[Patch 7.19] Post Worlds 2017 General Discussion - Page 6

On September 30 2017 10:49 GrandInquisitor wrote:

No one bans him now, but no one banned him before either. The point is whether the patch was a "massive fucking nerf". That's not gonna be something you can tell from changes in the ban rate, not when there's apparently only one guy on the server banning Skarner.

Same as to the play rate. Nobody plays Skarner. We all know that. But that was true pre-patch and post-patch. That doesn't mean anything as to whether he got nerfed. The only data you have is winrate data, and even though it has a low sample size, the important point is that it's consistent between patches.

On September 30 2017 11:16 nafta wrote:

When sample size is that low data might as well not exist. I am not saying anything about skarner since I haven't read patch notes. The general logic of looking at stats like this is pointless. Unless something is an outlier you might as well just ignore it.

On September 30 2017 13:50 cLutZ wrote:
I still don't see how soloq is properly related to balance...

On October 01 2017 01:29 Jek wrote:

What would be a better alternative?

On October 02 2017 00:22 Sent. wrote:
Finally picked up Ezreal jungle and have to say he's really strong. Buff -> wolves -> buff -> toplaner / jungler is a very effective route.The only problem is sometimes you get games with 4 glass cannons on your team, but these games are winnable too.

On October 02 2017 02:17 Sent. wrote:
I would say he should be kicked out of the jungle because he offers everything except hard cc while being extremely safe.

Supports typically aren't that most amazing tanks if the game is even or a little behind unless you on someone like Ali.

On September 30 2017 13:42 GrandInquisitor wrote:

So now we're getting deep into stats nerdery but your view is a common misconception.

Low sample sizes can be used all the time. There is no magic number where a sample size is large enough to be "OK". Instead, you evaluate the statistic's predictive power on a continuum through a confidence level. A large, well-distributed sample allows you to have stronger confidence to make stronger predictions. A small sample just means you have lower confidence in your findings. What confidence level, and size of sample, required is entirely up to the hypothesis you're testing and whether you're trying to prove it or disprove it. For example, the most extreme hypotheses (e.g., all white people have blonde hair) can be disproved with a n = 1 sample.

More generally, what you actually care about is how the sample was distributed. A truly randomly representative sample can theoretically be of any size and still have very strong predictive power. Meanwhile, even very large samples are useless if they are not representative of the overall population.

Finally, bad data is still data. So long as you're aware of the limitations and biases of the data, it remains useful to prove whatever it can prove. Surveys of extremely biased, small samples are still useful when answering questions where those limitations are irrelevant, particularly if there is no contradictory data.

So here, lolalytics Skarner data has relatively low predictive power because nobody plays Skarner. But that's OK, because:

1) We have limited reason to believe that the Skarner data was collected in a non-representative way that would significantly bias the relevant finding;
2) The hypothesis we wanted to test was helpfully very extreme, and so even stats with low confidence level are sufficient to disprove it;
3) The hypothesis is looking for a comparison between last patch and this patch, and there's no reason to believe that the sample was somehow corrupted between patches or that it's not comparable in some way.

In short I would consider win rate data to be generally reliable but dependent wholly on whatever you're trying to prove. Win rate data is very good at proving whether a champion got better or worse, provided no large changes in play rate. Win rate data is very bad at proving what champion is best for any given player, given biases in how we choose to play who we play. Win rate data is equally bad at determining whether a champion is "better" or "worse" than another, and moderately bad at determining even what counters what.

Item win rate data is hopelessly bad for a multitude of additional reasons, with the possible exception of first item win rate. But when the signal is strong enough (like how Janna used to have 70+% win rates with Ardent Censer and still nobody said anything about it) even it can tell a useful story.

On October 02 2017 04:59 nafta wrote:

How do you take into account play rate on other shit changing? Or people figuring out how to beat something when they encounter it often like "qss is good vs skarner"?

In short you are ignoring a lot of variables which matter a lot.

Hell I did a small "survey" and asked 20 people I know ranging from bronze to diamond to describe all 4 of skarner's abilities and estimate their numbers. Try to guess how many of them got it close? To make it even more fun I don't even know most of his numbers yet I regularly play with people who play him.

On October 02 2017 21:56 GrandInquisitor wrote:
To give an extreme example: if Skarner won 100% of games where QSS wasn't built, and 0% of games where QSS is built, and QSS is built 51% of the time - he's gonna have a winrate of 51%, which obviously is very stupid, but doesn't change the fact that on average, 51% of the time, having a Skarner on your team means you're going to win.