Flight delays are one of the most discussed, yet not fully understood, topics in the aviation industry. In this paper, we shed more light into propagation of flight delays by providing a spatio-temporal analysis of flight departure delays of the US domestic air network for the year 2017. The analysis focuses on four US air carriers (full-service and low-cost) and two time events characterized by extreme weather conditions, in addition to a baseline case free of extreme weather conditions. We constructed a Delay Propagation Network (DPN) for each (time event, airline) pair detecting patterns of causality between hourly delays in airports using a Granger Causality approach. In addition, we identified four (time event, airline) pairs with a volume of cancellations large enough to construct a Cancellation Propagation Network (CPN), analogously to DPNs. For the baseline case, we observed that central nodes of the airport network (i.e., hubs) usually act as absorbers or intermediary nodes in the DPN. DPNs were more homogeneously distributed in space for point-to-point than for hub-and-spoke networks. For extreme weather events, we observed that the size of a DPN increases with the percentage of canceled flights as long as this stays below 10%. Conversely, it suddenly decreases when the percentage exceeds such tipping point because most causal relationships among delays are lost due to the volume of cancellations. We also observed that some airports located in the region of the extreme weather event were among the central nodes of the DPN. Those airports, together with the hub airports, acted as the top generators, absorbers, or intermediary nodes of the DPN. On the other hand, CPNs monotonously increased in size with the proportion of canceled flights. CPNs are less noisy and therefore easier to interpret than DPNs, as cancellations stem primarily from the extreme weather event only. In CPNs, hubs act as cancellation absorbers, due to the larger volume of resources that airlines allocate there. @en