Source: Sigma Zone, by Philip Mayfield

The Binomial Distribution is commonly used in statistics in a variety of applications. Binomial data and statistics are presented to us daily. For example, in the election of political officials we may be asked to choose between two candidates. Polling organizations often take samples of “likely voters” in an attempt to predict who will be elected before the actual election occurs.

To illustrate this, let’s assume that two candidates are running in an election for Governor of California. This fictitious election pits Mr. Gubinator vs. Mr. Ventura. We would like to know who is winning the race, and therefore we conduct a poll of likely voters in California. If the poll gives the voters a choice between the two candidates, then the results can be reasonably modeled with the Binomial Distribution. In our poll of 50 likely voters, 58% indicate they intend to vote for Mr. Gubinator. Does this mean that 58% of all voters intend to vote for Mr. Gubinator? Probably not. If we were to repeat this poll several times in the same day (using a different group of 50 each time) we would find that the percentage that intends to vote for Mr. Gubinator would change with each poll.

The poll and most binomial samples come with some error. When polls are presented in the media, on the bottom of the screen or page you often see a small note with wording similar to “Margin of error +/- 5%”. This +/-5% indicates that if the poll was repeated multiple times, the result would likely fall in the range of 58% +/- 5%, or 53% to 63%. The margin of error is also called the confidence interval and is used to describe how much uncertainty we have in the sample estimate. There are several ways to estimate the Binomial Confidence Interval (CI); in this article we will focus on the Normal Approximation Method and the Clopper-Pearson Method.

Normal Approximation Method of the Binomial Confidence Interval

The equation for the Normal Approximation for the Binomial CI is shown below.

where:

p = proportion of interest

n = sample size

α = desired confidence

z1- α/2 = “z value” for desired level of confidence

z1- α/2 = 1.96 for 95% confidence

z1- α/2 = 2.57 for 99% confidence

z1- α/2 = 3 for 99.73% confidence

Using our previous example, if a poll of 50 likely voters resulted in 29 expressing their desire to vote for Mr. Gubinator, the resulting 95% CI would be calculated as follows.

Thus, we would be 95% confident that the proportion of the target population (all voters in California) who intend to vote for Mr. Gubernator falls between 44% and 72%.

While this method is very easy to teach and understand, you may have noticed that z1- α/2 is derived from the Normal Distribution and not the Binomial Distribution. The use of the z value from the Normal Distribution is where the method earns its moniker “Normal Approximation”. While the use of the Normal Distribution seems odd at first, it is supported by the central limit theorem and with sufficiently large n, the Normal Distribution is a good estimate of the Binomial Distribution.

However, there are times when the Normal Distribution is not a good estimator of the Binomial. When p is very small or very large, the Normal Approximation starts to suffer from increased inaccuracy. Specifically, when np > 5 or n(1-p)>5 the Normal Approximation method should not be used [1]. Additionally, if you try to calculate any CI with p=0 or p=1, you will find that it is not possible.

Normal Approximation Summary

  • Advantages

    Easy to teach and understand

    Easy to calculate by hand
  • Disadvantages

    Accuracy suffers when np < 5 or n(1-p)<5

    Calculation not possible when p =0 or p=1

Exact Confidence Interval

The deficiencies in the Normal Approximation were addressed by Clopper and Pearson when they developed the Clopper-Pearson method which is commonly referred to as the “Exact Confidence Interval” [3]. Instead of using a Normal Approximation, the Exact CI inverts two single-tailed Binomial test at the desired alpha. Specifically, the Exact CI is range from plb to pub that satisfies the following conditions [2].

The population proportion falls in the range

plb to pub where:

plb is the confidence interval lower bound

pub is the confidence interval upper bound

n is the number of trials

k is the number of successes in n trials

α is the percent chance of making a Type I error, 1-α is the confidence

While the Normal Approximation method is easy to teach and understand, I would rather deliver a lesson on quantum mechanics than attempt to explain the equations behind the Exact Confidence Interval. While the population proportion falls in the range plb to pub, the calculation of these values is non-trivial and for most requires the use of a computer. You may note that the equations above are based upon the Binomial Cumulative Distribution Function (cdf). The Beta Distribution can be used to calculate the Binomial cdf, and so a more common way to represent the Binomial Exact CI is using the equations below.

The F Distribution can also be used to estimate the Binomial cdf, and so alternative formulas use the F in lieu of the Beta Distribution.

Exact Binomial Confidence Interval Summary

  • Advantages

    Accurate when np > 5 or n(1-p)>5

    Calculation is possible when p =0 or p=1
  • Disadvantages

    Formulas are complex and require computers to calculate

Which to use

The Normal Approximation method serves as a simple way to introduce the idea of the confidence interval. The formula is easy to understand and calculate, which allows the student to easily grasp the concept. However, the inaccuracies with very small p or the inability handle p=0 is a somewhat severe limitation in business applications. For example, if a test of 10 cell phones reveals zero defects, what is the confidence interval of the defective phones in the total population? This question is commonly posed and yet the Normal Approximation cannot be used to find an answer. As personal computers with ample calculation power have become prevalent, there is a trend towards using the Exact CI in lieu the Normal Approximation. At SigmaZone.com, we believe that the best method is to teach the concept using the Normal Approximation method and then tell the students that it is just an approximation. We then point out that the software calculates the exact confidence interval which can handle p=0 or p=1.

Final Notes

The term “Exact Confidence Interval” is a bit of a misnomer. Neyman noted [4] that “exact probability statements are impossible in the case of the Binomial Distribution”. This stems from the fact that k, the number of successes in n trials, must be expressed as an integer. Various methods have been suggested as improvements to the Exact CI, including the Wilson Method and the Modified Wilson Method.

