Crit+% weapons and Critical Hit WS

[Kirschy]

I apologize in advance if this information is found easily. I was hoping someone could point me to discussion and/or tests involving the +Crit % rate on main weapon affecting critical for WS that can critical.

My other question is for people who know a good deal of statistics. Let's say I was to design the following test:

Test A - Mamushito+1(DMG:38)
Use Blade: Jin @100tp on too weak mobs so that the mob dies after 1 hit. Record criticals vs non-criticals

Test B - Senj(DMG:38 Crit+6%)
Use Blade: Jin @100tp on too weak mobs so that the mob dies after 1 hit.
Record criticals vs non-criticals.

How many times would each test need to be done to be reasonably sure that the critical+% on senj is or isn't affect the critical rate of WS?

[CelinaAlexander]

I think it was stated somewhere that the crit% only affects the first hit, which would make your test seem conclusive even after 100 WSs even though it wouldn't be.

I might be thinking of something else though

EDIT: To clarify, what I mean is that your test would only prove that the crit% affects the first hit, which may or may not be already known.

[Wenslydale]

Personally, when collecting my data sets I like to shoot for 100 of each. I know its probably overkill, but you get a really nice set of numbers to work with, and the math is easier cause 100 is so easily divisable. (maths make my head hurts!)

Usually after 25 tests or so, you'll have a fairly good idea of what the general mechanics are, but 100 will allow you to rule out anomalous data.

[CelinaAlexander]

Personally, when collecting my data sets I like to shoot for 100 of each. I know its probably overkill, but you get a really nice set of numbers to work with, and the math is easier cause 100 is so easily divisable. (maths make my head hurts!)

Usually after 25 tests or so, you'll have a fairly good idea of what the general mechanics are, but 100 will allow you to rule out anomalous data.

It's especially awesome to do it this way if you have a lot of different people testing 100 of each, because then you can (almost) eliminate the human error portion of it.

[Kirschy]

While it would great if 100 tests were enough, I don't think it's the case. Thats why I asked if anyone fluent in statistics could help me. Assuming the crit rate+6% from senj DID affect the crit rate of Jin, 100 tests might only show 6 more criticals. And that doesn't seem enough to account for luck. To put it into perspective, if I flipped a coin 100 times and it landed heads 60 times and tails 40, it doesn't mean heads is more likely, it's just random luck. I'm mostly looking for an estimate on how much testing I'd need for a signifigant confidence interval or margin of error or whatever =x.

I hadn't heard of the crit+% weapons only affecting the first hit of a critical WS. Like you said it wouldn't matter for my test, but if you have information about this, I'd be very interested. The only tests I had read about were on Rampage at different TPs. At 100TP Rampage's first hit was equal to your natural crit% against the mob. At 200 there was a 20% increase, and at 300 there was a 40% increase. But unfortunately I can't find this test or any other tests related to critical WS =x.

[Thorny]

100 is well within margin of error testing for a 6% difference...

Edit: Beaten by Kirschy.. I'd say 300-500 is enough to please the general community, ~700-800 to call it fact beyond a reasonable doubt. It sounds excessive, but 6% is a miniscule difference making statistical abnomalies far more likely.

[CelinaAlexander]

He didn't mean only 100 tests, he meant doing sets of 100 tests and then averaging them.

[Wenslydale]

He didn't mean only 100 tests, he meant doing sets of 100 tests and then averaging them.

This. I guess earlier I should have clarified. Do as many sets as you like, I usually go for at least 10 (10 sets of 100 for a total of 1000) Again, keep the math easy, 99% of math errors are from fractions and decimal places.

[CelinaAlexander]

This. I guess earlier I should have clarified. Do as many sets as you like, I usually go for at least 10 (10 sets of 100 for a total of 1000) Again, keep the math easy, 99% of math errors are from fractions and decimal places.

The other 1% are from baseless statistics

[Valyana]

http://dimensionresearch.com/resourc...mple_size.html

tells me that to get 95% confidence that your measurement for a particular sample is within 3% of 9% crits, you need at least 350 trials. For within 3% of 15%, at least 544 trials. (I chose within 3% to avoid overlapping error bars)

The Rampage/Jin crit rate testing was done by Nagamaki, and posted in the old torques/dex/crit rate thread.

[Max™]

I can kind of comment on this second hand, due to my Rampager.

If the 6% Crit didn't work during WS, on all hits, there is no way to explain why it sould average the same or better than another weapon with +7~9 Dmg on it, much less any other stats.

I've had it parsed, friend compared his Byakko damages over several parties with the same gear versus a few Rampager parses, and never went back.

Just 6% on the first hit alone wouldn't be enough to beat 7 Dmg 5 Att.

As for Jin, I always understood it applied to the three mainhand hits, but not the offhand, and I would assume the first double, but not the second since you can only get two, one for each hand.

[Kirschy]

Yea, Max, what you're saying is what I expect to see. Its would be nice to see conclusive proof though. Doesn't affect me at all, but I'm always curious on that kind of thing. If I need to perform about ~350 tests x2... I'd be WSing for like 8-12 hours =x. Doesn't sound like anything I'd like to test anytime soon. Thanks for the responses~

[Max™]

Yeah, sucks even more using lvl 1 Bunnies.

Though, gigas pets in lower delkfutts should be 1 shotted, and low enough stats to clearly distinguish crit from non-crit.

[Hanyoko]

I've been tempted to test the effect "Enhances rate of critical hits" for a pet hypothesis I have (that it needs less dex over agility to reach crit cap).

Since this thread is somewhat related I felt the need to ask here, might any of you tester types know where to locate mobs with an agility around 50, or how to determine what a mob's agility is?

[Max™]

Determine it's job and level, check to determine it's evasion, from that you should be able to work backwards to the Agi.

Like, most monsters are War/War, so find what the stats would for a war subbing war at say, lvl 40, check it's evasion with +/- individual points of accuracy from gear til you can tell what point exactly makes it go neutral > low or high, use the maths on the Wiki I think, though Gobhli should have it somewhere.

Basically just need to find the monster level, and check for eva/def/traits (or if someone has a handy guide, like all ****** monsters are War/War), cause there will only be one or two ways to add up the Agi/2 and the stats that job combo should have to produce the check you see.

[Tordall]

I think the answer is that you're not looking for a 6% difference in the crit rate, you're looking for a 6% increase in the crit rate.

As such you need to get the standard deviation in each sample size low enough so both samples don't completely overlap with eachother. The standard deviation will actually depend a bit on the chance for the crit hit.

Standard deviation is: sqrt((p * (1-p)) / n)

Where p is the crit probability and n is the sample size.

So if you do 100 tests and crit 20 times, then the crit rate would be 0.20 +- 0.04. Which probably isn't satisfactory.

If you did 300 tests and crit 60 times, then the crit rate would be 0.20 +- 0.023. Which is looking better for distinguising between .20 and .26.

For finding the difference you'll want to subtract the two and add their uncertainties. Let's consider a base crit rate of 20% (0.20 probability), two samples sizes (100 and 300), and landing right on the probabilities.

100 sample: 20 crits & 26 crits. .20 +- .04 and .26 +- .043. Difference is .06 +- .083. E.g. 6% +- 8.3%. -- so, definitely not conclusive.

300 sample: 60 crits & 78 crits. .20 +- .023 and .26 +- .0253. Difference is .06 +- 0.0483. E.g. 6% +- 4.8% -- so, enough to tell the difference, but you can only be about 80% confident of the results (conference intervals are another topic).