Dexterity's impact on critical hits

[Masamune]

Why didnot you simply use /checks to determine if yur mob is indeed "correct" lvl before engaging ?

and what were yur STR when yu engaged that mob who killed yu ? sound like yu got too much str

[pchan]

Well no test have shown that crit% was an integer yet, you would need for exemple 50K sample size on dDEX=7 and dDEX=8. If those tests yield the expected results (9% and 10% with low error ~ 0.2%), then you can assume crit% are integers, and when you find y% crit in a test with a confidence interval I, the actual crit% is the only onteger in I (with 95% "chance"), so it seems super high sample size is no longer needed (once again *if* we can confirm 1 or 2 jumps from 2 different integer values).

edit : I'm targetting the ones that check "low eva" at 285 acc, which are the second highest, it just seems that the dolphin I use is not killed by anyone, so it's always the same. Also confirming that 61 STR doesn't damage them while 62 does.

[Masamune]

Personnally i use the HighDef check, lot easier since only add+1 to get real mob def -> level.

[Byrthnoth]

Well no test have shown that crit% was an integer yet, you would need for exemple 50K sample size on dDEX=7 and dDEX=8. If those tests yield the expected results (9% and 10% with low error ~ 0.2%), then you can assume crit% are integers, and when you find y% crit in a test with a confidence interval I, the actual crit% is the only onteger in I (with 95% "chance"), so it seems super high sample size is no longer needed (once again *if* we can confirm 1 or 2 jumps from 2 different integer values).

Well, if you can show that some samples (like 3 and 6 or 7 and 8 ) aren't statistically different from each other, while other samples (like 6 and 7) are at the 95% confidence level, then that's evidence enough of flooring for me.

[Masamune]

Well, if you can show that some samples (like 3 and 6 or 7 and 8 ) aren't statistically different from each other, while other samples (like 6 and 7) are at the 95% confidence level, then that's evidence enough of flooring for me.

i think same as Byrth. :s

[Byrthnoth]

It'll be great if people can report their Accuracy and Hit rate too. If we're willing to assume that 2 (Acc-Eva) = 1% Hit rate then we could combine samples and get an ungodly huge comparison of odd (Acc - Eva) vs. even (Acc - Eva).

[CDF]

On level determination:

Again, I wasn't paying attention, so my shadows were down for like 15 seconds while I got hit by three straight hits (only 845 HP). In the future I'll use /check when I have enough space and gear to fine tune attack and/or accuracy.

On testing for two proportions:

Use the Z-test for two proportions, not a t-test. I hope that calculator works.

On determining if crit rate takes integer values (pchan):

Could you explain where you are getting n = 50,000 from? We have already showed statistically significant jumps for mere +1 changes in dDEX (or at least borderline in my case, but that was a result of not collecting enough data, I assert) with relatively small sample sizes, which points to discrete "jumps" in light of the long-range observation that changes in dDEX yield very slight changes in crit rate between 0 and 40.

To rule out x.5% values with high probability (not overkill like 99%), you wouldn't need 50,000 sample sizes for 9% and 10% each.

[pchan]

Can't really predict what the result would be but so far : 10.19 +/- 0.48% and 9.13% +/- 0.5% is what we have, so we can only say that the jump is at least ... 9.63% -> 9.71%, and that's not even with 95% probability as you take the intersection of 2 events that have 0.95 proba. At least I'd want the dDEX=6 to be with a lower error so that the "jump" be > 0.5% (well actually you need to lower both test's error probably ... 25 k maybe should be anough)

[Masamune]

i can "continue" the dDEX6 test, since i kept the parse.

But question is more: is it really that necessary? personnally im almost thinking its a waste of time :s

Atm whacking for dDEX=12, results tomorrow.
(fucked up dDEX=10 cuze Hpemde run away exactly when my char were auto recasting refueling... so parsed like 10k "out of range" messages rofl... )

[CDF]

I'm with the idea that you show two samples separated by 1 DEX have the "same" crit rate (just have sufficient sample sizes and hence power to detect any reasonable differences that may occur) while another set of two samples separated by 1 DEX have a significant difference in crit rate.

If anything, I would start looking at the high end to see if crit rate is capped at a lower dDEX than 50, but that would be a pain in the ass for me on ninja (can't really suppress large amounts of STR with equipment while lots of DEX equipment has STR on it too).

[Masamune]

CDF, before going look at higher end dDEXs, can you do 2 things please :
- continue more sampling of yur dDEX=22 point to confirm its "confidence"
- PM Shamaya which point he has already done, or plan to do, would be "silly" imo to see both of yu doing same point...

[pchan]

i can "continue" the dDEX6 test, since i kept the parse.

I will do that tonight, hopefully the total will be 25k, and then i'll rework on dDEX=7.

