Dealing with missing data in Canadian Longitudinal Studying on Aging (CLSA): a comparison of complete case analysis and multiple imputation method using DXA data as an example

Dealing with missing data in observation study is a challenge for researchers. If handled inappropriately, missing data may potentially biased results, decreased statistical power and underestimated uncertainty, and reduced the validity of conclusion. Multiple imputation (MI) has been accepted as the gold standard method in many areas of research, if appropriately implemented, provides unbiased and valid estimates of associations on the basis of information from full data, under the assumption of missing at random. However, appropriate imputation method may vary depending on the data type (i.e., continuous and categorical data), amount of missing data, and mechanisms of missing data. The present study will compare the results of the complete case analysis and MI at four levels of missingness (e.g., 5%, 10%, 20% and 40%) with DXA BMD T-scores used in a fracture risk model as an example.