Class: Statsample::Test::UMannWhitney

Inherits:
Object
  • Object
show all
Defined in:
lib/statsample/test/umannwhitney.rb

Overview

U Mann-Whitney test

Non-parametric test for assessing whether two independent samples of observations come from the same distribution.

Assumptions

  • The two samples under investigation in the test are independent of each other and the observations within each sample are independent.

  • The observations are comparable (i.e., for any two observations, one can assess whether they are equal or, if not, which one is greater).

  • The variances in the two groups are approximately equal.

Higher differences of distributions correspond to to lower values of U.

Constant Summary collapse

MAX_MN_EXACT =

Max for m*n allowed for exact calculation of probability

10000

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(v1, v2) ⇒ UMannWhitney

Create a new U Mann-Whitney test Params: Two Statsample::Vectors



114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
 
# File 'lib/statsample/test/umannwhitney.rb', line 114

def initialize(v1,v2)
  @n1=v1.valid_data.size
  @n2=v2.valid_data.size

  data=(v1.valid_data+v2.valid_data).to_scale
  groups=(([0]*@n1)+([1]*@n2)).to_vector
  ds={'g'=>groups, 'data'=>data}.to_dataset
  @t=nil
  @ties=data.data.size!=data.data.uniq.size        
  if(@ties)
    adjust_for_ties(ds['data'])
  end
  ds['ranked']=ds['data'].ranked(:scale)
  
  @n=ds.cases
    
  @r1=ds.filter{|r| r['g']==0}['ranked'].sum
  @r2=((ds.cases*(ds.cases+1)).quo(2))-r1
  @u1=r1-((@n1*(@n1+1)).quo(2))
  @u2=r2-((@n2*(@n2+1)).quo(2))
  @u=(u1<u2) ? u1 : u2
end

Instance Attribute Details

#r1 ⇒ Object (readonly)

Sample 1 Rank sum



99
100
101
 
# File 'lib/statsample/test/umannwhitney.rb', line 99

def r1
  @r1
end

#r2 ⇒ Object (readonly)

Sample 2 Rank sum



101
102
103
 
# File 'lib/statsample/test/umannwhitney.rb', line 101

def r2
  @r2
end

#t ⇒ Object (readonly)

Value of compensation for ties (useful for demostration)



109
110
111
 
# File 'lib/statsample/test/umannwhitney.rb', line 109

def t
  @t
end

#u ⇒ Object (readonly)

U Value



107
108
109
 
# File 'lib/statsample/test/umannwhitney.rb', line 107

def u
  @u
end

#u1 ⇒ Object (readonly)

Sample 1 U (useful for demostration)



103
104
105
 
# File 'lib/statsample/test/umannwhitney.rb', line 103

def u1
  @u1
end

#u2 ⇒ Object (readonly)

Sample 2 U (useful for demostration)



105
106
107
 
# File 'lib/statsample/test/umannwhitney.rb', line 105

def u2
  @u2
end

Class Method Details

.distribution_permutations(n1, n2) ⇒ Object

Generate distribution for permutations. Very expensive, but useful for demostrations



77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
 
# File 'lib/statsample/test/umannwhitney.rb', line 77

def self.distribution_permutations(n1,n2)
  base=[0]*n1+[1]*n2
  po=Statsample::Permutation.new(base)
  upper=0
  total=n1*n2
  req={}
  po.each do |perm|
    r0,s0=0,0
    perm.each_index {|c_i|
      if perm[c_i]==0
        r0+=c_i+1
        s0+=1
      end
    }
    u1=r0-((s0*(s0+1)).quo(2))
    u2=total-u1
    temp_u= (u1 <= u2) ? u1 : u2
    req[perm]=temp_u
  end
  req
end

.u_sampling_distribution_as62(n1, n2) ⇒ Object

U sampling distribution, based on Dinneen & Blakesley (1973) algorithm. This is the algorithm used on SPSS Parameters:

  • n1: group 1 size

  • n2: group 2 size

Reference:

