Well, I used as much hit as I could, and ran a short test for you, here are the results:
2270 Melee attacks
I am under the 9% soft cap here, and not 1 Mutilate( out of 500ish) missed, also the poison miss rates were all within a expected deviation of 17% which is the poison cap.
Expertise values were quite consistent with the +-6.5 Expertise cap.
Weird stuff:
8.78% hit as the char screen says( 5% precision)
Expected miss rate: 19.22%
Actual miss rate: 17.5%
Deviation: 1.72%
I was using a friends rogue to do the tests @ 77. I go to work a couple days and I come back and he decides that he wants to play Mut/Prep for pvp and levels to 80. I have another rogue lvl 72 that I can perform some tests on here in a little bit. I play Alliance on my main, the 72 horde is just for ganking friends and people that are stealing my ore
Either way, leveling the 2nd rogue isnt as much of a priority as gearing my raiding rogue, so I cant make any promises on a time frame.
Now, I suppose this has been accounted for previously, however if it has not, it has a slight impact on the poison hit cap. Namely it is lower because for the same hit rating, my poison hit % is higher then my melee one.
Yes this is well known and accounted for. It takes less hit RATING per % for spell hit than it does for melee. The caps posted in the pocket guide are correct.
Now, I suppose this has been accounted for previously, however if it has not, it has a slight impact on the poison hit cap. Namely it is lower because for the same hit rating, my poison hit % is higher then my melee one.
Yes it is accounted for already. Physical Hit is 32.79 for 1%, Spell Hit is 26.23 for 1%.
with regards to precision 5/5, is that % show in the hit rating calculation when you hover over your hit rating, or is that additional 5% on top of the # shown there?
with regards to precision 5/5, is that % show in the hit rating calculation when you hover over your hit rating, or is that additional 5% on top of the # shown there?
Today I managed to get my hit rating up high enough to perform some tests on the autoattack miss rate. All following parses were made by autoattacking the heroic training dummy.
First I equipped gear with a total of 722 hit rating (22.02%). Together with 5/5 precision, this brought me up to 27.02% hit. I then proceeded to take a few swings at the dummy. As suspected, after 10,000 swings, no misses at all.
I then unequipped 4 hit rating, leaving me at 718 hit rating (26.90%), and beat up the dummy some more. Ended this session after another 10,000 swings, this time with 7 misses.
The results looks pretty clear to me, but since calculating chance isn't exactly one of my strong points, I'll leave that to the people that know what they're doing.
Don't worry about hit caps; they're listed there but not as important as you think. The specials cap is the only one you need to worry about and I can hardly imagine a scenario in which you won't hit it... maybe if you wear white gear, maybe.
Edit: If after going through that, you still have more questions or comments please post in the Simple Questions/Answers thread in this forum.
Last edited by Tinwhisker : 01/09/09 at 11:44 AM.
Reason: redirecting any new questions or comments
So, Latito did some testing over the past few days, and running the numbers on the results, it appears we can conclusively say what has previously been speculated, to wit: Expertise, despite what the UI might lead you to believe, does not drop fractional points. That is, if you have rating sufficient to give you 25.5 expertise rating, you actually benefit from all 25.5 points of it, not only 25 as has previously been conjectured.
Edit: The analysis provided herein is incorrect, but will be left for posterity so the following discussion makes sense. Pretty sure the conclusion still holds, but additional statistical analysis will be required to prove it.
The details:
First, let us posit that one of the following 3 cases is true:
1) Expertise Rating does *not* truncate.
2) Expertise Rating *does* truncate, and the true dodge rate is less than 6.4.
3) Expertise Rating *does* truncate, and the true dodge rate is greater than or equal to 6.4.
So, Latito provided me with 2 data sets; the first consists of 12500 attacks performed with exactly 210 expertise, which gives 25.6.... expertise rating. If it does *not* truncate, this would provide 6.4044... dodge reduction; thus, if case 2 were true, we'd expect to see zero dodges. As it turns out, however, he observed 15 dodges in 12500 attacks; as such, we can eliminate option 2.
The second data set was taken at 213 expertise rating, and is 6000 attacks. No dodges were observed. If 3) were true, we would expect this to yield a dodge rate of at least .15%; as such, the probability of getting no dodges in 6000 attacks is less than .9985 ^ 6000 = .012% - in other words, option 3 is, statistically, strongly unlikely. Since we have, statistically speaking, thus eliminated options 2 and 3, our only remaining option is #1.
So, assuming #1, what do we estimate the dodge rate to be? Well, based on the first of the two data sets, we estimate the dodge rate with 210 expertise to be somewhere between 6.464 and 6.585; based on the second, all we can say with statistical confidence is that it's greater than 6.446% - i.e., no additional information.
As such: we can be reasonably confident that Expertise does not truncate, and it seems like the base dodge rate is 6.5; but we can't conclusively prove that it's exactly 6.5 with this data alone.
