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>
23 #include "data/case.h"
24 #include "data/casegrouper.h"
25 #include "data/casereader.h"
26 #include "data/dataset.h"
27 #include "data/dictionary.h"
28 #include "data/format.h"
29 #include "data/value.h"
30 #include "language/command.h"
31 #include "language/dictionary/split-file.h"
32 #include "language/lexer/lexer.h"
33 #include "language/lexer/value-parser.h"
34 #include "language/lexer/variable-parser.h"
35 #include "libpspp/ll.h"
36 #include "libpspp/message.h"
37 #include "libpspp/misc.h"
38 #include "libpspp/taint.h"
39 #include "linreg/sweep.h"
40 #include "math/categoricals.h"
41 #include "math/covariance.h"
42 #include "math/interaction.h"
43 #include "math/moments.h"
44 #include "output/tab.h"
47 #define _(msgid) gettext (msgid)
52 const struct variable **dep_vars;
55 const struct variable **factor_vars;
57 /* In the current implementation, design_vars will
58 normally be the same as factor_vars.
59 This will change once interactions, nested variables
60 and repeated measures become involved.
63 const struct variable **design_vars;
65 size_t n_interactions;
66 struct interaction **interactions;
68 enum mv_class exclude;
70 /* The weight variable */
71 const struct variable *wv;
73 const struct dictionary *dict;
83 struct moments *totals;
85 struct categoricals *cats;
88 Sums of squares due to different variables. Element 0 is the SSE
89 for the entire model. For i > 0, element i is the SS due to
95 static void output_glm (const struct glm_spec *,
96 const struct glm_workspace *ws);
97 static void run_glm (struct glm_spec *cmd, struct casereader *input,
98 const struct dataset *ds);
101 static bool parse_design_spec (struct lexer *lexer, struct glm_spec *glm);
105 cmd_glm (struct lexer *lexer, struct dataset *ds)
108 struct const_var_set *factors = NULL;
111 glm.dict = dataset_dict (ds);
113 glm.n_factor_vars = 0;
114 glm.n_design_vars = 0;
115 glm.n_interactions = 0;
116 glm.interactions = NULL;
118 glm.factor_vars = NULL;
119 glm.design_vars = NULL;
120 glm.exclude = MV_ANY;
121 glm.intercept = true;
122 glm.wv = dict_get_weight (glm.dict);
125 if (!parse_variables_const (lexer, glm.dict,
126 &glm.dep_vars, &glm.n_dep_vars,
127 PV_NO_DUPLICATE | PV_NUMERIC))
130 lex_force_match (lexer, T_BY);
132 if (!parse_variables_const (lexer, glm.dict,
133 &glm.factor_vars, &glm.n_factor_vars,
134 PV_NO_DUPLICATE | PV_NUMERIC))
137 if (glm.n_dep_vars > 1)
139 msg (ME, _("Multivariate analysis is not yet implemented"));
144 const_var_set_create_from_array (glm.factor_vars, glm.n_factor_vars);
146 while (lex_token (lexer) != T_ENDCMD)
148 lex_match (lexer, T_SLASH);
150 if (lex_match_id (lexer, "MISSING"))
152 lex_match (lexer, T_EQUALS);
153 while (lex_token (lexer) != T_ENDCMD
154 && lex_token (lexer) != T_SLASH)
156 if (lex_match_id (lexer, "INCLUDE"))
158 glm.exclude = MV_SYSTEM;
160 else if (lex_match_id (lexer, "EXCLUDE"))
162 glm.exclude = MV_ANY;
166 lex_error (lexer, NULL);
171 else if (lex_match_id (lexer, "INTERCEPT"))
173 lex_match (lexer, T_EQUALS);
174 while (lex_token (lexer) != T_ENDCMD
175 && lex_token (lexer) != T_SLASH)
177 if (lex_match_id (lexer, "INCLUDE"))
179 glm.intercept = true;
181 else if (lex_match_id (lexer, "EXCLUDE"))
183 glm.intercept = false;
187 lex_error (lexer, NULL);
192 else if (lex_match_id (lexer, "CRITERIA"))
194 lex_match (lexer, T_EQUALS);
195 if (lex_match_id (lexer, "ALPHA"))
197 if (lex_force_match (lexer, T_LPAREN))
199 if (! lex_force_num (lexer))
201 lex_error (lexer, NULL);
205 glm.alpha = lex_number (lexer);
207 if ( ! lex_force_match (lexer, T_RPAREN))
209 lex_error (lexer, NULL);
