So what is happening here is that you're basically noting that the relationship between FPVs and actual seats is a non-linear mapping. This means the predicting the number of percentage of Dail seats from a sample is what you might call a 'second quadrant' problem where the underlying distribution is well-behaved but the relationship between the distribution and the final outcome is somewhat complex.

Usually what statisticians would do here is try to employ the

'Delta Method' - approximate the payoff function by a straight line in a neighbourhood around the mean. The idea is that the population and sample mean will end up close enough that they'll both end up in the neighbourhood over which the straight line approximation is valid. Since a linear mapping of a normally distributed random quantity is itself is normally distributed and so on, this means you can extend the stuff you study in first year statistics to the case where the payoff is a differentiable function of the population mean. The

Strong Law of Large Numbers means that the sample average and population average will end up close enough together that you can get away with using secondary school-level calculus like this if your sample is big enough. You just need to make your sample big enough that the linear approximation holds.

Of course, the relationship between FPVs and percentage of seats won isn't a differentiable mapping. In such situation, if you can't get differentiability, you can try to construct other linear approximations, or bound the payoff function with a pair of linear functions. This requires a decent understanding of both Statistics and Real Analysis, but it's doable. At the bottom is one I whipped together using some of FG's recent election results (the red point is the joint sample average). This works because STV in Ireland is fairly proportional.

Of course the weakness here is that the bounding lines do mean you can't really predict FG's seat share, but it's still a confidence interval that's well behaved for statistical purpose atleast!!!!. Even if it's ridiculously wide. E.g if FG's vote is between 22% and 28% with 95% confidence for example, the system of approximations below gives a 95% confidence interval for the percentage of seats FG would get as being between 13.7% and 40.7%. This is a perfectly correct confidence interval for FG; don't forget that - FF got around 12% of seats in 2011 for example, and have won absolute majorities before, even if it tells us little. I'm pretty certain after the next election I can quote this post and tell everyone that I called it right using some pretty slick statistical tricks.

So you're probably right that STV and loads of 3-seaters makes predicting the number of seats hard, but predicting FPVs is much easier. But my original post only talked about FPVs. You can probably reduce the effects of the non-linearity by trying to estimate the FPVs for individual constituencies and increasing the sample size accordingly, but you'd have to do all kinds of multivariate wizardry as well in this case.