IPW vs Imputation for Non-response : a simulation study

Sammanfattning

The non-response is a growing issue and it is effecting the result of survey quality, it can there- fore not be ignored without substantial consequences. There are many methods on how to deal with non-response, both on how to avoid the non-response and how to minimise the con- sequences, like the non-response bias. This thesis aims to compare a weighting adjustment, IPW, with a imputation method, regression imputation. Further the result will be compared with poststratification, which is widely used. A Monte Carlo simulation is done with different response mechanisms for both a response rate of 50% and 30%. The Horvitz-Thompson esti- mator is used to estimate the total of the study variable. The result shows that the imputation method has a low bias and a low standard deviation when the response has the MCAR and MAR mechanism. The IPW method has a high bias, and the standard deviation is rather high. The poststratification has more reliable estimates in the sense that it performs rather well in all cases.