216 lex_error (lexer, NULL);
220 else if (lex_match_id (lexer, "METHOD"))
222 lex_match (lexer, T_EQUALS);
223 if ( !lex_force_match_id (lexer, "SSTYPE"))
225 lex_error (lexer, NULL);
229 if ( ! lex_force_match (lexer, T_LPAREN))
231 lex_error (lexer, NULL);
235 if ( ! lex_force_int (lexer))
237 lex_error (lexer, NULL);
241 if (3 != lex_integer (lexer))
243 msg (ME, _("Only type 3 sum of squares are currently implemented"));
249 if ( ! lex_force_match (lexer, T_RPAREN))
251 lex_error (lexer, NULL);
255 else if (lex_match_id (lexer, "DESIGN"))
257 lex_match (lexer, T_EQUALS);
259 if (! parse_design_spec (lexer, &glm))
262 if ( glm.n_interactions == 0)
264 msg (ME, _("One or more design variables must be given"));
272 lex_error (lexer, NULL);
279 lex_error (lexer, _("/DESIGN is mandatory in GLM"));
284 struct casegrouper *grouper;
285 struct casereader *group;
288 grouper = casegrouper_create_splits (proc_open (ds), glm.dict);
289 while (casegrouper_get_next_group (grouper, &group))
290 run_glm (&glm, group, ds);
291 ok = casegrouper_destroy (grouper);
292 ok = proc_commit (ds) && ok;
295 const_var_set_destroy (factors);
296 free (glm.factor_vars);
297 for (i = 0 ; i < glm.n_interactions; ++i)
298 interaction_destroy (glm.interactions[i]);
299 free (glm.interactions);
301 free (glm.design_vars);
308 const_var_set_destroy (factors);
309 free (glm.factor_vars);
310 for (i = 0 ; i < glm.n_interactions; ++i)
311 interaction_destroy (glm.interactions[i]);
313 free (glm.interactions);
315 free (glm.design_vars);
320 static void get_ssq (struct covariance *, gsl_vector *,
321 const struct glm_spec *);
324 not_dropped (size_t j, size_t * dropped, size_t n_dropped)
328 for (i = 0; i < n_dropped; i++)
337 get_ssq (struct covariance *cov, gsl_vector * ssq, const struct glm_spec *cmd)
339 const struct variable **vars;
340 gsl_matrix *small_cov = NULL;
341 gsl_matrix *cm = covariance_calculate_unnormalized (cov);
350 dropped = xcalloc (covariance_dim (cov), sizeof (*dropped));
351 vars = xcalloc (covariance_dim (cov), sizeof (*vars));
352 covariance_get_var_indices (cov, vars);
354 for (k = 0; k < cmd->n_design_vars; k++)
357 for (i = 1; i < covariance_dim (cov); i++)
359 if (vars[i] == cmd->design_vars[k])
361 dropped[n_dropped++] = i;
365 gsl_matrix_alloc (cm->size1 - n_dropped, cm->size2 - n_dropped);
366 gsl_matrix_set (small_cov, 0, 0, gsl_matrix_get (cm, 0, 0));
369 for (i = 0; i < cm->size1; i++)
371 if (not_dropped (i, dropped, n_dropped))
374 for (j = 0; j < cm->size2; j++)
376 if (not_dropped (j, dropped, n_dropped))
378 gsl_matrix_set (small_cov, n, m,
379 gsl_matrix_get (cm, i, j));
386 reg_sweep (small_cov, 0);
387 gsl_vector_set (ssq, k + 1,
388 gsl_matrix_get (small_cov, 0, 0)
389 - gsl_vector_get (ssq, 0));
390 gsl_matrix_free (small_cov);
395 gsl_matrix_free (cm);
398 //static void dump_matrix (const gsl_matrix *m);
401 run_glm (struct glm_spec *cmd, struct casereader *input,
402 const struct dataset *ds)
404 bool warn_bad_weight = true;
407 struct dictionary *dict = dataset_dict (ds);
408 struct casereader *reader;
411 struct glm_workspace ws;
412 struct covariance *cov;
414 ws.cats = categoricals_create (cmd->interactions, cmd->n_interactions,
415 cmd->wv, cmd->exclude,
416 NULL, NULL, NULL, NULL);
418 cov = covariance_2pass_create (cmd->n_dep_vars, cmd->dep_vars,
419 ws.cats, cmd->wv, cmd->exclude);
422 c = casereader_peek (input, 0);
425 casereader_destroy (input);
428 output_split_file_values (ds, c);
431 taint = taint_clone (casereader_get_taint (input));
433 ws.totals = moments_create (MOMENT_VARIANCE);
435 for (reader = casereader_clone (input);
436 (c = casereader_read (reader)) != NULL; case_unref (c))
