Another approach to selecting $K$

As Pritchard et al. point out, when assumptions underlying the model are violated, the approach I've just described may lead to overestimates of $K$. Evanno et al. [] suggested another approach based on how much $L(K)$ changes from one value of $K$ to the next. Fortunately, the procedure is pretty straightforward:

1. Calculate the mean, $L(K)$, and standard deviation, $S(K)$, of the log probability of the data for each $K$ you considered.

2. Calculate $L'(K) = L(K) - L(K-1)$ for $K=2$ up to the maximum value that you considered.

3. Calculate $L''(K) = \vert L'(K+1) - L'(K)\vert$ for values of $K=2$ up to one less than the maximum value that you considered.

4. Calculate $\Delta(K) = L''(K)/S(K)$.

5. Choose the $K$ that corresponds to the largest $\Delta(K)$.