Thread:Hayashi H/@comment-749631-20160105173522/@comment-26486243-20160105191658

Pretty much what it looks like from my point of view as well, when I saw the source. But one does not simply dump multiple variables from a random sample into a computer, ask it to find trends and use it as hard evidence.

If a test has a 95% confidence level for instance that an association can be found, then there is a 5% chance that an association is detected when there is actually none based on random chance alone. The larger the number of variables analysed at the same time, the chance that any one of them is not a true association, but a quirk of the sample taken, will increase [i]exponentially[/i].

This kind of investigation is not useless, mind, because it proposes possible variables that can then be tested. Based on the factors this test identified, separate tests that hold all variables constant except one with sufficient sample size can show if each of these identified variables actually affect the end-result of accuracy, and if they do, how much of an effect has.

To quantify the effect of intrinsic torpedo stat accuracy, the easiest way is to test a lvl 99 non-fully torpmodded submarine for a 4 digit number of samples, then torpedo-modernise the same submarine and test again on the same target for a 4 digit number of samples. If quantifying is not needed and we just want to check if the effect exists or not, then 100 samples will be enough.

For equipment torpedo stat accuracy, the idea is to use different torpedoes with same accuracy value against same target. Probably triple torpedo vs quad o2 torpedo.

For luck, the same sub before and after luck-modernisation with same equips on same target.

For upgrading, same sub, same torpedo type, except zero star mod vs max star mod.

Finally for evasion, to compare the effects between targets with different evasion values with the same sub, same equipment.

It goes without saying all tests need to be done under red morale conditions so that the accuracy soft cap effect cannot happen. This isn't that hard to test in 1-1, since the first node has a single destroyer in Line Ahead with possible EVA values of 14, 15 and 16, and Shioi Kai will never take any damage more than scratch in that node unless the enemy DD crits - and even that will only happen if Shioi misses the initial torpedo.

As it is it's only useful for idea generation, and isn't anywhere near conclusive enough - having only a three digit number of samples with a methodology that isn't clearly defined attempting to quantify effect sizes for >5 variables at once... It's as good as anecdotal observations where 'proving' anything is concerned.