The use of doubly robust estimators is a key for estimating the population mean response in the presence of incomplete data. Cao et al. (2009) proposed an alternative doubly robust estimator which exhibits strong performance compared to existing estimation methods. In this thesis, we apply the jackknife empirical likelihood, the jackknife empirical likelihood with nuisance parameters, the profile empirical likelihood, and an empirical likelihood method based on the influence function to make an inference for the population mean. We use these methods to construct confidence intervals for the population mean, and compare the coverage probabilities and interval lengths using both the ``usual'' doubly robust estimator and the alternative estimator proposed by Cao et al. (2009). An extensive simulation study is carried out to compare the different methods. Finally, the proposed methods are applied to two real data sets.

[pt] Uma ferramenta importante na avaliação de políticas econômicas é a estimação do efeito médio de um programa ou tratamento sobre uma variável de interesse. A principal dificuldade desse cálculo deve-se µa atribuição do tratamento aos potenciais participantes geralmente não ser aleatória, causando viés de seleção quando desconsiderada. Uma maneira de resolver esse problema é supor que o econometrista observa um conjunto de características determinantes, a menos de um componente estritamente aleatório, da participação. Sob esta hipótese, conhecida como Ignorabilidade, métodos semiparamétricos de estimação foram desenvolvidos, entre os quais a imputação de valores contrafactuais e a reponderação da amostra. Ambos são consistentes e capazes de atingir, assintoticamente, o limite de eficiência semiparamétrico. Entretanto, nas amostras frequentemente disponíveis, o desempenho desses métodos nem sempre é satisfatório. O objetivo deste trabalho é estudar como a combinação das duas estratégias pode produzir estimadores com melhores propriedades em amostras pequenas. Para isto, consideramos duas formas de integrar essas abordagens, tendo como referencial teórico a literatura de estimação duplamente robusta desenvolvida por James Robins e co-autores. Analisamos suas propriedades e discutimos por que podem superar o uso isolado de cada uma das técnicas que os compõem. Finalmente, comparamos, num exercício de Monte Carlo, o desempenho desses estimadores com os de imputação e reponderação. Os resultados mostram que a combinação de estratégias pode reduzir o viés e a variância, mas isso depende da forma como é implementada. Concluímos que a escolha dos parâmetros de suavização é decisiva para o desempenho da estimação em amostras de tamanho moderado. / [en] Estimation of mean treatment effect is an important tool for evaluating economic policy. The main difficulty in this calculation is caused by nonrandom assignment of potential participants to treatment, which leads to selection bias when ignored. A solution to this problem is to suppose the econometrician observes a set of covariates that determine participation, except for a strictly random component. Under this assumption, known as Ignorability, semiparametric methods were developed, including imputation of counterfactual outcomes and sample reweighing. Both are consistent and can asymptotically achieve the semiparametric efficiency bound. However, in sample sizes commonly available, their performance is not always satisfactory. The goal of this dissertation is to study how combining these strategies can lead to better estimation in small samples. We consider two different ways of merging these methods, based on Doubly Robust inference literature developed by James Robins and his co-authors, analyze their properties and discuss why they would overcome each of their components. Finally, we compare the proposed estimators to imputation and reweighing in a Monte Carlo exercise. Results show that while combined strategies may reduce bias and variance, it depends on the way it is implemented. We conclude that the choice of smoothness parameters is critical to obtain good estimates in moderate size samples.

Le sujet principal de cette thèse porte sur l'étude de l'estimation de la variance d'une statistique basée sur des données d'enquête imputées via le bootstrap (ou la méthode de Cyrano). L'application d'une méthode bootstrap conçue pour des données d'enquête complètes (en absence de non-réponse) en présence de valeurs imputées et faire comme si celles-ci étaient de vraies observations peut conduire à une sous-estimation de la variance. Dans ce contexte, Shao et Sitter (1996) ont introduit une procédure bootstrap dans laquelle la variable étudiée et l'indicateur de réponse sont rééchantillonnés ensemble et les non-répondants bootstrap sont imputés de la même manière qu'est traité l'échantillon original. L'estimation bootstrap de la variance obtenue est valide lorsque la fraction de sondage est faible. Dans le chapitre 1, nous commençons par faire une revue des méthodes bootstrap existantes pour les données d'enquête (complètes et imputées) et les présentons dans un cadre unifié pour la première fois dans la littérature. Dans le chapitre 2, nous introduisons une nouvelle procédure bootstrap pour estimer la variance sous l'approche du modèle de non-réponse lorsque le mécanisme de non-réponse uniforme est présumé. En utilisant seulement les informations sur le taux de réponse, contrairement à Shao et Sitter (1996) qui nécessite l'indicateur de réponse individuelle, l'indicateur de réponse bootstrap est généré pour chaque échantillon bootstrap menant à un estimateur bootstrap de la variance valide même pour les fractions de sondage non-négligeables. Dans le chapitre 3, nous étudions les approches bootstrap par pseudo-population et nous considérons une classe plus générale de mécanismes de non-réponse. Nous développons deux procédures bootstrap par pseudo-population pour estimer la variance d'un estimateur imputé par rapport à l'approche du modèle de non-réponse et à celle du modèle d'imputation. Ces procédures sont également valides même pour des fractions de sondage non-négligeables. / The aim of this thesis is to study the bootstrap variance estimators of a statistic based on imputed survey data. Applying a bootstrap method designed for complete survey data (full response) in the presence of imputed values and treating them as true observations may lead to underestimation of the variance. In this context, Shao and Sitter (1996) introduced a bootstrap procedure in which the variable under study and the response status are bootstrapped together and bootstrap non-respondents are imputed using the imputation method applied on the original sample. The resulting bootstrap variance estimator is valid when the sampling fraction is small. In Chapter 1, we begin by doing a survey of the existing bootstrap methods for (complete and imputed) survey data and, for the first time in the literature, present them in a unified framework. In Chapter 2, we introduce a new bootstrap procedure to estimate the variance under the non-response model approach when the uniform non-response mechanism is assumed. Using only information about the response rate, unlike Shao and Sitter (1996) which requires the individual response status, the bootstrap response status is generated for each selected bootstrap sample leading to a valid bootstrap variance estimator even for non-negligible sampling fractions. In Chapter 3, we investigate pseudo-population bootstrap approaches and we consider a more general class of non-response mechanisms. We develop two pseudo-population bootstrap procedures to estimate the variance of an imputed estimator with respect to the non-response model and the imputation model approaches. These procedures are also valid even for non-negligible sampling fractions.