/* PSPP - a program for statistical analysis.
- Copyright (C) 2005 Free Software Foundation, Inc.
+ Copyright (C) 2005, 2010, 2011, 2017 Free Software Foundation, Inc.
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
along with this program. If not, see <http://www.gnu.org/licenses/>. */
#include <config.h>
+
+#include "math/linreg.h"
+
#include <gsl/gsl_blas.h>
#include <gsl/gsl_cblas.h>
#include <gsl/gsl_errno.h>
#include <gsl/gsl_fit.h>
+#include <gsl/gsl_linalg.h>
#include <gsl/gsl_multifit.h>
-#include <linreg/sweep.h>
-#include <math/coefficient.h>
-#include <math/linreg.h>
-#include <math/coefficient.h>
-#include <math/covariance-matrix.h>
-#include <math/design-matrix.h>
-#include <src/data/category.h>
-#include <src/data/variable.h>
-#include <src/data/value.h>
-#include <gl/xalloc.h>
+
+#include "data/value.h"
+#include "data/variable.h"
+#include "linreg/sweep.h"
+
+#include "gl/xalloc.h"
/*
Find the least-squares estimate of b for the linear model:
Springer. 1998. ISBN 0-387-98542-5.
*/
-
-/*
- Get the mean and standard deviation of a vector
- of doubles via a form of the Kalman filter as
- described on page 32 of [3].
- */
-static int
-linreg_mean_std (gsl_vector_const_view v, double *mp, double *sp, double *ssp)
+struct linreg
{
- size_t i;
- double j = 0.0;
- double d;
- double tmp;
- double mean;
- double variance;
-
- mean = gsl_vector_get (&v.vector, 0);
- variance = 0;
- for (i = 1; i < v.vector.size; i++)
- {
- j = (double) i + 1.0;
- tmp = gsl_vector_get (&v.vector, i);
- d = (tmp - mean) / j;
- mean += d;
- variance += j * (j - 1.0) * d * d;
- }
- *mp = mean;
- *sp = sqrt (variance / (j - 1.0));
- *ssp = variance;
+ double n_obs; /* Number of observations. */
+ int n_indeps; /* Number of independent variables. */
+ int n_coeffs; /* The intercept is not considered a
+ coefficient here. */
- return GSL_SUCCESS;
-}
+ /*
+ Pointers to the variables.
+ */
+ const struct variable *depvar;
+ const struct variable **indep_vars;
-/*
- Set V to contain an array of pointers to the variables
- used in the model. V must be at least C->N_COEFFS in length.
- The return value is the number of distinct variables found.
- */
-int
-pspp_linreg_get_vars (const void *c_, const struct variable **v)
-{
- const pspp_linreg_cache *c = c_;
- const struct variable *tmp;
- int i;
- int j;
- int result = 0;
+ double *coeff;
+ double intercept;
+ /*
+ Means and standard deviations of the variables.
+ If these pointers are null when pspp_linreg() is
+ called, pspp_linreg() will compute their values.
+
+ Entry i of indep_means is the mean of independent
+ variable i, whose observations are stored in the ith
+ column of the design matrix.
+ */
+ double depvar_mean;
+ gsl_vector *indep_means;
+ gsl_vector *indep_std;
/*
- Make sure the caller doesn't try to sneak a variable
- into V that is not in the model.
+ Sums of squares.
*/
- for (i = 0; i < c->n_coeffs; i++)
- {
- v[i] = NULL;
- }
- for (j = 0; j < c->n_coeffs; j++)
- {
- tmp = pspp_coeff_get_var (c->coeff[j], 0);
- assert (tmp != NULL);
- /* Repeated variables are likely to bunch together, at the end
- of the array. */
- i = result - 1;
- while (i >= 0 && v[i] != tmp)
- {
- i--;
- }
- if (i < 0 && result < c->n_coeffs)
- {
- v[result] = tmp;
- result++;
- }
- }
- return result;
+ double ssm; /* Sums of squares for the overall model. */
+ double sst; /* Sum of squares total. */
+ double sse; /* Sum of squares error. */
+ double mse; /* Mean squared error. This is just sse /
+ dfe, but since it is the best unbiased
+ estimate of the population variance, it
+ has its own entry here. */
+ /*
+ Covariance matrix of the parameter estimates.
+ */
+ gsl_matrix *cov;
+ /*
+ Degrees of freedom.
