Here's another way to think about it: let's say you have two children, that you are the parent of two children and you do a good job raising them.
Well, 2 is a damn small sample size isn't it? How could we possibly know whether you are a good parent having raised only 2 children. A good sample should be nearer 1,000 to get a very small standard deviation no?
Except we are not trying to figure out if you would be a good parent on the scale of all 1 billion children in the world - that would make no sense. You have raised 2 children. That is the whole population of the study. So by definition it is an accurate "sample" since it coincides completely with the total population studied.
If we want to know how Joe Jackson did in year X and we find that in all his PA he ended up averaging .311 that IS how well he did. We are not comparing it as a proportion to ALL the PA in all of history by all batters. That is what the average itself does, as well as "normalization" - but that is an accurate sample of Jackson since it is the whole population of his PA, which are the unit of analysis.
If we were comparing it to other seasons by Joe Jackson, in any case how many would there be in any batter's career? 20 or so max? But the unit of analysis is now not PA but seasons, and so yes, here one season is a small sample, but certainly 3-5 would get us close to a significant result.