Methods to Incorporate Conditional Dependencies in Diagnostic Simulation Models

OBJECTIVES: Asthma is one of the most diagnosed diseases in adults and children. As there is no “gold standard” test, a diagnosis is often reached through the employment of various tests, none of which can be entirely depended upon. Where multiple tests are available, some degree of conditional dependency (i.e. correlation between test results within disease state) is expected. Such dependencies can arise due to tests measuring the same phenomenon or being affected by some characteristics of the patients. When setting up a clinical and cost-effective diagnostic algorithm, failure to account for conditional dependency could lead to erroneous conclusions and inefficient decision-making, as shown in previous literature.

METHODS:

A statistical economic model was developed to investigate the most cost-effective diagnostic pathway for asthma in the UK. Individual level data from a study where adults underwent a variety of tests for asthma was used to estimate the correlation between tests.

Two different approaches were used to incorporate conditional dependency in the model. With the direct approach, joint sensitivity and joint specificity for any test combinations of interests were estimated using the adults’ data. As this was not possible in children, an alternative indirect approach was employed. Here, pseudo individual level data were simulated using a multivariate probit model, with accuracy estimated from children but correlations borrowed from the adults’ data.

RESULTS: Incorporating conditional dependencies significantly affected the results of the diagnostic analysis, particularly for tests with a statistically significant correlation. We further demonstrated a multivariable probit modelling approach to allowing for conditional dependence, which could be more generally applicable.

CONCLUSIONS: The cost-effectiveness results of strategies with highly correlated tests were highly sensitive to allowing for conditional dependence, suggesting that this is an important feature to account for in future models of the accuracy of multiple tests used in combination.