In the study of initial value problems a lot is known. The examples I'm most familiar with out (nonlinear or linear) wave or Schrodinger equations. In such cases, one typically takes data in a Sobolev space X in the x variable and then looks for solutions which varies continuously in t in that space, i.e. in $C^0(X)$. Much is known about the limits of the level of regularity and how this depends on the nonlinearity. Many other equations can be studied in this framework, Benjamin-Ono, KdV, Navier-Stokes, et c. I'm only vaguelly aware of boundary initial value problems, but a similar type of analysis should apply. A problem where I am not aware of any study is where one specifies boundary data on a curve (or hypersurface) y(t), specifies initial data u(0,x)=f(x) and boundary data u(y(t),t))=g(t) and then looks for continuous dependence of the solution as a function of y.
(This problem arose in a mathematical finance talk by Borovkov. Essentially the Black-Scholes equation is a (backwards) heat equation with the terminal (instead of initial) data given by the pay-off of the option and the solution of the PDE giving the current value of the option. A "barrier option" has the additional constraint that the option becomes worthless if the price ever goes outside a (possibly time dependent) window. Thus, this is a (Dirichlet) boundary initial value problem. A question then arises, if one approximates the boundary by a piecewise continuous approximation, how much does this approximation distort the solution?)
For me, natural questions might be