In the epoch of precision cosmology, we often face the problem of deciding whether or not cosmological data
support the introduction of a new quantity in our model.  For instance, we might ask whether it is necessary to
consider a running of the spectral index, an extra isocurvature mode, or a non-constant dark energy equation of
state.  These kinds of questions should be considered in the framework of model selection, rather than as
parameter estimation problems.</br></br>

I show that the usual concept of "confidence interval" is not always informative in this respect and that
Bayesian evidence correctly takes into account the information content of the data.  Evaluation of Bayes factors
allows one to confirm theoretical predictions, instead of just disproving them as in sampling statistics, in
accordance with our intuition.</br></br>

I illustrate the power of some approximate methods by applying them to the spectral tilt for the perturbations 
and to the spatial curvature of the universe.  A new method to formulate predictions for Planck quality 
measurements is also discussed.