In a probabilistic safety assessment for nuclear power plants, an important issue is the treatment and quantification of the uncertainties involved in each step of the system safety or accident analysis. There are two main types of uncertainties that should be explicitly considered in the analysis, i.e., parameter uncertainties contained in the model describing the behavior of real systems or accidents, and modeling uncertainties due to the imperfect description of the model itself. The latter case indicates a representation of imprecision in the analyst’s knowledge about models or their predictions. Although the field of uncertainty analysis has progressed to the point that several studies have been carried out that maintain a distinction between parameter and model uncertainty, in recent times, the model uncertainty analysis has indeed been less complete than that of the former type. However, there are important advantages to explicit consideration of the modeling uncertainty in risk analysis. The most important advantage is that it mitigates the overconfidence that can occur when a single model is used to make predictions since uncertainty bounds tend to be more realistic when a range of plausible models is considered. The second advantage is that it facilitates scientific communication because scientifically defensible analyses that explicitly incorporate a range of models obviate the problem of arguing over whose model is correct. The third advantage is the enhancement of credibility in the predictions or final outcomes. For these reasons, the modeling uncertainty should be incorporated into the current context of uncertainty analysis. A formal approach on the expression of highly uncertain models and its assessment within a probabilistic framework are provided. The basic idea of the current procedure is that the quantification of modeling uncertainties can be made by combining all the uncertainties assigned to alternative models into a probability distribution (or a family of probability distributions) about a particular result of interest, conditional on all the modeling assumptions that have been made in the analysis.