| Description: | [DMS Seminar] Dr. Sumanta Adhya (West Bengal State University, Barasat, India) -- Model-Based Inference for Finite Population Distribution |
| Date: | Friday, Apr 20, 2018 |
| Time: | 4:15 p.m. - 5:15 p.m. |
| Venue: | G09, Lecture Hall Complex |
| Details: | This talk mainly focuses on efficient estimation of finite population distribution of a survey variable (both categorical and continuous) by adopting model-based (or, predictive) approach to survey inference when one or more auxiliary variables are known for the whole finite population. Traditional survey estimators are design-based and the basis of inference is repeated sampling from the finite population. When auxiliary information is available, a stochastic relation between response and auxiliary variables (called superpopulation model) is used to increase the precision of the estimators within design-based framework. The main feature of these estimators are their model-robust property; that is, they are either exactly or approximately design-unbiased irrespective of underlying models. Instead of design-based approach we consider a standard prediction problem that predicts unknown finite population quantity given all available information; that is, values of sample responses, complete auxiliary information and the underlying model. This approach is seriously criticized for its heavy model dependency. We find a way around by developing predictors which either depend on nonparametric regression models or are based on parametric regressions equipped with a simple model selection procedure. We also introduce bootstrap based hybrid variance estimators of the predictors. These hybrid estimators avoid computational burden of bootstrapping in finite population set-up by including positive features of analytical estimators partially. |
| Calendar: | Seminar Calendar (entered by sushil.gorai) |