1 | // $Id: Fisher.cc 616 2006-08-31 08:52:02Z jari $ |
---|
2 | |
---|
3 | #include <c++_tools/statistics/Fisher.h> |
---|
4 | #include <c++_tools/statistics/Score.h> |
---|
5 | #include <c++_tools/statistics/utility.h> |
---|
6 | #include <c++_tools/classifier/Target.h> |
---|
7 | |
---|
8 | #include <gsl/gsl_cdf.h> |
---|
9 | #include <gsl/gsl_randist.h> |
---|
10 | |
---|
11 | namespace theplu { |
---|
12 | namespace statistics { |
---|
13 | |
---|
14 | |
---|
15 | Fisher::Fisher(bool b) |
---|
16 | : Score(b), a_(0), b_(0), c_(0), d_(0), oddsratio_(1.0), value_cutoff_(0) |
---|
17 | { |
---|
18 | } |
---|
19 | |
---|
20 | |
---|
21 | double Fisher::Chi2() const |
---|
22 | { |
---|
23 | double a,b,c,d; |
---|
24 | expected(a,b,c,d); |
---|
25 | return (a-a_)*(a-a_)/a + (b-b_)*(b-b_)/b + |
---|
26 | (c-c_)*(c-c_)/c + (d-d_)*(d-d_)/d; |
---|
27 | } |
---|
28 | |
---|
29 | |
---|
30 | void Fisher::expected(double& a, double& b, double& c, double& d) const |
---|
31 | { |
---|
32 | double N = a_+b_+c_+d_; |
---|
33 | a =((a_+b_)*(a_+c_)) / N; |
---|
34 | b =((a_+b_)*(b_+d_)) / N; |
---|
35 | c =((c_+d_)*(a_+c_)) / N; |
---|
36 | d =((c_+d_)*(b_+d_)) / N; |
---|
37 | } |
---|
38 | |
---|
39 | double Fisher::oddsratio(const double a, |
---|
40 | const double b, |
---|
41 | const double c, |
---|
42 | const double d) |
---|
43 | { |
---|
44 | // If a column sum or a row sum is zero, the table is nonsense |
---|
45 | if ((a==0 || d==0) && (c==0 || b==0)){ |
---|
46 | //Peter, should throw exception |
---|
47 | std::cerr << "Warning: Fisher: Table is not valid\n"; |
---|
48 | return oddsratio_ = 1.0; |
---|
49 | } |
---|
50 | |
---|
51 | oddsratio_=(a*d)/(b*d); |
---|
52 | if (absolute_ && oddsratio_<1) |
---|
53 | return 1/oddsratio_; |
---|
54 | return oddsratio_; |
---|
55 | } |
---|
56 | |
---|
57 | |
---|
58 | double Fisher::p_value() const |
---|
59 | { |
---|
60 | double p=1; |
---|
61 | if ( ( a_<minimum_size_ || b_<minimum_size_ || |
---|
62 | c_<minimum_size_ || d_<minimum_size_) && !weighted_ ){ |
---|
63 | |
---|
64 | p=p_value_exact(); |
---|
65 | } |
---|
66 | else |
---|
67 | p=p_value_approximative(); |
---|
68 | |
---|
69 | if (!absolute_){ |
---|
70 | p=p/2; |
---|
71 | if (oddsratio_<0.5){ |
---|
72 | // last term because >= not equal to !(<=) |
---|
73 | u_int a = static_cast<u_int>(a_); |
---|
74 | u_int b = static_cast<u_int>(b_); |
---|
75 | u_int c = static_cast<u_int>(c_); |
---|
76 | u_int d = static_cast<u_int>(d_); |
---|
77 | return 1-p+gsl_ran_hypergeometric_pdf(a, a+b, c+d, a+c); |
---|
78 | } |
---|
79 | } |
---|
80 | return p; |
---|
81 | } |
---|
82 | |
---|
83 | |
---|
84 | double Fisher::p_value_approximative() const |
---|
85 | { |
---|
86 | return gsl_cdf_chisq_Q(Chi2(), 1.0); |
---|
87 | } |
---|
88 | |
---|
89 | double Fisher::p_value_exact() const |
---|
90 | { |
---|
91 | u_int a = static_cast<u_int>(a_); |
