Enhances the effect of...

[divisortheory]

Well, look at what causes the confidence interval to be wide in the first place. It's the -3*Deviation and +3*Deviation. Then look at the denominator inside the Square Root. It's /NumHits. When NumHits gets really big, the number inside the Sqrt gets really small. And when you take the square root of a smaller number, the result is smaller. So you end up subtracting less off of one end, and adding less to the other end of the confidence interval. How much less? Using the calculations earlier, Genome's 95% Confidence Interval had a range of about 8.27%.

Let's say he had done 2,000 trials though and gotten the same percentage. Then his confidence interval would have been like this:

CritPct = .163
TotalHits = 2000

95% Conf Interval Range = 6*Deviation = 6*Sqrt(.163*.837/2000) = .0496

Confidence Interval Range = 4.96%

If he had used 2500 samples, it would be

6*Sqrt(.163*.837/2500) = 4.43%

Epigram's real range will be different, but can't figure it out unless you know the actual crit% he parsed. It should give you an idea how many samples are needed for conclusive results though.

[divisortheory]

Back to the original topic though, lots of these items still need details!

[epigram]

95% CI btw = 1.96 standard deviations.

Pretty much everything is within 3 standard deviations of a normal distribution.

99% is usually defined as 2.54 standard deviations or something like that.

[genome]

I think the problem or what raises my eyebrow is that your data analysis rests on unfounded and possibly incorrect assumptions.

If you look at the conclusions you draw, it's pretty clear that SE did not implement critical hits simply as a set of independent bernoulli trials (calling a perfect random number generator and that's it without extra checks for example) and follows that same distribution curve. You are also assuming the distribution and deviation, which as I mentioned is nowhere near as wide. Or at least I would ask you to show test data to first prove that critical hits do in fact follow that distribution in the first place.

If you ask people who regularly parse in KRT what crit% they get (and KRT PTs have a lot more sources of error and variation, all of which are not even considered in the simplified assumption "it probably just follows a common bernoulli distribution"), they can tell you their crit% is actually very stable even on 700 hits or so.

If they were and everything was following every assumption in your analysis, repeating the experiment would also give you the same results as you suggested, that a confidence interval of 95% is that wide, anywhere within 12% to 20% to have a 95% chance to find crit% in that range.

Basically if you repeat the tests at least a few times, you will never see it fall that low.
If you try tests without p.body several times and it always ends up lower than AF, and osode+byakko ends up higher by a very repeatable amount, that's very convincing evidence, even before making statistical assumptions about distributions that might be incorrect (for example SE might implement a few extra "forced criticals" or "forced non-crit" hits when your random critical rate goes a bit too far off one end).

If you play around with the probability of this type of results happening by pure chance several times and repeatably is very low and it's much more likely your initial assumption was incorrect (or if you really want to use that statistical analysis you can add up trials from multiple runs to reduce your assumed deviation).

[Kriz]

I'm going to say that, as an educated guess, Blessed Briault +1 will give +6% MDB, if NQ give 5%.

This is due to the fact that all the other blessed pieces (which have haste) are boosted by one (well, for the fact, as are all the other stats). Like Blessed Mitts +1 are 6% Haste over NQ 5%.

I'll let you know the reality, you know, when I HQ the body. Sooner or later, gotta HQ the pants first!

[epigram]

I don't think you really understand the probability at hand with regard to a bernoulli distribution.

Let's take this scenario.

I go out and fight EM Goobbue. I don't change my gear, I don't use WS's, all I do is use the same stats against a Goobbue that I assume doesn't magically change it's stats during the fight.

Now, regardless of what formula SE uses, there is a probability of me landing a critical attack each attack. The only real assumption I make is that the probability of me landing a crit is the same for each hit if no stats change.

That is all I need to assume a bernoulli distribution.

The moment I have a bernoulli distribution and a large enough sample I can envoke CLT and use normality assumptions.

[Krandor]

Cover: 15 seconds unmerited with no Cover+

Will edit this post once I pull my Coronet out >.>

Are you sure it's 15 seconds? I tested it the other day without wearing a single piece of Cover+ (and no merits into it) and I swear I timed it as 30 seconds. >.>

Hell I'll Fraps it when I get home tonight.

Cover times likely seem variable because just about anything can fuck it up. Getting stunned, having the Target of cover no longer behind you, and all sorts of other assorted goodies can cancel the effect. I don't have any plans to use PLD today (Translation : Cerb and Tia aren't due) + AF1 head is in storage so I can't add any hard numbers to this discussion.

[genome]

I don't think you really understand the probability at hand with regard to a bernoulli distribution.

