209749 Multiple imputation via chained equations of SF-12 health scores for a longitudinal study of displaced population from Hurricane Katrina

Monday, November 9, 2009: 9:30 AM

Yoon Soo Park, MS , National Center for Disaster Preparedness, Mailman School of Public Health, Columbia University, New York, NY
Tasha Stehling-Ariza, MPH , National Center for Disaster Preparedness, Columbia University, New York, NY
David M. Abramson, PhD MPH , National Center for Disaster Preparedness, Columbia University, New York, NY
The SF-12 is a condensed 12-item survey that encompasses all mental and physical health components associated with the SF-36 (Ware et al., 2002). The purpose of this study is to assess the recovery of item-level missing values in a longitudinal study of displaced population from Hurricane Katrina with three waves of data (n=1,079). To date, there are no algorithms proposed by the survey author to deal with item-level missing data, which prevents calculating the SF-12 scores. Past imputation studies (Liu et al, 2005; Perneger & Burnand, 2004) using multiple imputation rely heavily on the multivariate normality assumption. Multiple imputation via Chained Equation (ICE) overcomes this limitation and imputes data using the conditional density of the variable given all others in the equation (Royston, 2005).

Two different approaches were used for this evaluation. The first compared standard error changes in internal consistency estimates using nonparametric bootstrap and jackknife methods when three waves of item-level missing data is imputed using ICE. The second examined cases with complete data on all three waves (n=613) by randomly removing them at selected conditions. The imputed values were studied by bias and proportionate variance comparisons to the true value. Preliminary results using ICE with 100 iterations and 5 cycles showed the standard errors of internal consistency calculated for the nonparametric bootstrap decreased from 0.166 (=0.905) to 0.003 (=0.904) with p<0.001 for the non-imputed and the imputed data, respectively, suggesting the effectiveness of this method in longitudinal data analysis.

Learning Objectives:
To analyze different methods of imputing longitudinal data. Identify the most effective imputation method.

Presenting author's disclosure statement:

Qualified on the content I am responsible for because: I am a Ph.D. Candidate in the Measurement, Evaluation, and Statistics program at Columbia University. I study psychometric inferences of missing data. I am also the Data Manager/Analyst at the National Center for Disaster Preparedness at Columbia University where the data for the current study was collected and analyzed.
Any relevant financial relationships? No

I agree to comply with the American Public Health Association Conflict of Interest and Commercial Support Guidelines, and to disclose to the participants any off-label or experimental uses of a commercial product or service discussed in my presentation.