new weight function has better sensitivity than that of Harrington-Fleming's weighted log-rank test, similar to that of the simple-log rank test. For late difference, all three tests have similar sensitivity. The new weight function can be used in testing the survival data with nonproportional hazards in public health relevance applications.

Well designed clinical studies theoretically produce accurate data from which a reasonable conclusion(s) may be drawn. Data accuracy may be hindered by the measurement tool or device. Additionally, the data may be in such a form that it is problematic from an analytic and interpretive point of view. An example of such a problematic form may be seen in censored, sample-selected, or truncated data. Clinical data may be particularly prone to censoring or truncation since various assays used to measure patient parameters have limited sensitivity. Lower and upper limits of assay sensitivity may have a direct impact on the clinical diagnosis and prognosis of the patient, especially if the patient is a high risk critical care patient. The aim of this report is to estimate mean cytokine levels using various approaches, including the arithmetic and geometric mean, and mean estimation from a tobit model. The data set is from the Department of Critical Care Medicine and contains values for several cytokines from 1753 patients (discharge status) or 1610 patients (follow-up status), including Interleukin 6 (IL-6), Interleukin 10 (IL-10), and Tumor Necrosis Factor (TNF). A brief overview of the immune system and its relationship to cytokine production will be presented prior to an explanation of the estimation procedures. Finally, recommendations for estimating a mean from the censored data set will be presented. Although not specific to Critical Care Medicine, the problem of censored data is evident in many areas of study, specifically Public Health. Guidelines for dealing with censored data would be a significant and valuable tool for Public Health professionals.

Organizational culture has been shown to be associated with intensive care unit job performance and patient outcomes. These findings have led to recommendations to improve the safety climate in ICUs. While ICUs within a single hospital may be expected to have similar climates, previous research has pointed to variations between ICUs. Also, ICU directors' assessments of their personnel's experiences may not be accurate. The purpose of this thesis was to determine whether variations in organizational culture exist between the ICUs of a single institution and between different types of personnel, as well as to assess the accuracy of ICU directors'perceptions of personnel attitudes. The personnel of four ICUs within a single hospital were surveyed using the Safety Attitudes Questionnaire - ICU, which was designed to assess organizational culture across six factors: teamwork climate, perceptions of management, safety climate, stress recognition, job satisfaction, and work environment. Mean and percent positive scores (percentage of scores greater than or equal to 75 on a 0-100 point scale) were calculated for each ICU and for each job type across ICUs. Generalized estimating equations were used to model each factor score by job type while accounting for a possible clustering effect due to ICU membership. Directors were asked to estimate their personnel's mean factor scores and differences between director estimates and actual scores were assessed using the Wilcoxon signed rank test. Scores were found to differ significantly across ICUs for all factors except stress recognition. Scores for job satisfaction, perceptions of management, and working conditions were found to differ significantly between physicians and nurses. ICU directors tended to overestimate the attitudes of their personnel, however the overestimation was not found to be significant. The results suggest that assessments based on hospital level analysis or director opinion may not be sufficient. It is seemingly important to account for differences between ICUs, as well as between personnel types, when creating policies affecting organizational culture. The public health relevance of this thesis is in determining a unit of analysis for organizational culture assessments to improve job performance of ICU personnel, and subsequently, to hopefully improve patient outcomes.

Longitudinal studies are common in many areas of public health. A usual method to analyze longitudinal data is by repeated-measures analysis of variance (ANOVA). A newer method, the mixed models approach, is gaining more acceptance due to the available use of computer programs. It is of public health importance to review the advantages of the recent mixed models approach to analyzing longitudinal data. The main characteristic of longitudinal studies is that the outcome of interest is measured on the same individual at several points in time. The standard approach to analyzing this type of data is the repeated-measures ANOVA, but this type of design assumes equal correlation between individuals and either includes data from individuals with complete observations only or imputes missing data, both of which suffer from the ineffective use of available data. Alternatively, the mixed model approach has the ability to model the data more accurately because it can take into account the correlation between repeated observations, as well as uses data from all individuals regardless of whether their data are complete. This thesis first reviews the literature on the repeated-measures ANOVA and mixed models techniques. Data from a placebo-controlled clinical trial of the drug methylphenidate (MPH) looking at the social/play behavior of children with attention deficit hyperactivity disorder (ADHD) and mental retardation (MR) are analyzed using repeated-measures ANOVA, repeated-measures ANOVA with the last observation carried forward (LOCF) and mixed models techniques. P-values and parameter estimates for the three methods are compared. MPH had a significant effect on the variables Withdrawn and Intensity in both of the repeated-measures analyses. With the repeated-measures with LOCF, MPH had a significant effect on the variables Activity Intensity Level and Sociability. The mixed models analysis found MPH to have a significant effect on the variables Intensity and Activity Intensity Level. The parameter estimates for the two repeated-measures ANOVA analyses were almost identical, but the mixed model parameter estimates were different. Mixed models should be used to analyze these data as assumptions of the repeated-measures ANOVA are violated. Mixed models also take into account the missing data and correlated outcomes.

Previous authors have combined tests for pairs and unpaired data so that population means can be compared using a paired study design with incomplete data. The primary object of my thesis is to determine the appropriate sample size and the appropriate proportion and configuration of complete data and incomplete data so that a normal approximation can be used to calculate p-values. The test statistic studied is one due to Wilson (1992) in which the sign test and rank sum test are combined to form of composite test statistic. To fulfill these objectives, the following approach is adopted: (1) Choose different data scenarios in terms of different sample sizes of paired data and different proportions of complete data. (2) Obtain the exact sampling distribution of the test statistic under each data scenario we study. (3) Obtain the normal approximation distribution under each data scenario we study. (4) Compare the exact and approximate cumulate distribution by their difference on each possible test statistic value. The results show that when the study groups are approximately balanced with respect to incomplete data, and have at least 9 observations in each group, the normal approximation appears to be useful when the number of complete pairs is as low as 5. However, when the groups are highly unbalanced with respect to incomplete data, using the normal approximation seems not to be appropriate, at least when the total sample size is 70 or less. These results may make public health studies easier to carry out when the data include both complete and incomplete pairs.

Estimated associations between an outcome variable and misclassified covariates tend to be biased when the methods of estimation that ignore the classification error are applied. Available methods to account for misclassification often require the use of a validation sample (i.e, a gold standard). But in practice, such gold standard may be unavailable or impractical. We propose a Bayesian approach to adjust for misclassification in a binary covariate in fixed and random effect logistic models when a gold standard is not available. This Markov Chain Monte Carto (MCMC) approach uses two imperfect measures of a dichotomous exposure under the assumption of conditional independence and non-differential misclassification. This approach is validated with several simulation studies. We illustrate the proposed approach to adjust for misclassification with respect to oxygenation status in a multi-center trial of patients with pneumonia, where 16 per cent of patients are classified discordantly by two assessments. The estimated log odds of inpatient care and the corresponding standard deviation are much larger in our proposed method compared to the models ignoring misclassification. We also applied the proposed Bayesian method to the EDCAP trial to assess the intervention effect allowing for misclassification with respect to risk status. Ignoring misclassification produces downwardly biased estimates and underestimates uncertainty. The public health significance of this study is that the proposed approach can correct for the bias of an estimated association when a covariate is misclassified and no gold standard is available, which is common problem in epidemiology studies.