31
for a particular toxicity type instead. The proposed approach provides alternative
model-based posterior inference. In accordance with the natural structure of the
data, our model treats toxicity grade as an ordinal data. The proposed model ac-
counts for the (high) dependence across different toxicities within the same patient.
The proposed model allows for extensions to more complicated designs by appropri-
ate changes in the linear model (2.3). For example, one could accommodate repeated
observations of adverse events by replacing the latent variable vij in (2.3) by vijh for
the hth repeated observation of type of toxicity j for patient i, and defining a new set
of random effects Rτj.
This model also has interesting applications in other areas such as health out-
comes research and clinical trial design. For example, some studies have shown that
even when treatments are known to be effective, many patients who could benefit
from them are not getting these treatments. Beta blocker medication, given after
heart attacks, can reduce mortality; blood-thinning medication can prevent stroke;
and thrombolytic therapy given immediately after a heart attack can reduce the dam-
age from the attack. The outcome instrument has focused on assessing the overall
level of functioning after receiving the treatment conditional on patients’ prognostic
characteristics. The overall level of functioning is a quantified variable on an ordinal
scale. Therefore, by assessing the ordinal outcomes within each category, health out-
come researchers will be able to identify and address the barriers to better care and,
eventually, translate these findings into practical strategies to improve care.
One critical issue is the choice of the size of the mixture in modeling ordinal
outcomes. We suggested as a rule of thumb to set the size of mixture, G, equal to
the number ordinal levels minus two (K — 1). Alternatively, one could treat G as an
unknown parameter and use reversible jump MCMC.
In summary, we have introduced an approach for flexible, model-based inference
for the adverse events reported in a Phase III clinical trial. The model includes
dependence across adverse events for the same patient. The computational effort of