[Shamaya]

http://img.photobucket.com/albums/v7...Rate41dDex.png
71fishy,another long parse accompanied with wonderful sleep, dDex41, 0.55CI.

On one hand, we have the TTaru blog (jp data) saying crit caps at dDex48. What we've thought from the Searain data is that it caps at 50, but looking at the colibri data it looks to be moreso above 50. If it's a 12 point spread of dDex for a 10% spread of crit, the deviation (is this the correct term?) would be 0.83.. There's not enough confidence in this sample though. But I'll go with CDF's wisdom from a few pages ago and just continue to parse different points. I wanted to try one closer to the cap next, but maybe I'll just do 42.

[Masamune]

Guys i got a surprising /shocked when i woke up to check up my dDEX12 results:

1733/18151=9.55% +- 0.428% (was even lower yesterday evening) and doesnot seem to raise or decrease from eyeballing KParser since 1hour...

Should i continue it ?

I'm gonna check again that mob's stats... my dex=79 and im pretty sure i checked it to be lvl71...EDIT: yes i double checked its lvl71...

For shamaya, yes that's why i suggested yu in PM to go 47 before 48+, just to be sure of the "trend" between points 40 and 47.

Below current progress we made so far, with Pchan, CDF, Shamaya and my results plugged in, and removed Searain's data (since they were based on assumed #totalhits).
http://www.bluegartr.com/forum/attac...1&d=1257097094

Btw i noticed something funny: after whacking a whole night at that UlHpemde, after i disengaged it this morning it followed me everywhere WITHOUT running away lol o.O was litterally stuck to my ass, exactly like a BST pet rofl.

[pchan]

Guys i got a surprising /shocked when i woke up to check up my dDEX12 results:

1700/17700~9.55% (was even lower yesterday evening) and doesnot seem to raise or decrease from eyeballing KParser since 1hour...

It's 9.60% with 0.43% error, nothing surprising *yet*, after all it's a "95%" CI, so there is a decent chance for the test to be off ; I remember that while doing dDEX=8 the first 3K sample gave me 8% ish.. 10% lies in your CI anyway, so I guess it is 10%.


Now for my testings about the first jump

dDEX=6 : 1183/13202=8.96% +/- 0.49% and when I add it to masamune's data for dDEX=6 we totalize
dDEX=6 : 2330/25765=9.04% +/- 0.35%

So yeah, I'm 80% convinced that the jump is between 9% and 10%, I will try the next jump later, with dDEX=14 or 15. If I can achieve it (I probably have to remove haste).

[CDF]

I wouldn't be that concerned about unusual results. (Is less really significantly less?) If you're collecting lots of individual samples and generating confidence intervals from them, you're bound to come up with a CI that doesn't actually cover the true parameter.

Hey, why didn't anyone call me on 91 - 67 = 23? Just kidding. Actually, I was comparing dDEX = 24 (total DEX 90) to dDEX = 23 (total DEX 89). Sorry about that and please update your data accordingly.

Anyway, I got some more data for dDEX = 23 (total DEX 90) and got 1190/10001 = 11.9% (edit: typo), for a total of 1708/14589 (11.7%) with approximate 95% CI (11.2%, 12.2%).

Recall that based on floor(dDEX/8 ), there is supposed to be a jump going from 23 to 24, but so far there isn't enough evidence to support that and we might just accept that there is no jump there, another strike against floor(dDEX/8 ).

As for hit rate, the updated total is 14589/15707 = 92.8% with 95% CI (92.5%, 93.3%). If I eventually establish that the hit rate is probably an integer percentage, I might consider adding +1 accuracy somewhere and see if it goes to 94% or not and rule out 93.5%. Yeah, technically not much work at all but it still ties up my game.

Aside: it would really be interesting to find evidence of random number generator failure (which would be really "LOL" since that's the driving force behind any MMORPG), but kparser can't record hits in order.

[Shamaya]

Yeah, that is very true Masa. The hpemde will follow, lol. I imagine a lot of people who have skilled up dark magic on them have noticed the same. I did the same thing earlier today. Dragged it really far to see if it would deaggro, and it didn't. Things like your report of 9.55% crit make this very frustrating but we'll just keep at it yeah.

[Byrthnoth]

Anyway, I got some more data for dDEX = 23 (total DEX 90) and got 1190/11001, for a total of 1708/14589 (11.7%) with approximate 95% CI (11.2%, 12.2%).

If I add your samples together (519+1190)/(4580+11001) I get 1708/15581 = 10.96%, which means there probably is a jump.

[CDF]

Sorry, that was a typo. 1190/10001 = 11.90%

[Byrthnoth]

Aaah, okay. Honestly, I'm falling out of love with the "integers onry" hypothesis.