when simulating a deterministic shock using check command I sometimes get an error, but when the check command is commented-out, Dynare still computes the impulse responses. They “look” normal in the sense that they seem to converge to the steady state (the procedure makes them converge to the initial values). The only thing is that when I compute deviations from the steady state, in the periods before the final value, they are on the order of magnitude 1e-8 away from zero. I am using Dynare 4.3.0 (troubles with check are not reported as one of the known bugs int his version).

The question is whether (1) the command check makes any sense at all for deterministic simulation, which takes into account nonlinearities, and (2) if this is a rounding thing.

The check command tells you something about the local stability properties around the steady state. For a linearized model, it is all you need to know. For the a nonlinear model, it is neither necessary nor sufficient. Nevertheless, you should check your convergence properties. Numerical inaccuracies of 1e-8 as you report are nothing to worry about, unless they signify a bigger issue. Often, you can simply test the behavior of your model by varying the simulation length via the periods statement. Try increasing it and check whether the convergence to the resting point continues.

Thank you very much!!! I checked the properties by varying the simulation length (200, 400, 600, 800, 1000 periods) and the results are the same (i.e., impulse responses lie on top of each other, with only tiny differences between 200 and 400 periods, but practically none for longer simulations).