1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2010, 2011 Free Software Foundation, Inc.
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation, either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <http://www.gnu.org/licenses/>. */
19 #include <gsl/gsl_cdf.h>
20 #include <gsl/gsl_matrix.h>
21 #include <gsl/gsl_combination.h>
24 #include "data/case.h"
25 #include "data/casegrouper.h"
26 #include "data/casereader.h"
27 #include "data/dataset.h"
28 #include "data/dictionary.h"
29 #include "data/format.h"
30 #include "data/value.h"
31 #include "language/command.h"
32 #include "language/dictionary/split-file.h"
33 #include "language/lexer/lexer.h"
34 #include "language/lexer/value-parser.h"
35 #include "language/lexer/variable-parser.h"
36 #include "libpspp/ll.h"
37 #include "libpspp/message.h"
38 #include "libpspp/misc.h"
39 #include "libpspp/taint.h"
40 #include "linreg/sweep.h"
41 #include "math/categoricals.h"
42 #include "math/covariance.h"
43 #include "math/interaction.h"
44 #include "math/moments.h"
45 #include "output/tab.h"
48 #define _(msgid) gettext (msgid)
53 const struct variable **dep_vars;
56 const struct variable **factor_vars;
58 size_t n_interactions;
59 struct interaction **interactions;
61 enum mv_class exclude;
63 /* The weight variable */
64 const struct variable *wv;
66 const struct dictionary *dict;
78 struct moments *totals;
80 struct categoricals *cats;
83 Sums of squares due to different variables. Element 0 is the SSE
84 for the entire model. For i > 0, element i is the SS due to
91 /* Default design: all possible interactions */
93 design_full (struct glm_spec *glm)
97 glm->n_interactions = (1 << glm->n_factor_vars) - 1;
99 glm->interactions = xcalloc (glm->n_interactions, sizeof *glm->interactions);
101 /* All subsets, with exception of the empty set, of [0, glm->n_factor_vars) */
102 for (sz = 1; sz <= glm->n_factor_vars; ++sz)
104 gsl_combination *c = gsl_combination_calloc (glm->n_factor_vars, sz);
108 struct interaction *iact = interaction_create (NULL);
110 for (e = 0 ; e < gsl_combination_k (c); ++e)
111 interaction_add_variable (iact, glm->factor_vars [gsl_combination_get (c, e)]);
113 glm->interactions[i++] = iact;
115 while (gsl_combination_next (c) == GSL_SUCCESS);
117 gsl_combination_free (c);
121 static void output_glm (const struct glm_spec *,
122 const struct glm_workspace *ws);
123 static void run_glm (struct glm_spec *cmd, struct casereader *input,
124 const struct dataset *ds);
127 static bool parse_design_spec (struct lexer *lexer, struct glm_spec *glm);
131 cmd_glm (struct lexer *lexer, struct dataset *ds)
134 struct const_var_set *factors = NULL;
137 glm.dict = dataset_dict (ds);
139 glm.n_factor_vars = 0;
140 glm.n_interactions = 0;
141 glm.interactions = NULL;
143 glm.factor_vars = NULL;
144 glm.exclude = MV_ANY;
145 glm.intercept = true;
146 glm.wv = dict_get_weight (glm.dict);
148 glm.dump_coding = false;
150 if (!parse_variables_const (lexer, glm.dict,
151 &glm.dep_vars, &glm.n_dep_vars,
152 PV_NO_DUPLICATE | PV_NUMERIC))
155 lex_force_match (lexer, T_BY);
157 if (!parse_variables_const (lexer, glm.dict,
158 &glm.factor_vars, &glm.n_factor_vars,
159 PV_NO_DUPLICATE | PV_NUMERIC))
162 if (glm.n_dep_vars > 1)
164 msg (ME, _("Multivariate analysis is not yet implemented"));
169 const_var_set_create_from_array (glm.factor_vars, glm.n_factor_vars);
171 while (lex_token (lexer) != T_ENDCMD)
173 lex_match (lexer, T_SLASH);
175 if (lex_match_id (lexer, "MISSING"))
177 lex_match (lexer, T_EQUALS);
178 while (lex_token (lexer) != T_ENDCMD
179 && lex_token (lexer) != T_SLASH)
181 if (lex_match_id (lexer, "INCLUDE"))
183 glm.exclude = MV_SYSTEM;
185 else if (lex_match_id (lexer, "EXCLUDE"))
187 glm.exclude = MV_ANY;
191 lex_error (lexer, NULL);
196 else if (lex_match_id (lexer, "INTERCEPT"))
198 lex_match (lexer, T_EQUALS);
199 while (lex_token (lexer) != T_ENDCMD
200 && lex_token (lexer) != T_SLASH)
202 if (lex_match_id (lexer, "INCLUDE"))
204 glm.intercept = true;
206 else if (lex_match_id (lexer, "EXCLUDE"))
208 glm.intercept = false;
