|
Abstract:
|
One approach to inference from sample surveys subject to unit nonresponse is to frame the problem in terms of the propensity to respond. But because the propensities are rarely known in practice, the nonresponse problem is transformed into one of estimation or prediction. Recently, a number of researchers have explored the possible application of techniques from the machine learning literature to the prediction of response propensities. For example, Rand researchers use generalized boosted models (GBM) to model response propensity for their 2014 RAND Military Workplace Study. We describe nonresponse adjustments that we provided for a parallel 2015 study of reservists, again using GBM. We then examine through simulation the relative merits of this approach compared to some other competitors, such as random forests, that have been recently studied.
|
