@Emptyset
2015-07-16T12:13:35.000000Z
字数 2489
阅读 3111
概率论札记
只有愚蠢的人才会相信眼睛看到的。
——安·兰德
故事要从一道贝叶斯定理的简单习题讲起。大意是艾滋病患病率为万分之一,误诊率为5%,患有艾滋病者被诊断出来的概率为99%,请问在这样的设定下如果你被诊断为艾滋病阳性,那么你患艾滋病的概率是多少,原题如下——
Problem Denoted blood is screened for AIDS. Suppose the test has 99% accuracy, and that one in ten thousand people in your age group are HIV positive. The test has a 5% false positive rating, as well. Suppose the test screens you as positive. What is the probability you have AIDS? Is it 99%?
Solution: E_1="test positive", E_2="test negative". A_1="You have AIDS", A_2="You don't have AIDS". Now we know
P(E1|A1)=99% , we need to findP(A1|E1) . Since "one in ten thousand people in your age group are HIV positive",P(A1)=1/10000 ."5% false positive rating" meansP(E1|A2)=5% . By Bayes' Theorem
P(A1|E1)==≈P(E1|A1)P(A1)P(E1|A1)P(A1)+P(E1|A2)P(A2)99%×11000099%×110000+5%×9999100000.198%
Note: 是不是看起来结论很不可思议?细细想来就知道是合理的,原因在于
讨论:联想到实际医疗中的误诊,逻辑上说,误诊可以简单分为两种——没病的看成有病了,或者有病的没有看出来。如果我们简单把前者定义为误诊,即定义误诊为“把没病的诊断成有病”,那么上面这道题目其实是在计算艾滋病诊断成阳性的可靠性,而题目中的误诊率是
(1)让我们把
在这个故事里,贝叶斯定理告诉我们一个稍微“反常识”的道理:即使误诊率从数字上看已经很低了,诊断结果的可靠性也依旧无法保证足够高,诊断可靠性受到发病率的约束。