The aim of the present work is to develop a tool for identification of Motor Third Party Liability insurance contracts which are at risk of cancellation. Methods for explorative data analysis, building a logistic regression model, comparing models and their validation and calibration are presented. Several models are developed on the real dataset using mentioned methods and then the final one is chosen. Behavior of the final model is verified by the validation on the out-of-time sample. Last step is calibration of the model to the expected value of the future portfolio cancellation rate.

Approved for public release, distribution is unlimited / The objective of this thesis is to estimate and explain the effects of graduate education and other factors on promotion to the rank of Lieutenant Colonel (O-5) in the US Army. Our focus was primarily on determining whether graduate education provides officers with higher promotion probabilities. Besides graduate education, data that were analyzed include basic demographic traits, the officers' prior enlisted status, and their commissioning source information. The data used in this study were taken from the Active Duty Military Master File for fiscal years 1981 through 2001. This study develops multivariate logit regression and classification tree models to examine and explore the structure of the data sets. Both the regression models and the classification trees yielded positive results for the effect of graduate education on promotion. According to the regression model results, the odds ratio associated with graduate education is between 1.79 and 2.25. Military Academy and ROTC/Scholarship graduates have higher promotion probabilities than those from other sources, and married officers have higher rates than single officers. Additionally, age has a negative effect on promotion; that is, promotion probability decreases with age. Prior enlisted status, number of dependents, gender, race, and DOD primary occupation code do not seem to have statistically significant effects on promotion. / First Lieutenant, Turkish Army

Approved for public release, distribution is unlimited / This thesis examines the process of selecting the midshipmen leadership, or "stripers," at the United States Naval Academy. Using a qualitative approach, it gathers data from the current cohort of decision makers who select the stripers each semester regarding what they believe to be the desirable and undesirable qualities of striper candidates. Shifting to a quantitative approach, those qualities are then used to create variables using data from the Naval Academy classes of 1999 through 2002. A logistic regression is then estimated with the purpose of gauging if those qualities are, in fact, represented in the selectees. A model is presented which indicates that, by and large, the goals of the selection process are being met. Recommendations for minor policy adjustments and for further research are made based on the findings of both the qualitative and quantitative data. / http://hdl.handle.net/10945/985 / Lieutenant, United States Navy

The research conducted in this thesis examines the relationship between forms of worklessness (both active unemployment and inactivity due to sickness and disability) and mortality for working age men. Previous research has shown that being out of work is associated with a greater risk of mortality relative to being in work. However, there remains debate as to whether this association is the result of a causal pathway leading from worklessness to mortality or whether it reflects the ‘selection' of individuals who are already at greater risk of mortality from pre-existing poor health or other characteristics. In the UK, many studies rely on the use of ‘wear-off' periods in which mortality events occurring within five years after the observation of employment status are ignored to allow the confounding effects of selection to diminish. Generally these studies concluded in support of a causal relationship. In contrast, more recent studies making use of innovative methodological designs such as natural experiments and linked register and health datasets have found less evidence for this explanation with many emphasising the role of confounding and selection. The thesis aims to firstly, examine the effectiveness of wear-off periods and secondly, to develop an alternative counterfactual approach to examine the relationship between worklessness (both active unemployment and health related inactivity) and mortality. These questions are addressed in three stand-alone papers. In the first paper, data from the Scottish Longitudinal Study and the England & Wales Longitudinal Study was used in logistic regression models which estimated the odds of death in a given time period after the 1991 Census for those aged 35–64 in 1991. The odds ratios for the different economic positions (in work, unemployed, retired, permanently sick and other inactive) were compared, as well as the changes in risk associated with cumulatively increasing the length of wear-off prior to follow-up. No evidence was found of health related selection for the unemployed in 1991 suggesting that the use of the five year wear-off period in many studies of mortality and unemployment may be an ineffective and unnecessary technique for mitigating the effects of health-related selection. The second paper examined men aged between 35 and 54 who were in work in 1991. Subsequent employment status in 2001 was observed (in work or unemployed) and the relative all-cause mortality risk of unemployment between 2001 and 2007 was estimated. To account for potential selection into unemployment of those in poor health, a counterfactual propensity score matching framework was used to construct unbiased and comparable samples of in work and unemployed individuals. Matching was based on a wide range of explanatory variables including health status prior to year of unemployment (hospital admissions and self-reported limiting long term illness) as well as measures of socio-economic position. The findings showed that unemployment was associated with a doubling (hazard ratio 2.1 95% CI 1.30 - 3.38) of the subsequent risk of mortality from all causes relative to employment. This scale of effect was consistent across different samples and was robust controlling for prior health and socio-demographic characteristics. These findings were interpreted as evidence that the often observed association between unemployment and mortality may contain a causal component. The second paper implemented a similar analytical design to address the lack of evidence for the independent mortality effect of inactivity due to sickness. The results showed that the mortality risk of economic inactivity due to sickness relative to active employment was significant (HR. 3.18, 95% CI 2.53-3.98) and suggest that economic inactivity due to sickness poses a mortality risk that is independent of prior health. The findings could be interpreted in two ways; either economic inactivity due to sickness is worse for health than actively seeking work or previous studies of unemployment and mortality have underestimated the true effect of being out of work generally. Across the three studies, the main contribution of the thesis is to reassert the importance of worklessness as a determinant of individual mortality. In doing so the studies also found little evidence of systematic confounding by either health or other characteristics. The thesis concludes with a comprehensive discussion of the wider implications of the findings in relation to both general methodological issues in observational epidemiology and possible policy interventions that could be implemented to tackle work-related inequalities in male mortality.

