No description available.

<p>Härtill 6 uppsatser</p> / digitalisering@umu

This thesis develops techniques for adjusting for selection bias using Gaussian process models. Selection bias is a key issue both in sample surveys and in observational studies for causal inference. Despite recently emerged techniques for dealing with selection bias in high-dimensional or complex situations, use of Gaussian process models and Bayesian hierarchical models in general has not been explored. Three approaches are developed for using Gaussian process models to estimate the population mean of a response variable with binary selection mechanism. The first approach models only the response with the selection probability being ignored. The second approach incorporates the selection probability when modeling the response using dependent Gaussian process priors. The third approach uses the selection probability as an additional covariate when modeling the response. The third approach requires knowledge of the selection probability, while the second approach can be used even when the selection probability is not available. In addition to these Gaussian process approaches, a new version of the Horvitz-Thompson estimator is also developed, which follows the conditionality principle and relates to importance sampling for Monte Carlo simulations. Simulation studies and the analysis of an example due to Kang and Schafer show that the Gaussian process approaches that consider the selection probability are able to not only correct selection bias effectively, but also control the sampling errors well, and therefore can often provide more efficient estimates than the methods tested that are not based on Gaussian process models, in both simple and complex situations. Even the Gaussian process approach that ignores the selection probability often, though not always, performs well when some selection bias is present. These results demonstrate the strength of Gaussian process models in dealing with selection bias, especially in high-dimensional or complex situations. These results also demonstrate that Gaussian process models can be implemented rather effectively so that the benefits of using Gaussian process models can be realized in practice, contrary to the common belief that highly flexible models are too complex to use practically for dealing with selection bias.

We study the consistency, robustness and efficiency of parameter estimation in different but related models via semiparametric approach. First, we revisit the second- order least squares estimator proposed in Wang and Leblanc (2008) and show that the estimator reaches the semiparametric efficiency. We further extend the method to the heteroscedastic error models and propose a semiparametric efficient estimator in this more general setting. Second, we study a class of semiparametric skewed distributions arising when the sample selection process causes sampling bias for the observations. We begin by assuming the anti-symmetric property to the skewing function. Taking into account the symmetric nature of the population distribution, we propose consistent estimators for the center of the symmetric population. These estimators are robust to model misspecification and reach the minimum possible estimation variance. Next, we extend the model to permit a more flexible skewing structure. Without assuming a particular form of the skewing function, we propose both consistent and efficient estimators for the center of the symmetric population using a semiparametric method. We also analyze the asymptotic properties and derive the corresponding inference procedures. Numerical results are provided to support the results and illustrate the finite sample performance of the proposed estimators.

