Title: Test of missing data mechanisms: An alternative to the Little test based on regression
Authors: Serguei Rouzinov - University of Lausanne (Switzerland) [presenting]
Andre Berchtold - University of Lausanne (Switzerland)
Abstract: Missing Data (MD) are common and occur in almost all surveys. The first step when working with MD is to determine whether they can be considered as missing completely at random (MCAR), missing at random (MAR), missing not at random (MNAR), or even a mixture of them (mixed MD mechanism). Nowadays, one of the most used methods for testing the MD mechanism is the Little likelihood ratio test. The aim of this method is to test whether a multivariate dataset has MCAR MD or not. However, this test has several limits. On the one hand, it is appropriate only for numerical data. On the other hand, it loses accuracy in presence of a mixed MD mechanism. An alternative approach based on a multiple linear regression model is proposed, whose dependent variable contains a certain amount of MD. The test draws conclusions from the comparison between the values predicted for the missing and the non-missing parts of the dependent variable. This test can be applied on categorical and numerical variables and it shows interesting properties in presence of a mixed MD mechanism. The test is described in details and numerical simulation results are provided.