"This project utilized generalized estimating equations and general linear modeling to model learning curves for sonographer performance in emergency ultrasonography. Performance was measured in two ways: image quality (interpretable vs. possible hindrance in interpretation) and agreement of findings between the sonographer and an expert reviewing sonographer. Records from 109 sonographers were split into two data sets-- training (n=50) and testing (n=59)--to conduct exploratory analysis and fit the final models for analysis, respectively. We determined that the number of scans of a particular exam type required for a sonographer to obtain quality images on that exam type with a predicted probability of 0.9 is highly dependent upon the person conducting the review, the indication of the scan (educational or medical), and the outcome of the scan (whether there is a pathology positive finding). Constructing family-wise 95% confidence intervals for each exam type demonstrated a large amount of variation for the number of scans required both between exam types and within exam types. It was determined that a sonographer's experience with a particular exam type is not a significant predictor of future agreement on that exam type and thus no estimates were made based on the agreement learning curves. In addition, we concluded based on a type III analysis that when already considering exam type related experience, the consideration of experience on other exam types does not significantly impact the learning curve for quality. However, the learning curve for agreement is significantly impacted by the additional consideration of experience on other exam types."

M.A. University of the Witwatersrand, Faculty of Humanities (Population Studies), 2012 / This study is based on secondary data analysis of the 1998 South African Demographic and Health Survey (SADHS) data set. The aim of the study was to identify socio-demographic, socio-economic and environmental variables that affect the survival of South African children under the age of five years. Descriptive analyses, frequency tables, Pearson’s chi-square tests of association and binary logistic regression analysis were used for data analysis in this study. Mothers who lost an under-five child were predominantly Black and rural. Such mothers were characterized by rural residential areas, relatively large family sizes, relatively poorer socioeconomic status, relatively poorer access to basic health services, relatively more child deliveries at home, and low level of education. The study showed that 269 of the 5, 066 children in the study died before celebrating their fifth birthday (5.31%). At the 5% level of significance, the survival of under-five children is significantly influenced by 2 of the 11 predictor variables found to be significantly associated in the univariate analysis and therefore included in the logistic regression analysis. These 2 predictor variables were: place of delivery of child [OR=0.97; P=0.000; CI = (0.96, 0.98)], and use of modern contraceptives by the mother [OR=0.73; P=0.002; CI = (0.59, 0.89)]. The study found that not using modern contraceptives gives a lower chance on death of a child under 5 as well as delivering at home, in the absence of a trained birth attendant. These findings are unexpected and contrary to what was found in the univariate analysis. No real explanation can be given for these findings and it would be interesting to see if the same results are found with more recent data. In order for the South African National Department of Health to fulfil its United Nations Millennium Development Goals, rural mothers and their under-five children must be provided with improved health as well as socioeconomic services.

