1 /* PSPP - linear regression.
2 Copyright (C) 2005 Free Software Foundation, Inc.
3 Written by Jason H Stover <jason@sakla.net>.
5 This program is free software; you can redistribute it and/or
6 modify it under the terms of the GNU General Public License as
7 published by the Free Software Foundation; either version 2 of the
8 License, or (at your option) any later version.
10 This program is distributed in the hope that it will be useful, but
11 WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with this program; if not, write to the Free Software
17 Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
22 #include <gsl/gsl_cdf.h>
23 #include <gsl/gsl_vector.h>
24 #include <gsl/gsl_matrix.h>
30 #include "dictionary.h"
32 #include "file-handle.h"
35 #include <linreg/pspp_linreg.h>
44 "REGRESSION" (regression_):
68 static struct cmd_regression cmd;
71 Array holding the subscripts of the independent variables.
76 Return value for the procedure.
78 int pspp_reg_rc = CMD_SUCCESS;
80 static void run_regression (const struct casefile *, void *);
82 STATISTICS subcommand output functions.
84 static void reg_stats_r (pspp_linreg_cache *);
85 static void reg_stats_coeff (pspp_linreg_cache *);
86 static void reg_stats_anova (pspp_linreg_cache *);
87 static void reg_stats_outs (pspp_linreg_cache *);
88 static void reg_stats_zpp (pspp_linreg_cache *);
89 static void reg_stats_label (pspp_linreg_cache *);
90 static void reg_stats_sha (pspp_linreg_cache *);
91 static void reg_stats_ci (pspp_linreg_cache *);
92 static void reg_stats_f (pspp_linreg_cache *);
93 static void reg_stats_bcov (pspp_linreg_cache *);
94 static void reg_stats_ses (pspp_linreg_cache *);
95 static void reg_stats_xtx (pspp_linreg_cache *);
96 static void reg_stats_collin (pspp_linreg_cache *);
97 static void reg_stats_tol (pspp_linreg_cache *);
98 static void reg_stats_selection (pspp_linreg_cache *);
99 static void statistics_keyword_output (void (*)(pspp_linreg_cache *),
100 int, pspp_linreg_cache *);
103 reg_stats_r (pspp_linreg_cache * c)
113 rsq = c->ssm / c->sst;
114 adjrsq = 1.0 - (1.0 - rsq) * (c->n_obs - 1.0) / (c->n_obs - c->n_indeps);
115 std_error = sqrt ((c->n_indeps - 1.0) / (c->n_obs - 1.0));
116 t = tab_create (n_cols, n_rows, 0);
117 tab_dim (t, tab_natural_dimensions);
118 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
119 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
120 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
121 tab_vline (t, TAL_0, 1, 0, 0);
123 tab_text (t, 1, 0, TAB_CENTER | TAT_TITLE, _("R"));
124 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("R Square"));
125 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Adjusted R Square"));
126 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Std. Error of the Estimate"));
127 tab_float (t, 1, 1, TAB_RIGHT, sqrt (rsq), 10, 2);
128 tab_float (t, 2, 1, TAB_RIGHT, rsq, 10, 2);
129 tab_float (t, 3, 1, TAB_RIGHT, adjrsq, 10, 2);
130 tab_float (t, 4, 1, TAB_RIGHT, std_error, 10, 2);
131 tab_title (t, 0, _("Model Summary"));
