Does Level Correction apply to everything?

[Tordall]

w/ regard to the statistics you need a lot to get uncertainty down. Also, one standard deviation is 68% uncertainty. About 1/3 of the existing data should be outside the measurement's uncertainty bars.

w/ regard to the actual topic. I was always under the impression that level correction applied to player's hit rate against mobs (my accuracy as ninja only really made sense if you assumed 5% level corrections). I'm also somewhat curious about the 2 handed acc forumla's accuracy, a lot of ares & shadow breastplate peeps seem to have at-or-near capped accuracy in places you wouldn't expect them to.

[Fwahm]

On topic of level correction, i would need some confirmation regarding /checks for mob's evasion/defense :



but, during my reads on various forums, i found also this :



Question :

Does the 1st quote formulas are "right to use" only for Even Match mobs ?

(Would mean i can already restart all my mob database -.-)

The checks only compare the raw stats; the level corrections are only applied later, possibly when you actually swing at a mob.

[pchan]

Thank you for your answer Pchan.

There is something i dont understand, and sound kinda illogical when i consider these 2 following situations :

Situation 1: parsing crit rates for a MobA, 1k sample size, assume i get p=15% => error=2.2%. Fine np.

Situation 2: parsing crit rates for a MobB, 1k sample size, keeping everything same testing conditions except mob, assume i get p=12% => error will be different if i follow the equation you proposed, since same sample size.
In reality, sounds logical to me if MobB = same mob but 4 levels higher (higher AGI), p will be lower obviously, so should get SAME error for SAME sample size ?
Did i miss something ?

Sorry but there my common sense fail. Please enlighten me

Edit: mmm maybe i found my mistake :

In Situation 1 (p=15%) i should have used N=1000*1.15=1150 total hits => 2% error
In Situation 2 (p=12%) i should have used N=1000*1.12=1120 total hits => 1.9% error
right or wrong ?

I'm not sure what you want to do. You can indeed have an approximate idea of the number of hits that you need to have a given error. IMO if you want to be able to statistically distinguish between 12% and 15% crit rates you need a very high sample size. As you see, with ~2% error the intervalls overlap ([10%,14%] and [13%,17%]). So the first thing needed is to have a sample size high enough (so that the error < 1.5%) but even then you can't be certain it's 2% difference. IMO a +/- 0.5% is acceptable but if you
do the math you are looking at 16000 sample size...

[Masamune]

As you see, with ~2% error the intervalls overlap ([10%,14%] and [13%,17%]).

erm... that 2% error shouldnot it be applied on the estimated p? ie the intervall would become : [p*(1-error);p*(1+error)]=[p*98%;p*102%]
so for p~12% => [11.76%;12.24%]
and for p~15% => [14.7%;15.3%] and that for N~1150hits
right or wrong ?

[pchan]

erm... that 2% error shouldnot it be applied on the estimated p? ie the intervall would become : [p*(1-error);p*(1+error)]=[p*98%;p*102%]
so for p~12% => [11.76%;12.24%]
and for p~15% => [14.7%;15.3%] and that for N~1150hits
right or wrong ?

it's not p*(1+error), it would be too easy ! The error is an absolute error, not a relative one. So it's [p-err,p+err].

[CDF]

w/ regard to the statistics you need a lot to get uncertainty down. Also, one standard deviation is 68% uncertainty. About 1/3 of the existing data should be outside the measurement's uncertainty bars.

Typical nonsense statement (ignore it) regarding statistics like all the ones in "multi-hit WS accuracy testing."

Situation 1: parsing crit rates for a MobA, 1k sample size, assume i get p=15% => error=2.2%. Fine np.

Situation 2: parsing crit rates for a MobB, 1k sample size, keeping everything same testing conditions except mob, assume i get p=12% => error will be different if i follow the equation you proposed, since same sample size.
In reality, sounds logical to me if MobB = same mob but 4 levels higher (higher AGI), p will be lower obviously, so should get SAME error for SAME sample size ?
Did i miss something ?

You can just think of it as a consequence of statistics and probability theory.

It's just that the actual standard error of the sample proportion (p) is itself a function of the actual proportion, so, for a given sample size n, the actual standard error will be different based on what the actual proportion is. However, when you take a sample, you don't know what the actual proportion (and therefore the actual standard error) is, so you have to use the estimated proportion, and the estimated standard error follows from that.