  • Dinneen, L., & Blakesley, B. (1973). Algorithm AS 62: A Generator for the Sampling Distribution of the Mann- Whitney U Statistic. Journal of the Royal Statistical Society, 22(2), 269-273



30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
 
# File 'lib/statsample/test/umannwhitney.rb', line 30

def self.u_sampling_distribution_as62(n1,n2)

  freq=[]
  work=[]
  mn1=n1*n2+1
  max_u=n1*n2
  minmn=n1<n2 ? n1 : n2
  maxmn=n1>n2 ? n1 : n2
  n1=maxmn+1
  (1..n1).each{|i| freq[i]=1}
  n1+=1
  (n1..mn1).each{|i| freq[i]=0}
  work[1]=0
  xin=maxmn
  (2..minmn).each do |i|
    work[i]=0
    xin=xin+maxmn
    n1=xin+2
    l=1+xin.quo(2)
    k=i
    (1..l).each do |j|
      k=k+1
      n1=n1-1
      sum=freq[j]+work[j]
      freq[j]=sum
      work[k]=sum-freq[n1]
      freq[n1]=sum
    end
  end
  
  # Generate percentages for normal U
  dist=(1+max_u/2).to_i
  freq.shift
  total=freq.inject(0) {|a,v| a+v }
  (0...dist).collect {|i|
    if i!=max_u-i
      ues=freq[i]*2
    else
      ues=freq[i]
    end
    ues.quo(total)
  }
end

Instance Method Details

#exact_probability ⇒ Object

Exact probability of finding values of U lower or equal to sample on U distribution. Use with caution with m*n>100000. Uses u_sampling_distribution_as62



155
156
157
158
159
160
161
162
 
# File 'lib/statsample/test/umannwhitney.rb', line 155

def exact_probability
  dist=UMannWhitney.u_sampling_distribution_as62(@n1,@n2)
  sum=0
  (0..@u.to_i).each {|i|
    sum+=dist[i]
  }
  sum
end

#report_building(generator) ⇒ Object

:nodoc:



150
151
152
 
# File 'lib/statsample/test/umannwhitney.rb', line 150

def report_building(generator) # :nodoc:
  generator.text(summary)
end

#summary ⇒ Object

Report results.



137
138
139
140
141
142
143
144
145
146
147
148
149
 
# File 'lib/statsample/test/umannwhitney.rb', line 137

def summary
  out="Mann-Whitney U\nSum of ranks v1: \#{@r1.to_f}\nSum of ranks v1: \#{@r2.to_f}\nU Value: \#{@u.to_f}\nZ: \#{sprintf(\"%0.3f\",z)} (p: \#{sprintf(\"%0.3f\",z_probability)})\n  HEREDOC\n  if @n1*@n2<MAX_MN_EXACT\n    out+=\"Exact p (Dinneen & Blakesley): \#{sprintf(\"%0.3f\",exact_probability)}\"\n  end\n    out\nend\n"

#z ⇒ Object

Z value for U, with adjust for ties. For large samples, U is approximately normally distributed. In that case, you can use z to obtain probabily for U. Reference:

  • SPSS Manual



180
181
182
183
184
185
186
187
188
189
190
191
 
# File 'lib/statsample/test/umannwhitney.rb', line 180

def z
  mu=(@n1*@n2).quo(2)
  if(!@ties)
    ou=Math::sqrt(((@n1*@n2)*(@n1+@n2+1)).quo(12))
  else
    n=@n1+@n2
    first=(@n1*@n2).quo(n*(n-1))
    second=((n**3-n).quo(12))-@t
    ou=Math::sqrt(first*second)
  end
  (@u-mu).quo(ou)
end

#z_probability ⇒ Object

Assuming H_0, the proportion of cdf with values of U lower than the sample. Use with more than 30 cases per group.



195
196
197
 
# File 'lib/statsample/test/umannwhitney.rb', line 195

def z_probability
  (1-Distribution::Normal.cdf(z.abs()))*2
end