212 lex_error (lexer, NULL);
217 else if (lex_match_id (lexer, "CRITERIA"))
219 lex_match (lexer, T_EQUALS);
220 if (lex_match_id (lexer, "ALPHA"))
222 if (lex_force_match (lexer, T_LPAREN))
224 if (! lex_force_num (lexer))
226 lex_error (lexer, NULL);
230 glm.alpha = lex_number (lexer);
232 if ( ! lex_force_match (lexer, T_RPAREN))
234 lex_error (lexer, NULL);
241 lex_error (lexer, NULL);
245 else if (lex_match_id (lexer, "METHOD"))
247 lex_match (lexer, T_EQUALS);
248 if ( !lex_force_match_id (lexer, "SSTYPE"))
250 lex_error (lexer, NULL);
254 if ( ! lex_force_match (lexer, T_LPAREN))
256 lex_error (lexer, NULL);
260 if ( ! lex_force_int (lexer))
262 lex_error (lexer, NULL);
266 if (3 != lex_integer (lexer))
268 msg (ME, _("Only type 3 sum of squares are currently implemented"));
274 if ( ! lex_force_match (lexer, T_RPAREN))
276 lex_error (lexer, NULL);
280 else if (lex_match_id (lexer, "DESIGN"))
282 lex_match (lexer, T_EQUALS);
284 if (! parse_design_spec (lexer, &glm))
287 if (glm.n_interactions > 0)
290 else if (lex_match_id (lexer, "SHOWCODES"))
291 /* Undocumented debug option */
293 lex_match (lexer, T_EQUALS);
295 glm.dump_coding = true;
299 lex_error (lexer, NULL);
310 struct casegrouper *grouper;
311 struct casereader *group;
314 grouper = casegrouper_create_splits (proc_open (ds), glm.dict);
315 while (casegrouper_get_next_group (grouper, &group))
316 run_glm (&glm, group, ds);
317 ok = casegrouper_destroy (grouper);
318 ok = proc_commit (ds) && ok;
321 const_var_set_destroy (factors);
322 free (glm.factor_vars);
323 for (i = 0 ; i < glm.n_interactions; ++i)
324 interaction_destroy (glm.interactions[i]);
325 free (glm.interactions);
333 const_var_set_destroy (factors);
334 free (glm.factor_vars);
335 for (i = 0 ; i < glm.n_interactions; ++i)
336 interaction_destroy (glm.interactions[i]);
338 free (glm.interactions);
344 static void get_ssq (struct covariance *, gsl_vector *,
345 const struct glm_spec *);
348 not_dropped (size_t j, const bool *ff)
354 fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f)
361 for (i = 0; i < cov->size1; i++)
363 if (not_dropped (i, dropped_f))
366 for (j = 0; j < cov->size2; j++)
368 if (not_dropped (j, dropped_f))
370 gsl_matrix_set (submatrix, n, m,
371 gsl_matrix_get (cov, i, j));
381 get_ssq (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd)
383 gsl_matrix *cm = covariance_calculate_unnormalized (cov);
386 bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped));
387 bool *submodel_dropped = xcalloc (covariance_dim (cov), sizeof (*submodel_dropped));
388 const struct categoricals *cats = covariance_get_categoricals (cov);
390 for (k = 0; k < cmd->n_interactions; k++)
392 gsl_matrix *model_cov = NULL;
393 gsl_matrix *submodel_cov = NULL;
394 size_t n_dropped_model = 0;
395 size_t n_dropped_submodel = 0;
396 for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++)
398 const struct interaction * x =
399 categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars);
401 model_dropped[i] = false;
402 submodel_dropped[i] = false;
403 if (interaction_is_subset (cmd->interactions [k], x))
405 assert (n_dropped_submodel < covariance_dim (cov));
406 n_dropped_submodel++;
407 submodel_dropped[i] = true;
409 if ( cmd->interactions [k]->n_vars < x->n_vars)
411 assert (n_dropped_model < covariance_dim (cov));
413 model_dropped[i] = true;
418 model_cov = gsl_matrix_alloc (cm->size1 - n_dropped_model, cm->size2 - n_dropped_model);
419 gsl_matrix_set (model_cov, 0, 0, gsl_matrix_get (cm, 0, 0));
420 submodel_cov = gsl_matrix_calloc (cm->size1 - n_dropped_submodel, cm->size2 - n_dropped_submodel);
422 fill_submatrix (cm, model_cov, model_dropped);
423 fill_submatrix (cm, submodel_cov, submodel_dropped);
425 reg_sweep (model_cov, 0);
426 reg_sweep (submodel_cov, 0);
428 gsl_vector_set (ssq, k + 1,
429 gsl_matrix_get (submodel_cov, 0, 0) - gsl_matrix_get (model_cov, 0, 0)
432 gsl_matrix_free (model_cov);
433 gsl_matrix_free (submodel_cov);
436 free (model_dropped);
437 free (submodel_dropped);
438 gsl_matrix_free (cm);
441 //static void dump_matrix (const gsl_matrix *m);
