Analysing Longitudinal Studies with Non-response: Issues and Statistical Methods

The impact of attrition and non-response in birth cohort studies: a need to incorporate missingness strategies

“…Both adjustment strategies assume MAR. Under MI, there are several different approaches appropriate to the type of data and/or data structure (Carpenter & Plewis, 2011). In this paper we demonstrate that analyses which draw upon the longitudinal history of the cohort require strategies to use all of the available data, including any partial information (incomplete wave responses).…”

“…In order to tackle these problems and make best use of all of the available data in any analysis it becomes necessary to make assumptions about how a proportion of our data came to be missing at all. Many authors (for example Carpenter & Plewis, 2011) term this process to be 'the missingness mechanism'. Following Rubin (1976), Little and Rubin (2002), Carpenter and Plewis (2011) and others we adopt a typology of missingness mechanisms described as 'missing completely at random' (MCAR), 'missing at random' (MAR) and 'missing not at random' (MNAR).…”

“…Many authors (for example Carpenter & Plewis, 2011) term this process to be 'the missingness mechanism'. Following Rubin (1976), Little and Rubin (2002), Carpenter and Plewis (2011) and others we adopt a typology of missingness mechanisms described as 'missing completely at random' (MCAR), 'missing at random' (MAR) and 'missing not at random' (MNAR). Briefly, MCAR implies that the probability of not answering a particular question is uncorrelated with the characteristics of the respondent, and in any longitudinal survey it means that the probability of dropping--out from any wave is uncorrelated with the characteristics of the CM.…”

“…In this way, it is proposed that the unobserved variables that simultaneously influence both the outcome and the missingness are captured by the residuals in the two models, which are allowed to correlate. Technically, the challenge for the analyst is to identify variables (or instruments) for the missingness model which predict the probability of missingness but do not correlate with the substantive outcome (see Carpenter & Plewis (2011) for an illustration using NCDS data).…”

“…for handling a multilevel structure see Goldstein, Carpenter & Browne (2014) and the type of variables (e.g. continuous, ordinal or multinomial variables) see Nathan (1983), Nathan & Holt (1980), Pfeffermann (2001) and Carpenter & Plewis (2011)). …”