Dynamic models are simplified representations of some real-world entities that change over time. They are essential analytical tools with significant applications, e.g., in environmental and social sciences. Due to physical constraints applied on the outputs, it happens that input parameters are confined to a non-rectangular domain. In order to perform sensitivity analysis in this setting, we introduce the notion of aggregated Shapley effects and we propose an algorithm to estimate them with associated bootstrap confidence intervals. Our procedure is applied to analyze the sensitivity of an avalanche flow dynamic model from an input/output sample obtained by considering only input combinations leading to avalanche events that are both realistic and of interest for risk purposes. More precisely, we analyze the sensitivity in two different settings: (i) little knowledge on the input parameter probability distribution, and (ii) well-calibrated input parameter distribution. This leads insightful results regarding avalanche dynamics and potential related hazard, which demonstrate the usefulness of our approach for practical problems.