444 run_glm (struct glm_spec *cmd, struct casereader *input,
445 const struct dataset *ds)
447 bool warn_bad_weight = true;
450 struct dictionary *dict = dataset_dict (ds);
451 struct casereader *reader;
454 struct glm_workspace ws;
455 struct covariance *cov;
457 ws.cats = categoricals_create (cmd->interactions, cmd->n_interactions,
458 cmd->wv, cmd->exclude,
459 NULL, NULL, NULL, NULL);
461 cov = covariance_2pass_create (cmd->n_dep_vars, cmd->dep_vars,
462 ws.cats, cmd->wv, cmd->exclude);
465 c = casereader_peek (input, 0);
468 casereader_destroy (input);
471 output_split_file_values (ds, c);
474 taint = taint_clone (casereader_get_taint (input));
476 ws.totals = moments_create (MOMENT_VARIANCE);
478 for (reader = casereader_clone (input);
479 (c = casereader_read (reader)) != NULL; case_unref (c))
481 double weight = dict_get_case_weight (dict, c, &warn_bad_weight);
483 for (v = 0; v < cmd->n_dep_vars; ++v)
484 moments_pass_one (ws.totals, case_data (c, cmd->dep_vars[v])->f,
487 covariance_accumulate_pass1 (cov, c);
489 casereader_destroy (reader);
491 if (cmd->dump_coding)
492 reader = casereader_clone (input);
497 (c = casereader_read (reader)) != NULL; case_unref (c))
499 double weight = dict_get_case_weight (dict, c, &warn_bad_weight);
501 for (v = 0; v < cmd->n_dep_vars; ++v)
502 moments_pass_two (ws.totals, case_data (c, cmd->dep_vars[v])->f,
505 covariance_accumulate_pass2 (cov, c);
507 casereader_destroy (reader);
510 if (cmd->dump_coding)
512 struct tab_table *t =
513 covariance_dump_enc_header (cov,
514 1 + casereader_count_cases (input));
516 (c = casereader_read (reader)) != NULL; case_unref (c))
518 covariance_dump_enc (cov, c, t);
520 casereader_destroy (reader);
525 gsl_matrix *cm = covariance_calculate_unnormalized (cov);
529 ws.total_ssq = gsl_matrix_get (cm, 0, 0);
534 Store the overall SSE.
536 ws.ssq = gsl_vector_alloc (cm->size1);
537 gsl_vector_set (ws.ssq, 0, gsl_matrix_get (cm, 0, 0));
538 get_ssq (cov, ws.ssq, cmd);
541 gsl_matrix_free (cm);
544 if (!taint_has_tainted_successor (taint))
545 output_glm (cmd, &ws);
547 gsl_vector_free (ws.ssq);
549 covariance_destroy (cov);
550 moments_destroy (ws.totals);
552 taint_destroy (taint);
556 output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws)
558 const struct fmt_spec *wfmt =
559 cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0;
561 double n_total, mean;
562 double df_corr = 0.0;
567 const int heading_columns = 1;
568 const int heading_rows = 1;
572 int nr = heading_rows + 4 + cmd->n_interactions;
576 msg (MW, "GLM is experimental. Do not rely on these results.");
577 t = tab_create (nc, nr);
578 tab_title (t, _("Tests of Between-Subjects Effects"));
580 tab_headers (t, heading_columns, 0, heading_rows, 0);
582 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, nc - 1, nr - 1);
584 tab_hline (t, TAL_2, 0, nc - 1, heading_rows);
585 tab_vline (t, TAL_2, heading_columns, 0, nr - 1);
587 tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Source"));
589 /* TRANSLATORS: The parameter is a roman numeral */
590 tab_text_format (t, 1, 0, TAB_CENTER | TAT_TITLE,
591 _("Type %s Sum of Squares"), "III");
592 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("df"));
593 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Mean Square"));
594 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("F"));
595 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("Sig."));
597 moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL);
602 df_corr += categoricals_df_total (ws->cats);
604 mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr);
607 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model"));
613 const double intercept = pow2 (mean * n_total) / n_total;
614 const double df = 1.0;
615 const double F = intercept / df / mse;
616 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept"));
617 tab_double (t, 1, r, 0, intercept, NULL);
618 tab_double (t, 2, r, 0, 1.00, wfmt);
619 tab_double (t, 3, r, 0, intercept / df, NULL);
620 tab_double (t, 4, r, 0, F, NULL);
621 tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
