3614 |
06 Feb 17 |
peter |
1 |
#ifndef theplu_yat_regression_negative_binomial |
3614 |
06 Feb 17 |
peter |
2 |
#define theplu_yat_regression_negative_binomial |
3614 |
06 Feb 17 |
peter |
3 |
|
3614 |
06 Feb 17 |
peter |
// $Id$ |
3614 |
06 Feb 17 |
peter |
5 |
|
3614 |
06 Feb 17 |
peter |
6 |
/* |
4207 |
26 Aug 22 |
peter |
Copyright (C) 2017, 2020, 2022 Peter Johansson |
3614 |
06 Feb 17 |
peter |
8 |
|
3614 |
06 Feb 17 |
peter |
This file is part of the yat library, http://dev.thep.lu.se/yat |
3614 |
06 Feb 17 |
peter |
10 |
|
3614 |
06 Feb 17 |
peter |
The yat library is free software; you can redistribute it and/or |
3614 |
06 Feb 17 |
peter |
modify it under the terms of the GNU General Public License as |
3614 |
06 Feb 17 |
peter |
published by the Free Software Foundation; either version 3 of the |
3614 |
06 Feb 17 |
peter |
License, or (at your option) any later version. |
3614 |
06 Feb 17 |
peter |
15 |
|
3614 |
06 Feb 17 |
peter |
The yat library is distributed in the hope that it will be useful, |
3614 |
06 Feb 17 |
peter |
but WITHOUT ANY WARRANTY; without even the implied warranty of |
3614 |
06 Feb 17 |
peter |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
3614 |
06 Feb 17 |
peter |
General Public License for more details. |
3614 |
06 Feb 17 |
peter |
20 |
|
3614 |
06 Feb 17 |
peter |
You should have received a copy of the GNU General Public License |
3614 |
06 Feb 17 |
peter |
along with yat. If not, see <http://www.gnu.org/licenses/>. |
3614 |
06 Feb 17 |
peter |
23 |
*/ |
3614 |
06 Feb 17 |
peter |
24 |
|
3980 |
24 Aug 20 |
peter |
25 |
#include "Multivariate.h" |
3980 |
24 Aug 20 |
peter |
26 |
|
3614 |
06 Feb 17 |
peter |
27 |
#include <yat/utility/Matrix.h> |
3614 |
06 Feb 17 |
peter |
28 |
#include <yat/utility/Vector.h> |
3614 |
06 Feb 17 |
peter |
29 |
|
3614 |
06 Feb 17 |
peter |
30 |
namespace theplu { |
3614 |
06 Feb 17 |
peter |
31 |
namespace yat { |
3614 |
06 Feb 17 |
peter |
32 |
namespace regression { |
3614 |
06 Feb 17 |
peter |
33 |
|
3614 |
06 Feb 17 |
peter |
34 |
/** |
3664 |
21 Jul 17 |
peter |
Negative Binomial regression models count data from a negative |
3664 |
21 Jul 17 |
peter |
binomial distribution, \f$ y \in NB(r;p) \f$, for which the mean |
3664 |
21 Jul 17 |
peter |
is \f$ \mu = \frac{pr}{1-p} \f$ is modeled as \f$ log(\mu) = \beta_0 |
3664 |
21 Jul 17 |
peter |
+ \beta_1 x_1 + ... + \beta_p x_p \f$ and the variance \f$ V = |
3664 |
21 Jul 17 |
peter |
\frac{pr}{(1-p)^2} = \alpha m(x) \f$ where \f$ \alpha \f$ is the |
3664 |
21 Jul 17 |
peter |
dispersion parameter \f$( \ge 1 )\f$ describing how wider the |
3664 |
21 Jul 17 |
peter |
distribution is compared to Poisson; for \f$ \alpha = 1 \f$ the |
3664 |
21 Jul 17 |
peter |
models equals Poisson regression. |
3614 |
06 Feb 17 |
peter |
43 |
|
3614 |
06 Feb 17 |
peter |
\since new in yat 0.15 |
3614 |
06 Feb 17 |
peter |
45 |
*/ |
3980 |
24 Aug 20 |
peter |
46 |
class NegativeBinomial : public Multivariate |
3614 |
06 Feb 17 |
peter |
47 |
{ |
3614 |
06 Feb 17 |
peter |
48 |
public: |
