Consider first a closed system divided into two subsystems A and B, separated by a fixed, adiabatic and impermeabler wall. These have a different amount of the same ideal gas, $n_{i,A}$ and $n_{i,B}$, $n_{i,A} > n_{i,B}$. Their temperatures and volumes are the same. Pressure is different, as per they different amounts of gas. Now change the wall so that it is permeable. Gas from subsystem A will quickly diffuse into subsystem B, until equilibrium is achieved. The entropy of each subsystem is straightforward to determine:
$$S_{A,B}=n_{A,B}[c_v\ln T+ R\ln(V/n_{A,B})+k]$$
Using this information, it's easy to determine that $\Delta S_A<0$ and $\Delta S_B > 0$ (knowing that $n_{f,A}=n_{f,B}=n_T/2$). In other words, the entropy of subsystem A decreases and the entropy of subsystem B increases. It's also possible to the determine that the overall entropy change is an increase. However, what I'm trying to point out here is that the entropy of subsystem A decreases.
Now consider a similar system, but this time with only 10 particles we can place either in the right or in the left (let's say that left is A and right is B). Originally, we have 7 particles in A and 3 in B. We let the system evolve until there are 5 particles in each side: equilibrium is achieved. The Botzmann postulate for entropy is $S=k_B \log W$. We can caclulate the mutiplicity with the binomial coefficient. It's easy to see that, now, both entropies increase (and obviously the total entropy of the system increases) because the macrostate with 5 particles in one side has more microstates than either the one with 3 or 7 (in fact, both entropies increase the same amount).
So, what's going on? How come in the purely thermodynamical version, one entropy decreases, while in the statistical version, both increase?