Missing data are ubiquitous in real world data analysis tasks and, since the mechanism leading to missingness is frequently unknown,
present a serious challenge to predictive modelling. For example, values could be missing due to a broken sensor or communication pipeline,
or missing in the responses to a formor questionnaire, and these situations could arise due to chance, due to the state of other values we can
measure, or due to the actions of an adversary or unwilling respondent. In particular values may be ‘missingcompletely at random’ (MCAR),
in which case the absence of a value depends on neither observed nor unobserved values, ‘missing at random’ (MAR) in which case absence can
depend only on observed values, and ‘missing not at random’ (MNAR) in which case absence depends on unobserved values.
The MCAR and MAR cases are often referred to as ‘ignorable’ cases, since the missing values can be estimated from
the sample when modelling the data generator. On the other hand when values are MNAR one cannot relyon sample information by itself,
since the presence or absence of a variable may depend on its own unobserved value, for example.
Various approaches have been proposed to manage the problem effectively, especially from the perspective of modelling.