Nonidentifiability in model calibration and its implications for medical decision making.

“

1) What does Nonidentifiability in model calibration mean? That there is more than unique solution to the calibration problem. That is, more than one parameter set could make your model replicate the target data. 2) Why should we care about Nonidentifiability in model calibration? Because when using the decision model to evaluate different policies/strategies/interventions on different but equally good fitting parameter sets could result on different recommendations! 3) How could we check if we face Nonidentifiability in model calibration? In our paper, we provide a few approaches to identify nonidentifiability, such as collinearity analysis and profiling the likelihood 4) What can we do to get rid of Nonidentifiability in model calibration? Either find new calibration targets, constrain the parameter space (e.g., based on expert knowledge) or adopt a Bayesian calibration approach. 5) What if we cannot get rid of nonidentifiability? Conduct a Bayesian calibration approach that will account for nonidentifiability (and propagate with PSA) or conduct a scenario analysis by evaluating your decision model on the different but equally good fitting parameter sets 6) Want to learn more about calibration and nonidentifiability in medical decision making and how to conduct such analyses in #stats? Join us in our short course in Model Calibration using #rstats at #SMDM18 @socmdm in Montreal!

”