package jsat.testing.goodnessoffit;
import jsat.distributions.ContinuousDistribution;
import jsat.distributions.Kolmogorov;
import jsat.linear.Vec;
/**
*
* @author Edward Raff
*/
public class KSTest
{
private static final Kolmogorov k = new Kolmogorov();
private Vec v;
/**
* Creates a new statistical test for testing. The 1 sample test, with v
* being the 1 sample. The 1 sample test compare the data to a given distribution,
* and see if it does not belong to the given distribution. The 2 sample test is
* designed to tell if the data is not from the same population.
*
* @param v the date to be one of the samples
*/
public KSTest(Vec v)
{
this.v = v.sortedCopy();
}
/**
* Change the original sample to v
* @param v the new original sample.
*/
public void setBaseData(Vec v)
{
this.v = v;
}
/**
* Calculates the D statistic for comparison against a continous distribution
* @param cd the distribution to compare against
* @return the max difference between the empirical CDF and the 'true' CDF of the given distribution
*/
protected double dCalc(ContinuousDistribution cd)
{
double max = 0;
for(int i = 0; i < v.length(); i++)
{
//ECDF(x) - F(x)
if(v.get(i) >= cd.min() && v.get(i) <= cd.max() )
{
double tmp = (i+1.0)/v.length() - cd.cdf(v.get(i));
max = Math.max(max, Math.abs(tmp));
}
else//The data dose not fit in the rang eof the distribution
{
max = Math.max(max, Math.abs((i+1.0)/v.length()));
}
}
return max;
}
private static double ECDF(Vec s, double x)
{
int min = 0;
int max = s.length()-1;
int mid = (min+max) /2;
do
{
if(x > s.get(mid))
min = mid+1;
else
max = mid-1;
}
while(s.get(mid) != x && min <= max);
return (mid+1.0)/s.length();
}
/**
* Calculates the D statistic for comparison against another data set
* @param o the other data set
* @return the max difrence in empirical CDF of the distributions.
*/
protected double dCaldO(Vec o)
{
double max = 0;
for(int i = 0; i < v.length(); i++)
{
//ECDF(x) - F(x)
double tmp = (i+1.0)/v.length() - ECDF(o, v.get(i));
max = Math.max(max, Math.abs(tmp));
}
for(int i = 0; i < o.length(); i++)
{
//ECDF(x) - F(x)
double tmp = (i+1.0)/o.length() - ECDF(v, o.get(i));
max = Math.max(max, Math.abs(tmp));
}
return max;
}
/**
* Returns the p-value for the KS Test against the given distribution cd.
* The null hypothesis of this test is that the given data set belongs to the given distribution.
* The alternative hypothesis is that the data set does not belong to the given distribution.
*
*
* @param cd the distribution to compare against
* @return the p-value of the test against this distribution
*/
public double testDist(ContinuousDistribution cd)
{
double d = dCalc(cd);
double n = v.length();
return pValue(n, d);
}
/**
* Returns the p-value for the 2 sample KS Test against the given data set data.
* The null hypothesis of this test is that the given data set is from the same population as data
* The alternative hypothesis is that the data set does not belong to the same population as data
* @param data the other distribution to compare against
* @return the p-value of the test against this data set
*/
public double testData(Vec data)
{
double d = dCaldO(data);
double n = v.length()*data.length() / ((double) v.length() +data.length());
return pValue(n, d);
}
private double pValue(double n, double d)
{
return 1 - k.cdf( (Math.sqrt(n) + 0.12 + 0.11/Math.sqrt(n)) * d);
}
}
Ads help maintain this website.