A statistical and numerical processing package with a simple and lightweight implementation for the basic stats: variance (sample and population), standard deviation (sample and population), covariance, Pearson (chi squared), range, median, least squares, root mean square, histogram, and mode.
Introduction
While working on another project, I found myself needing to calculate basic statistics on various sets of data of various underlying types. LINQ has Count, Min, Max, and Average, but no other statistical aggregates. As I always do in a case like this, I started with Google, figuring someone else must have written some handy extension methods for this already. There are plenty of statistical and numerical processing packages out there, but what I want is a simple and lightweight implementation for the basic stats: variance (sample and population), standard deviation (sample and population), covariance, Pearson (chi squared), range, median, least squares, root mean square, histogram, and mode.
Background
I've modeled the API on the various overloads of Enumerable.Average, so you are able to use these methods on the same types of collections that those methods accept. Hopefully, this will make the usage familiar and easy to use.
That means overloads for collections of the common numerical data types and their Nullable counter parts, as well as convenient selector overloads.
public static decimal? StandardDeviation(this IEnumerable<decimal?> source);
public static decimal StandardDeviation(this IEnumerable<decimal> source);
public static double? StandardDeviation(this IEnumerable<double?> source);
public static double StandardDeviation(this IEnumerable<double> source);
public static float? StandardDeviation(this IEnumerable<float?> source);
public static float StandardDeviation(this IEnumerable<float> source);
public static double? StandardDeviation(this IEnumerable<int?> source);
public static double StandardDeviation(this IEnumerable<int> source);
public static double? StandardDeviation(this IEnumerable<long?> source);
public static double StandardDeviation(this IEnumerable<long> source);
public static decimal? StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, decimal?> selector);
public static decimal StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, decimal> selector);
public static double? StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, double?> selector);
public static double StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, double> selector);
public static float? StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, float?> selector);
public static float StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, float> selector);
public static double? StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, int?> selector);
public static double StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, int> selector);
public static double? StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, long?> selector);
public static double StandardDeviation<TSource>
(this IEnumerable<TSource> source, Func<TSource, long> selector);
All of the overloads that take a collection of Nullable types only include actual values in the calculated result. For example:
public static double? StandardDeviation(this IEnumerable<double?> source)
{
IEnumerable<double> values = source.AllValues();
if (values.Any())
return values.StandardDeviation();
return null;
}
where the AllValues method is:
public static IEnumerable<T> AllValues<T>(this IEnumerable<T?> source) where T : struct
{
Debug.Assert(source != null);
return source.Where(x => x.HasValue).Select(x => (T)x);
}
A Note About Mode
Since a distribution of values may not have a mode, all of the Mode methods return a Nullable type. For instance, in the series { 1, 2, 3, 4 }, no single value appears more than once. In cases such as this, the return value will be null.
In the case where there are multiple modes, Mode returns the maximum mode (i.e., the value that appears the most times). If there is a tie for the maximum mode, it returns the smallest value in the set of maximum modes.
There are also two methods for calculating all modes in a series. These return an IEnumerable of all of the modes in descending order of modality.
The Statistics Calculations
Links, descriptions, and mathematical images from Wikipedia.
Variance
Variance is the measure of the amount of variation of all the scores for a variable (not just the extremes which give the range).
Sample variance is typically denoted by the lower case sigma squared: σ2.
public static double Variance(this IEnumerable<double> source)
{
int n = 0;
double mean = 0;
double M2 = 0;
foreach (double x in source)
{
n = n + 1;
double delta = x - mean;
mean = mean + delta / n;
M2 += delta * (x - mean);
}
return M2 / (n - 1);
}
Standard Deviation
The Standard Deviation of a statistical population, a data set, or a probability distribution is the square root of its variance.
