solow_multivariate/
mvstats.rs1use solow_distributions::f_sf;
10
11#[derive(Clone, Copy, Debug, PartialEq)]
15pub struct MvStat {
16 pub value: f64,
18 pub num_df: f64,
20 pub den_df: f64,
22 pub f_value: f64,
24 pub pr_f: f64,
26}
27
28#[derive(Clone, Copy, Debug, PartialEq)]
30pub struct MultivariateStats {
31 pub wilks_lambda: MvStat,
34 pub pillai_trace: MvStat,
36 pub hotelling_lawley: MvStat,
38 pub roy_greatest_root: MvStat,
40}
41
42pub fn multivariate_stats(
52 eigenvals: &[f64],
53 r_err_sscp: usize,
54 r_contrast: usize,
55 df_resid: f64,
56) -> MultivariateStats {
57 let tolerance = 1e-8;
58 let v = df_resid;
59 let p = r_err_sscp as f64;
60 let q = r_contrast as f64;
61 let s = p.min(q);
62
63 let eigv2: Vec<f64> = eigenvals
65 .iter()
66 .copied()
67 .filter(|&x| x > tolerance)
68 .collect();
69 let eigv1: Vec<f64> = eigv2.iter().map(|&e| e / (1.0 - e)).collect();
70
71 let m = (((p - q).abs()) - 1.0) / 2.0;
72 let n = (v - p - 1.0) / 2.0;
73
74 let wilks_value: f64 = eigv2.iter().map(|&e| 1.0 - e).product();
76 let r = v - (p - q + 1.0) / 2.0;
77 let u = (p * q - 2.0) / 4.0;
78 let mut df1 = p * q;
79 let t = if p * p + q * q - 5.0 > 0.0 {
80 ((p * p * q * q - 4.0) / (p * p + q * q - 5.0)).sqrt()
81 } else {
82 1.0
83 };
84 let df2 = r * t - 2.0 * u;
85 let lmd = wilks_value.powf(1.0 / t);
86 let f = (1.0 - lmd) / lmd * df2 / df1;
87 let wilks = MvStat {
88 value: wilks_value,
89 num_df: df1,
90 den_df: df2,
91 f_value: f,
92 pr_f: f_sf(f, df1, df2),
93 };
94
95 let pillai_value: f64 = eigv2.iter().sum();
97 df1 = s * (2.0 * m + s + 1.0);
98 let df2 = s * (2.0 * n + s + 1.0);
99 let f = df2 / df1 * pillai_value / (s - pillai_value);
100 let pillai = MvStat {
101 value: pillai_value,
102 num_df: df1,
103 den_df: df2,
104 f_value: f,
105 pr_f: f_sf(f, df1, df2),
106 };
107
108 let hl_value: f64 = eigv1.iter().sum();
110 let (df1, df2, f) = if n > 0.0 {
111 let b = (p + 2.0 * n) * (q + 2.0 * n) / 2.0 / (2.0 * n + 1.0) / (n - 1.0);
112 let df1 = p * q;
113 let df2 = 4.0 + (p * q + 2.0) / (b - 1.0);
114 let c = (df2 - 2.0) / 2.0 / n;
115 let f = df2 / df1 * hl_value / c;
116 (df1, df2, f)
117 } else {
118 let df1 = s * (2.0 * m + s + 1.0);
119 let df2 = s * (s * n + 1.0);
120 let f = df2 / df1 / s * hl_value;
121 (df1, df2, f)
122 };
123 let hotelling = MvStat {
124 value: hl_value,
125 num_df: df1,
126 den_df: df2,
127 f_value: f,
128 pr_f: f_sf(f, df1, df2),
129 };
130
131 let roy_value: f64 = eigv1.iter().copied().fold(f64::NEG_INFINITY, f64::max);
133 let rmax = p.max(q);
134 let df1 = rmax;
135 let df2 = v - rmax + q;
136 let f = df2 / df1 * roy_value;
137 let roy = MvStat {
138 value: roy_value,
139 num_df: df1,
140 den_df: df2,
141 f_value: f,
142 pr_f: f_sf(f, df1, df2),
143 };
144
145 MultivariateStats {
146 wilks_lambda: wilks,
147 pillai_trace: pillai,
148 hotelling_lawley: hotelling,
149 roy_greatest_root: roy,
150 }
151}