In order for you to know the dependence of the probability threshold with respect to the number of qubits for the planar code, which is the one considered in Fowler's review, which I will be refering in this answer, it is important to understand the concept of distance $d$.
As seen in FIG. 3, a number of $X$ operators in a horizontal line moving through the entire surface code yield a logical $X$ operator in the surface code. For the case of the figure, the minimum number of $X$ and $Z$ operators required for such an operation is 5, thus the distance of the code is 5. A distance of a code defines that it can corrects all errors of weight $(d-1)/2$ or lower. For the case of the MWPM decoding method, an example is provided in FIG. 5, where for a distance-4 repetition code, a syndrome can correspond to two weight two physical errors.
Given these definitions it may seem that increasing the number of physical qubits of the squared surface code and thus, increasing its distance, will always yield a better performance. Unfortunately, that is not the case. Increasing the size of the code will also increase the average number of errors within the code, if the physical error probability is high enough or above the critical probability threshold, increasing the size of the code will be detrimental to its performance, defying the necessity of the code itself.
For surface code decoders such as MWPM or the Union Find decoder, which seek to join non-trivial syndrome elements, having a sparse error means having non-trivial syndrome elements close together, which facilitates their matching and consequent decoding. Sparse errors are probable at low physical probabilities, as the physical probability increases, longer chains become more probable difficulting the correct decoding of the syndrome.
Up to my knowledge, the best way to find probability thresholds under arbitrary conditions (biases in noise models, planar/rotatedplanar/toric code, different decoders...) is to run simulations numerically. Choose a number of distances and compute their average probability of code error with respect to several physical error probabilities. The point where the curves cross will indicate the maximum probability at which increasing the size of the code improves its performance, that is, the probability threshold. If you are interested in computing said simulations I would recommend you David K. Tuckett's QECSIM package for data qubit errors. If you want to go to a deeper level considering circuit level noise, Craig Gidney's STIM package for noise simulation joined with Oscar Higgott's Pymatching will be more useful.