Let's take this scenario.

I go out and fight EM Goobbue. I don't change my gear, I don't use WS's, all I do is use the same stats against a Goobbue that I assume doesn't magically change it's stats during the fight.

Now, regardless of what formula SE uses, there is a probability of me landing a critical attack each attack. The only real assumption I make is that the probability of me landing a crit is the same for each hit if no stats change.

That is all I need to assume a bernoulli distribution.

The moment I have a bernoulli distribution and a large enough sample I can envoke CLT and use normality assumptions.

Yes, I understand what you are saying, critical is a yes/no event, and you assume every time you throw a punch is an independent event.

The point I was making is that it doesn't seem to work that way, it doesn't really match what you would expect the distribution to be, and you already know is a much more complex process since accuracy and a lot of other factors affect which hits land. Even the pseudo-random-number generator used can introduce a significant bias and make each "independent" event probability not independent at all.
I think it's too large of an oversimplification and I even gave you another example of some way it could be coded to actually change the distribution by enforcing hits every so many misses. All the data I have seen shows smaller variation values than you would predict with those assumptions.

I wouldn't discount your data, but i would like to see that instead of just a description of your analysis and derivation of the value 2-3%, which I think doesn't really match the data we are trying to study.
Making oversimplifying assumptions and choosing the wrong distribution of the data is understandable when we really don't know anything about something, but if the data doesn't reflect that distribution that's a flag you need to rethink some of your assumptions.

[Khamsin]

Even the pseudo-random-number generator used can introduce a significant bias and make each "independent" event probability not independent at all.

I thought gone were the days of using an Echo Screen before Joker Doom.

[BerenTebogo]

PLD NIN BRD trioing a NIN aern after other NIN DCed

PLD used cover on me while I threw on DD gear and went to town.

Cover did NOT last 30 seconds....<3 RR2

[Havok]

O: Drachen Brais+1 - Enhances "Ancient Circle" Effect
- Increases duration of [[Ancient Circle]] by 30 seconds.(Needs verification)

makes ancient circle last 1:30, not sure how much of an increase that is though

Conte Corazza - Jump TP Bonus
Conte Cosciales - Jump TP Bonus

i believe it's TP+3 instead of 2, but i dont own either piece so i can't verify that

and, I didn't see warrior's cuisses listed, does anyone have proof of what they actually do for double attack? i've heard different things

[aurik]

Yes, I understand what you are saying, critical is a yes/no event, and you assume every time you throw a punch is an independent event.

The point I was making is that it doesn't seem to work that way, it doesn't really match what you would expect the distribution to be, and you already know is a much more complex process since accuracy and a lot of other factors affect which hits land. Even the pseudo-random-number generator used can introduce a significant bias and make each "independent" event probability not independent at all.
I think it's too large of an oversimplification and I even gave you another example of some way it could be coded to actually change the distribution by enforcing hits every so many misses. All the data I have seen shows smaller variation values than you would predict with those assumptions.

I wouldn't discount your data, but i would like to see that instead of just a description of your analysis and derivation of the value 2-3%, which I think doesn't really match the data we are trying to study.
Making oversimplifying assumptions and choosing the wrong distribution of the data is understandable when we really don't know anything about something, but if the data doesn't reflect that distribution that's a flag you need to rethink some of your assumptions.

If you don't know anything about statistics, you really ought not talk about it. Epigram is completely right in that your tests, while interesting, failed to show to a significant degree that pln.body is less than or equal to the crit rate on the control piece.

Making the claim that there is bias introduced by the pseudoRNG is incredibly shortsighted; the quality of most RNGs are sufficiently random that they don't introduce any detectable bias. The largest complaint about pRNGs is that they often will never produce long strings of abnormal results, whereas with a true RNG it is theoretically possible but astronomically rare. However this effect is so minor that for all practical purposes it can be ignored. As a computer scientist, I'm pretty fed up with this quasi-argument repeated mostly by people who really don't have the foggiest clue what they are talking about.

As for critical hit variation, I did parse a lot in KRT and I often found that even after several thousand hits my critical rate varied from parse to parse by a fairly large margin, on the order of 1% or more. Of course, this is to be expected; it takes an incredibly huge number of samples before such a variation would be abnormal.

[epigram]

Making oversimplifying assumptions and choosing the wrong distribution of the data is understandable when we really don't know anything about something, but if the data doesn't reflect that distribution that's a flag you need to rethink some of your assumptions.

I think you are actually making things way too complex, and I think it's pretty safe to say that your most likely a CS person as opposed to a Statistician.