+ */
+ double dft;
+ double dfe;
+ double dfm;
+
+ int dependent_column; /* Column containing the dependent variable. Defaults to last column. */
+ int refcnt;
+
+ bool origin;
+};
+
+const struct variable **
+linreg_get_vars (const struct linreg *c)
+{
+ return c->indep_vars;
}
/*
- Allocate a pspp_linreg_cache and return a pointer
- to it. n is the number of cases, p is the number of
- independent variables.
+ Allocate a linreg and return a pointer to it. n is the number of
+ cases, p is the number of independent variables.
*/
-pspp_linreg_cache *
-pspp_linreg_cache_alloc (size_t n, size_t p)
+struct linreg *
+linreg_alloc (const struct variable *depvar, const struct variable **indep_vars,
+ double n, size_t p, bool origin)
{
- pspp_linreg_cache *c;
+ struct linreg *c;
+ size_t i;
- c = (pspp_linreg_cache *) malloc (sizeof (pspp_linreg_cache));
- c->depvar = NULL;
+ c = xmalloc (sizeof (*c));
+ c->depvar = depvar;
+ c->indep_vars = xnmalloc (p, sizeof (*indep_vars));
+ c->dependent_column = p;
+ for (i = 0; i < p; i++)
+ {
+ c->indep_vars[i] = indep_vars[i];
+ }
c->indep_means = gsl_vector_alloc (p);
c->indep_std = gsl_vector_alloc (p);
- c->ssx = gsl_vector_alloc (p); /* Sums of squares for the
- independent variables.
- */
- c->ss_indeps = gsl_vector_alloc (p); /* Sums of squares for the
- model parameters.
- */
- c->cov = gsl_matrix_alloc (p + 1, p + 1); /* Covariance matrix. */
+
c->n_obs = n;
c->n_indeps = p;
+ c->n_coeffs = p;
+ c->coeff = xnmalloc (p, sizeof (*c->coeff));
+ c->cov = gsl_matrix_calloc (c->n_coeffs + 1, c->n_coeffs + 1);
+ c->dft = n;
+ if (!origin)
+ c->dft--;
+
+ c->dfm = p;
+ c->dfe = c->dft - c->dfm;
+ c->intercept = 0.0;
+ c->depvar_mean = 0.0;
/*
Default settings.
*/
- c->method = PSPP_LINREG_SWEEP;
- c->predict = pspp_linreg_predict;
- c->residual = pspp_linreg_residual; /* The procedure to compute my
- residuals. */
- c->get_vars = pspp_linreg_get_vars; /* The procedure that returns
- pointers to model
- variables. */
- c->resid = NULL; /* The variable storing my residuals. */
- c->pred = NULL; /* The variable storing my predicted values. */
+
+ c->refcnt = 1;
+
+ c->origin = origin;
return c;
}
-bool
-pspp_linreg_cache_free (void *m)
+
+void
+linreg_ref (struct linreg *c)
{
- int i;
+ c->refcnt++;
+}
- pspp_linreg_cache *c = m;
- if (c != NULL)
+void
+linreg_unref (struct linreg *c)
+{
+ if (--c->refcnt == 0)
{
gsl_vector_free (c->indep_means);
gsl_vector_free (c->indep_std);
- gsl_vector_free (c->ss_indeps);
gsl_matrix_free (c->cov);
- gsl_vector_free (c->ssx);
- for (i = 0; i < c->n_coeffs; i++)
- {
- pspp_coeff_free (c->coeff[i]);
- }
+ free (c->indep_vars);
free (c->coeff);
free (c);
}
- return true;
-}
-static void
-cache_init (pspp_linreg_cache *cache, const struct design_matrix *dm)
-{
- assert (cache != NULL);
- cache->dft = cache->n_obs - 1;
- cache->dfm = cache->n_indeps;
- cache->dfe = cache->dft - cache->dfm;
- cache->n_coeffs = dm->m->size2;
- cache->intercept = 0.0;
}
static void
-post_sweep_computations (pspp_linreg_cache *cache, const struct design_matrix *dm,
- gsl_matrix *sw)
+post_sweep_computations (struct linreg *l, gsl_matrix *sw)
{
- gsl_matrix *xm;
- gsl_matrix_view xtx;
- gsl_matrix_view xmxtx;
double m;
- double tmp;
size_t i;
size_t j;
int rc;
-
+
assert (sw != NULL);
- assert (cache != NULL);
+ assert (l != NULL);
- cache->sse = gsl_matrix_get (sw, cache->n_indeps, cache->n_indeps);
- cache->mse = cache->sse / cache->dfe;
+ l->sse = gsl_matrix_get (sw, l->n_indeps, l->n_indeps);
+ l->mse = l->sse / l->dfe;
/*
Get the intercept.