In June 2016, the United Kingdom held its EU referendum, colloquially known as Brexit, in which the people of the island nation voted on whether their country should remain a member of or leave the European Union. This thesis investigates what economic variables may have lain behind the majority outcome of a given voting area (or district) and to what degree they may have impacted it. A logistic regression is conducted primarily on referendum and election data from the Electoral Commission, census data from the Office for National Statistics, and political leaning scores as quantified by the Manifesto Project. The resulting model, which exhibits a hit ratio of 92 percent correct predictions, shows that age, education, national identity, political leaning, irreligion, and unemployment have significant correlations with the majority Brexit outcome of a district. On the other hand, population, health, and income variables do not have statistically significant effects; however, poor health, on average, does seem to have a large positive effect on the odds when taking relative sample size into account.

The aim of the thesis is to investigate how the classification performance of random forest and logistic regression differ, given an imbalanced data set with MCAR missing data. The performance is measured in terms of accuracy and sensitivity. Two analyses are performed: one with a simulated data set and one application using data from the Swedish population registries. The simulation study is created to have the same class imbalance at 1:5. The missing values are handled using three different techniques: complete case analysis, predictive mean matching and mean imputation. The thesis concludes that logistic regression and random forest are on average equally accurate, with some instances of random forest outperforming logistic regression. Logistic regression consistently outperforms random forest with regards to sensitivity. This implies that logistic regression may be the best option for studies where the goal is to accurately predict outcomes in the minority class. None of the missing data techniques stood out in terms of performance.

Based on the curriculum of 2011 a new grading scale ranging from A-F was introduced in the Swedish upper secondary school system. Previous research on similar data have focused on the earlier grading scale, and its crucial that the new circumstances are addressed to understand the impact on grading. Using 348 essays from the national writing test this study investigates the use of automated essay scoring as a way of grading in this new setting. Using various classication methods the models for younger students outperform the corresponding models for older students. This implies that it is harder to predict grades on essays written by older students. Based on the current data the result shows that with the new grading scale the use of automated essay scoring should be used with caution.

Modelos de Credit Scoring são utilizados para estimar a probabilidade de um cliente proponente ao crédito se tornar inadimplente, em determinado período, baseadas em suas informações pessoais e financeiras. Neste trabalho, a técnica proposta em Credit Scoring é Redes Bayesianas (RB) e seus resultados foram comparados aos da Regressão Logística. As RB avaliadas foram as Bayesian Network Classifiers, conhecidas como Classificadores Bayesianos, com seguintes tipos de estrutura: Naive Bayes, Tree Augmented Naive Bayes (TAN) e General Bayesian Network (GBN). As estruturas das RB foram obtidas por Aprendizado de Estrutura a partir de uma base de dados real. Os desempenhos dos modelos foram avaliados e comparados através das taxas de acerto obtidas da Matriz de Confusão, da estatística Kolmogorov-Smirnov e coeficiente Gini. As amostras de desenvolvimento e de validação foram obtidas por Cross-Validation com 10 partições. A análise dos modelos ajustados mostrou que as RB e a Regressão Logística apresentaram desempenho similar, em relação a estatística Kolmogorov- Smirnov e ao coeficiente Gini. O Classificador TAN foi escolhido como o melhor modelo, pois apresentou o melhor desempenho nas previsões dos clientes maus pagadores e permitiu uma análise dos efeitos de interação entre variáveis. / Credit Scoring Models are used to estimate the insolvency probability of a customer, in a period, based on their personal and financial information. In this text, the proposed model for Credit Scoring is Bayesian Networks (BN) and its results were compared to Logistic Regression. The BN evaluated were the Bayesian Networks Classifiers, with structures of type: Naive Bayes, Tree Augmented Naive Bayes (TAN) and General Bayesian Network (GBN). The RB structures were developed using a Structure Learning technique from a real database. The models performance were evaluated and compared through the hit rates observed in Confusion Matrix, Kolmogorov-Smirnov statistic and Gini coefficient. The development and validation samples were obtained using a Cross-Validation criteria with 10-fold. The analysis showed that the fitted BN models have the same performance as the Logistic Regression Models, evaluating the Kolmogorov-Smirnov statistic and Gini coefficient. The TAN Classifier was selected as the best BN model, because it performed better in prediction of bad customers and allowed an interaction effects analysis between variables.