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/assertion.h"
37 #include "libpspp/ll.h"
38 #include "libpspp/message.h"
39 #include "libpspp/misc.h"
40 #include "libpspp/taint.h"
41 #include "linreg/sweep.h"
42 #include "math/categoricals.h"
43 #include "math/covariance.h"
44 #include "math/interaction.h"
45 #include "math/moments.h"
46 #include "output/tab.h"
49 #define _(msgid) gettext (msgid)
54 const struct variable **dep_vars;
57 const struct variable **factor_vars;
59 size_t n_interactions;
60 struct interaction **interactions;
62 enum mv_class exclude;
64 /* The weight variable */
65 const struct variable *wv;
67 const struct dictionary *dict;
80 struct moments *totals;
82 struct categoricals *cats;
85 Sums of squares due to different variables. Element 0 is the SSE
86 for the entire model. For i > 0, element i is the SS due to
93 /* Default design: all possible interactions */
95 design_full (struct glm_spec *glm)
99 glm->n_interactions = (1 << glm->n_factor_vars) - 1;
101 glm->interactions = xcalloc (glm->n_interactions, sizeof *glm->interactions);
103 /* All subsets, with exception of the empty set, of [0, glm->n_factor_vars) */
104 for (sz = 1; sz <= glm->n_factor_vars; ++sz)
106 gsl_combination *c = gsl_combination_calloc (glm->n_factor_vars, sz);
110 struct interaction *iact = interaction_create (NULL);
112 for (e = 0 ; e < gsl_combination_k (c); ++e)
113 interaction_add_variable (iact, glm->factor_vars [gsl_combination_get (c, e)]);
115 glm->interactions[i++] = iact;
117 while (gsl_combination_next (c) == GSL_SUCCESS);
119 gsl_combination_free (c);
123 static void output_glm (const struct glm_spec *,
124 const struct glm_workspace *ws);
125 static void run_glm (struct glm_spec *cmd, struct casereader *input,
126 const struct dataset *ds);
129 static bool parse_design_spec (struct lexer *lexer, struct glm_spec *glm);
133 cmd_glm (struct lexer *lexer, struct dataset *ds)
136 struct const_var_set *factors = NULL;
139 glm.dict = dataset_dict (ds);
141 glm.n_factor_vars = 0;
142 glm.n_interactions = 0;
143 glm.interactions = NULL;
145 glm.factor_vars = NULL;
146 glm.exclude = MV_ANY;
147 glm.intercept = true;
148 glm.wv = dict_get_weight (glm.dict);
150 glm.dump_coding = false;
153 if (!parse_variables_const (lexer, glm.dict,
154 &glm.dep_vars, &glm.n_dep_vars,
155 PV_NO_DUPLICATE | PV_NUMERIC))
158 lex_force_match (lexer, T_BY);
160 if (!parse_variables_const (lexer, glm.dict,
161 &glm.factor_vars, &glm.n_factor_vars,
162 PV_NO_DUPLICATE | PV_NUMERIC))
165 if (glm.n_dep_vars > 1)
167 msg (ME, _("Multivariate analysis is not yet implemented"));
172 const_var_set_create_from_array (glm.factor_vars, glm.n_factor_vars);
174 while (lex_token (lexer) != T_ENDCMD)
176 lex_match (lexer, T_SLASH);
178 if (lex_match_id (lexer, "MISSING"))
180 lex_match (lexer, T_EQUALS);
181 while (lex_token (lexer) != T_ENDCMD
182 && lex_token (lexer) != T_SLASH)
184 if (lex_match_id (lexer, "INCLUDE"))
186 glm.exclude = MV_SYSTEM;
188 else if (lex_match_id (lexer, "EXCLUDE"))
190 glm.exclude = MV_ANY;
194 lex_error (lexer, NULL);
199 else if (lex_match_id (lexer, "INTERCEPT"))
201 lex_match (lexer, T_EQUALS);
202 while (lex_token (lexer) != T_ENDCMD
203 && lex_token (lexer) != T_SLASH)
205 if (lex_match_id (lexer, "INCLUDE"))
207 glm.intercept = true;
209 else if (lex_match_id (lexer, "EXCLUDE"))
211 glm.intercept = false;
215 lex_error (lexer, NULL);
220 else if (lex_match_id (lexer, "CRITERIA"))
222 lex_match (lexer, T_EQUALS);
223 if (lex_match_id (lexer, "ALPHA"))
225 if (lex_force_match (lexer, T_LPAREN))
227 if (! lex_force_num (lexer))
229 lex_error (lexer, NULL);
233 glm.alpha = lex_number (lexer);
235 if ( ! lex_force_match (lexer, T_RPAREN))
237 lex_error (lexer, NULL);
244 lex_error (lexer, NULL);
248 else if (lex_match_id (lexer, "METHOD"))
250 lex_match (lexer, T_EQUALS);
251 if ( !lex_force_match_id (lexer, "SSTYPE"))
253 lex_error (lexer, NULL);
