Online Program
Thursday, February 18 | |
PS1 Poster Session 1 & Opening Mixer sponsored by SAS |
Thu, Feb 18, 5:30 PM - 7:00 PM
Ballroom Foyer |
Missing Value Assumption in Modeling Repeated Measures Using Generalized Estimating Equations (303233)*Michael P Chen, U.S. Centers for Disease Control and PreventionKeywords: Model, GEE, repeated measures Generalized Estimating Equations(GEE) methodology can accommodate observations missing completely at random(MCAR) in repeated measures. However, its performance for a large number of missing repeated measures is unclear. In tuberculosis contact investigations, if we apply GEE, modeling contacts as repeated measures, missing observations range from 0 into the hundreds. We built a data set comprising 123 patients with 1-8 contacts(312 total) and several covariates and ran a GEE model. We increased the 8 contacts of one patient to 64, 124, and 160 completely at random (simulating MCAR), and ran the corresponding GEE models with the same correlation structure. The p-values of coefficients from the four models were compared. In the first model, p-values for 2 coefficients were significant at the 0.05 level (p1=.027 and p2=.039). In the other models, when missing observations relative to the maximum repeated measures (368,432,and 464 contacts) increase, the first coefficient remains significant (p1=.030,.029,.029) while the other does not (p2=.041,.062,.075). When most missing observations arise due to a few cases having a large number of repeated measures, GEE estimates may be biased.
|