This probably only addresses half of your question...
Most of the time, estimates of the error probability are not so important. If you're using a non-degenerate error correcting code for a single logical qubit, all you need to know (effectively) is enough to resolve a choice between two different error corrections (i.e. you want to know that the number of errors is smaller than half the distance of the code with high probability). Small mistakes in estimation are not going to change that, unless you're right up at the boundary of whether it works or not (and then, whichever side of the boundary you're on, the 'with high probability' rider I just put on it is unlikely to hold. You want to know you're at least a few standard deviations away from that threshold).
The easiest example is the classical repetition code. At the time of decoding, if you see a bunch of 0s and 1s, which do you decode it to? You assume that the per-bit probability of error is less than 1/2, and you decode to whichever bit value you see more of. It doesn't matter if that error probability is 0.3 or 0.31, you proceed in the same way.
Where your estimate of the error probability does make a difference is if you're using error correcting codes out beyond the literal distance value of the code. The typical example of this is with the Toric Code. This is defined for $O(N^2)$ qubits on an $N\times N$ grid. The code has distance $N$, i.e. there are combinations of $N/2+1$ errors that would lead to a logical error when you do the syndrome measurement and correction. However, people still try to use the Toric Code when there are $\epsilon N^2$ errors, for some small enough $\epsilon$. That is based heavily on an assumption about the noise model and its parameters to determine which errors are most likely. I did some work at some point (not quite for depolarising noise) looking at the Toric code and how the error correction threshold changes depending on how good an estimate you have of the noise parameters. For example, if you have different rates of X and Z errors, but you assume they're the same rate, as compared to knowing what their rates are. See Figure 2 of that paper.