published_or_final_version / Medical Sciences / Master / Master of Medical Sciences

published_or_final_version / Statistics / Master / Master of Social Sciences

Randomized clinical trials provide the best evidence on the effect of treatment studied. There are different types of measures on the treatment effect, depending on the endpoints of the trials. For a given measure, based on the data from clinical trials, various statistical procedures are available for the inference of the treatment effect in terms of this measure. In a cancer clinical trial with a time to an event as the endpoint, hazard ratio is a popular measure for the relative difference between treatment groups. Most current statistical inference procedures for hazard ratio rely on the proportional hazard assumption, which may not be applicable to practice when it does not hold. Nonparametric confidence intervals for the hazard ratio have been proposed based on the asymptotic normality of the kernel estimate for the hazard ratio, but they were found not very satisfactory in the simulation studies. In the first part of this thesis, the empirical likelihood method is used to construct the confidence interval for the time-dependent hazard ratio. The asymptotic distribution of the empirical likelihood ratio is derived and simulation studies are conducted to evaluate the proposed method. It was also argued that the measure of the relative treatment effect based on the hazard ratio may be difficult to understand by clinicians. An alternative measure called probabilistic index was suggested and the C-index was proposed to estimate this index. However, it was pointed out recently that the expected value of the estimate based on the C-index may be far removed from the true index. In the second part of this thesis, assuming a semi-parametric density ratio model, two new estimates based on respectively the conditional likelihood and weighted empirical likelihood are proposed. Associated confidence intervals are also derived based on the bootstrap re-sampling method. The proposed inference procedures are evaluated by Monte-Carlo simulations and applied to the analysis of data from a clinical trial on early breast cancer. After primary analysis including all patients is completed in clinical trials, analysis by subgroups defined based on covariates of patients is often of interest to assess the homogeneity of treatment effects over these subgroups. The treatment-covariate interaction is usually used for this assessment. In the last part of this thesis, a non-parametric measure is used to quantify the interaction between treatments and binary covariates in the presence of censoring. Asymptotic distribution of the interaction estimates are derived and the bootstrap method is applied to construct the confidence intervals. The proposed approaches are also evaluated and compared by Monte-Carlo simulations and applied to a real data set from clinical trial. / Thesis (Ph.D, Mathematics & Statistics) -- Queen's University, 2011-05-21 11:07:52.992

Response adaptive designs are developed for ethical considerations, which sequentially modify the treatment allocations based on the accumulating information in the trial so more patients receive the potentially better treatment. Yi and Wang (2009) proposed a variance-penalized criterion to evaluate the performance of a response adaptive design based on the expected number of patients assigned to the better treatment and the power of statistical test. We use the variance-penalized criterion to examine different response adaptive randomization procedures for normally distributed responses. We propose a new target allocation proportion which increases the chance that more patients receive the better treatment. Simulation results indicate that our proposed design has the advantage of assigning more patients to the potentially better treatment with minimum loss in statistical power, and our design performs better than the designs in the literature based on the variance-penalized criterion.

Clinical laboratory safety data consist of a wide range of biochemical and haematological variables which are collected to monitor the safety of a new treatment during a clinical trial. Although the data are multivariate, testing for abnormal measurements is usually done for only one variable at a time. A Monte Carlo simulation study is described, which compares 16 methods, some of which are new, for detecting multivariate outliers with a view to finding patients with an unusual set of laboratory measurements at a follow-up assessment. Multivariate normal and bootstrap simulations are used to create data sets of various dimensions. Both symmetrical and asymmetrical contamination are considered in this study. The results indicate that in addition to the routine univariate methods, it is desirable to run a battery of multivariable methods on laboratory safety data in an attempt to highlight possible outliers. Mahalanobis distance is a well-known criterion which is included in the study. Appropriate critical values when testing for a single multivariate outlier using Mahalanobis Distance are derived in this thesis, and the jack-knifed Mahalanobis distance is also discussed. Finally, the presence of missing data in laboratory safety data sets is the motivation behind a study which compares eight multiple imputation methods. The multiple imputation study is described, and the performance of two outlier detection methods in the presence of three different proportions of missing data are discussed. Measures are introduced for assessing the accuracy of the missing data results, depending on which method of analysis is used.

