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>
29 #include "cat-routines.h"
31 #include "design-matrix.h"
32 #include "dictionary.h"
34 #include "file-handle.h"
37 #include <linreg/pspp_linreg.h>
38 #include "missing-values.h"
43 #define REG_LARGE_DATA 1000
48 "REGRESSION" (regression_):
73 static struct cmd_regression cmd;
76 Array holding the subscripts of the independent variables.
81 File where the model will be saved if the EXPORT subcommand
84 struct file_handle *model_file;
87 Return value for the procedure.
89 int pspp_reg_rc = CMD_SUCCESS;
91 static void run_regression (const struct casefile *, void *);
93 STATISTICS subcommand output functions.
95 static void reg_stats_r (pspp_linreg_cache *);
96 static void reg_stats_coeff (pspp_linreg_cache *);
97 static void reg_stats_anova (pspp_linreg_cache *);
98 static void reg_stats_outs (pspp_linreg_cache *);
99 static void reg_stats_zpp (pspp_linreg_cache *);
100 static void reg_stats_label (pspp_linreg_cache *);
101 static void reg_stats_sha (pspp_linreg_cache *);
102 static void reg_stats_ci (pspp_linreg_cache *);
103 static void reg_stats_f (pspp_linreg_cache *);
104 static void reg_stats_bcov (pspp_linreg_cache *);
105 static void reg_stats_ses (pspp_linreg_cache *);
106 static void reg_stats_xtx (pspp_linreg_cache *);
107 static void reg_stats_collin (pspp_linreg_cache *);
108 static void reg_stats_tol (pspp_linreg_cache *);
109 static void reg_stats_selection (pspp_linreg_cache *);
110 static void statistics_keyword_output (void (*)(pspp_linreg_cache *),
111 int, pspp_linreg_cache *);
114 reg_stats_r (pspp_linreg_cache * c)
124 rsq = c->ssm / c->sst;
125 adjrsq = 1.0 - (1.0 - rsq) * (c->n_obs - 1.0) / (c->n_obs - c->n_indeps);
126 std_error = sqrt ((c->n_indeps - 1.0) / (c->n_obs - 1.0));
127 t = tab_create (n_cols, n_rows, 0);
128 tab_dim (t, tab_natural_dimensions);
129 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
130 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
131 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
132 tab_vline (t, TAL_0, 1, 0, 0);
134 tab_text (t, 1, 0, TAB_CENTER | TAT_TITLE, _("R"));
135 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("R Square"));
136 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Adjusted R Square"));
137 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Std. Error of the Estimate"));
138 tab_float (t, 1, 1, TAB_RIGHT, sqrt (rsq), 10, 2);
139 tab_float (t, 2, 1, TAB_RIGHT, rsq, 10, 2);
140 tab_float (t, 3, 1, TAB_RIGHT, adjrsq, 10, 2);
141 tab_float (t, 4, 1, TAB_RIGHT, std_error, 10, 2);
142 tab_title (t, 0, _("Model Summary"));
147 Table showing estimated regression coefficients.
150 reg_stats_coeff (pspp_linreg_cache * c)
165 n_rows = c->n_coeffs + 2;
167 t = tab_create (n_cols, n_rows, 0);
168 tab_headers (t, 2, 0, 1, 0);
169 tab_dim (t, tab_natural_dimensions);
170 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
171 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
172 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
173 tab_vline (t, TAL_0, 1, 0, 0);
175 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("B"));
176 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Std. Error"));
177 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Beta"));
178 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("t"));
179 tab_text (t, 6, 0, TAB_CENTER | TAT_TITLE, _("Significance"));
180 tab_text (t, 1, 1, TAB_LEFT | TAT_TITLE, _("(Constant)"));
181 coeff = c->coeff[0].estimate;
182 tab_float (t, 2, 1, 0, coeff, 10, 2);
183 std_err = sqrt (gsl_matrix_get (c->cov, 0, 0));
184 tab_float (t, 3, 1, 0, std_err, 10, 2);
185 beta = coeff / c->depvar_std;
186 tab_float (t, 4, 1, 0, beta, 10, 2);
187 t_stat = coeff / std_err;
188 tab_float (t, 5, 1, 0, t_stat, 10, 2);
189 pval = 2 * gsl_cdf_tdist_Q (fabs (t_stat), 1.0);
190 tab_float (t, 6, 1, 0, pval, 10, 2);
191 for (j = 1; j <= c->n_indeps; j++)
194 label = var_to_string (c->coeff[j].v);
195 tab_text (t, 1, j + 1, TAB_CENTER, label);
197 Regression coefficients.
199 coeff = c->coeff[j].estimate;
200 tab_float (t, 2, j + 1, 0, coeff, 10, 2);
202 Standard error of the coefficients.
204 std_err = sqrt (gsl_matrix_get (c->cov, j, j));
205 tab_float (t, 3, j + 1, 0, std_err, 10, 2);
207 'Standardized' coefficient, i.e., regression coefficient
208 if all variables had unit variance.