*/
- m = cache->depvar_mean;
- for (i = 0; i < cache->n_indeps; i++)
+ m = l->depvar_mean;
+ for (i = 0; i < l->n_indeps; i++)
{
- tmp = gsl_matrix_get (sw, i, cache->n_indeps);
- cache->coeff[i]->estimate = tmp;
- m -= tmp * pspp_linreg_get_indep_variable_mean (cache, design_matrix_col_to_var (dm, i));
+ double tmp = gsl_matrix_get (sw, i, l->n_indeps);
+ l->coeff[i] = tmp;
+ m -= tmp * linreg_get_indep_variable_mean (l, i);
}
/*
Get the covariance matrix of the parameter estimates.
Only the upper triangle is necessary.
*/
-
+
/*
The loops below do not compute the entries related
to the estimated intercept.
*/
- for (i = 0; i < cache->n_indeps; i++)
- for (j = i; j < cache->n_indeps; j++)
+ for (i = 0; i < l->n_indeps; i++)
+ for (j = i; j < l->n_indeps; j++)
{
- tmp = -1.0 * cache->mse * gsl_matrix_get (sw, i, j);
- gsl_matrix_set (cache->cov, i + 1, j + 1, tmp);
+ double tmp = -1.0 * l->mse * gsl_matrix_get (sw, i, j);
+ gsl_matrix_set (l->cov, i + 1, j + 1, tmp);
}
- /*
- Get the covariances related to the intercept.
- */
- xtx = gsl_matrix_submatrix (sw, 0, 0, cache->n_indeps, cache->n_indeps);
- xmxtx = gsl_matrix_submatrix (cache->cov, 0, 1, 1, cache->n_indeps);
- xm = gsl_matrix_calloc (1, cache->n_indeps);
- for (i = 0; i < xm->size2; i++)
- {
- gsl_matrix_set (xm, 0, i,
- pspp_linreg_get_indep_variable_mean (cache, design_matrix_col_to_var (dm, i)));
- }
- rc = gsl_blas_dsymm (CblasRight, CblasUpper, cache->mse,
- &xtx.matrix, xm, 0.0, &xmxtx.matrix);
- gsl_matrix_free (xm);
- if (rc == GSL_SUCCESS)
- {
- tmp = cache->mse / cache->n_obs;
- for (i = 1; i < 1 + cache->n_indeps; i++)
- {
- tmp -= gsl_matrix_get (cache->cov, 0, i)
- * pspp_linreg_get_indep_variable_mean (cache, design_matrix_col_to_var (dm, i - 1));
- }
- gsl_matrix_set (cache->cov, 0, 0, tmp);
-
- cache->intercept = m;
- }
- else
- {
- fprintf (stderr, "%s:%d:gsl_blas_dsymm: %s\n",
- __FILE__, __LINE__, gsl_strerror (rc));
- exit (rc);
- }
-}
-
-/*
- Fit the linear model via least squares. All pointers passed to pspp_linreg
- are assumed to be allocated to the correct size and initialized to the
- values as indicated by opts.
- */
-int
-pspp_linreg (const gsl_vector * Y, const struct design_matrix *dm,
- const pspp_linreg_opts * opts, pspp_linreg_cache * cache)
-{
- int rc;
- gsl_matrix *design = NULL;
- gsl_matrix_view xtx;
- gsl_vector_view xty;
- gsl_vector_view xi;
- gsl_vector_view xj;
- gsl_vector *param_estimates;
- struct pspp_coeff *coef;
- const struct variable *v;
- const union value *val;
- size_t i;
- size_t j;
- double tmp;
- double m;
- double s;
- double ss;
-
- if (cache == NULL)
- {
- return GSL_EFAULT;
- }
- if (opts->get_depvar_mean_std)
+ if (! l->origin)
{
- linreg_mean_std (gsl_vector_const_subvector (Y, 0, Y->size),
- &m, &s, &ss);
- cache->depvar_mean = m;
- cache->depvar_std = s;
- cache->sst = ss;
- }
- cache_init (cache, dm);
- for (i = 0; i < dm->m->size2; i++)
- {
- if (opts->get_indep_mean_std[i])
- {
- linreg_mean_std (gsl_matrix_const_column (dm->m, i), &m, &s, &ss);
- v = design_matrix_col_to_var (dm, i);
- val = NULL;
- if (var_is_alpha (v))
- {
- j = i - design_matrix_var_to_column (dm, v);
- val = cat_subscript_to_value (j, v);
- }
- coef = pspp_linreg_get_coeff (cache, v, val);
- pspp_coeff_set_mean (coef, m);
- pspp_coeff_set_sd (coef, s);
- gsl_vector_set (cache->ssx, i, ss);
-
- }
- }
-
- if (cache->method == PSPP_LINREG_SWEEP)
- {
- gsl_matrix *sw;
- /*
- Subtract the means to improve the condition of the design
- matrix. This requires copying dm->m and Y. We do not divide by the
- standard deviations of the independent variables here since doing
- so would cause a miscalculation of the residual sums of
- squares. Dividing by the standard deviation is done GSL's linear
- regression functions, so if the design matrix has a poor
- condition, use QR decomposition.