Dans la théorie de l'apprentissage supervisé, l'hypothèse selon laquelle l'échantillon de d'apprentissage et de test proviennent de la même distribution de probabilité, joue un rôle crucial. Malheureusement, cette hypothèse essentielle est souvent violée en présence d'un biais de sélection. Dans ce contexte, les algorithmes d'apprentissage supervisés standards peuvent souffrir d'un biais significatif. Dans cette thèse, nous abordons le problème du biais de sélection en apprentissage supervisé en utilisant la méthode de pondération de l'importance ("importance weighting" en anglais).Dans un premier temps, nous présentons le cadre formel de l'apprentissage supervisé et discutons des effets potentiellement néfastes du biais sur les performances prédictives. Nous étudions ensuite en détail comment les techniques de pondération de l'importance permettent, sous certaines hypothèses, de corriger le biais de sélection durant l'apprentissage de modèles génératifs et discriminants. Nous étudions enfin le potentiel des réseaux bayésiens comme outils de représentation graphique des relations d'indépendances conditionnelles entre les variables du problème et celles liées au mécanisme de sélection lui-même. Nous illustrons sur des exemples simples comment la graphe, construit avec de la connaissance experte, permet d'identifier a posteriori un sous-ensemble restreint de variables sur lesquelles « agir » pour réduire le biais.Dans un second temps, nous accordons une attention particulière au « covariate shift », i.e. un cas particulier de biais de sélection où la distribution conditionnelle P(y|x) est invariante entre l'échantillon d'apprentissage et de test. Nous proposons deux méthodes pour améliorer la pondération de l'importance en présence de covariate shift. Nous montrons d'abord que le modèle non pondéré est localement moins biaisé que le modèle pondéré sur les échantillons faiblement pondérés, puis nous proposons une première méthode combinant les modèles pondérés et non pondérés afin d'améliorer les performances prédictives dans le domaine cible. Enfin, nous étudions la relation entre le covariate shift et le problème des données manquantes dans les jeux de données de petite taille et proposons une seconde méthode qui utilise des techniques d'imputation de données manquantes pour corriger le covariate shift dans des scénarios simples mais réalistes. Ces méthodes sont validées expérimentalement sur de nombreux jeux de données / In the theory of supervised learning, the identical assumption, i.e. the training and test samples are drawn from the same probability distribution, plays a crucial role. Unfortunately, this essential assumption is often violated in the presence of selection bias. Under such condition, the standard supervised learning frameworks may suffer a significant bias. In this thesis, we address the problem of selection bias in supervised learning using the importance weighting method. We first introduce the supervised learning frameworks and discuss the importance of the identical assumption. We then study the importance weighting framework for generative and discriminative learning under a general selection scheme and investigate the potential of Bayesian Network to encode the researcher's a priori assumption about the relationships between the variables, including the selection variable, and to infer the independence and conditional independence relationships that allow selection bias to be corrected.We pay special attention to covariate shift, i.e. a special class of selection bias where the conditional distribution P(y|x) of the training and test data are the same. We propose two methods to improve importance weighting for covariate shift. We first show that the unweighted model is locally less biased than the weighted one on low importance instances, and then propose a method combining the weighted and the unweighted models in order to improve the predictive performance in the target domain. Finally, we investigate the relationship between covariate shift and the missing data problem for data sets with small sample sizes and study a method that uses missing data imputation techniques to correct the covariate shift in simple but realistic scenarios

Labor intensive sectors such as the specialty crop sector have historically had strong reliance on foreign labor, constituting roughly oneifth of all U.S. farms while incurring roughly two-thirds of direct-hire expenses. It is estimated that more than half unauthorized of the foreign-born labor force in the specialty crop sector are unauthorized for US employment. Using data from the National Agricultural Workers Survey for 1989-2014, this study uses a treatment effects approach (via propensity score matching and minimum-biased estimation) to evaluate the farm wage implications of legalization of foreign-born specialty crop farm workers nationally, as well as specifically in California. Positive wage effects are estimated in nationally and in California, with higher magnitude effects observed in California.

This research explores the effects of playing interscholastic sports on labor market income in the United States for males (n=5782) and females (n=6266) who participated in the National Education Longitudinal Study of 1988. Previous research has explored the effects of human capital and social capital on positive life outcomes for interscholastic athletes, but little research has looked into possible cultural capital advantages gained through interscholastic sports participation. Using multiple regression analysis and controlling for the effects of human and social capital, I examine whether participation in interscholastic sports operate as cultural capital. Results indicate that after net of controls the relationship between interscholastic sports participation and labor market income remains positive and significant for males who play sports and females who play the culturally popular sport of basketball. Results also indicate that those male student athletes who play culturally popular sports (football, basketball, or baseball) report more income six years after high school graduation than their counterparts who play a less culturally popular sport. This article provides evidence that cultural capital theory is a useful tool in exploring the relationship between interscholastic sports and labor market income.