257 if ( ! lex_force_match (lexer, T_LPAREN))
259 lex_error (lexer, NULL);
263 if ( ! lex_force_int (lexer))
265 lex_error (lexer, NULL);
269 glm.ss_type = lex_integer (lexer);
270 if (1 != glm.ss_type && 2 != glm.ss_type )
272 msg (ME, _("Only types 1 & 2 sum of squares are currently implemented"));
278 if ( ! lex_force_match (lexer, T_RPAREN))
280 lex_error (lexer, NULL);
284 else if (lex_match_id (lexer, "DESIGN"))
286 lex_match (lexer, T_EQUALS);
288 if (! parse_design_spec (lexer, &glm))
291 if (glm.n_interactions > 0)
294 else if (lex_match_id (lexer, "SHOWCODES"))
295 /* Undocumented debug option */
297 lex_match (lexer, T_EQUALS);
299 glm.dump_coding = true;
303 lex_error (lexer, NULL);
314 struct casegrouper *grouper;
315 struct casereader *group;
318 grouper = casegrouper_create_splits (proc_open (ds), glm.dict);
319 while (casegrouper_get_next_group (grouper, &group))
320 run_glm (&glm, group, ds);
321 ok = casegrouper_destroy (grouper);
322 ok = proc_commit (ds) && ok;
325 const_var_set_destroy (factors);
326 free (glm.factor_vars);
327 for (i = 0 ; i < glm.n_interactions; ++i)
328 interaction_destroy (glm.interactions[i]);
329 free (glm.interactions);
337 const_var_set_destroy (factors);
338 free (glm.factor_vars);
339 for (i = 0 ; i < glm.n_interactions; ++i)
340 interaction_destroy (glm.interactions[i]);
342 free (glm.interactions);
348 static void get_ssq (struct covariance *, gsl_vector *,
349 const struct glm_spec *);
352 not_dropped (size_t j, const bool *ff)
358 fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f)
365 for (i = 0; i < cov->size1; i++)
367 if (not_dropped (i, dropped_f))
370 for (j = 0; j < cov->size2; j++)
372 if (not_dropped (j, dropped_f))
374 gsl_matrix_set (submatrix, n, m,
375 gsl_matrix_get (cov, i, j));
385 get_ssq (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd)
387 gsl_matrix *cm = covariance_calculate_unnormalized (cov);
390 bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped));
391 bool *submodel_dropped = xcalloc (covariance_dim (cov), sizeof (*submodel_dropped));
392 const struct categoricals *cats = covariance_get_categoricals (cov);
394 for (k = 0; k < cmd->n_interactions; k++)
396 gsl_matrix *model_cov = NULL;
397 gsl_matrix *submodel_cov = NULL;
398 size_t n_dropped_model = 0;
399 size_t n_dropped_submodel = 0;
400 for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++)
402 const struct interaction * x =
403 categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars);
405 model_dropped[i] = false;
406 submodel_dropped[i] = false;
407 if (interaction_is_subset (cmd->interactions [k], x))
409 assert (n_dropped_submodel < covariance_dim (cov));
410 n_dropped_submodel++;
411 submodel_dropped[i] = true;
413 if ( cmd->interactions [k]->n_vars < x->n_vars)
415 assert (n_dropped_model < covariance_dim (cov));
417 model_dropped[i] = true;
422 model_cov = gsl_matrix_alloc (cm->size1 - n_dropped_model, cm->size2 - n_dropped_model);
423 submodel_cov = gsl_matrix_alloc (cm->size1 - n_dropped_submodel, cm->size2 - n_dropped_submodel);
425 fill_submatrix (cm, model_cov, model_dropped);
426 fill_submatrix (cm, submodel_cov, submodel_dropped);
428 reg_sweep (model_cov, 0);
429 reg_sweep (submodel_cov, 0);
431 gsl_vector_set (ssq, k + 1,
432 gsl_matrix_get (submodel_cov, 0, 0) - gsl_matrix_get (model_cov, 0, 0)
435 gsl_matrix_free (model_cov);
436 gsl_matrix_free (submodel_cov);
439 free (model_dropped);
440 free (submodel_dropped);
441 gsl_matrix_free (cm);
444 //static void dump_matrix (const gsl_matrix *m);
447 run_glm (struct glm_spec *cmd, struct casereader *input,
448 const struct dataset *ds)
450 bool warn_bad_weight = true;
453 struct dictionary *dict = dataset_dict (ds);
454 struct casereader *reader;
457 struct glm_workspace ws;
458 struct covariance *cov;
460 ws.cats = categoricals_create (cmd->interactions, cmd->n_interactions,
461 cmd->wv, cmd->exclude,
462 NULL, NULL, NULL, NULL);