438 double weight = dict_get_case_weight (dict, c, &warn_bad_weight);
440 for (v = 0; v < cmd->n_dep_vars; ++v)
441 moments_pass_one (ws.totals, case_data (c, cmd->dep_vars[v])->f,
444 covariance_accumulate_pass1 (cov, c);
446 casereader_destroy (reader);
448 categoricals_done (ws.cats);
451 (c = casereader_read (reader)) != NULL; case_unref (c))
453 double weight = dict_get_case_weight (dict, c, &warn_bad_weight);
455 for (v = 0; v < cmd->n_dep_vars; ++v)
456 moments_pass_two (ws.totals, case_data (c, cmd->dep_vars[v])->f,
459 covariance_accumulate_pass2 (cov, c);
461 casereader_destroy (reader);
464 gsl_matrix *cm = covariance_calculate_unnormalized (cov);
468 ws.total_ssq = gsl_matrix_get (cm, 0, 0);
473 Store the overall SSE.
475 ws.ssq = gsl_vector_alloc (cm->size1);
476 gsl_vector_set (ws.ssq, 0, gsl_matrix_get (cm, 0, 0));
477 get_ssq (cov, ws.ssq, cmd);
480 gsl_matrix_free (cm);
483 if (!taint_has_tainted_successor (taint))
484 output_glm (cmd, &ws);
486 gsl_vector_free (ws.ssq);
488 covariance_destroy (cov);
489 moments_destroy (ws.totals);
491 taint_destroy (taint);
495 output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws)
497 const struct fmt_spec *wfmt =
498 cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0;
500 double n_total, mean;
501 double df_corr = 0.0;
506 const int heading_columns = 1;
507 const int heading_rows = 1;
511 int nr = heading_rows + 4 + cmd->n_interactions;
515 t = tab_create (nc, nr);
516 tab_title (t, _("Tests of Between-Subjects Effects"));
518 tab_headers (t, heading_columns, 0, heading_rows, 0);
520 tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, nc - 1, nr - 1);
522 tab_hline (t, TAL_2, 0, nc - 1, heading_rows);
523 tab_vline (t, TAL_2, heading_columns, 0, nr - 1);
525 tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Source"));
527 /* TRANSLATORS: The parameter is a roman numeral */
528 tab_text_format (t, 1, 0, TAB_CENTER | TAT_TITLE,
529 _("Type %s Sum of Squares"), "III");
530 tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("df"));
531 tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Mean Square"));
532 tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("F"));
533 tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("Sig."));
535 moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL);
540 for (f = 0; f < cmd->n_interactions; ++f)
541 df_corr += categoricals_n_count (ws->cats, f) - 1.0;
543 mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr);
546 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model"));
552 const double intercept = pow2 (mean * n_total) / n_total;
553 const double df = 1.0;
554 const double F = intercept / df / mse;
555 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept"));
556 tab_double (t, 1, r, 0, intercept, NULL);
557 tab_double (t, 2, r, 0, 1.00, wfmt);
558 tab_double (t, 3, r, 0, intercept / df, NULL);
559 tab_double (t, 4, r, 0, F, NULL);
560 tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
565 for (f = 0; f < cmd->n_interactions; ++f)
567 struct string str = DS_EMPTY_INITIALIZER;
568 const double df = categoricals_n_count (ws->cats, f) - 1.0;
569 const double ssq = gsl_vector_get (ws->ssq, f + 1);
570 const double F = ssq / df / mse;
571 interaction_to_string (cmd->interactions[f], &str);
572 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, ds_cstr (&str));
575 tab_double (t, 1, r, 0, ssq, NULL);
576 tab_double (t, 2, r, 0, df, wfmt);
577 tab_double (t, 3, r, 0, ssq / df, NULL);
578 tab_double (t, 4, r, 0, F, NULL);
580 tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
586 /* Corrected Model */
587 const double df = df_corr - 1.0;