3614 |
06 Feb 17 |
peter |
49 |
/** |
3663 |
20 Jul 17 |
peter |
\brief alpha parameter |
3663 |
20 Jul 17 |
peter |
51 |
|
4149 |
01 Mar 22 |
peter |
The alpha parameter describes the dispersion of the |
3663 |
20 Jul 17 |
peter |
data. Greater alpha implies greater dispersion and unity alpha |
3664 |
21 Jul 17 |
peter |
means the model is same as Poisson regression. |
3663 |
20 Jul 17 |
peter |
55 |
*/ |
3663 |
20 Jul 17 |
peter |
56 |
double alpha(void) const; |
3663 |
20 Jul 17 |
peter |
57 |
|
3663 |
20 Jul 17 |
peter |
58 |
/** |
3663 |
20 Jul 17 |
peter |
\brief Covariance of parameters |
3664 |
21 Jul 17 |
peter |
60 |
|
3664 |
21 Jul 17 |
peter |
Covariance matrix is inferred as \f$ \alpha \left( X' W X |
3664 |
21 Jul 17 |
peter |
\right)^{-1} \f$ where \f$ W \f$ is a diagonal matrix with |
3664 |
21 Jul 17 |
peter |
element \f$ W_{ii} = \mu_i \f$ |
3663 |
20 Jul 17 |
peter |
64 |
*/ |
3663 |
20 Jul 17 |
peter |
65 |
const utility::Matrix& covariance(void) const; |
3663 |
20 Jul 17 |
peter |
66 |
|
3663 |
20 Jul 17 |
peter |
67 |
/** |
3614 |
06 Feb 17 |
peter |
\brief fit model parameters |
3664 |
21 Jul 17 |
peter |
69 |
|
3664 |
21 Jul 17 |
peter |
The parameters are tuned to minimise deviation between data and model |
3664 |
21 Jul 17 |
peter |
\f$ 2 \sum_i y_i \ln (y_i / \mu_i) \f$ |
3982 |
26 Aug 20 |
peter |
72 |
|
3982 |
26 Aug 20 |
peter |
Each row in \c x represents a sample. |
4130 |
20 Jan 22 |
peter |
74 |
|
4130 |
20 Jan 22 |
peter |
\since New in yat 0.20 |
3614 |
06 Feb 17 |
peter |
76 |
*/ |
4130 |
20 Jan 22 |
peter |
77 |
void fit2(const utility::MatrixBase& x, const utility::VectorBase& y); |
3614 |
06 Feb 17 |
peter |
78 |
|
3614 |
06 Feb 17 |
peter |
79 |
/** |
4130 |
20 Jan 22 |
peter |
Just kept for back compatibility with yat 0.19. Exactly the |
4130 |
20 Jan 22 |
peter |
same behaviour as for fit2. |
4130 |
20 Jan 22 |
peter |
82 |
*/ |
4130 |
20 Jan 22 |
peter |
83 |
void fit(const utility::Matrix& X, const utility::VectorBase& y); |
4130 |
20 Jan 22 |
peter |
84 |
|
4130 |
20 Jan 22 |
peter |
85 |
/** |
3614 |
06 Feb 17 |
peter |
\return model parameters |
3614 |
06 Feb 17 |
peter |
87 |
*/ |
3614 |
06 Feb 17 |
peter |
88 |
const utility::Vector& fit_parameters(void) const; |
3614 |
06 Feb 17 |
peter |
89 |
|
3614 |
06 Feb 17 |
peter |
90 |
/** |
3663 |
20 Jul 17 |
peter |
Predicted value given \a x. The prediction is calculated as |
3614 |
06 Feb 17 |
peter |
\f$ \exp{\beta_0 + \beta_1 x_1 +...+\beta_p x_p} \f$ |
3614 |
06 Feb 17 |
peter |
93 |
*/ |
3614 |
06 Feb 17 |
peter |
94 |
double predict(const utility::VectorBase& x) const; |
3659 |
13 Jul 17 |
peter |
95 |
|
3659 |
13 Jul 17 |
peter |
96 |
private: |
3659 |
13 Jul 17 |
peter |
97 |
double alpha_; |
3659 |
13 Jul 17 |
peter |
98 |
utility::Vector beta_; |
3663 |
20 Jul 17 |
peter |
99 |
utility::Matrix covariance_; |
3614 |
06 Feb 17 |
peter |
100 |
}; |
3614 |
06 Feb 17 |
peter |
101 |
|
3614 |
06 Feb 17 |
peter |
102 |
}}} |
3614 |
06 Feb 17 |
peter |
103 |
|
3614 |
06 Feb 17 |
peter |
104 |
#endif |