-
- The design matrix here does not include a column for the intercept
- (i.e., a column of 1's). If using PSPP_LINREG_QR, we need that column,
- so design is allocated here when sweeping, or below if using QR.
- */
- design = gsl_matrix_alloc (dm->m->size1, dm->m->size2);
- for (i = 0; i < dm->m->size2; i++)
- {
- v = design_matrix_col_to_var (dm, i);
- m = pspp_linreg_get_indep_variable_mean (cache, v);
- for (j = 0; j < dm->m->size1; j++)
- {
- tmp = (gsl_matrix_get (dm->m, j, i) - m);
- gsl_matrix_set (design, j, i, tmp);
- }
- }
- sw = gsl_matrix_calloc (cache->n_coeffs + 1, cache->n_coeffs + 1);
- xtx = gsl_matrix_submatrix (sw, 0, 0, cache->n_coeffs, cache->n_coeffs);
-
- for (i = 0; i < xtx.matrix.size1; i++)
- {
- tmp = gsl_vector_get (cache->ssx, i);
- gsl_matrix_set (&(xtx.matrix), i, i, tmp);
- xi = gsl_matrix_column (design, i);
- for (j = (i + 1); j < xtx.matrix.size2; j++)
- {
- xj = gsl_matrix_column (design, j);
- gsl_blas_ddot (&(xi.vector), &(xj.vector), &tmp);
- gsl_matrix_set (&(xtx.matrix), i, j, tmp);
- }
- }
-
- gsl_matrix_set (sw, cache->n_coeffs, cache->n_coeffs, cache->sst);
- xty = gsl_matrix_column (sw, cache->n_coeffs);
+ gsl_matrix *xm;
+ gsl_matrix_view xtx;
+ gsl_matrix_view xmxtx;
/*
- This loop starts at 1, with i=0 outside the loop, so we can get
- the model sum of squares due to the first independent variable.
- */
- xi = gsl_matrix_column (design, 0);
- gsl_blas_ddot (&(xi.vector), Y, &tmp);
- gsl_vector_set (&(xty.vector), 0, tmp);
- tmp *= tmp / gsl_vector_get (cache->ssx, 0);
- gsl_vector_set (cache->ss_indeps, 0, tmp);
- for (i = 1; i < cache->n_coeffs; i++)
+ Get the covariances related to the intercept.
+ */
+ xtx = gsl_matrix_submatrix (sw, 0, 0, l->n_indeps, l->n_indeps);
+ xmxtx = gsl_matrix_submatrix (l->cov, 0, 1, 1, l->n_indeps);
+ xm = gsl_matrix_calloc (1, l->n_indeps);
+ for (i = 0; i < xm->size2; i++)
{
- xi = gsl_matrix_column (design, i);
- gsl_blas_ddot (&(xi.vector), Y, &tmp);
- gsl_vector_set (&(xty.vector), i, tmp);
+ gsl_matrix_set (xm, 0, i,
+ linreg_get_indep_variable_mean (l, i));
}
-
- /*
- Sweep on the matrix sw, which contains XtX, XtY and YtY.
- */
- reg_sweep (sw);
- post_sweep_computations (cache, dm, sw);
- gsl_matrix_free (sw);
- }
- else if (cache->method == PSPP_LINREG_CONDITIONAL_INVERSE)
- {
- /*
- Use the SVD of X^T X to find a conditional inverse of X^TX. If
- the SVD is X^T X = U D V^T, then set the conditional inverse
- to (X^T X)^c = V D^- U^T. D^- is defined as follows: If entry
- (i, i) has value sigma_i, then entry (i, i) of D^- is 1 /
- sigma_i if sigma_i > 0, and 0 otherwise. Then solve the normal
- equations by setting the estimated parameter vector to
- (X^TX)^c X^T Y.
