Dynamic predictors of and corrections for attrition in Understanding Society
When variables of substantive interest in a study are related to drop-out, this threatens the validity of inferences drawn unless appropriate corrections are made. Identifying predictors of attrition informs these corrections, as well as retention strategies. The project evaluated whether variables related to key research themes in Understanding Society, including family separation, changes in job status, changes in health status, and changes in employment status are associated with attrition. Weights provided to users of the Understanding Society datasets provide a correction for non-random attrition; however, their performance requires evaluation. The project compared weighting against an alternative method of addressing non-random non-response: multiple imputation (MI). Parameter estimates and standard errors (indexing efficiency) from the two methods were compared in a series of example ‘main analyses’. Although the true parameter estimates are impossible to access, sensitivity analyses such as these, which use missing data strategies based on different assumptions, can help identify the extent to which missing data modelling choices can affect substantive conclusions and inform choices about how best to model drop-out. Collectively, the results informed recruitment and retention strategies by identifying subgroups vulnerable to attrition who may require a tailored approach. They are also be important for establishing whether and how current methods of addressing non-response in Understanding Society may benefit from elaboration or revision.
Read more about Aja’s work on her profile page.