Standard deviation is typically denoted by the lower case sigma: σ.
public static double StandardDeviation(this IEnumerable<double> source)
{
return Math.Sqrt(source.Variance());
}
Median
Median is the number separating the higher half of a sample, a population, or a probability distribution, from the lower half.
public static double Median(this IEnumerable<double> source)
{
var sortedList = from number in source
orderby number
select number;
int count = sortedList.Count();
int itemIndex = count / 2;
if (count % 2 == 0)
return (sortedList.ElementAt(itemIndex) +
sortedList.ElementAt(itemIndex - 1)) / 2;
return sortedList.ElementAt(itemIndex);
}
Mode
Mode is the value that occurs the most frequently in a data set or a probability distribution.
public static T? Mode<T>(this IEnumerable<T> source) where T : struct
{
var sortedList = from number in source
orderby number
select number;
int count = 0;
int max = 0;
T current = default(T);
T? mode = new T?();
foreach (T next in sortedList)
{
if (current.Equals(next) == false)
{
current = next;
count = 1;
}
else
{
count++;
}
if (count > max)
{
max = count;
mode = current;
}
}
if (max > 1)
return mode;
return null;
}
Histogram
A Histogram is a representation of a continuous distribution of data. Given a continuous data set, the histogram counts how many occurrences of its data points fall into a set of contiguous ranges of values (aka bins). There is no single approach for determining the number of bins as this is dependent on the data and analysis being performed. There are some standard mechanisms for calculating bin size based on the number of data points. Three of these are included in a set of BinCount extension methods. There are also different approaches to determining the range of each bin. These are indicated with the BinningMode enumeration. In all cases except one, bin ranges include the values the >= the range minimum and < the range maximum; [min, max). When the BinningMode is MaxValueInclusive, the maximum bin range will include the max value rather than exclude it: [min, max].
public enum BinningMode
{
Unbounded,
MaxValueInclusive,
ExpandRange
}
Creating the histogram involves creating an array of Bins with the appropriate ranges and then determining how many data points fall into each range.
public static IEnumerable<Bin> Histogram
(this IEnumerable<double> source, int binCount, BinningMode mode = BinningMode.Unbounded)
{
if (source == null)
throw new ArgumentNullException("source");
if (!source.Any())
throw new InvalidOperationException("source sequence contains no elements");
var bins = BinFactory.CreateBins(source.Min(), source.Max(), binCount, mode);
source.AssignBins(bins);
return bins;
}
Range
Range is the length of the smallest interval which contains all the data.
public static double Range(this IEnumerable<double> source)
{
return source.Max() - source.Min();
}
Covariance
Covariance is a measure of how much two variables change together.
public static double Covariance(this IEnumerable<double> source, IEnumerable<double> other)
{
int len = source.Count();
double avgSource = source.Average();
double avgOther = other.Average();
double covariance = 0;
for (int i = 0; i < len; i++)
covariance += (source.ElementAt(i) - avgSource) * (other.ElementAt(i) - avgOther);
return covariance / len;
}
Pearson's Chi Square Test
Pearson's chi square test is used to assess two types of comparisons: tests of goodness of fit, and tests of independence.
In other words, it is a measure of how well a sample distribution matches a predicted distribution or the degree of correlation between two sample distributions. Pearson's is often used in scientific applications to test the validity of hypotheses.
public static double Pearson(this IEnumerable<double> source,
IEnumerable<double> other)
{
return source.Covariance(other) / (source.StandardDeviationP() *
other.StandardDeviationP());
}
Linear Least Squares
Least Squares is an approach for determining the approximate solution for a distribution of data used in regression analysis. Said another way, given a distribution of 2 dimensional data, what is the equation that best predicts y as a function of x in the form y = mx + b, where m is the slope of the line and b is where it intercepts the y axis on a 2d graph.