- */
- }
- else
- {
- gsl_multifit_linear_workspace *wk;
- /*
- Use QR decomposition via GSL.
- */
-
- param_estimates = gsl_vector_alloc (1 + dm->m->size2);
- design = gsl_matrix_alloc (dm->m->size1, 1 + dm->m->size2);
-
- for (j = 0; j < dm->m->size1; j++)
+ rc = gsl_blas_dsymm (CblasRight, CblasUpper, l->mse,
+ &xtx.matrix, xm, 0.0, &xmxtx.matrix);
+ gsl_matrix_free (xm);
+ if (rc == GSL_SUCCESS)
{
- gsl_matrix_set (design, j, 0, 1.0);
- for (i = 0; i < dm->m->size2; i++)
+ double tmp = l->mse / l->n_obs;
+ for (i = 1; i < 1 + l->n_indeps; i++)
{
- tmp = gsl_matrix_get (dm->m, j, i);
- gsl_matrix_set (design, j, i + 1, tmp);
+ tmp -= gsl_matrix_get (l->cov, 0, i)
+ * linreg_get_indep_variable_mean (l, i - 1);
}
- }
+ gsl_matrix_set (l->cov, 0, 0, tmp);
- wk = gsl_multifit_linear_alloc (design->size1, design->size2);
- rc = gsl_multifit_linear (design, Y, param_estimates,
- cache->cov, &(cache->sse), wk);
- for (i = 0; i < cache->n_coeffs; i++)
- {
- cache->coeff[i]->estimate = gsl_vector_get (param_estimates, i + 1);
- }
- cache->intercept = gsl_vector_get (param_estimates, 0);
- if (rc == GSL_SUCCESS)
- {
- gsl_multifit_linear_free (wk);
- gsl_vector_free (param_estimates);
+ l->intercept = m;
}
else
{
- fprintf (stderr, "%s:%d: gsl_multifit_linear returned %d\n",
- __FILE__, __LINE__, rc);
+ fprintf (stderr, "%s:%d:gsl_blas_dsymm: %s\n",
+ __FILE__, __LINE__, gsl_strerror (rc));
+ exit (rc);
}
}
-
-
- cache->ssm = cache->sst - cache->sse;
- /*
- Get the remaining sums of squares for the independent
- variables.
- */
- m = 0;
- for (i = 1; i < cache->n_indeps; i++)
- {
- j = i - 1;
- m += gsl_vector_get (cache->ss_indeps, j);
- tmp = cache->ssm - m;
- gsl_vector_set (cache->ss_indeps, i, tmp);
- }
-
- gsl_matrix_free (design);
- return GSL_SUCCESS;
}
/*
- Is the coefficient COEF contained in the list of coefficients
- COEF_LIST?
- */
-static int
-has_coefficient (const struct pspp_coeff **coef_list, const struct pspp_coeff *coef,
- size_t n)
-{
- size_t i = 0;
-
- while (i < n)
- {
- if (coef_list[i] == coef)
- {
- return 1;
- }
- i++;
- }
- return 0;
-}
-/*
- Predict the value of the dependent variable with the
- new set of predictors. PREDICTORS must point to a list
- of variables, each of whose values are stored in VALS,
- in the same order.
+ Predict the value of the dependent variable with the new set of
+ predictors. VALS are assumed to be in the order corresponding to the
+ order of the coefficients in the linreg struct.
*/
double
-pspp_linreg_predict (const struct variable **predictors,
- const union value **vals, const void *c_, int n_vals)
+linreg_predict (const struct linreg *c, const double *vals, size_t n_vals)
{
- const pspp_linreg_cache *c = c_;
- int j;
- size_t next_coef = 0;
- const struct pspp_coeff **coef_list;
- const struct pspp_coeff *coe;
+ size_t j;
double result;
- double tmp;
- if (predictors == NULL || vals == NULL || c == NULL)
+ if (vals == NULL || c == NULL)
{
return GSL_NAN;
}
+ assert (n_vals == c->n_coeffs);
if (c->coeff == NULL)
{
/* The stupid model: just guess the mean. */
return c->depvar_mean;
}
- coef_list = xnmalloc (c->n_coeffs, sizeof (*coef_list));
result = c->intercept;
- /*
- The loops guard against the possibility that the caller passed us
- inadequate information, such as too few or too many values, or
- a redundant list of variable names.
