Grant: Discovery Grants Program - Individual
University of Waterloo
$15,000.00
Solving Missing Data Problems Using the Empirical Likelihood Method
FY 2016-17
Summary
Data collected from statistical studies are often incomplete. Missing data can raise considerable challenges to statistical analysis. This research program will develop novel statistical methodologies to solve several important missing data problems via the innovative use of the empirical likelihood (EL) theory.
The first aim is to obtain multiply robust estimators in regression analysis when data are incomplete. In the missing data literature, doubly robust estimators have been studied extensively. They allow a single model for the missingness mechanism and a single model for the data distribution. Estimation consistency is achieved under the assumption that one of these two models is correctly specified. This assumption is restrictive in practice, as the true data generating process is usually unknown. I will develop methods that allow multiple models for both the missingness mechanism and the data distribution, yielding estimators that are consistent if any one of those multiple models is correctly specified. The idea is to assign the complete cases with proper weights that summarize the information contained in all models. The EL method will play a central role in calculating the weights. The new estimators will be more robust against model misspecification than the existing doubly robust estimators. Regression analysis in the setting of both cross-sectional and longitudinal data will be considered.
The second aim is to refine the EL method under highly correlated moment restrictions so that it can better serve this research program. It is well known that the EL method often has unstable performance when the moment restrictions are highly correlated, especially if the sample size is small. This shortcoming may undermine the capacity of EL in dealing with missing data problems when the multiple models postulated result in high collinearity among the moment restrictions. There is no formal solution currently available in the literature. I plan to exploit the EL method by selecting the most informative components to use after a principal component decomposition of the moment conditions. Extensive simulation experiments will be conducted to study the numerical behavior of different selection criteria, and theoretical properties will be established.
The third aim is to develop a general test for the missing completely at random (MCAR) mechanism that can be generally applied instead of targeting a specific data type. The MCAR mechanism ensures that data analysis can be done by simply ignoring subjects with missing observations. However, it is often assumed without being formally tested, leading to questionable conclusions. There has been limited literature on testing the MCAR assumption, and each existing test only targets a specific data type, making it difficult to find the appropriate test to use in general. I plan to propose a test employing the empirical weights on the complete cases calculated by the EL method. A large deviation of these weights from being uniformly distributed reflects evidence against the MCAR assumption. The proposed test will not assume any parametric model for the data, and thus its application is not restricted to a specific data type.
This research program will provide a unified framework for testing the MCAR assumption and for estimation in regression analysis with missing data. Due to the empirical likelihood framework I plan to adopt, the new estimators will have high robustness against model misspecification, and the new tests will have a much broader range of applications than the existing ones. This research will benefit many scientific fields in Canada where missing data problems exist, including the fields of life sciences, sociology, engineering, economics, and finance.
Program
Discovery Grants Program - Individual
Recipient
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2014-05055
Name nom
Han, Peisong
Department département
Statistics and Actuarial Science
Organization id
33
Institution établissement
University of Waterloo
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2016
Competition year année de concours
2014
Award amount
15000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
Discovery grants programs
Group fr
Programmes de subventions à la découverte
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1100
Area of application group en
Health, education and social services
Area of application group fr
Santé publique, éducation et services sociaux
Area of application en
Health, education and social services
Area of application fr
Santé publique, éducation et services sociaux
Research subject code
3004
Research subject group en
Statistics and probability
Research subject group fr
Statistique et probabilité
Research subject en
Applied statistics
Research subject fr
Statistique appliquée
Partie
3
Nb partie
5
Application title
Solving Missing Data Problems Using the Empirical Likelihood Method
Keyword
NSERC
Application summary
Data collected from statistical studies are often incomplete. Missing data can raise considerable challenges to statistical analysis. This research program will develop novel statistical methodologies to solve several important missing data problems via the innovative use of the empirical likelihood (EL) theory.
The first aim is to obtain multiply robust estimators in regression analysis when data are incomplete. In the missing data literature, doubly robust estimators have been studied extensively. They allow a single model for the missingness mechanism and a single model for the data distribution. Estimation consistency is achieved under the assumption that one of these two models is correctly specified. This assumption is restrictive in practice, as the true data generating process is usually unknown. I will develop methods that allow multiple models for both the missingness mechanism and the data distribution, yielding estimators that are consistent if any one of those multiple models is correctly specified. The idea is to assign the complete cases with proper weights that summarize the information contained in all models. The EL method will play a central role in calculating the weights. The new estimators will be more robust against model misspecification than the existing doubly robust estimators. Regression analysis in the setting of both cross-sectional and longitudinal data will be considered.
The second aim is to refine the EL method under highly correlated moment restrictions so that it can better serve this research program. It is well known that the EL method often has unstable performance when the moment restrictions are highly correlated, especially if the sample size is small. This shortcoming may undermine the capacity of EL in dealing with missing data problems when the multiple models postulated result in high collinearity among the moment restrictions. There is no formal solution currently available in the literature. I plan to exploit the EL method by selecting the most informative components to use after a principal component decomposition of the moment conditions. Extensive simulation experiments will be conducted to study the numerical behavior of different selection criteria, and theoretical properties will be established.
The third aim is to develop a general test for the missing completely at random (MCAR) mechanism that can be generally applied instead of targeting a specific data type. The MCAR mechanism ensures that data analysis can be done by simply ignoring subjects with missing observations. However, it is often assumed without being formally tested, leading to questionable conclusions. There has been limited literature on testing the MCAR assumption, and each existing test only targets a specific data type, making it difficult to find the appropriate test to use in general. I plan to propose a test employing the empirical weights on the complete cases calculated by the EL method. A large deviation of these weights from being uniformly distributed reflects evidence against the MCAR assumption. The proposed test will not assume any parametric model for the data, and thus its application is not restricted to a specific data type.
This research program will provide a unified framework for testing the MCAR assumption and for estimation in regression analysis with missing data. Due to the empirical likelihood framework I plan to adopt, the new estimators will have high robustness against model misspecification, and the new tests will have a much broader range of applications than the existing ones. This research will benefit many scientific fields in Canada where missing data problems exist, including the fields of life sciences, sociology, engineering, economics, and finance.