Three essays on doubly robust estimation methods
Başlık çevirisi mevcut değil.
- Tez No: 401079
- Danışmanlar: PROF. DR. WINFRIED POHLMEIER, PROF. DR. FRIEDHELM PFEIFFER
- Tez Türü: Doktora
- Konular: Ekonomi, Economics
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2011
- Dil: İngilizce
- Üniversite: Universität Konstanz
- Enstitü: Yurtdışı Enstitü
- Ana Bilim Dalı: Belirtilmemiş.
- Bilim Dalı: Belirtilmemiş.
- Sayfa Sayısı: 242
Özet
Özet yok.
Özet (Çeviri)
This dissertation addresses the doubly robust estimation of treatment effects in evaluation econometrics. The doubly robust methods combine two methods which estimate the same parameter of interest based on two different model specifications. The consistency of these single methods hinges upon the correct model specification. The doubly robust methods require both model specifications, but stay consistent even if one of the models is misspecified. In that sense, doubly robust methods provide more protection against model misspecification. This thesis aims at analyzing various doubly robust methods and at comparing them with other methods. It comprises three stand-alone research papers that all deal with doubly robust estimation methods, both in a theoretical and an empirical framework. More precisely, this dissertation investigates the properties of several doubly robust estimators, applies these estimators to research questions from the field of labour economics and extends the existing doubly robust methods for different treatment effect parameters. The thesis is organized as follows: the first chapter investigates the doubly robust estimators of the average treatment effect in the binary treatment effect framework. The second chapter concentrates on a special treatment effect, namely the local average treatment effect. The third chapter deals with the doubly robust estimation in a multivalued treatment effect framework. The first chapter of this thesis provides an overview of the parametric average treatment effect estimation methods. The methods which are investigated belong to three main groups: regression, weighting and doubly robust methods. The existing methods are presented in a Generalized Method of Moments (GMM) estimation framework. For one of the doubly robust methods, adjustments which increase the efficiency of the method are proposed. After the theoretical review of the methods, the properties in finite samples are explored by means of a comprehensive Monte Carlo experiment. The goal of the Monte Carlo experiment is twofold: the first goal is to demonstrate the doubly robustness property. The setup is constructed to analyze the finite sample properties of the methods under correct specifications, as well as under misspecifications of the relevant models. The results clearly show that even in small samples the doubly robustness property is observable. Second, the consequences of a violation of the strict overlap assumption for finite sample properties are examined. The results of the Monte Carlo experiment indicate that the weighting methods are strongly affected by violation of strict overlap; however, the regression and the doubly robust methods are not harmed by this kind of violation. For both considerations, I also investigate the properties under different propensity score distributions. The results suggest that all methods perform worse if the sample is unbalanced in terms of treated-control ratio. However, the weighting methods are affected at most. Moreover, I provide two empirical examples where the reviewed methods are used. As a first application, I estimate the causal return to higher education on earnings using the rich data set British National Child Development Study (NCDS). The estimates indicate an average wage return of 20% from obtaining higher education degree versus anything less. The second empirical example is the estimation of the causal effect of grade retention on having an upper secondary school degree (Abitur). The effect of grade retention on high-school drop out is an important question in education research. However, the problem has not been studied as an econometric evaluation problem. The estimation results suggest that grade retention decreases the probability of getting upper secondary school degree. The effect is more evident for males than for females. The second chapter reviews the parametric estimation methods of the Local Average Treatment Effect (LATE). The contribution of this chapter is manifold. First, I propose an easy to implement regression and doubly robust method to estimate the LATE, which controls for the covariates easily. Second, this paper clarifies the relation between the Average Treatment Effect (ATE) and the LATE. The connection between these two treatment effects eases the derivation of the asymptotic properties of the LATE estimators. Third, I investigate the finite sample properties of various LATE estimators under different scenarios. The goals of the simulation study are the following: (i) compare the properties of several methods under correct specifications; (ii) demonstrate the robustness of the doubly robust methods under several misspecification scenarios; (iii) examine the effect of a violation of the strict overlap assumption; (iv) examine the effectiveness of some trimming rules, which are commonly used when estimating