I have seen good results from dark-frame subtraction (in MLV App) using either frames generated immediately before or after the clip to be processed or using frames generated some significant time afterwards, so this is something of a theoretical aside, but what are the factors that determine the pattern of noise on the sensor, such that using different conditions for dark-frame generation from those of the clip of interest will produce suboptimal results?
I imagine area of sensor being sampled (resolution and centering), shutter speed, and ISO will be the most important factors (all possible to reproduce at any time), but what influence do factors such as temperature or long-term variation in noise patterns over time play, such that long-after-the-fact darkframe generation and subtraction might produce suboptimal results?