136 Table showing estimated regression coefficients.
139 reg_stats_coeff (pspp_linreg_cache * c)
154 n_rows = c->n_coeffs + 2;
156 t = tab_create (n_cols, n_rows, 0);
157 tab_headers (t, 2, 0, 1, 0);
158 tab_dim (t, tab_natural_dimensions);
159 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
160 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
161 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
162 tab_vline (t, TAL_0, 1, 0, 0);
164 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("B"));
165 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Std. Error"));
166 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Beta"));
167 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("t"));
168 tab_text (t, 6, 0, TAB_CENTER | TAT_TITLE, _("Significance"));
169 tab_text (t, 1, 1, TAB_LEFT | TAT_TITLE, _("(Constant)"));
170 coeff = c->coeff[0].estimate;
171 tab_float (t, 2, 1, 0, coeff, 10, 2);
172 std_err = sqrt (gsl_matrix_get (c->cov, 0, 0));
173 tab_float (t, 3, 1, 0, std_err, 10, 2);
174 beta = coeff / c->depvar_std;
175 tab_float (t, 4, 1, 0, beta, 10, 2);
176 t_stat = coeff / std_err;
177 tab_float (t, 5, 1, 0, t_stat, 10, 2);
178 pval = 2 * gsl_cdf_tdist_Q (fabs (t_stat), 1.0);
179 tab_float (t, 6, 1, 0, pval, 10, 2);
180 for (j = 1; j <= c->n_indeps; j++)
183 label = var_to_string (c->coeff[j].v);
184 tab_text (t, 1, j + 1, TAB_CENTER, label);
186 Regression coefficients.
188 coeff = c->coeff[j].estimate;
189 tab_float (t, 2, j + 1, 0, coeff, 10, 2);
191 Standard error of the coefficients.
193 std_err = sqrt (gsl_matrix_get (c->cov, j, j));
194 tab_float (t, 3, j + 1, 0, std_err, 10, 2);
196 'Standardized' coefficient, i.e., regression coefficient
197 if all variables had unit variance.
199 beta = gsl_vector_get (c->indep_std, j);
200 beta *= coeff / c->depvar_std;
201 tab_float (t, 4, j + 1, 0, beta, 10, 2);
204 Test statistic for H0: coefficient is 0.
206 t_stat = coeff / std_err;
207 tab_float (t, 5, j + 1, 0, t_stat, 10, 2);
209 P values for the test statistic above.
211 pval = 2 * gsl_cdf_tdist_Q (fabs (t_stat), 1.0);
212 tab_float (t, 6, j + 1, 0, pval, 10, 2);
214 tab_title (t, 0, _("Coefficients"));
219 Display the ANOVA table.
222 reg_stats_anova (pspp_linreg_cache * c)
226 const double msm = c->ssm / c->dfm;
227 const double mse = c->sse / c->dfe;
228 const double F = msm / mse;
229 const double pval = gsl_cdf_fdist_Q (F, c->dfm, c->dfe);
234 t = tab_create (n_cols, n_rows, 0);
235 tab_headers (t, 2, 0, 1, 0);
236 tab_dim (t, tab_natural_dimensions);
238 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
240 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
241 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
242 tab_vline (t, TAL_0, 1, 0, 0);
244 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("Sum of Squares"));
245 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("df"));
246 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Mean Square"));
247 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("F"));
248 tab_text (t, 6, 0, TAB_CENTER | TAT_TITLE, _("Significance"));
250 tab_text (t, 1, 1, TAB_LEFT | TAT_TITLE, _("Regression"));
251 tab_text (t, 1, 2, TAB_LEFT | TAT_TITLE, _("Residual"));
252 tab_text (t, 1, 3, TAB_LEFT | TAT_TITLE, _("Total"));
254 /* Sums of Squares */
255 tab_float (t, 2, 1, 0, c->ssm, 10, 2);
256 tab_float (t, 2, 3, 0, c->sst, 10, 2);
257 tab_float (t, 2, 2, 0, c->sse, 10, 2);
260 /* Degrees of freedom */
