Contributions and limits of mathematical models in epidemy modelling - Are models good?

From the literature, it can be stated that numerous and important advantages may be drawn from a good model: explanations of the effects of medical or-biological factors, forecast of results of public health policies, assessments of sociological features. Models are now necessary tools, essential or just useful for bur understanding or our investigations in epidemiology, especially of communicable diseases. The semantic shift from "good model" to "model" is neither insignificant-nor fortuitous. The difference is worth of some attention. A quick survey of models shows that model building relies on hypotheses and that methods of building suffer front discrete bur fundamental flaws. The difficulties stem in the choice of parameters whose values define and govern the model. The estimation of parameters and results cannot be as straightforward as in regression-for instance. The kind of complexity encountered in dynamic modelling does not allow to instinctively identify the probably right and the obviously wrong. Hence the capital role of validation steps. The most common type of validation is the reproduction of already observed values. This is doubly unsatisfactory, first because an apparently correct conclusion is not the proof of a correct reasoning, then because the obtained results may be mechanical consequences of premises, a fact hidden behind the complexity of the model. Improvements of model quality can be achieved, without exclusion; by two ways actively worked out in the current literature: improvement of parameter estimation methods and development of stability, precision and sensitivity of forecasts. A third way is to make models more intellectually tractable,by clarifying the models but also giving more insight to the users. (C) 2003 Editions scientifiques et medicales Elsevier SAS. Tous droits reserves.