It strikes me that the desertion component defeat and make invalid any attempt to assert something about the whole using 25% of the men with the unit in September. This breaks the chain to go on to June.
This means for one to say something about the entire unit in September, one would need to be able to make a defensible assertion that the distribution of the qualities of the 25% had the same distribution as all 100% (or that they vary in some acceptable range). Then, to take this back to June, one would need to make a defensible assertion that the distribution of the qualities of the 100% in September have the same distribution as the 100% in June (or whatever other date).
Those two hurdles seem difficult to clear. Even when one does "proper" statistical sampling, there is a need to "prove" the sample is respresentative of the whole population being sampled.
To bore you further, in *proper* sampling, there is also the fun part of making sure the "frame" - the source from which one makes the selection - is the same as the population - the intended whole from which one is seeking to make comments about. Thus, if one wants opinions of all people, to make one's selection from a phone book would invalidate the sample because it leaves out - at least - (1) those people without telephones and (2) those who request their name not be listed. Those two groups might well have characteristics that would change the result, had they been included for sampling.
The deserters might surely have some other sort of characteristics than those who stayed (like the people without telephone, but with mobile telephones).
This said, one of my professors used the what remained of the early 1800's voting results of an American election (very patchy data, to say the least) to present all manner of conclusions - properly qualified, of course.
Good luck. - R