The state of Virginia has been a leader in the design and implementation of welfare reform measures. State welfare reforms were enacted in 1996 and federal reforms followed shortly after in 1996. Initial decreases in program caseloads and the movement of former recipients from unemployment to employment have led initial reform measures to be widely heralded as successes. Significant concerns remain, however, about the ability of non-metropolitan labor markets to absorb female household heads currently on welfare. This thesis addresses potential differences in the impact of welfare reform measures in non-metropolitan and metropolitan labor markets by estimating wage and reservation wage equations for female household heads in Northern and Southwest Virginia. The results suggest young children and lack of access to automobiles create significantly greater barriers to employment in non-metropolitan than metropolitan labor markets. Estimated potential earnings in Southwest Virginia were lower than in Northern Virginia and suggest that female household heads will have trouble escaping poverty through employment. In fact, initial reported earnings for both areas have fallen below estimates of living wages needed to escape poverty. The results also suggest traditional labor market characteristics do not explain all of the differences in earnings, particularly the differences in the observed wages of persons exiting welfare as compared to the general population. If this is the case, policies that only address child care and transportation costs may have little impact as to the ability of welfare recipients to get and keep jobs that enable them to become economically self-sufficient. / Master of Science

This thesis explores the use of observational microdata in cost-effectiveness analysis. The application of econometric methods adjusting for selection bias is first reviewed and critically appraised in the economic evaluation literature using a structured template. Limitations of identified studies include lack of good quality evidence regarding the performance of different analytical approaches; inadequate assessment of the sensitivity of their results to violations of fundamental assumptions or variations to crucial estimator parameters; failure to combine the cost and effectiveness outcomes in a summary measure; and no consideration of stochastic uncertainty for the purpose of evaluating cost-effectiveness. Data from the Birthplace national cohort study are used in an attempt to address these limitations in the context of an empirical comparison of estimators relying on regression, matching, as well as the propensity score. It is argued that although these methods cannot address the potential impact of unobservable confounding, a novel approach to bias-corrected matching, combining entropy balancing with seemingly unrelated regression, still has the potential to offer important advantages in terms of analytical robustness. The net economic benefit is proposed as a straightforward way to exploit the strengths of rigorous econometric methodology in the development of reliable and informative cost-effectiveness analyses.

O objetivo desse trabalho é analisar a relação entre o desempenho escolar e os salários dos jovens brasileiros. Examinamos se a qualidade do ensino, mensurada pelas notas obtidas por uma geração nos exames de proficiência realizados pelo Instituto Nacional de Estudos e Pesquisas Anísio Teixeira, o INEP, em seu Estado ao término do ensino médio, afeta os salários a serem recebidos por esta geração quando ela estiver no mercado de trabalho, cinco e seis anos depois, respectivamente com dados do Censo Demográfico de 2000 e da Pesquisa Nacional por Amostra de Domicílios, a PNAD, de 2001. A partir de um modelo de pseudo-painel, corrigimos os problemas de viés de seleção gerados pelas migrações e pelo alto nível educacional da amostra selecionada, através do modelo de Roy (1951) aplicado em Dahl (2001). Os determinantes da proficiência escolar, tais como “background" familiar, a infra-estrutura escolar e o perfil de docentes e diretores também foram analisados. A partir da estimação do modelo em dois estágios, os resultados do primeiro estágio mostram que o desempenho dos alunos em exames de proficiência é positivamente relacionado aos investimentos em educação, como melhor remuneração a professores e diretores, critérios mais rigorosos de seleção desses profissionais, assim como investimentos em infra-estrutura, que melhoram a qualidade da escola. Isso aponta para a existência de uma relação entre os recursos destinados à educação e a qualidade da mesma no Brasil. Adicionalmente, os resultados do segundo estágio do modelo mostram que as notas obtidas por uma geração nos testes de proficiência são significantes para explicar os salários futuros da mesma. Dessa forma o estudo confirma a importância de políticas públicas que invistam na qualidade da escola ao invés de políticas destinadas apenas a aumentar os anos de estudo da população. / The aim of this paper is to examine the relationship between the school performance and the wages of young Brazilians workers. We examine if school quality, measured by test scores of a generation in SAEB at the end of high school, affect the earnings of this generation when they enter the labor force, five and six years later, with sample of Censo (2000) and PNAD (2001). We use a pseudo-panel model to correct the problems of selection bias, created by migrations and by the high education level of the selected sample, through a Roy model (1951) applied in Dahl (2001). The determinants of school performance, like familiar background, school structure, teacher and director profiles also were analyzed. We conclude that school characteristics are responsible for a good performance of students in tests scores, and that the school performance explains the differences of the earnings of young Brazilians workers.