Wang, Zexi. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2007. / Includes bibliographical references (leaves 63-64). / Abstract --- p.i / Acknowledgement --- p.iii / Chapter 0 --- Introduction --- p.1 / Chapter 1 --- The Design of the Experiments --- p.4 / Chapter 1.1 --- Flies and the Devices --- p.5 / Chapter 1.2 --- Parameter Settings and Interested Information --- p.8 / Chapter 2 --- Video Processing --- p.11 / Chapter 2.1 --- "Videos, Computer Vision and Image Processing" --- p.11 / Chapter 2.2 --- Details in Video Processing --- p.14 / Chapter 3 --- Data Analysis --- p.20 / Chapter 3.1 --- Background --- p.20 / Chapter 3.2 --- Outline of Data Analysis in Our Project --- p.22 / Chapter 4 --- Effect of the Medicine --- p.25 / Chapter 4.1 --- Hypothesis Testing --- p.26 / Chapter 4.2 --- Two-sample t Test --- p.28 / Chapter 5 --- Significance of the Two Factors --- p.32 / Chapter 5.1 --- Background of ANOVA --- p.33 / Chapter 5.2 --- The Model of ANOVA --- p.35 / Chapter 5.3 --- Two-way ANOVA in Our Data Analysis --- p.42 / Chapter 6 --- Regression Model --- p.45 / Chapter 6.1 --- Background of Regression Analysis --- p.47 / Chapter 6.2 --- Polynomial Regression Models --- p.52 / Chapter 6.2.1 --- Background --- p.52 / Chapter 6.2.2 --- R2 and adjusted R2 --- p.53 / Chapter 6.3 --- Model Verification --- p.58 / Chapter 6.4 --- A Simpler Model As the Other Choice --- p.59 / Chapter 6.5 --- Conclusions --- p.60 / Chapter 7 --- Further Studies --- p.61 / Bibliography --- p.62

The last two decades have seen several large-scale epidemics of international impact, including human, animal and plant epidemics. Policy makers face health challenges that require epidemic predictions based on limited information. There is therefore a pressing need to construct models that allow us to frame all available information to predict an emerging outbreak and to control it in a timely manner. The aim of this thesis is to develop an early-warning modelling approach that can predict emerging disease outbreaks. Based on Bayesian techniques ideally suited to combine information from different sources into a single modelling and estimation framework, I developed a suite of approaches to epidemiological data that can deal with data from different sources and of varying quality. The SEIR model, particle filter algorithm and a number of influenza-related datasets were utilised to examine various models and methodologies to predict influenza outbreaks. The data included a combination of consultations and diagnosed influenza-like illness (ILI) cases for five influenza seasons. I showed that for the pandemic season, different proxies lead to similar behaviour of the effective reproduction number. For influenza datasets, there exists a strong relationship between consultations and diagnosed datasets, especially when considering time-dependent models. Individual parameters for different influenza seasons provided similar values, thereby offering an opportunity to utilise such information in future outbreaks. Moreover, my findings showed that when the temperature drops below 14°C, this triggers the first substantial rise in the number of ILI cases, highlighting that temperature data is an important signal to trigger the start of the influenza epidemic. Further probing was carried out among Maltese citizens and estimates on the under-reporting rate of the seasonal influenza were established. Based on these findings, a new epidemiological model and framework were developed, providing accurate real-time forecasts with a clear early warning signal to the influenza outbreak. This research utilised a combination of novel data sources to predict influenza outbreaks. Such information is beneficial for health authorities to plan health strategies and control epidemics.

Statistical research can be more difficult to plan than other kinds of projects, since the research must adapt as knowledge is gained. This dissertation establishes a formal language and methodology for designing experimental research strategies with limited resources. It is a mathematically rigorous extension of a sequential and adaptive form of statistical research called response surface methodology. It uses sponsor-given information, conditions, and resource constraints to decompose an overall project into individual stages. At each stage, a "parent" decision-maker determines what design of experimentation to do for its stage of research, and adapts to the feedback from that research's potential "children", each of whom deal with a different possible state of knowledge resulting from the experimentation of the "parent". The research of this dissertation extends the real-world rigor of the statistical field of design of experiments to develop an deterministic, adaptive algorithm that produces deterministically generated, reproducible, testable, defendable, adaptive, resource-constrained multi-stage experimental schedules without having to spend physical resource.

