B1600
Title: Fitting marginalized two- part joint models to semi-continuous medical cost and survival from complex surveys
Authors: Mohadeseh Shojaei Shahrokhabadi - University of Pretoria (South Africa) [presenting]
Din Chen - University of Pretoria (South Africa)
Abstract: To address the medical costs data problems including right skewness, clumping at zero, and censoring due to death and incomplete follow-up, Marginalized Two-part Joint Models (MTJM) have been developed. When the primary interest is to estimate covariate effects on the average costs amongst the entire population of both users and non-users, MTJM may be most useful. In the original formulation of MTJM, a Log-normal distribution with a constant variance parameter was assumed for the positive values. We extend this model, allowing the positive values to follow a more flexible distribution- Generalized Gamma- which takes the Log-normal distribution as a special case. We use a simulation study to compare the performance of these two models with respect to bias, coverage, and efficiency. In addition, the performance of these methods is compared through application to a set of real electronic health record (EHRs) data collected in Iran. The simulation results show when the response distribution is unknown or mis-specified, that the Generalized Gamma provides a potentially more robust alternative estimator to the log-normal. For analyzing semi-continuous data with clumping at zero, researchers should consider which method is consistent with research objectives, and simultaneously appropriate for the data available.
