Evaluating effectiveness of preoperative testing procedure: some notes on modelling strategies in multi-centre...

Rationale In technology assessment in health-related fields the construction of a model for interpreting the economic implications of the introduction of a technology is only a part of the problem. The most important part is often the formulation of a model that can be used for selecting patients to submit to the new cost-saving procedure or medical strategy. The model is usually complicated by the fact that data are often non-homogeneous with respect to some uncontrolled variables and are correlated. The most typical example is the so-called hospital effect in multi-centre studies.

We show the implications derived by different choices in modelling strategies when evaluating the usefulness of preoperative chest radiography, an exam performed before surgery, usually with the aim to detect unsuspected abnormalities that could influence the anaesthetic management and/or surgical plan.

We analyze the data from a multi-centre study including more than 7000 patients. We use about 6000 patients to fit regression models using both a population averaged and a subject-specific approach. We explore the limitations of these models when used for predictive purposes using a validation set of more than 1000 patients.

We show the importance of taking into account the heterogeneity among observations and the correlation structure of the data and propose an approach for integrating a population-averaged and subject specific approach into a single modeling strategy. We find that the hospital represents an important variable causing heterogeneity that influences the probability of a useful POCR.

We find that starting with a marginal model, evaluating the shrinkage effect and eventually move to a more detailed model for the heterogeneity is preferable. This kind of flexible approach seems to be more informative at various phases of the model-building strategy.