...one of the most highly
regarded and expertly designed C++ library projects in the
world.

— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards

Imagine you have a process that follows a negative binomial distribution: for each trial conducted, an event either occurs or does it does not, referred to as "successes" and "failures". The frequency with which successes occur is variously referred to as the success fraction, success ratio, success percentage, occurrence frequency, or probability of occurrence.

If, by experiment, you want to measure the the best estimate of success
fraction is given simply by *k* / *N*,
for *k* successes out of *N* trials.

However our confidence in that estimate will be shaped by how many trials
were conducted, and how many successes were observed. The static member
functions `negative_binomial_distribution<>::find_lower_bound_on_p`

and `negative_binomial_distribution<>::find_upper_bound_on_p`

allow you to calculate the confidence intervals for your estimate of
the success fraction.

The sample program neg_binom_confidence_limits.cpp illustrates their use.

First we need some includes to access the negative binomial distribution (and some basic std output of course).

#include <boost/math/distributions/negative_binomial.hpp> using boost::math::negative_binomial; #include <iostream> using std::cout; using std::endl; #include <iomanip> using std::setprecision; using std::setw; using std::left; using std::fixed; using std::right;

First define a table of significance levels: these are the probabilities that the true occurrence frequency lies outside the calculated interval:

double alpha[] = { 0.5, 0.25, 0.1, 0.05, 0.01, 0.001, 0.0001, 0.00001 };

Confidence value as % is (1 - alpha) * 100, so alpha 0.05 == 95% confidence
that the true occurrence frequency lies **inside**
the calculated interval.

We need a function to calculate and print confidence limits for an observed frequency of occurrence that follows a negative binomial distribution.

void confidence_limits_on_frequency(unsigned trials, unsigned successes) { // trials = Total number of trials. // successes = Total number of observed successes. // failures = trials - successes. // success_fraction = successes /trials. // Print out general info: cout << "______________________________________________\n" "2-Sided Confidence Limits For Success Fraction\n" "______________________________________________\n\n"; cout << setprecision(7); cout << setw(40) << left << "Number of trials" << " = " << trials << "\n"; cout << setw(40) << left << "Number of successes" << " = " << successes << "\n"; cout << setw(40) << left << "Number of failures" << " = " << trials - successes << "\n"; cout << setw(40) << left << "Observed frequency of occurrence" << " = " << double(successes) / trials << "\n"; // Print table header: cout << "\n\n" "___________________________________________\n" "Confidence Lower Upper\n" " Value (%) Limit Limit\n" "___________________________________________\n";

And now for the important part - the bounds themselves. For each value
of *alpha*, we call `find_lower_bound_on_p`

and `find_upper_bound_on_p`

to obtain lower and upper bounds respectively. Note that since we are
calculating a two-sided interval, we must divide the value of alpha in
two. Had we been calculating a single-sided interval, for example: *"Calculate
a lower bound so that we are P% sure that the true occurrence frequency
is greater than some value"* then we would **not**
have divided by two.

// Now print out the upper and lower limits for the alpha table values. for(unsigned i = 0; i < sizeof(alpha)/sizeof(alpha[0]); ++i) { // Confidence value: cout << fixed << setprecision(3) << setw(10) << right << 100 * (1-alpha[i]); // Calculate bounds: double lower = negative_binomial::find_lower_bound_on_p(trials, successes, alpha[i]/2); double upper = negative_binomial::find_upper_bound_on_p(trials, successes, alpha[i]/2); // Print limits: cout << fixed << setprecision(5) << setw(15) << right << lower; cout << fixed << setprecision(5) << setw(15) << right << upper << endl; } cout << endl; } // void confidence_limits_on_frequency(unsigned trials, unsigned successes)

And then call confidence_limits_on_frequency with increasing numbers of trials, but always the same success fraction 0.1, or 1 in 10.

int main() { confidence_limits_on_frequency(20, 2); // 20 trials, 2 successes, 2 in 20, = 1 in 10 = 0.1 success fraction. confidence_limits_on_frequency(200, 20); // More trials, but same 0.1 success fraction. confidence_limits_on_frequency(2000, 200); // Many more trials, but same 0.1 success fraction. return 0; } // int main()

Let's see some sample output for a 1 in 10 success ratio, first for a mere 20 trials:

______________________________________________ 2-Sided Confidence Limits For Success Fraction ______________________________________________ Number of trials = 20 Number of successes = 2 Number of failures = 18 Observed frequency of occurrence = 0.1 ___________________________________________ Confidence Lower Upper Value (%) Limit Limit ___________________________________________ 50.000 0.04812 0.13554 75.000 0.03078 0.17727 90.000 0.01807 0.22637 95.000 0.01235 0.26028 99.000 0.00530 0.33111 99.900 0.00164 0.41802 99.990 0.00051 0.49202 99.999 0.00016 0.55574

As you can see, even at the 95% confidence level the bounds (0.012 to 0.26) are really very wide, and very asymmetric about the observed value 0.1.

Compare that with the program output for a mass 2000 trials:

______________________________________________ 2-Sided Confidence Limits For Success Fraction ______________________________________________ Number of trials = 2000 Number of successes = 200 Number of failures = 1800 Observed frequency of occurrence = 0.1 ___________________________________________ Confidence Lower Upper Value (%) Limit Limit ___________________________________________ 50.000 0.09536 0.10445 75.000 0.09228 0.10776 90.000 0.08916 0.11125 95.000 0.08720 0.11352 99.000 0.08344 0.11802 99.900 0.07921 0.12336 99.990 0.07577 0.12795 99.999 0.07282 0.13206

Now even when the confidence level is very high, the limits (at 99.999%, 0.07 to 0.13) are really quite close and nearly symmetric to the observed value of 0.1.