You also asked

if one can get the same results by doing a (first or higher order) perturbation directly on the model in Dynare without log-linearizing it first

The answer is yes, with one extra step. The endogenous variables in log-linearized models will have the interpretation as â€śpercent deviation from steady stateâ€ť, while those from a â€śdirectly linearizedâ€ť model will have the interpretation as â€śdeviation from steady stateâ€ť. So by itself you will *not* get the same results. But you can obtain them by defining additional auxiliary endogenous variables which will then have the interpretation as percent deviation from steady state e.g. if Y_{t} is the level of output then define some new variable \hat{Y}_{t} \equiv \frac{Y_{t}-\bar{Y}}{Y_{t}} \approx \log\left(\frac{Y_{t}}{\bar{Y}}\right) which will then have the desired interpretation.

So if you write down the nonlinear model from that paper supplemented by auxiliary variables as above, calibrate at the same parameter values, and ask Dynare to solve/simulate, you should be able to easily reproduce IRFâ€™s from that original paper.