Updating for multiple settings

Background

Updating of a prediction model can be considered for a single new setting, but also for a range of settings, such as multiple hospitals. We can consider such settings as parts of an underlying superpopulation, making them to some extent related. We first quantify the distribution of differences between settings, and subsequently update the model to setting-specific values considering this distribution. This approach is well possible with random effects models or Empirical Bayes estimation. We illustrate the approach for logistic regression models.

We may specifically be interested in differences between centres in the context of quality assessment. We illustrate modern methods for estimation of differences and rank ordering between centres for patients with stroke (“provider profiling”).