261 tab_float (t, 3, 1, 0, c->dfm, 4, 0);
262 tab_float (t, 3, 2, 0, c->dfe, 4, 0);
263 tab_float (t, 3, 3, 0, c->dft, 4, 0);
267 tab_float (t, 4, 1, TAB_RIGHT, msm, 8, 3);
268 tab_float (t, 4, 2, TAB_RIGHT, mse, 8, 3);
270 tab_float (t, 5, 1, 0, F, 8, 3);
272 tab_float (t, 6, 1, 0, pval, 8, 3);
274 tab_title (t, 0, _("ANOVA"));
278 reg_stats_outs (pspp_linreg_cache * c)
283 reg_stats_zpp (pspp_linreg_cache * c)
288 reg_stats_label (pspp_linreg_cache * c)
293 reg_stats_sha (pspp_linreg_cache * c)
298 reg_stats_ci (pspp_linreg_cache * c)
303 reg_stats_f (pspp_linreg_cache * c)
308 reg_stats_bcov (pspp_linreg_cache * c)
321 n_cols = c->n_indeps + 1 + 2;
322 n_rows = 2 * (c->n_indeps + 1);
323 t = tab_create (n_cols, n_rows, 0);
324 tab_headers (t, 2, 0, 1, 0);
325 tab_dim (t, tab_natural_dimensions);
326 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
327 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
328 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
329 tab_vline (t, TAL_0, 1, 0, 0);
330 tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Model"));
331 tab_text (t, 1, 1, TAB_CENTER | TAT_TITLE, _("Covariances"));
332 for (i = 1; i < c->n_indeps + 1; i++)
334 j = indep_vars[(i - 1)];
335 struct variable *v = cmd.v_variables[j];
336 label = var_to_string (v);
337 tab_text (t, 2, i, TAB_CENTER, label);
338 tab_text (t, i + 2, 0, TAB_CENTER, label);
339 for (k = 1; k < c->n_indeps + 1; k++)
341 col = (i <= k) ? k : i;
342 row = (i <= k) ? i : k;
343 tab_float (t, k + 2, i, TAB_CENTER,
344 gsl_matrix_get (c->cov, row, col), 8, 3);
347 tab_title (t, 0, _("Coefficient Correlations"));
351 reg_stats_ses (pspp_linreg_cache * c)
356 reg_stats_xtx (pspp_linreg_cache * c)
361 reg_stats_collin (pspp_linreg_cache * c)
366 reg_stats_tol (pspp_linreg_cache * c)
371 reg_stats_selection (pspp_linreg_cache * c)
377 statistics_keyword_output (void (*function) (pspp_linreg_cache *),
378 int keyword, pspp_linreg_cache * c)
387 subcommand_statistics (int *keywords, pspp_linreg_cache * c)
390 The order here must match the order in which the STATISTICS
391 keywords appear in the specification section above.
418 Set everything but F.
420 for (i = 0; i < f; i++)
427 for (i = 0; i < all; i++)
435 Default output: ANOVA table, parameter estimates,
436 and statistics for variables not entered into model,
439 if (keywords[defaults] | d)
447 statistics_keyword_output (reg_stats_r, keywords[r], c);
448 statistics_keyword_output (reg_stats_anova, keywords[anova], c);
449 statistics_keyword_output (reg_stats_coeff, keywords[coeff], c);
450 statistics_keyword_output (reg_stats_outs, keywords[outs], c);
451 statistics_keyword_output (reg_stats_zpp, keywords[zpp], c);
452 statistics_keyword_output (reg_stats_label, keywords[label], c);
453 statistics_keyword_output (reg_stats_sha, keywords[sha], c);
454 statistics_keyword_output (reg_stats_ci, keywords[ci], c);
455 statistics_keyword_output (reg_stats_f, keywords[f], c);
456 statistics_keyword_output (reg_stats_bcov, keywords[bcov], c);
457 statistics_keyword_output (reg_stats_ses, keywords[ses], c);
458 statistics_keyword_output (reg_stats_xtx, keywords[xtx], c);
459 statistics_keyword_output (reg_stats_collin, keywords[collin], c);
460 statistics_keyword_output (reg_stats_tol, keywords[tol], c);
461 statistics_keyword_output (reg_stats_selection, keywords[selection], c);
465 cmd_regression (void)
467 if (!parse_regression (&cmd))
471 multipass_procedure_with_splits (run_regression, &cmd);