464 cov = covariance_2pass_create (cmd->n_dep_vars, cmd->dep_vars,
465 ws.cats, cmd->wv, cmd->exclude);
468 c = casereader_peek (input, 0);
471 casereader_destroy (input);
474 output_split_file_values (ds, c);
477 taint = taint_clone (casereader_get_taint (input));
479 ws.totals = moments_create (MOMENT_VARIANCE);
481 for (reader = casereader_clone (input);
482 (c = casereader_read (reader)) != NULL; case_unref (c))
484 double weight = dict_get_case_weight (dict, c, &warn_bad_weight);
486 for (v = 0; v < cmd->n_dep_vars; ++v)
487 moments_pass_one (ws.totals, case_data (c, cmd->dep_vars[v])->f,
490 covariance_accumulate_pass1 (cov, c);
492 casereader_destroy (reader);
494 if (cmd->dump_coding)
495 reader = casereader_clone (input);
500 (c = casereader_read (reader)) != NULL; case_unref (c))
502 double weight = dict_get_case_weight (dict, c, &warn_bad_weight);
504 for (v = 0; v < cmd->n_dep_vars; ++v)
505 moments_pass_two (ws.totals, case_data (c, cmd->dep_vars[v])->f,
508 covariance_accumulate_pass2 (cov, c);
510 casereader_destroy (reader);
513 if (cmd->dump_coding)
515 struct tab_table *t =
516 covariance_dump_enc_header (cov,
517 1 + casereader_count_cases (input));
519 (c = casereader_read (reader)) != NULL; case_unref (c))
521 covariance_dump_enc (cov, c, t);
523 casereader_destroy (reader);
528 gsl_matrix *cm = covariance_calculate_unnormalized (cov);
532 ws.total_ssq = gsl_matrix_get (cm, 0, 0);
537 Store the overall SSE.
539 ws.ssq = gsl_vector_alloc (cm->size1);
540 gsl_vector_set (ws.ssq, 0, gsl_matrix_get (cm, 0, 0));
541 switch (cmd->ss_type)
547 get_ssq (cov, ws.ssq, cmd);
555 gsl_matrix_free (cm);
558 if (!taint_has_tainted_successor (taint))
559 output_glm (cmd, &ws);
561 gsl_vector_free (ws.ssq);
563 covariance_destroy (cov);
564 moments_destroy (ws.totals);
566 taint_destroy (taint);
569 static const char *roman[] =
571 "", /* The Romans had no concept of zero */
579 output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws)
581 const struct fmt_spec *wfmt =
582 cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0;
584 double n_total, mean;
585 double df_corr = 0.0;
590 const int heading_columns = 1;
591 const int heading_rows = 1;
595 int nr = heading_rows + 4 + cmd->n_interactions;
599 msg (MW, "GLM is experimental. Do not rely on these results.");
600 t = tab_create (nc, nr);
601 tab_title (t, _("Tests of Between-Subjects Effects"));
603 tab_headers (t, heading_columns, 0, heading_rows, 0);
605 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, nc - 1, nr - 1);
607 tab_hline (t, TAL_2, 0, nc - 1, heading_rows);
608 tab_vline (t, TAL_2, heading_columns, 0, nr - 1);
610 tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Source"));
612 /* TRANSLATORS: The parameter is a roman numeral */
613 tab_text_format (t, 1, 0, TAB_CENTER | TAT_TITLE,
614 _("Type %s Sum of Squares"),
615 roman[cmd->ss_type]);
616 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("df"));
617 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Mean Square"));
618 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("F"));
619 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("Sig."));
621 moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL);
626 df_corr += categoricals_df_total (ws->cats);
628 mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr);
631 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model"));
637 const double intercept = pow2 (mean * n_total) / n_total;
638 const double df = 1.0;
639 const double F = intercept / df / mse;
640 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept"));
641 tab_double (t, 1, r, 0, intercept, NULL);
642 tab_double (t, 2, r, 0, 1.00, wfmt);
643 tab_double (t, 3, r, 0, intercept / df, NULL);
644 tab_double (t, 4, r, 0, F, NULL);
645 tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
650 for (f = 0; f < cmd->n_interactions; ++f)
652 struct string str = DS_EMPTY_INITIALIZER;
653 const double df = categoricals_df (ws->cats, f);
