As long as you're trying to think in terms of Skada DPS rankings you need to pick a fixed point from which to draw your conclusions. That's problematic even if you know the shape of the underlying distributions (which we don't) and greatly influences the results of your ranking. If I had to give a recommendation for that approach I'd suggest to take a look at any percentile lower or equal than p90, with constant ilvl, for fights with all specs having at least a few hundred data points, for each fight, for each difficulty. You could then reach a conclusion such as
"Under the assumption of a combination of RNG and skill resulting in (pick percentile) relative performance, the DPS of spec X is A and the DPS of spec Y is B for this particular fight on this particular difficulty, given that both players work with a set of gear in a comparable range."
Fixed percentile rankings are available on warcraftlogs.com, but the assumptions are incredibly impactful. Just play around with the settings to see what I mean. The idea behind the boxplot is to get away from this, if only a little.
Rankings are insufficient to draw balancing conclusions.
I'm attaching more details.* If this isn't news to you just ignore it.
The distributions are configured to fit the percentiles of the boxplots as well as possible. The shape of the distributions can differ substantially.**
To take your main point, I agree that showing the maximum values could be misleading and it does affect the scale. Since they should not be considered anything else besides a theoretical possibility, I excluded them and rescaled the graph. For now I'm leaving this here, but I'm planning to do a general update as soon as we have more 7.2 data.
My personal take on this would be to play Arcane on Spellblade and Tichondrius, play Frost on everything else and pay attention to what happens to Fire in the near future. But as I've said above, with 7.2 this conclusion is somewhat up in the air until we have more reliable data.
* I have to apologize, this was all done without a second revision. Unfortunately I somewhat forgot to label the axes. X is DPS, Y(lower), not marked, is density, Y(upper) is cumulative density. I also somewhat mislabeled the right X axis - the scale is the same, but the ticks are smaller. Also keep in mind that I've slightly altered the boxplot from what's the standard because the data didn't allow for more details. Typically you'd show outliers as dots outside of the whiskers, which provides a much better impression of their weight in the data set.
** That's somewhat reflected in the boxplots, and although difficult to see you can draw some conclusions from the distances between the percentiles. For example, an asymmetric plot corresponds to a skewed distribution. It's probably also what's going to drive the in-house balancing discussions. Since you want to maximize player happiness (~ time spent in the game), left-side RNG-based outliers are problematic. I'd really be curious to see their data, because you could very clearly disentangle RNG and player ability (analysis of apm, time played, reaction time etc.).
blizzard and frozen orb and icelance split does actualy alot of aoe specially wen the adds arent up that long 
