### A/B Split Test Calculator / G-test Calculator - aTool在线工具

• G-test Calculator支持Yates' continuity correction，可以用于检测多个方案统计数据的优胜情况以及结果可信度。一般大于95%才算可信

#### G-test说明 | Introduce

• 在一组测量值时，有时会出现个别数据与其他数据相差较大的可疑值。在无法证明是实验失误引起的误差时，我们可以用统计学方法予以取舍。G检验（G-test)就是其中一种常用的方法。
• 在统计中，G-测试正在越来越多地被使用的情况，其中卡方检验先前建议的似然比或最大似然的统计显着性检验。
• 对于一个合理的规模的样品中，G-检验和卡方检验会导致同样的结论。然而，近似为G-测试的理论卡方分布比为Pearson卡方检验更好。

#### G-test | Introduce

• This form does a G-test calculation to determine the outcome of an A/B test. The G-test statistic is a measure of how much overall variation there is from an ideal prediction that you would expect if all versions were the same. In the case where all versions are the same and there are many trials, the distribution of this statistic is known. If the statistic we have is very unlikely, then we have good evidence that we are seeing real differences of some sort. Our confidence that there is a real difference is 100% minus the odds of a less likely G-test statistic than the one we saw.
• At what confidence do we end the test? There is no hard rule, however a common benchmark is 99% confidence. Please note that 99% confidence does not mean you are right 99% of the time. How often you're right depends in complicated ways on information you can't get, like what the size of the real difference is. Instead it means that if there was no real difference, then there was only a 1% chance of your making a mistake at the point in time where you made your decision.
• The fields should be self-explanatory except the Yates' continuity correction. The Yates' continuity correction adjusts the numbers slightly to take into account the fact that the test results must be a discrete number. This makes the test slightly more accurate for small sample sizes, but the difference is seldom material. It defaults to true.