Ladder-Balance-Data - Page 20

On July 13 2012 04:53 monkybone wrote:


Didn't you read the whole post?

On July 11 2012 01:34 skeldark wrote:

[...]

Result
TIME Filter: only between 1 Jan 1970 00:00:00 GMT - 12 Jul 2012 16:52:47 GMT
Datasize 10063
Average MMR: 1593.1
Min Difference to be significant:
90% : +-16
99% : +-24
99,99% : +-36
[...]

TIME Filter: only between 1 Jan 1970 00:00:00 GMT - 12 Jul 2012 16:52:47 GMT
MMR Filter: Only Master+
Datasize 2278
Average MMR: 2278.03
Min Difference to be significant:
90% : +-15
99% : +-23
99,99% : +-35
[...]

On July 13 2012 05:00 hunts wrote:
Sorry if you answered this before somewhere, but do you have a rough estimate of the MMR ranges of each league? I'd really like to see that as it would be very interesting.

On July 13 2012 05:00 BBS wrote:
I failed to read about how you canceld out effects like disconnects or other cases in which the game is not representative :o

On July 13 2012 05:03 xian_ wrote:


The significance intervall is lower for the larger sample? That doesn't make sense to me.

Someone help me out here please.

I dont think that disconnects are race depending. Why should one race have more disconnects than another?

On July 13 2012 04:53 lolcanoe wrote:

Yes but what's the statistical value of creating random data groups?

And what do you mean you didn't mix them together - aren't you data values calculated by average(t) - average(t,z,p)? And you did use a player-weighted average right?

On July 13 2012 05:08 BBS wrote:


Indeed, concerning the sample size, it may not be relevant, that not all races are played equally distributed in different countries, which have their own respective disconnect-rates. So even if there were like 70% of the Russians playing Terran and suffer from an increased disconnect-rate of say 100%, that might not really bias the sample at all. Might become a point, when you crack the trillion-observation-sample-size :D

On July 13 2012 05:29 lolcanoe wrote:

I find it pretty disturbing that I have to requote myself everytime i ask questions about the data itself. It so hard to get to the bottom of how you calculated the confidence interval that I'm suspecting that you don't understand what's going on yourself.

First we argue about normality, the next moment you tell me it's not based on normality at all. I press further to ask how you are getting the intervals then, you tell me that 99.99% of "random groups" fall are in a certain range. Now there's good reason to ask why these random groups exist, what they are, and what value they add but once again you seem to be dodging what's important here.

Just by looking at the MMR graph alone, it seems like we have a pretty high standard deviation regardless of what assumptions we are making. I suppose what you really want (since you are dealing with averages) is the standard error of the mean adjusted by the sqrt(n).

But I don't see any of that happenning here...

Finally for those of you who keep asking us to do the questions ourselves, the complete data package doesn't seem to be readily available. It also seems to be constantly changing...

On July 13 2012 05:03 skeldark wrote:

depends on the server
us / eu
#START PROMOTE_OFFSETS
bronce - master
0,754,1050,1280,1536,1993,



I dont think that disconnects are race depending. Why should one race have more disconnects than another?

On July 13 2012 05:47 hunts wrote:


TY, is there any information on how high masters MMR goes or about where GM starts? Or is that not available due to the nature of GM. I believe there's still a cap on masters MMR though, wonder if that number is available somewhere.

On July 13 2012 05:38 skeldark wrote:
[I add the random groups to show that my data is not just a random point, because it is outsite of the range that random points would produce.
I try to come up with simple examples and i just dont know how to explain it different than i did.

On July 13 2012 05:38 skeldark wrote:
Your questions are kind of random. I cant explain you every sentence i wrote in the op and than you want me to explain every sentence of the explanation. I think i explain decent what i did. I dont see any way to explain it better.
I even made an simple example of the calculation.

On July 13 2012 05:29 lolcanoe wrote:

I find it pretty disturbing that I have to requote myself everytime i ask questions about the data itself. It so hard to get to the bottom of how you calculated the confidence interval that I'm suspecting that you don't understand what's going on yourself.

First we argue about normality, the next moment you tell me it's not based on normality at all. I press further to ask how you are getting the intervals then, you tell me that 99.99% of "random groups" fall are in a certain range. Now there's good reason to ask why these random groups exist, what they are, and what value they add but once again you seem to be dodging what's important here.

Just by looking at the MMR graph alone, it seems like we have a pretty high standard deviation regardless of what assumptions we are making. I suppose what you really want (since you are dealing with averages) is the standard error of the mean adjusted by the sqrt(n).

But I don't see any of that happenning here...

Finally for those of you who keep asking us to do the questions ourselves, the complete data package doesn't seem to be readily available. It also seems to be constantly changing...

On July 13 2012 05:52 lolcanoe wrote:

Jesus christ I don't think an explanation could be more obfuscated than that.




The question are not random. What's random is your procedure. Your tests are inconsistent with any conventional test and the conclusions are based on pseudo-stats at best.

In light of the bullshit here, let me post very very concisely what you can do to the data to make this test statistically more sound.

Assume normality - but add a disclaimer that normality is an assumption that you do not take lightly..

Take T, Z, P data pools and calculate the following:

Mean: Average MMR of each race.
SD: Use a SD calculator to find the standard deviations of each race.

Test two races at a time (start with T vs Z, etc), using the 2 sample t test. See http://ccnmtl.columbia.edu/projects/qmss/the_ttest/twosample_ttest.html

Test the t-statistic against a P of .01.

Post the data with caveats about the conclusion and caveats about the normality assumption.

- Done.

On July 13 2012 05:53 VediVeci wrote:
Im all for the rigorous application of statistics, but I think you may be a bit too high up in your ivory tower here. This is clearly not publishable work, on that everyone seems to agree. It is however, clever manipulation of the data which returned some interesting results. There comes a point where you need to accept this for what it is because you aren't getting anywhere with what you're doing.

On July 13 2012 06:10 lolcanoe wrote:


If by clever manipulation you mean random incoherent manipulation than fine. But there is no ivory tower here - all these basic tests are taught in every college-level stats course.

The t-test, SD, and mean, and normality caveats aren't complex entities that require much explanation.

As far running the test, I've only been avoiding it because the data itself has gone through so many changes and critiques over the past days. No point in running the test on data that is still being accumulated or filtered.

Skeledark, assure me that the datafile in the OP is up to date with the most recent corrections and I'll do the test as soon as I get back from work.

On July 13 2012 05:03 skeldark wrote:

depends on the server
us / eu
#START PROMOTE_OFFSETS
bronce - master
0,754,1050,1280,1536,1993,



I dont think that disconnects are race depending. Why should one race have more disconnects than another?

On July 13 2012 06:19 furerkip wrote:


Because the sample size of your Terran/Protoss/Zerg groups aren't really close to equal, so if Terran (the smaller group) has the same amount of disconnects as Protoss or Zerg, then it affects it more?