Response adaptive designs are developed for ethical considerations, which sequentially modify the treatment allocations based on the accumulating information in the trial so more patients receive the potentially better treatment. Yi and Wang (2009) proposed a variance-penalized criterion to evaluate the performance of a response adaptive design based on the expected number of patients assigned to the better treatment and the power of statistical test. We use the variance-penalized criterion to examine different response adaptive randomization procedures for normally distributed responses. We propose a new target allocation proportion which increases the chance that more patients receive the better treatment. Simulation results indicate that our proposed design has the advantage of assigning more patients to the potentially better treatment with minimum loss in statistical power, and our design performs better than the designs in the literature based on the variance-penalized criterion.

No description available.

<b><u>Purpose</u></b>: Some randomized trials are reported widely, while others remain unpublished. It is essential to systematic reviewers and meta-analysts that factors leading to publication bias in the form of delayed or non-publication of an eligible study are identified. This thesis is an attempt to do this. <br></br><br></br> <b><u>Data</u></b>: The set of randomized trials identified by the Childhood Acute Lymphoblastic Leukaemia (ALL) Collaborative Group was used. This consists of 149 trials comprising 243 randomized comparisons (randomizations), starting prior to 1 January 1988, reported in 257 articles, published prior to 1 January 2000. Each mention of a randomization in an article (irrespective of whether results are given) generates a publication record, of which there are 610. <br></br><br></br> <b><u>Methods</u></b>: The main focus is on identifying which trial characteristics lead to a delay in publication of a randomization. Time to the first mention of a randomization in an article (irrespective of whether any results are given) and to the first reporting of its results are both modelled using ordinary linear regression (the independence model). However, when these analyses are extended to include all mentions and all reportings of results respectively, non-independence necessitates the use of techniques for dealing with repeated measures. In such cases the independence model is the starting point, the residuals from which are used to form the covariance matrix, which in turn is used to suggest plausible correlation structures for repeated measures models. Generalised estimating equation (GEE) analysis is used to select an appropriate correlation structure, and a linear mixed effects model serves to confirm this. The conclusions are then discussed in the context of other studies identified. Finally logistic regression is used to identify trial characteristics associated with a randomization remaining unpublished, and Poisson and negative binomial models to identify those affecting frequency of reporting. <br></br><br></br> <b><u>Results</u></b>: Evidence was found of ‘pipeline bias’ in the reporting of first results since, although direction of effect was not found to be significant, highly statistically significant results are published faster than others. However this is not so for first mentions. Negative results (i.e. those in favour of the standard/control) arm were submitted for first publication faster than all others, although this did not effect time to publication. In addition, geographic location is an important predictor of whether a randomization is ever mentioned in an article, frequency of mentions and of time to first publication and results from single-centre trials are published more frequently than those with multi-centre participation. <br></br><br></br> <b><u>Conclusions</u></b>: Although ‘pipeline bias’ was identified in the analysis of time first reporting of results, it was not present in the analysis of time to first mention, and so not a problem for those wishing only to identify randomized trials for inclusion in meta-analyses. The importance of geographic location suggests that the practice of contacting known trialists is worthwhile in addition to the computerised literature searches and should be continued. <br></br><br></br>

Thesis (M. Med. Sc.)--University of Hong Kong, 2000. / Includes bibliographical references (leaves 46-52).

Thesis (M. Sc.)--University of Alberta, 2009. / Title from pdf file main screen (viewed on Aug. 27, 2009). "A thesis submitted to the Faculty of Graduate Studies and Research in partial fulfillment of the requirements for the degree of Master of Science in Statistics, Department of Mathematical and Statistical Sciences, University of Alberta." Includes bibliographical references.