Wintertime Arctic sea ice extent has been declining since the late 20^th century, particularly over the Atlantic sector that encompasses the Barents-Kara Seas and Baffin Bay. This sea-ice decline is attributable to various Arctic environmental changes, such as enhanced downward infrared radiation (IR), preseason sea-ice reduction, enhanced inflow of warm Atlantic water into the Arctic Ocean, and sea-ice export. However, their relative contributions are uncertain. Utilizing ERA-Interim reanalysis and satellite-based data, it is shown here that a positive trend of downward IR accounts for nearly half of the sea-ice concentration (SIC) decline during the 1979-2011 winter over the Atlantic sector. Furthermore, we find that the Arctic downward IR increase is driven by horizontal atmospheric water flux and warm air advection into the Arctic, and not by evaporation and surface heat flux from the Arctic Ocean. These horizontal heat fluxes are linked to La-Nina-like tropical convection. In all CMIP5 climate models that are analyzed here, high pattern correlations are found between the surface air temperature trend and downward IR trend. However, there are two groups of CMIP5 models: one with small correlations between the Arctic surface air temperature trend and the surface heat flux trend (Group 1), and the other with large correlations (Group 2) between the same two variables. There is evidence that the Group 1 models are consistent with the aforementioned observation-based finding that the Arctic warming is closely related to large-scale circulation changes. In contrast, the Group 2 models are at odds with this observation in that their Arctic warming is more closely tied to surface heat fluxes than with the large-scale circulation change. Interestingly, while Group 1 models have a warm or weak bias, Group 2 models have large cold biases in the Arctic. This difference suggests that deficiencies that cause the cold bias of the mean state may contribute to the surface heat flux being an important warming factor in the Group 2 models, and that while ensemble model projections are widely used, the underlying warming mechanism may widely differ from model to model. These findings also underscore the need to improve our understanding of the mechanism that maintains the mean state of the Arctic in both nature and models.