- */
for (j = 0; j < n_vals; j++)
{
- coe = pspp_linreg_get_coeff (c, predictors[j], vals[j]);
- if (!has_coefficient (coef_list, coe, next_coef))
- {
- tmp = pspp_coeff_get_est (coe);
- if (var_is_numeric (predictors[j]))
- {
- tmp *= vals[j]->f;
- }
- result += tmp;
- coef_list[next_coef++] = coe;
- }
+ result += linreg_coeff (c, j) * vals[j];
}
- free (coef_list);
return result;
}
double
-pspp_linreg_residual (const struct variable **predictors,
- const union value **vals,
- const union value *obs, const void *c, int n_vals)
+linreg_residual (const struct linreg *c, double obs, const double *vals, size_t n_vals)
{
- double pred;
- double result;
-
- if (predictors == NULL || vals == NULL || c == NULL || obs == NULL)
+ if (vals == NULL || c == NULL)
{
return GSL_NAN;
}
- pred = pspp_linreg_predict (predictors, vals, c, n_vals);
-
- result = isnan (pred) ? GSL_NAN : (obs->f - pred);
- return result;
+ return (obs - linreg_predict (c, vals, n_vals));
}
/*
- Which coefficient is associated with V? The VAL argument is relevant
- only to categorical variables.
+ Mean of the independent variable.
*/
-struct pspp_coeff *
-pspp_linreg_get_coeff (const pspp_linreg_cache * c,
- const struct variable *v, const union value *val)
+double
+linreg_get_indep_variable_mean (const struct linreg *c, size_t j)
{
- if (c == NULL)
- {
- return NULL;
- }
- if (c->coeff == NULL || c->n_indeps == 0 || v == NULL)
- {
- return NULL;
- }
- return pspp_coeff_var_to_coeff (v, c->coeff, c->n_coeffs, val);
+ assert (c != NULL);
+ return gsl_vector_get (c->indep_means, j);
}
-/*
- Return the standard deviation of the independent variable.
- */
-double pspp_linreg_get_indep_variable_sd (pspp_linreg_cache *c, const struct variable *v)
+
+void
+linreg_set_indep_variable_mean (struct linreg *c, size_t j, double m)
{
- if (var_is_numeric (v))
- {
- const struct pspp_coeff *coef;
- coef = pspp_linreg_get_coeff (c, v, NULL);
- return pspp_coeff_get_sd (coef);
- }
- return GSL_NAN;
+ assert (c != NULL);
+ gsl_vector_set (c->indep_means, j, m);
}
-void pspp_linreg_set_indep_variable_sd (pspp_linreg_cache *c, const struct variable *v,
- double s)
+#if 0
+static void
+linreg_fit_qr (const gsl_matrix *cov, struct linreg *l)
{
- if (var_is_numeric (v))
+ double intcpt_coef = 0.0;
+ double intercept_variance = 0.0;
+ gsl_matrix *xtx;
+ gsl_matrix *q;
+ gsl_matrix *r;
+ gsl_vector *xty;
+ gsl_vector *tau;
+ gsl_vector *params;
+ size_t i;
+ size_t j;
+
+ xtx = gsl_matrix_alloc (cov->size1 - 1, cov->size2 - 1);
+ xty = gsl_vector_alloc (cov->size1 - 1);
+ tau = gsl_vector_alloc (cov->size1 - 1);
+ params = gsl_vector_alloc (cov->size1 - 1);
+
+ for (i = 0; i < xtx->size1; i++)
{
- struct pspp_coeff *coef;
- coef = pspp_linreg_get_coeff (c, v, NULL);
- pspp_coeff_set_sd (coef, s);
+ gsl_vector_set (xty, i, gsl_matrix_get (cov, cov->size2 - 1, i));
+ for (j = 0; j < xtx->size2; j++)
+ {
+ gsl_matrix_set (xtx, i, j, gsl_matrix_get (cov, i, j));
+ }
}
-}
+ gsl_linalg_QR_decomp (xtx, tau);
+ q = gsl_matrix_alloc (xtx->size1, xtx->size2);
+ r = gsl_matrix_alloc (xtx->size1, xtx->size2);
-/*
- Mean of the independent variable.
