Tez No	İndirme	Tez Künye	Durumu
402350		Dealing with selection bias in multilevel observational studies: An evaluation of propensity score and direct estimation procedures / Yazar:SUNGUR Danışman: DR. WALTER LEITE Yer Bilgisi: UNIVERSITY OF FLORIDA / YURTDIŞI ENSTİTÜ Konu:Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol = Computer Engineering and Computer Science and Control Anahtar Kelime:	Onaylandı Doktora İngilizce 2015 108 s.
Secondary data analyses are popular among educational researchers because of the increased availability of nationwide longitudinal studies. One issue with using secondary data to evaluate program effectiveness is potential selection bias due to a lack of random assignment of participants to conditions. Rubin's potential outcomes framework is commonly used to address the lack of random assignment issue. Another complication is that a hierarchical data structure is common for large scale datasets and cluster effects should be accounted for in the analysis. Both of those issues should be addressed properly, so that the analysis results would be unbiased. Propensity score methods are commonly used to mitigate the selection bias problem when data is not clustered, but little research has been done on how to account for cluster effects in propensity score estimation. In addition, no research has been performed on the direct estimation of average treatment effects in multilevel observational studies yet. I conducted a Monte Carlo simulation study to understand the performance of various hierarchical logistic regression models and data mining methods when the treatment assignment mechanism is complex and varies across clusters. I specifically focused on propensity score based procedures in Chapter 2 and direct estimation based procedures in Chapter 3. Results (i) guide the applied researcher in terms of how to implement propensity score and direct treatment effect estimation procedures, (ii) provide acceptable levels of model misspecifications in propensity score and direct treatment effect estimation methods, and (iii) suggest a preferable approach to select most appropriate propensity score and direct treatment effect estimation procedure under specific conditions. The results indicated that the parametric modeling is preferable over data mining in both propensity score and direct estimation procedures. Dealing with selection bias through propensity scores requires more observations than dealing with selection bias through direct estimation. In the propensity score procedures, the effect of model misspecification was not apparent in the relative bias of the ATT estimates but standard errors were inflated. Finally, standard errors that are obtained from direct estimation tend to be overestimated when the dependence between observed and counterfactual outcomes is ignored.