Now, what are the implications of this? In addition to the obvious point about not needing to worry about breakpoints other than the cap, the one that springs to mind is that the 214th expertise point only gives about 14% of the benefit of the preceding ones; as such, while the first 213 points may well be worth roughly 2 EP (depending on which spreadsheet you wind up using), bu the 214th is only worth about .3 EP - i.e., not really worth using except for those that believe a one in 24300 chance of being dodged is simply too disruptive to one's cycle, in ways not well modeled by the spreadsheets.
This, of course, assumes you can get your expertise to any arbitrary value you might happen to want to achieve, which obviously isn't the case. But it is worth keeping in mind when selecting items moving forward.
Also something to note, my Miss, Crit and Glance %'s all lined up with our already accepted values. I don't think any of that was in question, but always comforting to see tests line up with known mechanics.
So, Latito did some testing over the past few days, and running the numbers on the results, it appears we can conclusively say what has previously been speculated, to wit: Expertise, despite what the UI might lead you to believe, does not drop fractional points. That is, if you have rating sufficient to give you 25.5 expertise rating, you actually benefit from all 25.5 points of it, not only 25 as has previously been conjectured.
The details:
First, let us posit that one of the following 3 cases is true:
1) Expertise Rating does *not* truncate.
2) Expertise Rating *does* truncate, and the true dodge rate is less than 6.4.
3) Expertise Rating *does* truncate, and the true dodge rate is greater than or equal to 6.4.
So, Latito provided me with 2 data sets; the first consists of 12500 attacks performed with exactly 210 expertise, which gives 25.6.... expertise rating. If it does *not* truncate, this would provide 6.4044... dodge reduction; thus, if case 2 were true, we'd expect to see zero dodges. As it turns out, however, he observed 15 dodges in 12500 attacks; as such, we can eliminate option 2.
The second data set was taken at 213 expertise rating, and is 6000 attacks. No dodges were observed. If 3) were true, we would expect this to yield a dodge rate of at least .15%; as such, the probability of getting no dodges in 6000 attacks is less than .9985 ^ 6000 = .012% - in other words, option 3 is, statistically, strongly unlikely. Since we have, statistically speaking, thus eliminated options 2 and 3, our only remaining option is #1.
So, assuming #1, what do we estimate the dodge rate to be? Well, based on the first of the two data sets, we estimate the dodge rate with 210 expertise to be somewhere between 6.464 and 6.585; based on the second, all we can say with statistical confidence is that it's greater than 6.446% - i.e., no additional information.
As such: we can be reasonably confident that Expertise does not truncate, and it seems like the base dodge rate is 6.5; but we can't conclusively prove that it's exactly 6.5 with this data alone.
Now, what are the implications of this? In addition to the obvious point about not needing to worry about breakpoints other than the cap, the one that springs to mind is that the 214th expertise point only gives about 14% of the benefit of the preceding ones; as such, while the first 213 points may well be worth roughly 2 EP (depending on which spreadsheet you wind up using), bu the 214th is only worth about .3 EP - i.e., not really worth using except for those that believe a one in 24300 chance of being dodged is simply too disruptive to one's cycle, in ways not well modeled by the spreadsheets.
This, of course, assumes you can get your expertise to any arbitrary value you might happen to want to achieve, which obviously isn't the case. But it is worth keeping in mind when selecting items moving forward.
I'm pretty sure the wording here is still a bit messed up. Latito's first data set (15 dodges in 12500 attacks with 210 expertise rating, sufficient for 25 expertise or ~6.40% dodge reduction) disproves the hypothesis that expertise does not truncate and the base dodge chance is 6.40% or less. His second data set (no dodges in 6000 attacks with 213 expertise rating, sufficient for 25 expertise or ~6.49% dodge reduction) strongly disproves the hypothesis that expertise does truncate and the base dodge chance is 6.40% or greater.
The conjunction of these two pieces of information tells us that if expertise doesn't truncate, the base dodge chance must be greater than 6.40%, and that if expertise does truncate, the base dodge chance must be less than 6.40%. However, this does not allow us to make any specific conclusions on whether expertise truncates or not.
First, let us posit that one of the following 3 cases is true:
1) Expertise Rating does *not* truncate.
2) Expertise Rating *does* truncate, and the true dodge rate is less than 6.4.
3) Expertise Rating *does* truncate, and the true dodge rate is greater than or equal to 6.4.
So, Latito provided me with 2 data sets; the first consists of 12500 attacks performed with exactly 210 expertise, which gives 25.6.... expertise rating. If it does *not* truncate, this would provide 6.4044... dodge reduction; thus, if case 2 were true, we'd expect to see zero dodges. As it turns out, however, he observed 15 dodges in 12500 attacks; as such, we can eliminate option 2.
This is confusing me. If expertise truncates, then 210 expertise rating awards 25.6 expertise, rounded down to 25 because of the truncation. 25 expertise removes 6.25% dodge. Therefore, in case 2, it is plausible to have an observable dodge rate of anywhere between 0% to 0.15%. The data indicated 0.12%. Therefore, case 2 was NOT disproved at all.
Oh, sorry. Yes, I botched the wording here. Let me try this again, hopefully with less suckage.
