|
Direct estimates of sampling variance in complex surveys are typically calculated with some form of replication, such as Balanced Repeated Replication, but these estimates can be volatile. Generalized Variance Functions (GVFs) are regression models fit to existing direct estimates of sampling variance to improve estimates of those variances (Wolter 2007).If multiple items exist then similar items may be grouped together to improve the model's strength, but proper item groupings may not be clear.
The Bureau of Labor Statistics Current Employment Statistics (CES) survey measures the monthly change in US establishments' jobs. The CES makes estimates for detailed industries as well as geographic domains such as States and Metropolitan areas. This research attempts to improve on GVFs with multiple linear regression models by using the regression tree algorithms implemented within GUIDE (Loh 2002); fitting linear regression models in each node. We use GUIDE to create specific models for groups of domains, or piecewise regression models for the covariates. We compare results from GUIDE to predictions using linear regression models with no explicit groupings.
|