the ATE. Given the results of the Monte Carlo study, three points can be made in favor of doubly robust methods: first, they are unbiased even if one of the relevant models is misspecified. Second, the efficiency loss due to the increased estimation error is negligible. Third, the doubly robust methods are not as much harmed by violation of the strict overlap assumption as the weighting methods. Furthermore, I provide two empirical application examples where the proposed doubly robust estimation approach and the other methods reviewed are applied. The first empirical study analyzes the causal effects of JTPA (Job Training Partnership Act) on the earnings for the subpopulation whose participation is manipulated by the assignment. Since the proportion of individuals who get the treatment without being assigned is very small, this effect is approximately equal to the treatment effect on the treated. The results show that the female participants gain more from the training than male participants. In the second empirical example, I estimate the causal effect of upper secondary school graduation on future earnings for the subpopulation whose graduation is induced by the binary instrument grade retention. The LATE estimates with the proposed instrument are larger than the standard OLS estimates of returns to schooling and even larger than other IV estimates of returns to schooling. The implication of this result is that those individuals whose graduation from upper secondary school is induced by grade retention are affected by the treatment more than other subpopulations. In the third chapter, I study the estimation of the causal effects of a multivalued treatment variable under Conditional Independence Assumption (CIA). I propose a combination of the regression adjustment with the weights to get doubly robust estimators of the treatment parameters of interest. This is a generalization of the doubly robust estimators of the average treatment effect and the treatment effect on treated. Following the existing literature, I provide the asymptotic distribution of the proposed estimator. In a small Monte Carlo experiment, I demonstrate the doubly robustness property of the proposed method. By means of the proposed doubly robust method, I estimate returns to schooling using the rich data set British Cohort Study (BCS), where the schooling variable is a (ordered) multivalued treatment variable. It is very common to estimate returns to education as a binary treatment effect, although education is multivalued in nature. Taking into account this multivalued nature can provide further insights on the returns to education. Although the estimation of the returns to education is a commonly investigated research question in empirical labour economics, there are very few studies which estimate the returns relying on CIA. The reason is that CIA requires a very rich data set. This, however, does not create any problems for this study, because the data set includes many other variable besides the standard controls, such as several IQ measures, noncognitive skill measures and behavioral measures. Since all the variables are measured during childhood, there is no danger of any reverse causality problem either. The analysis is done for female and male samples separately to capture possible gender differences. The average returns of most of the education levels are higher for males than females. Interestingly, selection into higher education versus no education or O-level is less efficient for males than for females. This means that there are males who only get O-level or no education, although they would have benefited more on average from getting higher education than those who did actually get the higher education. For both genders, selection into higher education versus A-level seems to be inefficient.
Benzer Tezler
- Attilâ İlhan'ın 'Aynanın İçindekiler' roman serisinde ulusal kimlik meselesi
The issue of national identity in Attilâ İlhan's 'Aynanın İçindekiler' novel series
MUSTAFA GENÇLER
Yüksek Lisans
Türkçe
2023
Sosyolojiİstanbul ÜniversitesiTürk Dili ve Edebiyatı Ana Bilim Dalı
DR. ÖĞR. ÜYESİ FIRAT KARAGÜLLE
- Virginia Woolf & the emergence of Europen feminism
Virginia Woolf ve Avrupa feminizminin ortaya çıkışı
CENGİZ KOÇ
Yüksek Lisans
İngilizce
2004
İngiliz Dili ve EdebiyatıFatih Üniversitesiİngiliz Dili ve Edebiyatı Ana Bilim Dalı
DOÇ. DR. WİSAM MANSOUR
- Une Approche Autofıctionnelle de L'œuvre de Drieu La Rochelle : Exemples de L'Homme couvert de femmes et de Gilles
Drieu la Rochelle'in yapıtına özkurgusal bir yaklaşım : L'Homme couvert de femmes ve Gilles örnekleri
BERAT ERBAŞ
Yüksek Lisans
Fransızca
2024
Batı Dilleri ve EdebiyatıGalatasaray ÜniversitesiFransız Dili ve Edebiyatı Ana Bilim Dalı
DOÇ. DR. ENGİN BEZCİ
- Causalité chez Hume et chez Kant et ses effets sur leurs philosophies pratiques
Hume'a ve Kant'a göre nedensellik ve pratik felsefelerine etkileri
DENİZ NESİN GÜRKAN
Yüksek Lisans
Fransızca
2022
FelsefeGalatasaray ÜniversitesiFelsefe Ana Bilim Dalı
PROF. DR. ALİYE KOVANLIKAYA
- L'articulation des stratégies de résistance et des processus de consentement chez les cols blancs
Beyaz yakalılarda direniş stratejileri ve rıza süreçlerinin eklemlenmesi
AYÇA YILMAZ
Yüksek Lisans
Fransızca
2012
SosyolojiGalatasaray ÜniversitesiSosyoloji Ana Bilim Dalı
PROF. DR. ALİ ERGUR