210 beta = gsl_vector_get (c->indep_std, j);
211 beta *= coeff / c->depvar_std;
212 tab_float (t, 4, j + 1, 0, beta, 10, 2);
215 Test statistic for H0: coefficient is 0.
217 t_stat = coeff / std_err;
218 tab_float (t, 5, j + 1, 0, t_stat, 10, 2);
220 P values for the test statistic above.
222 pval = 2 * gsl_cdf_tdist_Q (fabs (t_stat), 1.0);
223 tab_float (t, 6, j + 1, 0, pval, 10, 2);
225 tab_title (t, 0, _("Coefficients"));
230 Display the ANOVA table.
233 reg_stats_anova (pspp_linreg_cache * c)
237 const double msm = c->ssm / c->dfm;
238 const double mse = c->sse / c->dfe;
239 const double F = msm / mse;
240 const double pval = gsl_cdf_fdist_Q (F, c->dfm, c->dfe);
245 t = tab_create (n_cols, n_rows, 0);
246 tab_headers (t, 2, 0, 1, 0);
247 tab_dim (t, tab_natural_dimensions);
249 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
251 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
252 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
253 tab_vline (t, TAL_0, 1, 0, 0);
255 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("Sum of Squares"));
256 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("df"));
257 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Mean Square"));
258 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("F"));
259 tab_text (t, 6, 0, TAB_CENTER | TAT_TITLE, _("Significance"));
261 tab_text (t, 1, 1, TAB_LEFT | TAT_TITLE, _("Regression"));
262 tab_text (t, 1, 2, TAB_LEFT | TAT_TITLE, _("Residual"));
263 tab_text (t, 1, 3, TAB_LEFT | TAT_TITLE, _("Total"));
265 /* Sums of Squares */
266 tab_float (t, 2, 1, 0, c->ssm, 10, 2);
267 tab_float (t, 2, 3, 0, c->sst, 10, 2);
268 tab_float (t, 2, 2, 0, c->sse, 10, 2);
271 /* Degrees of freedom */
272 tab_float (t, 3, 1, 0, c->dfm, 4, 0);
273 tab_float (t, 3, 2, 0, c->dfe, 4, 0);
274 tab_float (t, 3, 3, 0, c->dft, 4, 0);
278 tab_float (t, 4, 1, TAB_RIGHT, msm, 8, 3);
279 tab_float (t, 4, 2, TAB_RIGHT, mse, 8, 3);
281 tab_float (t, 5, 1, 0, F, 8, 3);
283 tab_float (t, 6, 1, 0, pval, 8, 3);
285 tab_title (t, 0, _("ANOVA"));
289 reg_stats_outs (pspp_linreg_cache * c)
294 reg_stats_zpp (pspp_linreg_cache * c)
299 reg_stats_label (pspp_linreg_cache * c)
304 reg_stats_sha (pspp_linreg_cache * c)
309 reg_stats_ci (pspp_linreg_cache * c)
314 reg_stats_f (pspp_linreg_cache * c)
319 reg_stats_bcov (pspp_linreg_cache * c)
332 n_cols = c->n_indeps + 1 + 2;
333 n_rows = 2 * (c->n_indeps + 1);
334 t = tab_create (n_cols, n_rows, 0);
335 tab_headers (t, 2, 0, 1, 0);
336 tab_dim (t, tab_natural_dimensions);
337 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
338 tab_hline (t, TAL_2, 0, n_cols - 1, 1);
339 tab_vline (t, TAL_2, 2, 0, n_rows - 1);
340 tab_vline (t, TAL_0, 1, 0, 0);
341 tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Model"));
342 tab_text (t, 1, 1, TAB_CENTER | TAT_TITLE, _("Covariances"));
343 for (i = 1; i < c->n_indeps + 1; i++)
345 j = indep_vars[(i - 1)];
346 struct variable *v = cmd.v_variables[j];
347 label = var_to_string (v);
348 tab_text (t, 2, i, TAB_CENTER, label);
349 tab_text (t, i + 2, 0, TAB_CENTER, label);
350 for (k = 1; k < c->n_indeps + 1; k++)
352 col = (i <= k) ? k : i;
353 row = (i <= k) ? i : k;
354 tab_float (t, k + 2, i, TAB_CENTER,
355 gsl_matrix_get (c->cov, row, col), 8, 3);
358 tab_title (t, 0, _("Coefficient Correlations"));
362 reg_stats_ses (pspp_linreg_cache * c)
367 reg_stats_xtx (pspp_linreg_cache * c)
372 reg_stats_collin (pspp_linreg_cache * c)
377 reg_stats_tol (pspp_linreg_cache * c)
382 reg_stats_selection (pspp_linreg_cache * c)
388 statistics_keyword_output (void (*function) (pspp_linreg_cache *),
389 int keyword, pspp_linreg_cache * c)
398 subcommand_statistics (int *keywords, pspp_linreg_cache * c)
401 The order here must match the order in which the STATISTICS
402 keywords appear in the specification section above.