654 const double ssq = gsl_vector_get (ws->ssq, f + 1);
655 const double F = ssq / df / mse;
656 interaction_to_string (cmd->interactions[f], &str);
657 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, ds_cstr (&str));
660 tab_double (t, 1, r, 0, ssq, NULL);
661 tab_double (t, 2, r, 0, df, wfmt);
662 tab_double (t, 3, r, 0, ssq / df, NULL);
663 tab_double (t, 4, r, 0, F, NULL);
665 tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
671 /* Corrected Model */
672 const double df = df_corr - 1.0;
673 const double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0);
674 const double F = ssq / df / mse;
675 tab_double (t, 1, heading_rows, 0, ssq, NULL);
676 tab_double (t, 2, heading_rows, 0, df, wfmt);
677 tab_double (t, 3, heading_rows, 0, ssq / df, NULL);
678 tab_double (t, 4, heading_rows, 0, F, NULL);
680 tab_double (t, 5, heading_rows, 0,
681 gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL);
685 const double df = n_total - df_corr;
686 const double ssq = gsl_vector_get (ws->ssq, 0);
687 const double mse = ssq / df;
688 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Error"));
689 tab_double (t, 1, r, 0, ssq, NULL);
690 tab_double (t, 2, r, 0, df, wfmt);
691 tab_double (t, 3, r++, 0, mse, NULL);
696 const double intercept = pow2 (mean * n_total) / n_total;
697 const double ssq = intercept + ws->total_ssq;
699 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total"));
700 tab_double (t, 1, r, 0, ssq, NULL);
701 tab_double (t, 2, r, 0, n_total, wfmt);
706 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total"));
709 tab_double (t, 1, r, 0, ws->total_ssq, NULL);
710 tab_double (t, 2, r, 0, n_total - 1.0, wfmt);
717 dump_matrix (const gsl_matrix * m)
720 for (i = 0; i < m->size1; ++i)
722 for (j = 0; j < m->size2; ++j)
724 double x = gsl_matrix_get (m, i, j);
737 If the match succeeds, the variable will be placed in VAR.
738 Returns true if successful */
740 lex_match_variable (struct lexer *lexer, const struct glm_spec *glm, const struct variable **var)
742 if (lex_token (lexer) != T_ID)
745 *var = parse_variable_const (lexer, glm->dict);
752 /* An interaction is a variable followed by {*, BY} followed by an interaction */
754 parse_design_interaction (struct lexer *lexer, struct glm_spec *glm, struct interaction **iact)
756 const struct variable *v = NULL;
759 switch (lex_next_token (lexer, 1))
773 if (! lex_match_variable (lexer, glm, &v))
775 interaction_destroy (*iact);
783 *iact = interaction_create (v);
785 interaction_add_variable (*iact, v);
787 if ( lex_match (lexer, T_ASTERISK) || lex_match (lexer, T_BY))
789 return parse_design_interaction (lexer, glm, iact);
796 parse_nested_variable (struct lexer *lexer, struct glm_spec *glm)
798 const struct variable *v = NULL;
799 if ( ! lex_match_variable (lexer, glm, &v))
802 if (lex_match (lexer, T_LPAREN))
804 if ( ! parse_nested_variable (lexer, glm))
807 if ( ! lex_force_match (lexer, T_RPAREN))
811 lex_error (lexer, "Nested variables are not yet implemented"); return false;
815 /* A design term is an interaction OR a nested variable */
817 parse_design_term (struct lexer *lexer, struct glm_spec *glm)
819 struct interaction *iact = NULL;
820 if (parse_design_interaction (lexer, glm, &iact))
822 /* Interaction parsing successful. Add to list of interactions */
823 glm->interactions = xrealloc (glm->interactions, sizeof *glm->interactions * ++glm->n_interactions);
824 glm->interactions[glm->n_interactions - 1] = iact;
828 if ( parse_nested_variable (lexer, glm))
836 /* Parse a complete DESIGN specification.
837 A design spec is a design term, optionally followed by a comma,
838 and another design spec.
841 parse_design_spec (struct lexer *lexer, struct glm_spec *glm)
843 if (lex_token (lexer) == T_ENDCMD || lex_token (lexer) == T_SLASH)
846 if ( ! parse_design_term (lexer, glm))
849 lex_match (lexer, T_COMMA);
851 return parse_design_spec (lexer, glm);