I just checked another 1000 lines of combatlog for dodges/parries on Drek'Thar during his whirlwind phase. Luckily it's in the combatlog with "gain" and "fade" and it is also a channeled ability.
The only instances I could find were after the whirlwind actually hit somebody and a few hundredths of a second (or tenth of a second) before the actual fade message.
After I deleted the part between the whirlwind hit and the fade message, I have only 1 instance of dodge (and none for parry), and this is Drek'Thar hitting no one with his whirlwind, and the fade message is only 0.021 seconds later, so I guess whirlwind was already ended at that point.
The channeling itself is 2 seconds, the time between the gain message and the actual hit is very very close to that 2 seconds. But the fade message sometimes came 0.5 seconds later, so I just went with the actual hits for whirlwind.

Miss is unaffected by such channeled abilities, just like you can miss mobs that are otherwise immune to your attacks.


Therefore I think we should stay away from any fights with targets that include a channeled ability (or an ability with a long cast time / which is often cast). Unfortunately I cannot remember which boss fights had channeled abilities. Karathress does, Prince as well, Nightbane also IIRC, but I don't know for the rest.

Next step would be to find out if other abilities that do not have a channel bar also prevent dodges/parries. Does anyone have an idea how to test this?

// Edit
Next test: Morogrim Tidewalker's Tidal Wave also prevents dodges/parries (in 2000 lines of log). He has a "begins to cast" message, but auto attacks during that time just normal.

For The Lurker Below, he seems to be able to dodge/parry while casting Spout. There is no "gain" or "begins to cast" message for that spell (maybe an indicator?), so I looked at a section of about 11 seconds with multiple Spout hits (one of our learning attempts), and there were quite some dodges and parries.
Interestingly, I only have a 4.02% dodge chance on that boss during 900 attacks...

If it's a channeled event, couldn't you cheat and just disregard any data that occurs during that channeled period? e.g. Thott lists the casting time as 1.5 seconds, you ignore any event that occurred 1.5 seconds before that spell's entry in the combat log. I understand the combatlog is not entirely clean, but would that be an acceptable margin of error?

I would think it's more important to make sure positioning is sound during the test period. It's much harder to determine if a boss turned at the completion of the casting time (based on the combat log data you would be using) than it is to determine if he was channeling a spell.

Karathress has an entry in the log for his sear nova, you know he's facing the tank when he casts it, so that should not be a big issue. The shadowbolt is a bigger issue because he could turn 180 degrees from the tank when it's launched (eliminating his ability to parry for a split-second). You have no way of knowing this based on the combatlog data, you just know it hit someone nearby.

What do you guys think of testing this on Garr? His only "cast" ability is that periodic buff purge he does.

Teron Gorefiend is an excellent fight to test on. Naj'entus possibly..I'm not sure how his immunity shield affects the log.

In this log I see misses, a dodge, resist and parry in between Tidal Shield heals on High Warlord Naj'entus.

FAO: sp00n re: Lurker data.

The Lurker precedes each Spout with "The Lurker takes a Deep Breath" or similar. Is that not something you can pickup in your logs?

here is a post i wrote on official forum based on the study here:
WoW Forums -> Weapon skill in raid tanking

There is Transcriptor.
Usual used to find out how bossencounters work to code mods for BigWigs.

I am an orc warrior using The Brutalizer and I have the logs from my guild for all BT/Hyjal fights (up to mother shahraz) for the past month and a half. If anyone would like any/all of them to dissect, I don't mind sharing them.

I unfortunatly havn't read the entire thread. However at this point it seems alot of it would be redundant anyway since most of it was focused on figuring out weapon skill to miss percentage.

I do have a sugestion though. Has anyone thought about using a private server to run some tests? (please read further before flames >.< )

My line of thinking here is that testing is being done the PTR where we know for a fact that things are being tweaked. (maybe not directly to the effects of weapon skill, but still) So data from there may be corupt anyway. So whats the differance between that, and using a private server to create a pure data set of 100,000 hits on Huntsman?

You could set the parry rating on your toon to a god like ammount to make sure you parried every hit, there by increasing the speed at which you obtained the data. If using a palidan to do this, you could just increase the spell damage to an insaine ammount for an instant clear of the trash just before him with one consecration.

Private servers have no clue whatsoever a real server does (unless someone hacked a blizzard server), they can only emulate what has been observerd or can be observerd.
I'm pretty sure no private server uses the correct weapon skill formula (or combat in general).

Assuming all the stuff sp00n posted (big love for this!) is right, then [Belt of One-Hundred Deaths] would have like...ehm... 49 hit rating for a fury warrior? WTF?
This is awesome!


The one thing I'm still thinking about:
So due to the Belt you need 48.88 less hit rating in order the reach the cap. Right?
But the weapon skill also reduces the dodge rate of the mob, how much dodge does the Mob loose with a weapon skill of 356?

Yes. Or in other words, you're missing 3.1% less than before.

Unknown, but I'm guessing 0.6% less dodge.

Also, as a warrior, you do want to pass this belt to your rogues.
Still the best for rogues, even if you're hit capped.