- */
-double pspp_linreg_get_indep_variable_mean (pspp_linreg_cache *c, const struct variable *v)
-{
- if (var_is_numeric (v))
+ gsl_linalg_QR_unpack (xtx, tau, q, r);
+ gsl_linalg_QR_solve (xtx, tau, xty, params);
+ for (i = 0; i < params->size; i++)
{
- struct pspp_coeff *coef;
- coef = pspp_linreg_get_coeff (c, v, NULL);
- return pspp_coeff_get_mean (coef);
+ l->coeff[i] = gsl_vector_get (params, i);
}
- return 0.0;
-}
-
-void pspp_linreg_set_indep_variable_mean (pspp_linreg_cache *c, const struct variable *v,
- double m)
-{
- if (var_is_numeric (v))
+ l->sst = gsl_matrix_get (cov, cov->size1 - 1, cov->size2 - 1);
+ l->ssm = 0.0;
+ for (i = 0; i < l->n_indeps; i++)
{
- struct pspp_coeff *coef;
- coef = pspp_linreg_get_coeff (c, v, NULL);
- pspp_coeff_set_mean (coef, m);
+ l->ssm += gsl_vector_get (xty, i) * l->coeff[i];
}
-}
+ l->sse = l->sst - l->ssm;
-/*
- Make sure the dependent variable is at the last column, and that
- only variables in the model are in the covariance matrix.
- */
-static struct design_matrix *
-rearrange_covariance_matrix (const struct design_matrix *cov, pspp_linreg_cache *c)
-{
- struct variable **v;
- struct variable **model_vars;
- struct variable *tmp;
- struct design_matrix *result;
- int n_vars;
- int found;
- size_t *columns;
- size_t i;
- size_t j;
- size_t k;
- size_t dep_col;
-
- assert (cov != NULL);
- assert (c != NULL);
- assert (cov->m->size1 > 0);
- assert (cov->m->size2 == cov->m->size1);
- v = xnmalloc (c->n_coeffs, sizeof (*v));
- model_vars = xnmalloc (c->n_coeffs, sizeof (*model_vars));
- columns = xnmalloc (cov->m->size2, sizeof (*columns));
- n_vars = pspp_linreg_get_vars (c, (const struct variable **) v);
- dep_col = 0;
- k = 0;
- for (i = 0; i < cov->m->size2; i++)
+ gsl_blas_dtrsm (CblasLeft, CblasLower, CblasNoTrans, CblasNonUnit, linreg_mse (l),
+ r, q);
+ /* Copy the lower triangle into the upper triangle. */
+ for (i = 0; i < q->size1; i++)
{
- tmp = design_matrix_col_to_var (cov, i);
- found = 0;
- j = 0;
- while (!found && j < n_vars)
+ gsl_matrix_set (l->cov, i + 1, i + 1, gsl_matrix_get (q, i, i));
+ for (j = i + 1; j < q->size2; j++)
{
- if (tmp == v[j])
- {
- found = 1;
- if (tmp == c->depvar)
- {
- dep_col = j;
- }
- else
- {
- columns[k] = j;
- k++;
- }
- }
- j++;
+ intercept_variance -= 2.0 * gsl_matrix_get (q, i, j) *
+ linreg_get_indep_variable_mean (l, i) *
+ linreg_get_indep_variable_mean (l, j);
+ gsl_matrix_set (q, i, j, gsl_matrix_get (q, j, i));
}
}
- k++;
- columns[k] = dep_col;
- /*
- K should now be equal to C->N_INDEPS + 1. If it is not, then
- either the code above is wrong or the caller didn't send us the
- correct values in C.
- */
- assert (k == c->n_indeps + 1);
- /*
- Put the model variables in the right order in MODEL_VARS.
- */
- for (i = 0; i < k; i++)
- {
- model_vars[i] = v[columns[i]];
- }
- result = covariance_matrix_create (k, model_vars);
- for (i = 0; i < result->m->size1; i++)
+ if (!l->origin)
{
- for (j = 0; j < result->m->size2; j++)
+ l->intercept = linreg_get_depvar_mean (l);
+ for (i = 0; i < l->n_indeps; i++)
+ {
+ double tmp = linreg_get_indep_variable_mean (l, i);
+ l->intercept -= l->coeff[i] * tmp;
+ intercept_variance += tmp * tmp * gsl_matrix_get (q, i, i);
+ }
+
+ /* Covariances related to the intercept. */
+ intercept_variance += linreg_mse (l) / linreg_n_obs (l);
+ gsl_matrix_set (l->cov, 0, 0, intercept_variance);
+ for (i = 0; i < q->size1; i++)
{
- gsl_matrix_set (result->m, i, j, gsl_matrix_get (cov->m, columns[i], columns[j]));
+ for (j = 0; j < q->size2; j++)
+ {
+ intcpt_coef -= gsl_matrix_get (q, i, j)
+ * linreg_get_indep_variable_mean (l, j);
+ }
+ gsl_matrix_set (l->cov, 0, i + 1, intcpt_coef);
+ gsl_matrix_set (l->cov, i + 1, 0, intcpt_coef);
+ intcpt_coef = 0.0;
}
}
- free (columns);
- free (v);
- return result;
+
+ gsl_matrix_free (q);
+ gsl_matrix_free (r);
+ gsl_vector_free (xty);
+ gsl_vector_free (tau);
+ gsl_matrix_free (xtx);
+ gsl_vector_free (params);
}
+#endif
+
+#define REG_LARGE_DATA 1000
+
/*
- Estimate the model parameters from the covariance matrix only. This
- method uses less memory than PSPP_LINREG, which requires the entire
- data set to be stored in memory.
