Class BoundsOnBinomialProportions

• public final class BoundsOnBinomialProportions
extends Object
Confidence intervals for binomial proportions.

This class computes an approximation to the Clopper-Pearson confidence interval for a binomial proportion. Exact Clopper-Pearson intervals are strictly conservative, but these approximations are not.

The main inputs are numbers n and k, which are not the same as other things that are called n and k in our sketching library. There is also a third parameter, numStdDev, that specifies the desired confidence level.

• n is the number of independent randomized trials. It is given and therefore known.
• p is the probability of a trial being a success. It is unknown.
• k is the number of trials (out of n) that turn out to be successes. It is a random variable governed by a binomial distribution. After any given batch of n independent trials, the random variable k has a specific value which is observed and is therefore known.
• pHat = k / n is an unbiased estimate of the unknown success probability p.

Alternatively, consider a coin with unknown heads probability p. Where n is the number of independent flips of that coin, and k is the number of times that the coin comes up heads during a given batch of n flips. This class computes a frequentist confidence interval [lowerBoundOnP, upperBoundOnP] for the unknown p.

Conceptually, the desired confidence level is specified by a tail probability delta.

Ideally, over a large ensemble of independent batches of trials, the fraction of batches in which the true p lies below lowerBoundOnP would be at most delta, and the fraction of batches in which the true p lies above upperBoundOnP would also be at most delta.

Setting aside the philosophical difficulties attaching to that statement, it isn't quite true because we are approximating the Clopper-Pearson interval.

Finally, we point out that in this class's interface, the confidence parameter delta is not specified directly, but rather through a "number of standard deviations" numStdDev. The library effectively converts that to a delta via delta = normalCDF (-1.0 * numStdDev).

It is perhaps worth emphasizing that the library is NOT merely adding and subtracting numStdDev standard deviations to the estimate. It is doing something better, that to some extent accounts for the fact that the binomial distribution has a non-gaussian shape.

In particular, it is using an approximation to the inverse of the incomplete beta function that appears as formula 26.5.22 on page 945 of the "Handbook of Mathematical Functions" by Abramowitz and Stegun.

Author:
Kevin Lang
• Method Summary

All Methods
Modifier and Type Method and Description
static double approximateLowerBoundOnP(long n, long k, double numStdDevs)
Computes lower bound of approximate Clopper-Pearson confidence interval for a binomial proportion.
static double approximateUpperBoundOnP(long n, long k, double numStdDevs)
Computes upper bound of approximate Clopper-Pearson confidence interval for a binomial proportion.
static double erf(double x)
Computes an approximation to the erf() function.
static double estimateUnknownP(long n, long k)
Computes an estimate of an unknown binomial proportion.
static double normalCDF(double x)
Computes an approximation to normalCDF(x).
• Method Detail

• approximateLowerBoundOnP

public static double approximateLowerBoundOnP(long n,
long k,
double numStdDevs)
Computes lower bound of approximate Clopper-Pearson confidence interval for a binomial proportion.

Implementation Notes:
The approximateLowerBoundOnP is defined with respect to the right tail of the binomial distribution.

• We want to solve for the p for which sumj,k,nbino(j;n,p) = delta.
• We now restate that in terms of the left tail.
• We want to solve for the p for which sumj,0,(k-1)bino(j;n,p) = 1 - delta.
• Define x = 1-p.
• We want to solve for the x for which Ix(n-k+1,k) = 1 - delta.
• We specify 1-delta via numStdDevs through the right tail of the standard normal distribution.
• Smaller values of numStdDevs correspond to bigger values of 1-delta and hence to smaller values of delta. In fact, usefully small values of delta correspond to negative values of numStdDevs.
• return p = 1-x.
Parameters:
n - is the number of trials. Must be non-negative.
k - is the number of successes. Must be non-negative, and cannot exceed n.
numStdDevs - the number of standard deviations defining the confidence interval
Returns:
the lower bound of the approximate Clopper-Pearson confidence interval for the unknown success probability.
• approximateUpperBoundOnP

public static double approximateUpperBoundOnP(long n,
long k,
double numStdDevs)
Computes upper bound of approximate Clopper-Pearson confidence interval for a binomial proportion.

Implementation Notes:
The approximateUpperBoundOnP is defined with respect to the left tail of the binomial distribution.

• We want to solve for the p for which sumj,0,kbino(j;n,p) = delta.
• Define x = 1-p.
• We want to solve for the x for which Ix(n-k,k+1) = delta.
• We specify delta via numStdDevs through the right tail of the standard normal distribution.
• Bigger values of numStdDevs correspond to smaller values of delta.
• return p = 1-x.
Parameters:
n - is the number of trials. Must be non-negative.
k - is the number of successes. Must be non-negative, and cannot exceed n.
numStdDevs - the number of standard deviations defining the confidence interval
Returns:
the upper bound of the approximate Clopper-Pearson confidence interval for the unknown success probability.
• estimateUnknownP

public static double estimateUnknownP(long n,
long k)
Computes an estimate of an unknown binomial proportion.
Parameters:
n - is the number of trials. Must be non-negative.
k - is the number of successes. Must be non-negative, and cannot exceed n.
Returns:
the estimate of the unknown binomial proportion.
• erf

public static double erf(double x)
Computes an approximation to the erf() function.
Parameters:
x - is the input to the erf function
Returns:
returns erf(x), accurate to roughly 7 decimal digits.
• normalCDF

public static double normalCDF(double x)
Computes an approximation to normalCDF(x).
Parameters:
x - is the input to the normalCDF function
Returns:
returns the approximation to normalCDF(x).