For this calculation, a struct is returned that indicates m and b.
public static LeastSquares LeastSquares(this IEnumerable<Tuple<double, double>> source)
{
int numPoints = 0;
double sumX = 0;
double sumY = 0;
double sumXX = 0;
double sumXY = 0;
foreach (var tuple in source)
{
numPoints++;
sumX += tuple.Item1;
sumY += tuple.Item2;
sumXX += tuple.Item1 * tuple.Item1;
sumXY += tuple.Item1 * tuple.Item2;
}
if (numPoints < 2)
throw new InvalidOperationException("Source must have at least 2 elements");
double b = (-sumX * sumXY + sumXX * sumY) / (numPoints * sumXX - sumX * sumX);
double m = (-sumX * sumY + numPoints * sumXY) / (numPoints * sumXX - sumX * sumX);
return new LeastSquares(m, b);
}
Root Mean Square
Root Mean Square is the measure of the magnitude of a varying series. This is particularly useful for waveforms.
public static double RootMeanSquare(this IEnumerable<double> source)
{
if (source.Count() < 2)
throw new InvalidOperationException("Source must have at least 2 elements");
double s = source.Aggregate(0.0, (x, d) => x += Math.Pow(d, 2));
return Math.Sqrt(s / source.Count());
}
Using the Code
The included Unit Tests should provide plenty of examples for how to use these methods, but at its simplest, they behave like other enumerable extension methods. The following program...
static void Main(string[] args)
{
IEnumerable<int> data = new int[] { 1, 2, 5, 6, 6, 8, 9, 9, 9 };
Console.WriteLine("Count = {0}", data.Count());
Console.WriteLine("Average = {0}", data.Average());
Console.WriteLine("Median = {0}", data.Median());
Console.WriteLine("Mode = {0}", data.Mode());
Console.WriteLine("Sample Variance = {0}", data.Variance());
Console.WriteLine("Sample Standard Deviation = {0}", data.StandardDeviation());
Console.WriteLine("Population Variance = {0}", data.VarianceP());
Console.WriteLine("Population Standard Deviation = {0}",
data.StandardDeviationP());
Console.WriteLine("Range = {0}", data.Range());
}
... produces:
Count = 9
Average = 6.11111111111111
Median = 6
Mode = 9
Sample Variance = 9.11111111111111
Sample Standard Deviation = 3.01846171271247
Population Variance = 8.09876543209877
Population Standard Deviation = 2.8458329944146
Range = 8
Points of Interest
I didn't spend much time optimizing the calculations, so be careful if you are evaluating extremely large data sets. If you come up with an optimization in any of the attached code, drop me a note and I'll update the source.
Hopefully, you'll find this code handy the next time you need some simple statistics calculation.
A Note about the T4 Templates
I've never found much use for code generation templates but in developing this library, they greatly simplified one use case: namely arithmetic operations cannot be directly expressed in C# generics. Because operators are implemented as static methods, and there is no mechanism to require a Type to have a particular static method, the compiler has no way of generically resolving "-" in this chunk of code:
public static T Range<T>(this IEnumerable<T> source)
{
return source.Max() - source.Min();
}
If you look at the pattern set by Average and Sum in the framework classes (which I've tried to emulate here), they operate on enumerations of int, long, float, double and decimal. In order to avoid much "copy, paste, modify operand types in code and comments" T4 templates came in very handy.
Basically for each operation that can operate on a set of intrinsic types, the template:
- declares a list of the types supported
- iterates over the list and generates the code commenting, method signature and body for all of the overloads supporting the given intrinsic type
public static partial class EnumerableStats
{
<# var types = new List<string>()
{
"int", "long", "float", "double", "decimal"
};
foreach(var type in types)
{#>
public static <#= type #>? Range(this IEnumerable<<#= type #>?> source)
{
IEnumerable<<#= type #>> values = source.AllValues();
if (values.Any())
return values.Range();
return null;
}
public static <#= type #> Range(this IEnumerable<<#= type #>> source)
{
return source.Max() - source.Min();
}
...
etc etc
...
<# } #>
}
This is nice as it ensures that all types support the same set of overloads, and have identical implementations and code commenting.
History
- 19th September, 2009: Version 1.0 - Initial upload
- 26th October, 2009: Version 1.1 - Added Covariance and Pearson as well as a couple of fixes/optimizations
- 3rd December, 2013: Version 1.2 - Updated variance implementation and added GitHub and NuGet links
- 30th August, 2014: Version 1.3 - Added description of Least Squares and Histogram