429 Set everything but F.
431 for (i = 0; i < f; i++)
438 for (i = 0; i < all; i++)
446 Default output: ANOVA table, parameter estimates,
447 and statistics for variables not entered into model,
450 if (keywords[defaults] | d)
458 statistics_keyword_output (reg_stats_r, keywords[r], c);
459 statistics_keyword_output (reg_stats_anova, keywords[anova], c);
460 statistics_keyword_output (reg_stats_coeff, keywords[coeff], c);
461 statistics_keyword_output (reg_stats_outs, keywords[outs], c);
462 statistics_keyword_output (reg_stats_zpp, keywords[zpp], c);
463 statistics_keyword_output (reg_stats_label, keywords[label], c);
464 statistics_keyword_output (reg_stats_sha, keywords[sha], c);
465 statistics_keyword_output (reg_stats_ci, keywords[ci], c);
466 statistics_keyword_output (reg_stats_f, keywords[f], c);
467 statistics_keyword_output (reg_stats_bcov, keywords[bcov], c);
468 statistics_keyword_output (reg_stats_ses, keywords[ses], c);
469 statistics_keyword_output (reg_stats_xtx, keywords[xtx], c);
470 statistics_keyword_output (reg_stats_collin, keywords[collin], c);
471 statistics_keyword_output (reg_stats_tol, keywords[tol], c);
472 statistics_keyword_output (reg_stats_selection, keywords[selection], c);
475 subcommand_export (int export, pspp_linreg_cache *c)
479 struct pspp_linreg_coeff coeff;
484 assert (model_file != NULL);
486 fp = fopen (handle_get_filename (model_file), "w");
487 fprintf (fp, "#include <string.h>\n\n");
488 fprintf (fp, "/*\n Estimate the mean of Y, the dependent variable for\n");
489 fprintf (fp, " the linear model of the form \n\n");
490 fprintf (fp, " Y = b0 + b1 * X1 + b2 * X2 + ... + bk * X2 + error\n\n");
491 fprintf (fp, " where X1, ..., Xk are the independent variables\n");
492 fprintf (fp, " whose values are stored in var_vals and whose names, \n");
493 fprintf (fp, " as known by PSPP, are stored in var_names. The estimated \n");
494 fprintf (fp, " regression coefficients (i.e., the estimates of b0,...,bk) \n");
495 fprintf (fp, " are stored in model_coeffs.\n*/\n");
496 fprintf (fp, "double\npspp_reg_estimate (const double *var_vals, const char *var_names[])\n{\n\tchar *model_depvars[%d] = {", c->n_indeps);
497 for (i = 1; i < c->n_indeps; i++)
500 fprintf (fp, "\"%s\",\n\t\t", coeff.v->name);
503 fprintf (fp, "\"%s\"};\n\t", coeff.v->name);
504 fprintf (fp, "double model_coeffs[%d] = {", c->n_indeps);
505 for (i = 1; i < c->n_indeps; i++)
508 fprintf (fp, "%.15e,\n\t\t", coeff.estimate);
511 fprintf (fp, "%.15e};\n\t", coeff.estimate);
513 fprintf (fp, "double estimate = %.15e;\n\t", coeff.estimate);
514 fprintf (fp, "int i;\n\tint j;\n\n\t");
515 fprintf (fp, "for (i = 0; i < %d; i++)\n\t", c->n_indeps);
516 fprintf (fp, "{\n\t\tfor (j = 0; j < %d; j++)\n\t\t", c->n_indeps);
517 fprintf (fp, "{\n\t\t\tif (strcmp (var_names[i], model_depvars[j]) == 0)\n");
518 fprintf (fp, "\t\t\t{\n\t\t\t\testimate += var_vals[i] * model_coeffs[j];\n");
519 fprintf (fp, "\t\t\t}\n\t\t}\n\t}\n\treturn estimate;\n}\n");
524 regression_custom_export (struct cmd_regression *cmd)
526 /* 0 on failure, 1 on success, 2 on failure that should result in syntax error */
527 if (!lex_force_match ('('))
534 model_file = fh_parse ();
535 if (model_file == NULL)
539 if (!lex_force_match (')'))
546 cmd_regression (void)
548 if (!parse_regression (&cmd))
552 multipass_procedure_with_splits (run_regression, &cmd);