477 Is variable k one of the dependent variables?
483 for (j = 0; j < cmd.n_dependent; j++)
486 compare_var_names returns 0 if the variable
489 if (!compare_var_names (cmd.v_dependent[j], cmd.v_variables[k], NULL))
496 run_regression (const struct casefile *cf, void *cmd_ UNUSED)
504 const union value *val;
505 struct casereader *r;
506 struct casereader *r2;
508 const struct variable *v;
509 struct recoded_categorical_array *ca;
510 struct recoded_categorical *rc;
511 struct design_matrix *X;
513 pspp_linreg_cache *lcache;
514 pspp_linreg_opts lopts;
516 n_data = casefile_get_case_cnt (cf);
517 n_indep = cmd.n_variables - cmd.n_dependent;
518 indep_vars = xnmalloc (n_indep, sizeof *indep_vars);
520 Y = gsl_vector_alloc (n_data);
521 lopts.get_depvar_mean_std = 1;
522 lopts.get_indep_mean_std = xnmalloc (n_indep, sizeof (int));
524 lcache = pspp_linreg_cache_alloc (n_data, n_indep);
525 lcache->indep_means = gsl_vector_alloc (n_indep);
526 lcache->indep_std = gsl_vector_alloc (n_indep);
529 Read from the active file. The first pass encodes categorical
532 ca = cr_recoded_cat_ar_create (cmd.n_variables, cmd.v_variables);
533 for (r = casefile_get_reader (cf);
534 casereader_read (r, &c); case_destroy (&c))
536 for (i = 0; i < ca->n_vars; i++)
538 v = (*(ca->a + i))->v;
539 val = case_data (&c, v->fv);
540 cr_value_update (*(ca->a + i), val);
543 cr_create_value_matrices (ca);
545 design_matrix_create (n_indep, (const struct variable **) cmd.v_variables,
549 The second pass creates the design matrix.
551 for (r2 = casefile_get_reader (cf); casereader_read (r2, &c);
553 /* Iterate over the cases. */
556 row = casereader_cnum (r2) - 1;
557 for (i = 0; i < cmd.n_variables; ++i) /* Iterate over the variables
558 for the current case.
561 v = cmd.v_variables[i];
562 val = case_data (&c, v->fv);
564 Independent/dependent variable separation. The
565 'variables' subcommand specifies a varlist which contains
566 both dependent and independent variables. The dependent
567 variables are specified with the 'dependent'
568 subcommand. We need to separate the two.
572 if (v->type != NUMERIC)
574 msg (SE, gettext ("Dependent variable must be numeric."));
575 pspp_reg_rc = CMD_FAILURE;
578 lcache->depvar = (const struct var *) v;
579 gsl_vector_set (Y, row, val->f);
583 if (v->type == ALPHA)
585 rc = cr_var_to_recoded_categorical (v, ca);
586 design_matrix_set_categorical (X, row, v, val, rc);
588 else if (v->type == NUMERIC)
590 design_matrix_set_numeric (X, row, v, val);
595 lopts.get_indep_mean_std[i] = 1;
600 Now that we know the number of coefficients, allocate space
601 and store pointers to the variables that correspond to the
604 lcache->coeff = xnmalloc (X->m->size2 + 1, sizeof (*lcache->coeff));
605 for (i = 0; i < X->m->size2; i++)
607 j = i + 1; /* The first coeff is the intercept. */
609 (const struct variable *) design_matrix_col_to_var (X, i);
610 assert (lcache->coeff[j].v != NULL);
613 Find the least-squares estimates and other statistics.
615 pspp_linreg ((const gsl_vector *) Y, X->m, &lopts, lcache);
616 subcommand_statistics (cmd.a_statistics, lcache);
618 design_matrix_destroy (X);
619 pspp_linreg_cache_free (lcache);
620 free (lopts.get_indep_mean_std);
622 casereader_destroy (r);