Longitudinal Data Observed Over Unequal Time Intervals

This monograph extends previous work on the
evaluation of GEE performance in parameter
estimation with correlation structure
misspecification. It examines a new structure to
model the correlation between binary responses
observed over unequal time intervals. The proposed
unequal-time correlation structure accounts for
unequal intervals by fixing the correlation
coefficient and interval length between the first 2
observations and by adjusting the correlations for
subsequent pairs by their relative interval length.

To evaluate this structure, correlated binary
responses under the logistic model were generated
with the unequal time correlation structure for a
full range of positive correlations and strength of
association under different scenarios of number of
measurements, sample size and prevalence of
exposure. The model was evaluated when the working
correlation structure was misspecified as
exchangeable, 1st order autoregressive,
unstructured, or when alternating logistic
regression was employed.

Autorentext
Soe Soe Thwin received her MS in Epidemiology from University
of Washington and PhD in Biostatistics from Boston University,
USA. She is currently a statistician at the Massachusetts
Veterans Epidemiologic Research and Information Center and
consults for Boston University School of Medicine.

Klappentext
This monograph extends previous work on the evaluation of GEE performance in parameter estimation with correlation structure misspecification. It examines a new structure to model the correlation between binary responses observed over unequal time intervals. The proposed unequal-time correlation structure accounts for unequal intervals by fixing the correlation coefficient and interval length between the first 2 observations and by adjusting the correlations for subsequent pairs by their relative interval length. To evaluate this structure, correlated binary responses under the logistic model were generated with the unequal time correlation structure for a full range of positive correlations and strength of association under different scenarios of number of measurements, sample size and prevalence of exposure. The model was evaluated when the working correlation structure was misspecified as exchangeable, 1st order autoregressive, unstructured, or when alternating logistic regression was employed.