The authors examine the dependence of the energy levels of a classically chaotic system on a parameter. They present numerical results which justify the use of a random matrix model for the statistical properties of this dependence. They illustrate the application of their model by calculating both the number of avoided crossings as a function of gap size and the distribution of curvatures of energy levels for a chaotic billiard: the distribution of large curvatures is determined by the density of avoided crossings. Their results confirm that the matrix elements are Gaussian distributed in the semiclassical limit, but they characterize significant deviations from the Gaussian distribution at finite energies.