558 Is variable k one of the dependent variables?
564 for (j = 0; j < cmd.n_dependent; j++)
567 compare_var_names returns 0 if the variable
570 if (!compare_var_names (cmd.v_dependent[j], cmd.v_variables[k], NULL))
577 run_regression (const struct casefile *cf, void *cmd_ UNUSED)
586 Keep track of the missing cases.
588 int *is_missing_case;
589 const union value *val;
590 struct casereader *r;
591 struct casereader *r2;
594 struct variable **indep_vars;
595 struct design_matrix *X;
597 pspp_linreg_cache *lcache;
598 pspp_linreg_opts lopts;
600 n_data = casefile_get_case_cnt (cf);
602 is_missing_case = xnmalloc (n_data, sizeof (*is_missing_case));
603 for (i = 0; i < n_data; i++)
604 is_missing_case[i] = 0;
606 n_indep = cmd.n_variables - cmd.n_dependent;
607 indep_vars = xnmalloc (n_indep, sizeof *indep_vars);
609 lopts.get_depvar_mean_std = 1;
610 lopts.get_indep_mean_std = xnmalloc (n_indep, sizeof (int));
614 Read from the active file. The first pass encodes categorical
615 variables and drops cases with missing values.
618 for (i = 0; i < cmd.n_variables; i++)
622 v = cmd.v_variables[i];
625 if (v->type == ALPHA)
627 /* Make a place to hold the binary vectors
628 corresponding to this variable's values. */
629 cat_stored_values_create (v);
631 for (r = casefile_get_reader (cf);
632 casereader_read (r, &c); case_destroy (&c))
634 row = casereader_cnum (r) - 1;
636 val = case_data (&c, v->fv);
637 cat_value_update (v, val);
638 if (mv_is_value_missing (&v->miss, val))
640 if (!is_missing_case[row])
642 /* Now it is missing. */
644 is_missing_case[row] = 1;
651 Y = gsl_vector_alloc (n_data);
653 design_matrix_create (n_indep, (const struct variable **) indep_vars,
655 lcache = pspp_linreg_cache_alloc (X->m->size1, X->m->size2);
656 lcache->indep_means = gsl_vector_alloc (X->m->size2);
657 lcache->indep_std = gsl_vector_alloc (X->m->size2);
660 The second pass creates the design matrix.
663 for (r2 = casefile_get_reader (cf); casereader_read (r2, &c);
665 /* Iterate over the cases. */
667 case_num = casereader_cnum (r2) - 1;
668 if (!is_missing_case[case_num])
670 for (i = 0; i < cmd.n_variables; ++i) /* Iterate over the variables
671 for the current case.
674 v = cmd.v_variables[i];
675 val = case_data (&c, v->fv);
677 Independent/dependent variable separation. The
678 'variables' subcommand specifies a varlist which contains
679 both dependent and independent variables. The dependent
680 variables are specified with the 'dependent'
681 subcommand. We need to separate the two.
685 if (v->type != NUMERIC)
688 gettext ("Dependent variable must be numeric."));
689 pspp_reg_rc = CMD_FAILURE;
692 lcache->depvar = (const struct variable *) v;
693 gsl_vector_set (Y, row, val->f);
697 if (v->type == ALPHA)
699 design_matrix_set_categorical (X, row, v, val);
701 else if (v->type == NUMERIC)
703 design_matrix_set_numeric (X, row, v, val);
706 lopts.get_indep_mean_std[i] = 1;
713 Now that we know the number of coefficients, allocate space
714 and store pointers to the variables that correspond to the
717 lcache->coeff = xnmalloc (X->m->size2 + 1, sizeof (*lcache->coeff));
718 for (i = 0; i < X->m->size2; i++)
720 j = i + 1; /* The first coeff is the intercept. */
722 (const struct variable *) design_matrix_col_to_var (X, i);
723 assert (lcache->coeff[j].v != NULL);
726 For large data sets, use QR decomposition.
728 if (n_data > sqrt (n_indep) && n_data > REG_LARGE_DATA)
730 lcache->method = PSPP_LINREG_SVD;
733 Find the least-squares estimates and other statistics.
735 pspp_linreg ((const gsl_vector *) Y, X->m, &lopts, lcache);
736 subcommand_statistics (cmd.a_statistics, lcache);
737 subcommand_export (cmd.sbc_export, lcache);
739 design_matrix_destroy (X);
740 pspp_linreg_cache_free (lcache);
741 free (lopts.get_indep_mean_std);
743 free (is_missing_case);
744 casereader_destroy (r);