+ Estimate the model parameters from the covariance matrix. This
+ function assumes the covariance entries corresponding to the
+ dependent variable are in the final row and column of the covariance
+ matrix.
*/
-int
-pspp_linreg_with_cov (const struct design_matrix *full_cov,
- pspp_linreg_cache * cache)
+void
+linreg_fit (const gsl_matrix *cov, struct linreg *l)
{
- struct design_matrix *cov;
-
+ assert (l != NULL);
assert (cov != NULL);
- assert (cache != NULL);
- cov = rearrange_covariance_matrix (full_cov, cache);
- cache_init (cache, cov);
- reg_sweep (cov->m);
- post_sweep_computations (cache, cov, cov->m);
- covariance_matrix_destroy (cov);
+ l->sst = gsl_matrix_get (cov, cov->size1 - 1, cov->size2 - 1);
+
+#if 0
+ /* This QR decomposition path seems to produce the incorrect
+ values. See https://savannah.gnu.org/bugs/?51373 */
+ if ((l->n_obs * l->n_obs > l->n_indeps) && (l->n_obs > REG_LARGE_DATA))
+ {
+ /*
+ For large data sets, use QR decomposition.
+ */
+ linreg_fit_qr (cov, l);
+ }
+ else
+#endif
+ {
+ gsl_matrix *params = gsl_matrix_calloc (cov->size1, cov->size2);
+ gsl_matrix_memcpy (params, cov);
+ reg_sweep (params, l->dependent_column);
+ post_sweep_computations (l, params);
+ gsl_matrix_free (params);
+ }
}
-double pspp_linreg_mse (const pspp_linreg_cache *c)
+double
+linreg_mse (const struct linreg *c)
{
assert (c != NULL);
return (c->sse / c->dfe);
}
+
+double
+linreg_intercept (const struct linreg *c)
+{
+ return c->intercept;
+}
+
+const gsl_matrix *
+linreg_cov (const struct linreg *c)
+{
+ return c->cov;
+}
+
+double
+linreg_coeff (const struct linreg *c, size_t i)
+{
+ return (c->coeff[i]);
+}
+
+const struct variable *
+linreg_indep_var (const struct linreg *c, size_t i)
+{
+ return (c->indep_vars[i]);
+}
+
+int
+linreg_n_indeps (const struct linreg *c)
+{
+ return c->n_indeps;
+}
+
+
+const struct variable *
+linreg_dep_var (const struct linreg *c)
+{
+ return c->depvar;
+}
+
+
+size_t
+linreg_n_coeffs (const struct linreg *c)
+{
+ return c->n_coeffs;
+}
+
+double
+linreg_n_obs (const struct linreg *c)
+{
+ return c->n_obs;
+}
+
+double
+linreg_sse (const struct linreg *c)
+{
+ return c->sse;
+}
+
+double
+linreg_ssreg (const struct linreg *c)
+{
+ return (c->sst - c->sse);
+}
+
+double linreg_sst (const struct linreg *c)
+{
+ return c->sst;
+}
+
+double
+linreg_dfmodel ( const struct linreg *c)
+{
+ return c->dfm;
+}
+
+double
+linreg_dferror ( const struct linreg *c)
+{
+ return c->dfe;
+}
+
+double
+linreg_dftotal ( const struct linreg *c)
+{
+ return c->dft;
+}
+
+void
+linreg_set_depvar_mean (struct linreg *c, double x)
+{
+ c->depvar_mean = x;
+}
+
+double
+linreg_get_depvar_mean (const struct linreg *c)
+{
+ return c->depvar_mean;
+}