This is actually a specific question, within the broader topic of 'rate adaptation algorithms'. The duration field is decided prior to the transmission of the frame, based on the modulation type, code rate, symbol duration etc. (collectively given by the modulation and coding scheme). In other words, 802.11 operates with a range of MCS levels, each level describing a unique communication scheme and hence a unique duration.
A rate adaptation algorithm (RAA) decides which MCS is to be used. As you mentioned, if the algorithm takes into account the SNR, then naturally, higher the SNR, shorter the duration (because the rate is higher) and vice-versa. However, if your algorithm is more naive and simply works with a constant rate (a fixed MCS) for all frames, then the duration field will remain the same as well.
In short, the answer to your question is - it depends on the RAA. Note that the duration field simply describes the duration of that particular frame, it is independent of whether the frame will be delivered successfully or not.