ABSTRACT
In design-based inference that underpins much of probability-based sample surveys, it may seem counter-intuitive that models can be used during data collection to improve the inferential properties of the survey data. Most models can be characterized as having two major goals, error reduction and cost reduction. There are several main reasons why the explicit use of models can be beneficial: they allow the formalizing of data-driven decisions, they facilitate the use of a wider array of data, and they provide the means to balance multiple objectives. Models for survival analysis are particularly appropriate for call record data, multilevel models can be suitable for data with nested structure, and even both together. There are two main types of estimate-level measures: the fraction of missing information, and the association between a survey variable and response propensity. The survey presents the ignorable non-response bias, assuming that the information that was used to estimate the response propensities will be used in post-survey adjustments.
