Laney Light presented a paper that evaluated several missing data imputation approaches.
Light, L., Harris, K. & Hubbard, F. (2017, August). Balancing Bias, Precision, and Sample Size Recovered in Determining a Practical Missing Data Imputation Approach. Presentation at the Joint Statistical Meetings, Baltimore, MD.
The KDQOL-36 is a short form version of the Kidney Disease Quality of Life Instrument, including 12 items comprising the SF-12 Physical and Mental Component Summary scales (PCS and MCS). Researchers administered this survey to 9,071 Medicare beneficiaries with end-stage renal disease participating in a quality improvement demonstration program.
In a recent presentation, Laney Light described this study, explaining that at least one SF-12 item was missing for 21% of respondents, resulting in missing PCS and MCS scores. In this presentation, Light compared three data imputation approaches: item substitution (substituting a correlated item from the same respondent); multiple imputation; and mean imputation, repeated for a maximum number of missing items ranging from 1 to 12.
Researchers also described how they simulated missing values from surveys with complete SF-12 data, replicating missing data patterns observed in the original data. They also imputed missing values and compares true MCS and PCS scores against those calculated from imputed data. Evaluation criteria included bias, precision measured by mean square error, and sample size recovered. In this presentation, researchers evaluated the pros and cons of each approach and also considered the ease of implementation and replicability.