...one of the most highly
regarded and expertly designed C++ library projects in the
world.
— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards
Imagine that you want to compare the standard deviations of two sample to determine if they differ in any significant way, in this situation you use the F distribution and perform an F-test. This situation commonly occurs when conducting a process change comparison: "is a new process more consistent that the old one?".
In this example we'll be using the data for ceramic strength from http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm. The data for this case study were collected by Said Jahanmir of the NIST Ceramics Division in 1996 in connection with a NIST/industry ceramics consortium for strength optimization of ceramic strength.
The example program is f_test.cpp, program output has been deliberately made as similar as possible to the DATAPLOT output in the corresponding NIST EngineeringStatistics Handbook example.
We'll begin by defining the procedure to conduct the test:
void f_test( double sd1, // Sample 1 std deviation double sd2, // Sample 2 std deviation double N1, // Sample 1 size double N2, // Sample 2 size double alpha) // Significance level {
The procedure begins by printing out a summary of our input data:
using namespace std; using namespace boost::math; // Print header: cout << "____________________________________\n" "F test for equal standard deviations\n" "____________________________________\n\n"; cout << setprecision(5); cout << "Sample 1:\n"; cout << setw(55) << left << "Number of Observations" << "= " << N1 << "\n"; cout << setw(55) << left << "Sample Standard Deviation" << "= " << sd1 << "\n\n"; cout << "Sample 2:\n"; cout << setw(55) << left << "Number of Observations" << "= " << N2 << "\n"; cout << setw(55) << left << "Sample Standard Deviation" << "= " << sd2 << "\n\n";
The test statistic for an F-test is simply the ratio of the square of the two standard deviations:
F = s12 / s22
where s1 is the standard deviation of the first sample and s2 is the standard deviation of the second sample. Or in code:
double F = (sd1 / sd2); F *= F; cout << setw(55) << left << "Test Statistic" << "= " << F << "\n\n";
At this point a word of caution: the F distribution is asymmetric, so we have to be careful how we compute the tests, the following table summarises the options available:
Hypothesis |
Test |
---|---|
The null-hypothesis: there is no difference in standard deviations (two sided test) |
Reject if F <= F(1-alpha/2; N1-1, N2-1) or F >= F(alpha/2; N1-1, N2-1) |
The alternative hypothesis: there is a difference in means (two sided test) |
Reject if F(1-alpha/2; N1-1, N2-1) <= F <= F(alpha/2; N1-1, N2-1) |
The alternative hypothesis: Standard deviation of sample 1 is greater than that of sample 2 |
Reject if F < F(alpha; N1-1, N2-1) |
The alternative hypothesis: Standard deviation of sample 1 is less than that of sample 2 |
Reject if F > F(1-alpha; N1-1, N2-1) |
Where F(1-alpha; N1-1, N2-1) is the lower critical value of the F distribution with degrees of freedom N1-1 and N2-1, and F(alpha; N1-1, N2-1) is the upper critical value of the F distribution with degrees of freedom N1-1 and N2-1.
The upper and lower critical values can be computed using the quantile function:
F(1-alpha; N1-1, N2-1) = quantile(fisher_f(N1-1, N2-1),
alpha)
F(alpha; N1-1, N2-1) = quantile(complement(fisher_f(N1-1, N2-1),
alpha))
In our example program we need both upper and lower critical values for alpha and for alpha/2:
double ucv = quantile(complement(dist, alpha)); double ucv2 = quantile(complement(dist, alpha / 2)); double lcv = quantile(dist, alpha); double lcv2 = quantile(dist, alpha / 2); cout << setw(55) << left << "Upper Critical Value at alpha: " << "= " << setprecision(3) << scientific << ucv << "\n"; cout << setw(55) << left << "Upper Critical Value at alpha/2: " << "= " << setprecision(3) << scientific << ucv2 << "\n"; cout << setw(55) << left << "Lower Critical Value at alpha: " << "= " << setprecision(3) << scientific << lcv << "\n"; cout << setw(55) << left << "Lower Critical Value at alpha/2: " << "= " << setprecision(3) << scientific << lcv2 << "\n\n";
The final step is to perform the comparisons given above, and print out whether the hypothesis is rejected or not:
cout << setw(55) << left << "Results for Alternative Hypothesis and alpha" << "= " << setprecision(4) << fixed << alpha << "\n\n"; cout << "Alternative Hypothesis Conclusion\n"; cout << "Standard deviations are unequal (two sided test) "; if((ucv2 < F) || (lcv2 > F)) cout << "ACCEPTED\n"; else cout << "REJECTED\n"; cout << "Standard deviation 1 is less than standard deviation 2 "; if(lcv > F) cout << "ACCEPTED\n"; else cout << "REJECTED\n"; cout << "Standard deviation 1 is greater than standard deviation 2 "; if(ucv < F) cout << "ACCEPTED\n"; else cout << "REJECTED\n"; cout << endl << endl;
Using the ceramic strength data as an example we get the following output:
F test for equal standard deviations ____________________________________ Sample 1: Number of Observations = 240 Sample Standard Deviation = 65.549 Sample 2: Number of Observations = 240 Sample Standard Deviation = 61.854 Test Statistic = 1.123 CDF of test statistic: = 8.148e-001 Upper Critical Value at alpha: = 1.238e+000 Upper Critical Value at alpha/2: = 1.289e+000 Lower Critical Value at alpha: = 8.080e-001 Lower Critical Value at alpha/2: = 7.756e-001 Results for Alternative Hypothesis and alpha = 0.0500 Alternative Hypothesis Conclusion Standard deviations are unequal (two sided test) REJECTED Standard deviation 1 is less than standard deviation 2 REJECTED Standard deviation 1 is greater than standard deviation 2 REJECTED
In this case we are unable to reject the null-hypothesis, and must instead reject the alternative hypothesis.
By contrast let's see what happens when we use some different sample data:, once again from the NIST Engineering Statistics Handbook: A new procedure to assemble a device is introduced and tested for possible improvement in time of assembly. The question being addressed is whether the standard deviation of the new assembly process (sample 2) is better (i.e., smaller) than the standard deviation for the old assembly process (sample 1).
____________________________________ F test for equal standard deviations ____________________________________ Sample 1: Number of Observations = 11.00000 Sample Standard Deviation = 4.90820 Sample 2: Number of Observations = 9.00000 Sample Standard Deviation = 2.58740 Test Statistic = 3.59847 CDF of test statistic: = 9.589e-001 Upper Critical Value at alpha: = 3.347e+000 Upper Critical Value at alpha/2: = 4.295e+000 Lower Critical Value at alpha: = 3.256e-001 Lower Critical Value at alpha/2: = 2.594e-001 Results for Alternative Hypothesis and alpha = 0.0500 Alternative Hypothesis Conclusion Standard deviations are unequal (two sided test) REJECTED Standard deviation 1 is less than standard deviation 2 REJECTED Standard deviation 1 is greater than standard deviation 2 ACCEPTED
In this case we take our null hypothesis as "standard deviation 1 is less than or equal to standard deviation 2", since this represents the "no change" situation. So we want to compare the upper critical value at alpha (a one sided test) with the test statistic, and since 3.35 < 3.6 this hypothesis must be rejected. We therefore conclude that there is a change for the better in our standard deviation.