In-depth Model Comparison

Eaton JW, et al. PLoS Med 2012.

There were many questions raised in the comprehensive model comparison undertaken in work package on Treatment as Prevention when models did not produce similar outcomes even when simulated interventions were standardized. There are several hypotheses for these remaining differences and identifying the true root causes will be of substantial theoretical interest (in terms of identifying the influence of unsupported priors) as well as practical interest (so that results of simpler models can be confidently extrapolated). Building on sustained enthusiasm for the project, the Consortium has initiated a second concerted phase of model comparison which will be designed to tackle these issues.


Previous work by the HIV Modelling Consortium found that there is substantial variation between models’ predictions about the impact of the same HIV treatment intervention, especially over the long-term, and that there were structural differences between models’ representations of all of the processes involved, including the natural history of HIV infection, the determinants of HIV transmission, population structure, and sexual mixing. This raises the question of which aspects of the models are the most important drivers of variation in model predictions. Longstanding theoretical work in HIV and STI modelling has demonstrated that different representations of sexual mixing can affect the predicted impact of an intervention—in particular, that models which represent the spread of HIV on networks of individuals with homogenous risk behaviours tend to show greater impact of interventions than models which represent the spread of HIV on sexual networks of individuals stratified into risk groups with large differences in the propensity to form new partnerships.


This work package aims to better characterise what aspects contribute to the variation in models’ predictions about the impact of interventions in generalised heterosexual HIV epidemics. The approach will be to standardise aspects of model structure and parameters in order to quantify the influence of remaining components that are allowed to vary. We will particularly focus on isolating the effect of different approaches to modelling sexual behaviour and sexual mixing.