---|
92 | u_int b = static_cast<u_int>(b_); |
---|
93 | u_int c = static_cast<u_int>(c_); |
---|
94 | u_int d = static_cast<u_int>(d_); |
---|
95 | |
---|
96 | // Since the calculation is symmetric and cdf_hypergeometric_P |
---|
97 | // loops up to k we choose the smallest number to be k and mirror |
---|
98 | // the matrix. This choice makes the p-value two-sided. |
---|
99 | |
---|
100 | if (a<b && a<c && a<d) |
---|
101 | return statistics::cdf_hypergeometric_P(a,a+b,c+d,a+c); |
---|
102 | else if (b<a && b<c && b<d) |
---|
103 | return statistics::cdf_hypergeometric_P(b,a+b,c+d,b+d); |
---|
104 | else if (c<a && c<b && c<d) |
---|
105 | return statistics::cdf_hypergeometric_P(c,c+d,a+b,a+c); |
---|
106 | |
---|
107 | return statistics::cdf_hypergeometric_P(d,c+d,a+b,b+d); |
---|
108 | |
---|
109 | } |
---|
110 | |
---|
111 | |
---|
112 | double Fisher::score(const classifier::Target& target, |
---|
113 | const utility::vector& value) |
---|
114 | { |
---|
115 | weighted_=false; |
---|
116 | a_=b_=c_=d_=0; |
---|
117 | for (size_t i=0; i<target.size(); i++) |
---|
118 | if (target.binary(i)) |
---|
119 | if (value(i)>value_cutoff_) |
---|
120 | a_++; |
---|
121 | else |
---|
122 | c_++; |
---|
123 | else |
---|
124 | if (value(i)>value_cutoff_) |
---|
125 | b_++; |
---|
126 | else |
---|
127 | d_++; |
---|
128 | |
---|
129 | // If a column sum or a row sum is zero, the table is non-sense |
---|
130 | if ((a_==0 || d_==0) && (c_==0 || b_==0)){ |
---|
131 | std::cerr << "Warning: Fisher: Table is not valid\n"; |
---|
132 | return 1; |
---|
133 | } |
---|
134 | |
---|
135 | return oddsratio(a_,b_,c_,d_); |
---|
136 | } |
---|
137 | |
---|
138 | double Fisher::score(const classifier::Target& target, |
---|
139 | const utility::vector& value, |
---|
140 | const utility::vector& weight) |
---|
141 | { |
---|
142 | weighted_=true; |
---|
143 | a_=b_=c_=d_=0; |
---|
144 | for (size_t i=0; i<target.size(); i++) |
---|
145 | if (target.binary(i)) |
---|
146 | if (value(i)>value_cutoff_) |
---|
147 | a_+=weight(i); |
---|
148 | else |
---|
149 | c_+=weight(i); |
---|
150 | else |
---|
151 | if (value(i)>value_cutoff_) |
---|
152 | b_+=weight(i); |
---|
153 | else |
---|
154 | d_+=weight(i); |
---|
155 | |
---|
156 | // If a column sum or a row sum is zero, the table is non-sense |
---|
157 | if ((a_==0 || d_==0) && (c_==0 || b_==0)){ |
---|
158 | std::cerr << "Warning: Fisher: Table is not valid\n"; |
---|
159 | return 1; |
---|
160 | } |
---|
161 | |
---|
162 | return oddsratio(a_,b_,c_,d_); |
---|
163 | } |
---|
164 | |
---|
165 | double Fisher::score(const u_int a, const u_int b, |
---|
166 | const u_int c, const u_int d) |
---|
167 | { |
---|
168 | a_=a; |
---|
169 | b_=b; |
---|
170 | c_=c; |
---|
171 | d_=d; |
---|
172 | return oddsratio(a,b,c,d); |
---|
173 | } |
---|
174 | |
---|
175 | }} // of namespace statistics and namespace theplu |
---|