The probability of observed data under a model, computed by integrating over all parameter values; used for model comparison and goodness-of-fit testing.