大阪大学　数理・データ科学教育研究センター
Center for Mathematical Modeling and Data Science,Osaka University

Open seminar on Data Science

Partha Lahiri (University of Maryland), Masayo Y. Hirose (Kyushu University), Shonosuke Sugasawa (The University of Tokyo)

1. ホーム
2. セミナー・集中講義
3. セミナー
4. Open seminar on Data Science

大阪大学 数理・データ科学セミナー データ科学セミナーシリーズ 第44回

Open seminar on Data Science

Partha Lahiri (University of Maryland), Masayo Y. Hirose (Kyushu University), Shonosuke Sugasawa (The University of Tokyo)

13:30-14:30
Presenter: Partha Lahiri (University of Maryland)
Title: Small Area Estimation in Presence of Linkage Errors
Abstract: Small Area Estimation is in high demand due to its usefulness in policy making and regional planning. Availability of good auxiliary data is an important component of Small Area Estimation. In the age of big data, we can get access to unprecedented amount of data. The challenge is how to combine data from multiple sources. One potential solution to this problem is record linkage, which is a statistical methodology to link records from multiple sources representing the same population unit. Huge literature is available for both small area estimation and record linkage. But until my keynote address at the 2017 SAE meeting in Paris, there was no suggestion to combine these two important research areas. In this paper, we suggest a unified way to integrate small area estimation with record linkage. Linkage errors are inevitable in the linked data sets because of the unavailability of error-free unique identifiers and because of possible errors in measuring or recording. In presence of linkage errors, small area estimators are subject to a substantial increase in mean square error. Hence, linkage errors should not be ignored when conducting small area estimation. To correct for linkage errors in small area estimation, we propose a general model that integrates small area estimation model, linkage error model and record linkage latent class model. The proposed model accounts for uncertainty in both record linkage process and small area estimation. We propose an estimating-equation-based approach to parameter estimation and and jackknife method of mean square error calculation. A Monte Carlo simulation study is performed to evaluate our proposed methods. Several methods for computation simplification are also provided. The talk is based on collaborative research with my former PhD student Dr. Ying Han.

14:30-15:30
Presenter: Masayo Y. Hirose (Kyushu University)
Title: A Parametric Empirical Bayes Confidence Interval in the Presence of High Leverage for Area Level Model
Abstract: Parametric empirical Bayes confidence interval is widely used especially when the sample size within each area is not large enough to make reliable direct estimates. Especially, there are already exist several second-order corrected parametric Bayes confidence intervals which asymptotically achieves a smaller length than that of the confidence interval based on the direct estimates. However, these intervals may have several practical issues with finite number of areas. In this talk, we will introduce a new parametric empirical Bayes confidence interval which achieves several properties even in the presence of high leverage and small number of areas. Moreover, we will also mention our confidence interval being more tractable, compared with some existing intervals. Furthermore, we will also report the results of our simulation study for showing overall superiority of our confidence interval method over the other methods.

15:45-16:45
Presenter: Shonosuke Sugasawa (The University of Tokyo)
Title: An Approximate Bayesian Approach to Regression Estimation with Many Auxiliary Variables
Abstract: Model-assisted estimation with complex survey data is challenging because the sampling design can be informative, and ignoring it may produce misleading results. Also, when there are many auxiliary variable, selecting significant variables associated with a response variable would be taken into account to achieve efficient estimation of population parameters of interest. In this paper, we formulate the survey regression estimator in the framework of Bayesian inference and incorporate shrinkage prior for regression coefficients to precisely estimate regression coefficients under many auxiliary variables, which enables us to get not only efficient point estimates but also reasonable credible intervals for population means.