Handling missing data 1
|Convenor:||Professor George Ploubidis|
|Affiliation:||University College London|
Selection bias, in the form of incomplete or missing data is unavoidable in surveys. It results in smaller samples, incomplete histories, lower statistical power and bias in sample composition if missingness is related to the observed and unobserved characteristics of respondents. It is well known that unbiased estimates cannot be obtained without properly addressing the implications of incompleteness. In this session we focus on item missingness, survey non-response, and attrition over time in longitudinal surveys. We aim to identify best practices when dealing with missing data.
Under Rubin’s framework, three types of missingness exist: Missing completely at ransom (MCAR) where the likelihood of response is unrelated to the respondents’ characteristics. Missing at random (MAR) where the likelihood of response is explained by the observed characteristics of respondents, and missing not at random (MNAR) where the likelihood of response is related to both observed and unobserved characteristics of respondents.
The objective of our session is to examine the principled techniques commonly used to deal with missing data. These include, inverse probability weights, multiple imputation, and full information maximum likelihood (FIML). All techniques rely on the MAR assumption, and therefore, their plausibility depends on the ability of the researcher to identify the predictors of response.
Contributors are welcomed to contrast these techniques with other procedures such as case-wise deletion, mean replacement, regression imputations, selection models (e.g. Heckman selection models), and others. Moreover, theoretical, empirical, and substantive applications of these techniques will be considered for presentation.