So do we know for sure yet whether weapon skill has an impact on dodge (if not the exact amount per skill point)? Where my gear level currently is, it doesn't matter, but for planning ahead, that would be cool to know. I did read the entire thread, and while there was a lot of evidence pointing towards an effect on dodge, nobody showed any conclusive proof of an effect (unless I just flat out missed a post somewhere).

the post directly above yours gives an estimate...but no one knows for sure yet :-) Its much harder to test.

So far the only controlled tests of dodge that has been done is the paladin tests, reported on page 3-4. Aldriana did a calculation for the dodge per weapon skill based on those tests and found that the 95% confidence limit on dodge/weapon skill is [0.068,0.144], so 0.1 per weapon skill seems rather likely (but from this test the interval is the only sure thing).

Aldrianas analysis: http://elitistjerks.com/361558-post71.html

The caveat I would throw in is that that range assumed that each point of +skill provides the same benefit, which is fairly likely untrue. In practice, what that means is that the difference between 350 and 365 lies somewhere between 1.02% and 2.16%, which could mean .1% per skill... but it could also mean, for instance, .2% for the first few points and .04% for the later ones. It's just really hard to say.

0.622% is the standard deviation (represented by sigma). It appears in the mathematical formula describing the shape of a normally distributed function. For such a function (any large enough data set we work with is approximated very well by a normally distributed function), roughly 68% of all experiments conducted will fall within one standard deviation of the true value you're testing for.

The 95% confidence interval means that given a large distribution of experiments, 95% of them will report confidence intervals that contain the true value. The width of the interval is +- 1.96 times the standard deviation.

That's a little confusing, so I'll try to be a little more explicit. If you repeated your 4744 swing test 100 times, you'd expect 95 of those tests to give a crit rate inside the 95% confidence interval of the true value. Roughly 5 of the tests would report a crit rate further away from the true value.

Here your 95% confidence interval would be 24.199% +- 0.622% * 1.96 = 24.199% +- 1.219% = [22.980, 25.418]

That's about 2 standard deviations away, so it's still possible that it's just a statistical variation. It's a little outside of the 95% confidence interval.

It's been a long time since I've taken a formal statistics class, so there might be a few minor mistakes in there. I think I got most of the important points right, though.

Ok, let me see if I got that correctly.
So the 0.622% here is "one" standard deviation (for this data set)?
Meaning that 68% of all tests will result in 24.199% +- 0.622%, correct?

I suppose the 68% value was proven by some smart guys and are now just standard values, and it doesn't change for another data set?
And that also the 1.96 multiplicator to receive a 95% confidence is a standard value and doesn't have to be recalculated every time.

So there is a table where can I find these standard values, e.g. the multiplactor used to calculate confidence value of 97% or 99% from a given "one" standard deviation?


Let's use the newly acquired knowledge to do that for dodge.
sqrt(254 * (4744 - 254)/4744)/4744 = 0.327%
95% confidence: 0.327 * 1.96 = 0.641%

So with 95% confidence the true dodge rate lies within 5.354% +- 0.641% = [4.713, 5.995].

The distribution of all 4744 swing tests has a set standard deviation and a mean value -- the one the servers use to check if you crit or not. Strictly speaking, you approximate this standard deviation using the data you take. 68% of all tests will be within one standard deviation of the distribution's mean value, not 24.199%.

Correct.

You could look at the table here to start. Alternatively, you could do some calculus using the formulas above. Aside from that, I think the multiplier for 90% is 1.645, and 99% is 2.576.

Edit: Err... pretty much exactly the table you're looking for is further down that page. Maybe I should read all the way down the links I post...


That looks good to me.

Your calculations lead to the same result but somehow mix the two steps together.

The standarddeviation is computed as
(e.g Standardabweichung - Wikipedia or Standard deviation - Wikipedia, the free encyclopedia at the bottom "for running sums")
Now for computing the confidence interval you need to adjust the stddev further:
(see e.g. Konfidenzintervall - Wikipedia or Confidence interval - Wikipedia, the free encyclopedia)
NOTE: This computation of the confidence interval for the expected mean is only valid if the underlying distribution is a normal distribution!
In our tests it is not (binominal distribution) but with the huge samplesize an approximation with a normaldistribution is justified.

Actual there are forumulas to compute the confidence interval for binomal distributed data
but this is more complex than above.

For the these values:
sqrt (254 * (4744 - 254)/ (4744*4743))=0.2251
Now the confidence interval:
0.2251 * 1.96 / sqrt(4744) = 0.641
5.354% +- 0.641%: 4.713 - 5.995%

You're calculating a confidence interval for a mean, while Rezarel (Spoon) does it for a proportion. You're anyway transforming your mean into a proportion (percentage) in the end so the end result is practically the same. The only difference is the N*(N-1) factor instead of N*N which won't have much effect for thousands of swings.

I usually use a web calculator to find confidence intervals:
Exact Confidence Interval for a Proportion

Using that I get a 95% confidence interval of 4.75% to 6.03% for the dodge rate above.

Argh, I told you I cannot cope with Wikipedia and it "explanations".
Like that:

Could be Chinese as well, I would understand just as much.

So... it's the same results and only a slightly different formula?
Why would I use a more complicated one if the easier is just as exact (besides from a mathematical point of view).