Polynomial chaos is used as an alternative to Monte Carlo methods for the propagation of uncertainty through dynamical systems. By truncating the infinite series of the polynomial chaos expansion to a finite order, the positivity of the approximate solution may be lost. We show in general how the positivity can be preserved by introducing weights into the finite polynomial approximation, where the polynomial systems are assumed to have compact support. In the case of Legendre polynomials examples of such weights are given explicitly.