588 const double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0);
589 const double F = ssq / df / mse;
590 tab_double (t, 1, heading_rows, 0, ssq, NULL);
591 tab_double (t, 2, heading_rows, 0, df, wfmt);
592 tab_double (t, 3, heading_rows, 0, ssq / df, NULL);
593 tab_double (t, 4, heading_rows, 0, F, NULL);
595 tab_double (t, 5, heading_rows, 0,
596 gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL);
600 const double df = n_total - df_corr;
601 const double ssq = gsl_vector_get (ws->ssq, 0);
602 const double mse = ssq / df;
603 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Error"));
604 tab_double (t, 1, r, 0, ssq, NULL);
605 tab_double (t, 2, r, 0, df, wfmt);
606 tab_double (t, 3, r++, 0, mse, NULL);
611 const double intercept = pow2 (mean * n_total) / n_total;
612 const double ssq = intercept + ws->total_ssq;
614 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total"));
615 tab_double (t, 1, r, 0, ssq, NULL);
616 tab_double (t, 2, r, 0, n_total, wfmt);
621 tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total"));
624 tab_double (t, 1, r, 0, ws->total_ssq, NULL);
625 tab_double (t, 2, r, 0, n_total - 1.0, wfmt);
632 dump_matrix (const gsl_matrix * m)
635 for (i = 0; i < m->size1; ++i)
637 for (j = 0; j < m->size2; ++j)
639 double x = gsl_matrix_get (m, i, j);
652 If the match succeeds, the variable will be placed in VAR.
653 Returns true if successful */
655 lex_match_variable (struct lexer *lexer, const struct glm_spec *glm, const struct variable **var)
657 if (lex_token (lexer) != T_ID)
660 *var = parse_variable_const (lexer, glm->dict);
667 /* An interaction is a variable followed by {*, BY} followed by an interaction */
669 parse_design_interaction (struct lexer *lexer, struct glm_spec *glm, struct interaction **iact)
671 const struct variable *v = NULL;
674 switch (lex_next_token (lexer, 1))
688 if (! lex_match_variable (lexer, glm, &v))
690 interaction_destroy (*iact);
698 *iact = interaction_create (v);
700 interaction_add_variable (*iact, v);
702 if ( lex_match (lexer, T_ASTERISK) || lex_match (lexer, T_BY))
704 lex_error (lexer, "Interactions are not yet implemented"); return false;
705 return parse_design_interaction (lexer, glm, iact);
708 glm->n_design_vars++;
709 glm->design_vars = xrealloc (glm->design_vars, sizeof (*glm->design_vars) * glm->n_design_vars);
710 glm->design_vars[glm->n_design_vars - 1] = v;
716 parse_nested_variable (struct lexer *lexer, struct glm_spec *glm)
718 const struct variable *v = NULL;
719 if ( ! lex_match_variable (lexer, glm, &v))
722 if (lex_match (lexer, T_LPAREN))
724 if ( ! parse_nested_variable (lexer, glm))
727 if ( ! lex_force_match (lexer, T_RPAREN))
731 lex_error (lexer, "Nested variables are not yet implemented"); return false;
735 /* A design term is an interaction OR a nested variable */
737 parse_design_term (struct lexer *lexer, struct glm_spec *glm)
739 struct interaction *iact = NULL;
740 if (parse_design_interaction (lexer, glm, &iact))
742 /* Interaction parsing successful. Add to list of interactions */
743 glm->interactions = xrealloc (glm->interactions, sizeof *glm->interactions * ++glm->n_interactions);
744 glm->interactions[glm->n_interactions - 1] = iact;
748 if ( parse_nested_variable (lexer, glm))
756 /* Parse a complete DESIGN specification.
757 A design spec is a design term, optionally followed by a comma,
758 and another design spec.
761 parse_design_spec (struct lexer *lexer, struct glm_spec *glm)
763 if (lex_token (lexer) == T_ENDCMD || lex_token (lexer) == T_SLASH)
766 if ( ! parse_design_term (lexer, glm))
769 lex_match (lexer, T_COMMA);
771 return parse_design_spec (lexer, glm);