The second data set proves that either a) Expertise does not truncate, or b) if it does truncate, the base dodge rate is at most 6.35% (as that's the 95% confidence interval). The first data set proves that either a) it does not truncate, or b) if it does truncate, the base dodge rate is between 6.31% and 6.43%. Thus, from this, we know that either a) Expertise does not truncate, or b) Expertise does truncate, and the base dodge rate is between 6.31 and 6.35%. Now, I'm pretty sure there's a statistical test that we can apply to eliminate the second data set by testing the null hypothesis "there is no difference in dodge rate between the two data sets" - alas, it's been a while since I've taken statistics so I don't recall how one would do such a test. But I'm pretty sure the ultimate conclusion is sound. If someone who's statistics is more current would like to jump in here, that would be great.
The conjunction of these two pieces of information tells us that if expertise doesn't truncate, the base dodge chance must be greater than 6.40%, and that if expertise does truncate, the base dodge chance must be less than 6.40%. However, this does not allow us to make any specific conclusions on whether expertise truncates or not.
I don't think so. The tables below is the format in which I originally PM'ed Aldriana and I think illustrates the possibilities well.
Below is the expected number of attacks per dodge.
Cap
Truncate
Precise
6.5
400
1046
6.4
667
Inf
Actual results: 12.5k attacks, 15 dodges. 1 in 833 attacks.
Below is the expected number of attacks per dodge.
Cap
Truncate
Precise
6.5
400
24301
6.4
667
Inf
Actual results: 6000 attacks, no dodges.
The only other possibly answer I can come up with is that expertise is truncated and the dodge cap is ~6.3%. This however meant that instead of 6.25 dodges in 12500 attacks, I got over double at 15. Also, I would of had a 0.9995 ^ 6000 = 5% chance to never dodge at all in the second data set. I'm not sure how to calculate the chance of getting 15 dodges in 12.5k attacks with a specific expected dodge chance, but it can't be that great.. and then that would need to be multiplied by a 5% chance of achieving no dodges in the following 6k attacks.
Okay, lets try this one more time. No guarantee that I haven't bungled the statistics, but I figure I'll take a crack at it and see what happens:
Null hypothesis: there is no difference between the dodge rate with 210 expertise and the dodge rate with 213 expertise (i.e., expertise truncates).
We perform Fisher's Exact Test to check the probability that this hypothesis is correct. In the wikipedia article notation, we have a = 15, b = 0, c = 12485, and d = 6000. Note that this is the most extreme table possible; hence, per the test, the chance that this data would be generated under our null hypothesis is simply the probability of this table - which is 12500! * 18485! / (12485! * 18500!) = .002785. (Note that a 1-tailed test suffices in this case, as our alternative hypothesis is that 210 expertise will be dodged *more*, not merely that they will be dodged different amounts). That is, our null hypothesis will generate this data only .3% of the time - well below the commonly accepted statistical threshhold of 5%. As such, it is exceedingly unlikely that our null hypothesis is correct; thus, Expertise appears to truncate.
Assuming this, as shown above, the 95% confidence interval on the true dodge rate is 6.464% to 6.585%; hence, it is highly likely (though not certain) that expertise does *not* truncate, and the true dodged rate is 6.5%.
However, this does not allow us to make any specific conclusions on whether expertise truncates or not.
Correct me if I'm wrong but in my opinion it is pretty clear that expertise does *not* truncate, thus Aldriana's hypothesis is accurate. Perhaps I did missunderstand something but as far as I can see this assumption is not confirmed yet, although there is noticeable difference between 209 and 213 ratings. Therefore, I can't understand why can't we make any specific conclusions on whether expertise truncates or not.
Correct me if I'm wrong but in my opinion it is pretty clear that expertise does *not* truncate, thus Aldriana's hypothesis is accurate. Perhaps I did missunderstand something but as far as I can see this assumption is not confirmed yet, although there is noticeable difference between 209 and 213 ratings. Therefore, I can't understand why can't we make any specific conclusions on whether expertise truncates or not.
I think based on your data, we now can conclusively say that there is no truncation.
The one reason I personally was holding out was because I wanted to see some data confirming at least 1 dodge at 213 expertise rating (which your data gives us). I wanted to eliminate the possibility of the game simply rounding 213 expertise rating incorrectly to 26 expertise. (Yes, it may seem unlikely to be the issue, but I want to cover all bases when claiming a Blizzard tooltip is simply wrong.)
Your data also proves that the dodge cap is greater than 6.495%, which is equally as useful.
So, using the same Fischer's Exact Test we applied before, we find that the probability that the data can be observed by random chance is .00000000021 - aka, tremendously improbably. Thus, we can conclusively say that there is a difference in the dodge rate between the two data sets, and, as such, that expertise does not truncate.
In terms of the actual dodged rate, the second data set suffices to show that the true dodged rate lies between 6.496% and 6.516%, which, for my purposes, is good enough to say that the dodged rate is exactly 6.5%.
As a random aside: the data also suffices to show that the glancing rate is between 23.75% and 24.50%, so the accepted glancing rate of 24% is looking pretty good; it also suffices to show that the miss rate is between 28.57% and 26.45% and 27.37%, which lines up well with the accepted number of 27%.
12/13/08, 10:29 AM
Similar Threads
