As far as I could see, it's when a tx input is using a tx input in the same block and the fee of the first tx is lower than the next tx. With this I get an approximation, at least the few I checked. I'm not sure if there are other heuristics involved?