The modern-day abundance of both data and computational resources has expanded the potential for statistical methods to improve our understanding of social phenomena. Such insights can also serve as a basis for policy decisions and resource allocation. This thesis explores pragmatic approaches to statistical inference in select applications related to social issues. We begin by considering underreporting of sexual assault on college campuses. Because many instances of sexual assault are never reported to authorities, the number of reported assaults does not fully reflect the true total number of assaults that occurred, and could arise from many combinations of reporting rate and true incidence. We estimate these quantities via a hierarchical Bayesian model of the reported data, drawing prior information from national crime statistics to help distinguish between reporting rates and incidence. Next, we consider the task of detecting election interference, focusing on Ghana’s 2020 presidential election. Using locally-aggregated vote tabulation data, we construct a randomization test to screen for potential ballot box stuffing. After identifying regions with suspicious results, we estimate alternative vote totals under a counterfactual scenario with no election interference. Next we turn our attention to synthetic control methods for causal inference. Using a motivating example of estimating the effects of exogenous economic shocks on support for political incumbents, we propose methods for leveraging shared information across multiple related studies. Finally, we address the issue of data privacy. Motivated by the need to balance the utility of data analysis with protection of individuals’ privacy, we develop randomized optimization algorithms for differentially private statistical inference. These procedures allow the construction of private confidence regions, and employ a bias correction that improves empirical performance in small samples.

Publicly available datasets in health science are often large and observational, in contrast to experimental datasets where a small number of data are collected in controlled experiments. Variables' causal relationships in the observational dataset are yet to be determined. However, there is a significant interest in health science to discover and analyze causal relationships from health data since identified causal relationships will greatly facilitate medical professionals to prevent diseases or to mitigate the negative effects of the disease. Recent advances in Computer Science, particularly in Bayesian networks, has initiated a renewed interest for causality research. Causal relationships can be possibly discovered through learning the network structures from data. However, the number of candidate graphs grows in a more than exponential rate with the increase of variables. Exact learning for obtaining the optimal structure is thus computationally infeasible in practice. As a result, heuristic approaches are imperative to alleviate the difficulty of computations. This research provides effective and efficient learning tools for local causal discoveries and novel methods of learning causal structures with a combination of background knowledge. Specifically in the direction of constraint based structural learning, polynomial-time algorithms for constructing causal structures are designed with first-order conditional independence. Algorithms of efficiently discovering non-causal factors are developed and proved. In addition, when the background knowledge is partially known, methods of graph decomposition are provided so as to reduce the number of conditioned variables. Experiments on both synthetic data and real epidemiological data indicate the provided methods are applicable to large-scale datasets and scalable for causal analysis in health data. Followed by the research methods and experiments, this dissertation gives thoughtful discussions on the reliability of causal discoveries computational health science research, complexity, and implications in health science research.

The Bayesian network (BN) formalism is a powerful representation for encoding domains characterised by uncertainty. However, before it can be used it must first be constructed, which is a major challenge for any real-life problem. There are two broad approaches, namely the hand-crafted approach, which relies on a human expert, and the data-driven approach, which relies on data. The former approach is useful, however issues such as human bias can introduce errors into the model. We have conducted a literature review of the expert-driven approach, and we have cherry-picked a number of common methods, and engineered a framework to assist non-BN experts with expert-driven construction of BNs. The latter construction approach uses algorithms to construct the model from a data set. However, construction from data is provably NP-hard. To solve this problem, approximate, heuristic algorithms have been proposed; in particular, algorithms that assume an order between the nodes, therefore reducing the search space. However, traditionally, this approach relies on an expert providing the order among the variables --- an expert may not always be available, or may be unable to provide the order. Nevertheless, if a good order is available, these order-based algorithms have demonstrated good performance. More recent approaches attempt to ''learn'' a good order then use the order-based algorithm to discover the structure. To eliminate the need for order information during construction, we propose a search in the entire space of Bayesian network structures --- we present a novel approach for carrying out this task, and we demonstrate its performance against existing algorithms that search in the entire space and the space of orders. Finally, we employ the hand-crafting framework to construct models for the task of diagnosis in a ''real-life'' medical domain, dementia diagnosis. We collect real dementia data from clinical practice, and we apply the data-driven algorithms developed to assess the concordance between the reference models developed by hand and the models derived from real clinical data.