There are simple tests that can be run to ensure independence, they aren't hard, and there are even some non-parametric versions. Now, you can start looking into the prng, but all the methods I have used, including the very limited distribution assumptions, are all very conservative and would not be affected to any large degree.

Now, I could work from a non-parametric approach but you are only going to see greater variance. Regardless of the method, worst case is that the pln. body has unknown critical hit rate. You still can't argue that it is broken since their exists no data to prove either way.

[genome]

I think the prng is getting people sidetracked, I just used prng as a side point/example, like people have mentioned with the old trick of echo screens with rngs with short periods.
The main point is that the process is not random because it is coded with a specific algorithm, just because you are not sure how it is actually coded or what pattern it follows it doesn't mean that it's a set of independent events with a fixed probability equal to the average you see over large samples, you could say the same for crafting done at any random uniformly distributed crafting time, but if you know about time moon etc the distribution is quite different, because they coded it that way, instead of being just a series of random rolls. Heck for all we know they might even use the total numbers of HQ on a server to calibrate your synth results (just a random example).
It's not that you are or are not a statistician or know/don't know about CS, it's that it's very easy for SE to add lines like that, I see 1% deviation that aurik mentions even with less than 1000 punches while you are predicting a variation range much greater than that, and the data doesn't really support that (again if you find data supporting that it would help).
If you want to do independence tests and compare distribution predictions, that would help your claim too.

[Eulore]

Dunno if it's been answered but it wasn't on the first page, but Valor Coronet's Enhances Rampart Effect adds 15 seconds (30->45) to Rampart's Duration.

[epigram]

Your data doesn't support anything, and you can't claim anything without a statistically valid test. Now if you are going to believe in some greater conspiracy with the prng then it's impossible for you to accept anything anyway, and similarly disproove anything.

Honestly, I think your fishing way too hard. I'll do a run again this weekend, even on the same mob for 2000 hits with and w/o pln. I'll even do it for Sipahi Jawshan as well.

[BerenTebogo]

Coulda sworn this was a thread on effects of traits on gear

Not on the math that I tried so hard to sleep through in college.

[genome]

Your data doesn't support anything, and you can't claim anything without a statistically valid test. Now if you are going to believe in some greater conspiracy with the prng then it's impossible for you to accept anything anyway, and similarly disproove anything.

Honestly, I think your fishing way too hard. I'll do a run again this weekend, even on the same mob for 2000 hits with and w/o pln. I'll even do it for Sipahi Jawshan as well.

I think if you can come up with any solid data for your claims it would help more than you coming up with conspiracies about prng conspiracies...

Statistical analysis is more than just applying the simplest model you can think of, finding a confidence interval, and just saying it's a number between 1 and 5% so you guess it's 2-3%.
If you don't use the right distribution and the right assumptions about your data, your results and analysis are going to be incorrect and inconsistent with the actual reality, that's all I am saying.
If you were to do statistics for any scientific prpose, the first thing that people will question is why you chose a specific distribution and analysis for that process.

I stayed clear of that sort of biased interpretation by just presenting detailed data, so people can keep using it for future analysis and discussion. Then I pointed out that the fact that other items like dex+ gear have a clear consistent increase in critical hit rates while p.body actually has a lower critical rate compared to even baseline is quite strange and significant, it could be explained away if my variance was much higher but that was not the case.
When repeating with different body pieces (AF1 vs AF2) is still within 1% of the same value, while p.body is 6% less than predicted by the null hypothesis is significant and raises a flag, suggesting to me that 2-3% might not be correct or not hold in all circumstances and needs further testing.

Btw sorry about the derail, all this started by me just mentioning 2-3% crit% effect p.body effect doesn't seem right from what I tested so far, which as a brief comment was still in-topic with the thread =P

[Devek]

I fliped a coin 10 times and it landed on tails 8 of those 10 times.

Anyway, you don't need anything else to come up with a pretty good conclusion here. Most arguments are if pieces of armor are better than existing pieces of armor. The fact there is so much testing and argument about if that body piece does ANYTHING is proof that it is worthless.

I'll take shura togi where I know each hit will be stronger and I don't have to deal with some bullshit probability crap (which always screws me over.. 0/38 on RSE sachet..) that determines if I do or do not crit and get take advantage of p.body.

I think that settles it, now lets look at the rest of the gear on the list.

[NynJa]

I'll take shura togi where I know each hit will be stronger and I don't have to deal with some bullshit probability crap (which always screws me over.. 0/38 on RSE sachet..) that determines if I do or do not crit and get take advantage of p.body.

Too bad every time you punch your attack is put through some probability crap to see how high your roll is.