Finally, to avoid a flood of emails I should note that the binomial distribution is a discrete probability distribution used to model the number of successes in n independent binomial experiments that have a constant probability of success p. The election example may not be applicable in that during the poll someone might indicate that they neither want to vote for Mr. Gubinator or Mr. Ventura or put another way, they have no preference. If this is the case, there are now three options, Mr. Gubinator, Mr. Ventura, and No Preference and the experiment is no longer binomial as there are three choices instead of two.

References

[1] Brown, L. D., Cai, T. T., and DasGupta, A. Interval Estimation for a Binomial Proportion. Statistical Science 16: 101-117, 2001.

[2] Gnedenko, B.V., Ushakov I.A., Pavlov I.V.. Statistical Reliability Engineering. Wiley, John & Sons, April 1999.

[3] Clopper, C. and Pearson, S. The use of confidence or fiducial limits illustrated in the case of the Binomial. Biometrika 26: 404-413, 1934.

[4] Neyman, J. On the problem of confidence intervals. The Annals of Mathematical Statistics, 6, 116, 1935.

Understanding Binomial Confidence Intervals 二项分布的置信区间的更多相关文章

  1. confidence intervals and precision|The One-Mean z-Interval Procedure|When to Use the One-Mean z-Interval Procedure

    Confidence Intervals for One Population Mean When σ Is Known Obtaining Confidence Intervals for a Po ...

  2. Permutation test: p, CI, CI of P 置换检验相关统计量的计算

    For research purpose, I've read a lot materials on permutation test issue. Here is a summary. Should ...

  3. 【概率论】5-2:伯努利和二项分布(The Bernoulli and Binomial Distributions)

    title: [概率论]5-2:伯努利和二项分布(The Bernoulli and Binomial Distributions) categories: - Mathematic - Probab ...

  4. frequentism-and-bayesianism-chs-ii

    frequentism-and-bayesianism-chs-ii 频率主义 vs 贝叶斯主义 II:当结果不同时   这个notebook出自Pythonic Perambulations的博文  ...

  5. R语言错误的提示(中英文翻译)

    # Chinese translations for R package # Copyright (C) 2005 The R Foundation # This file is distribute ...

  6. C4.5算法(摘抄)

    1. C4.5算法简介 C4.5是一系列用在机器学习和数据挖掘的分类问题中的算法.它的目标是监督学习:给定一个数据集,其中的每一个元组都能用一组属性值来描述,每一个元组属于一个互斥的类别中的某一类.C ...

  7. scipy.stats

    scipy.stats Scipy的stats模块包含了多种概率分布的随机变量,随机变量分为连续的和离散的两种.所有的连续随机变量都是rv_continuous的派生类的对象,而所有的离散随机变量都是 ...

  8. 小马哥课堂-统计学-t分布

    T distribution 定义 在概率论和统计学中,学生t-分布(t-distribution),可简称为t分布,用于根据小样本来估计 呈正态分布且方差未知的总体的均值.如果总体方差已知(例如在样 ...

  9. Bayesian Statistics for Genetics | 贝叶斯与遗传学

    Common sense reduced to computation - Pierre-Simon, marquis de Laplace (1749–1827) Inventor of Bayes ...

随机推荐

  1. python基础之数据类型(二)

    Python3 元组 Python 的元组与列表类似,不同之处在于元组的元素不能修改. 元组使用小括号,列表使用方括号. 元组创建很简单,只需要在括号中添加元素,并使用逗号隔开即可. 不可变的tupl ...

  2. java多线程-信号量

    Semaphore(信号量)是一个线程同步结构,用于在线程间传递信号,以避免出现信号丢失,或者像锁一样用于保护一个关键区域.自从 5.0 开始,jdk 在 java.util.concurrent 包 ...

  3. javascript中DOM部分基础知识总结

    1.DOM介绍      1.1 DOM概念      文档对象模型(Document Object Model),它定义了访问和处理HTML文档的标准方法.现在我们主要接触到的是HTML DOM. ...

  4. .NET 各种框架

    基于.NET平台常用的框架整理 分布式缓存框架: Microsoft Velocity:微软自家分布式缓存服务框架. Memcahed:一套分布式的高速缓存系统,目前被许多网站使用以提升网站的访问速度 ...

  5. git mv 命令 移动或重命名

    最基本的代码: git mv file_old file_new 手册: NAME git-mv - Move or rename a file, a directory, or a symlink ...

  6. JavaScript中this指向的简单理解

    首先必须要说的是,this的指向在函数定义的时候是确定不了的,只有函数执行的时候才能确定this到底指向谁,实际上this的最终指向的是那个调用它的对象(这句话有些问题,后面会解释为什么会有问题,虽然 ...

  7. arcgis server10.2.2发布地图基础服务的具体步骤

    1.直接打开制作好的.mxd文档,比如这里: 2.打开mxd文档之后,打开菜单:file-share as -services 弹出地图发布服务的界面: 点击publish之后,耐心的等待一段时间,地 ...

  8. SharePoint 2013 工作流之使用Designer配置示例篇

    在SharePoint 2013中,支持SharePoint Designer 2013(以下简称SPD)配置简单的工作流,完成我们的业务需要.下面,我们就举一个小例子,实现SPD配置工作流. 1. ...

  9. 解决xcode升级之后安装的插件失效

    title: 解决xcode升级之后安装的插件失效date: 2015-08-23 11:07:53categories: 编辑工具 tags: xcode 我的博客:http://daycoding ...

  10. Google C++单元测试框架GoogleTest---TestFixture使用

    一.测试夹具(Test Fixtures):对多个测试使用相同的数据配置 如果你发现自己写了两个或更多的测试来操作类似的数据,你可以使用测试夹具.它允许您为几个不同的测试重复使用相同的对象配置. 要创 ...