Future censuses and surveys are almost certain, for reasons of cost-savings, to make heavy use of observational administrative-records databases and maybe of databases generated by web searches or optional (self-selected) web-respondents. To make Census Bureau data products based on such observational data fully representative of the general population (in mathematically demonstrable ways, as our current surveys are), new sampling designs and analysis methods will have to be developed. This talk describes new research concerning probability sample survey designs for data collection which will yield consistent population estimates for survey attributes under various model assumptions (i) on the propensities for population members to be in an observational database; (ii) on the coverage by a population frame on the population members not captured by the database; and (iii) on nonresponse in the designed probability sample.
The formal machinery developed here is intended to generalize models currently in use for nonresponse and for frame non-coverage.
|