626 for (f = 0; f < cmd->n_interactions; ++f)
628 struct string str = DS_EMPTY_INITIALIZER;
629 const double df = categoricals_df (ws->cats, f);
630 const double ssq = gsl_vector_get (ws->ssq, f + 1);
631 const double F = ssq / df / mse;
632 interaction_to_string (cmd->interactions[f], &str);
633 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, ds_cstr (&str));
636 tab_double (t, 1, r, 0, ssq, NULL);
637 tab_double (t, 2, r, 0, df, wfmt);
638 tab_double (t, 3, r, 0, ssq / df, NULL);
639 tab_double (t, 4, r, 0, F, NULL);
641 tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
647 /* Corrected Model */
648 const double df = df_corr - 1.0;
649 const double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0);
650 const double F = ssq / df / mse;
651 tab_double (t, 1, heading_rows, 0, ssq, NULL);
652 tab_double (t, 2, heading_rows, 0, df, wfmt);
653 tab_double (t, 3, heading_rows, 0, ssq / df, NULL);
654 tab_double (t, 4, heading_rows, 0, F, NULL);
656 tab_double (t, 5, heading_rows, 0,
657 gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL);
661 const double df = n_total - df_corr;
662 const double ssq = gsl_vector_get (ws->ssq, 0);
663 const double mse = ssq / df;
664 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Error"));
665 tab_double (t, 1, r, 0, ssq, NULL);
666 tab_double (t, 2, r, 0, df, wfmt);
667 tab_double (t, 3, r++, 0, mse, NULL);
672 const double intercept = pow2 (mean * n_total) / n_total;
673 const double ssq = intercept + ws->total_ssq;
675 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total"));
676 tab_double (t, 1, r, 0, ssq, NULL);
677 tab_double (t, 2, r, 0, n_total, wfmt);
682 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total"));
685 tab_double (t, 1, r, 0, ws->total_ssq, NULL);
686 tab_double (t, 2, r, 0, n_total - 1.0, wfmt);
693 dump_matrix (const gsl_matrix * m)
696 for (i = 0; i < m->size1; ++i)
698 for (j = 0; j < m->size2; ++j)
700 double x = gsl_matrix_get (m, i, j);
713 If the match succeeds, the variable will be placed in VAR.
714 Returns true if successful */
716 lex_match_variable (struct lexer *lexer, const struct glm_spec *glm, const struct variable **var)
718 if (lex_token (lexer) != T_ID)
721 *var = parse_variable_const (lexer, glm->dict);
728 /* An interaction is a variable followed by {*, BY} followed by an interaction */
730 parse_design_interaction (struct lexer *lexer, struct glm_spec *glm, struct interaction **iact)
732 const struct variable *v = NULL;
735 switch (lex_next_token (lexer, 1))
749 if (! lex_match_variable (lexer, glm, &v))
751 interaction_destroy (*iact);
759 *iact = interaction_create (v);
761 interaction_add_variable (*iact, v);
763 if ( lex_match (lexer, T_ASTERISK) || lex_match (lexer, T_BY))
765 return parse_design_interaction (lexer, glm, iact);
772 parse_nested_variable (struct lexer *lexer, struct glm_spec *glm)
774 const struct variable *v = NULL;
775 if ( ! lex_match_variable (lexer, glm, &v))
778 if (lex_match (lexer, T_LPAREN))
780 if ( ! parse_nested_variable (lexer, glm))
783 if ( ! lex_force_match (lexer, T_RPAREN))
787 lex_error (lexer, "Nested variables are not yet implemented"); return false;
791 /* A design term is an interaction OR a nested variable */
793 parse_design_term (struct lexer *lexer, struct glm_spec *glm)
795 struct interaction *iact = NULL;
796 if (parse_design_interaction (lexer, glm, &iact))
798 /* Interaction parsing successful. Add to list of interactions */
799 glm->interactions = xrealloc (glm->interactions, sizeof *glm->interactions * ++glm->n_interactions);
800 glm->interactions[glm->n_interactions - 1] = iact;
804 if ( parse_nested_variable (lexer, glm))
812 /* Parse a complete DESIGN specification.
813 A design spec is a design term, optionally followed by a comma,
814 and another design spec.
817 parse_design_spec (struct lexer *lexer, struct glm_spec *glm)
819 if (lex_token (lexer) == T_ENDCMD || lex_token (lexer) == T_SLASH)
822 if ( ! parse_design_term (lexer, glm))
825 lex_match (lexer, T_COMMA);
827 return parse_design_spec (lexer, glm);