We present an analysis of different sources of impact model uncertainty and combine this with probabilistic projections of climate change. Climatic envelope models describing the spatial distribution of palsa mires (mire complexes with permafrost peat hummocks) in northern Fennoscandia were calibrated for three baseline periods, eight state-of-the-art modelling techniques and 25 versions sampling the parameter uncertainty of each technique – a total of 600 models. The sensitivity of these models to changes in temperature and precipitation was analysed to construct impact response surfaces. These were used to assess the behaviour of models when extrapolated into changed climate conditions, so that new criteria, in addition to conventional model evaluation statistics, could be defined for determining model reliability. Impact response surfaces were also combined with climate change projections to estimate the risk of areas suitable for palsas disappearing during the 21st century. Structural differences in impact models appeared to be a major source of uncertainty, with 60% of the models giving implausible projections. Generalized additive modelling (GAM) was judged to be the most reliable technique for model extrapolation. Using GAM, it was estimated as <i>very likely</i> (>90% probability) that the area suitable for palsas is reduced to less than half the baseline area by the period 2030–2049 and as <i>likely</i> (>66% probability) that the entire area becomes unsuitable by 2080–2099 (A1B emission scenario). The risk of total loss of palsa area was reduced for a mitigation scenario under which global warming was constrained to below 2 °C relative to pre-industrial climate, although it too implied a considerable reduction in area suitable for palsas.