X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fglm.c;h=5df61c3e0e7b2721f010a1b597e56441a452e0bf;hb=6e5736d55e61f5dca727428213a835998eeacead;hp=b54934783beb12023886adeccfb3eb19f1c4a0a8;hpb=b2be34cda6c4adf4cf396bf1805059eb6ed13646;p=pspp-builds.git diff --git a/src/language/stats/glm.c b/src/language/stats/glm.c index b5493478..5df61c3e 100644 --- a/src/language/stats/glm.c +++ b/src/language/stats/glm.c @@ -18,6 +18,7 @@ #include #include +#include #include #include "data/case.h" @@ -32,6 +33,7 @@ #include "language/lexer/lexer.h" #include "language/lexer/value-parser.h" #include "language/lexer/variable-parser.h" +#include "libpspp/assertion.h" #include "libpspp/ll.h" #include "libpspp/message.h" #include "libpspp/misc.h" @@ -64,9 +66,12 @@ struct glm_spec const struct dictionary *dict; + int ss_type; bool intercept; double alpha; + + bool dump_coding; }; struct glm_workspace @@ -84,6 +89,37 @@ struct glm_workspace gsl_vector *ssq; }; + +/* Default design: all possible interactions */ +static void +design_full (struct glm_spec *glm) +{ + int sz; + int i = 0; + glm->n_interactions = (1 << glm->n_factor_vars) - 1; + + glm->interactions = xcalloc (glm->n_interactions, sizeof *glm->interactions); + + /* All subsets, with exception of the empty set, of [0, glm->n_factor_vars) */ + for (sz = 1; sz <= glm->n_factor_vars; ++sz) + { + gsl_combination *c = gsl_combination_calloc (glm->n_factor_vars, sz); + + do + { + struct interaction *iact = interaction_create (NULL); + int e; + for (e = 0 ; e < gsl_combination_k (c); ++e) + interaction_add_variable (iact, glm->factor_vars [gsl_combination_get (c, e)]); + + glm->interactions[i++] = iact; + } + while (gsl_combination_next (c) == GSL_SUCCESS); + + gsl_combination_free (c); + } +} + static void output_glm (const struct glm_spec *, const struct glm_workspace *ws); static void run_glm (struct glm_spec *cmd, struct casereader *input, @@ -111,6 +147,8 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) glm.intercept = true; glm.wv = dict_get_weight (glm.dict); glm.alpha = 0.05; + glm.dump_coding = false; + glm.ss_type = 3; if (!parse_variables_const (lexer, glm.dict, &glm.dep_vars, &glm.n_dep_vars, @@ -228,9 +266,10 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) goto error; } - if (3 != lex_integer (lexer)) + glm.ss_type = lex_integer (lexer); + if (1 > glm.ss_type && 3 < glm.ss_type ) { - msg (ME, _("Only type 3 sum of squares are currently implemented")); + msg (ME, _("Only types 1, 2 & 3 sums of squares are currently implemented")); goto error; } @@ -248,14 +287,16 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) if (! parse_design_spec (lexer, &glm)) goto error; - - if ( glm.n_interactions == 0) - { - msg (ME, _("One or more design variables must be given")); - goto error; - } - - design = true; + + if (glm.n_interactions > 0) + design = true; + } + else if (lex_match_id (lexer, "SHOWCODES")) + /* Undocumented debug option */ + { + lex_match (lexer, T_EQUALS); + + glm.dump_coding = true; } else { @@ -266,8 +307,7 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) if ( ! design ) { - lex_error (lexer, _("/DESIGN is mandatory in GLM")); - goto error; + design_full (&glm); } { @@ -305,78 +345,234 @@ error: return CMD_FAILURE; } -static void get_ssq (struct covariance *, gsl_vector *, - const struct glm_spec *); +static inline bool +not_dropped (size_t j, const bool *ff) +{ + return ! ff[j]; +} -static bool -not_dropped (size_t j, const size_t *dropped, size_t n_dropped) +static void +fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f) { size_t i; - - for (i = 0; i < n_dropped; i++) + size_t j; + size_t n = 0; + size_t m = 0; + + for (i = 0; i < cov->size1; i++) { - if (j == dropped[i]) - return false; + if (not_dropped (i, dropped_f)) + { + m = 0; + for (j = 0; j < cov->size2; j++) + { + if (not_dropped (j, dropped_f)) + { + gsl_matrix_set (submatrix, n, m, + gsl_matrix_get (cov, i, j)); + m++; + } + } + n++; + } } - return true; } + +/* + Type 1 sums of squares. + Populate SSQ with the Type 1 sums of squares according to COV + */ static void -get_ssq (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) +ssq_type1 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) { gsl_matrix *cm = covariance_calculate_unnormalized (cov); size_t i; - size_t j; size_t k; - size_t *dropped = xcalloc (covariance_dim (cov), sizeof (*dropped)); + bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped)); + bool *submodel_dropped = xcalloc (covariance_dim (cov), sizeof (*submodel_dropped)); const struct categoricals *cats = covariance_get_categoricals (cov); + size_t n_dropped_model = 0; + size_t n_dropped_submodel = 0; + + for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) + { + n_dropped_model++; + n_dropped_submodel++; + model_dropped[i] = true; + submodel_dropped[i] = true; + } + for (k = 0; k < cmd->n_interactions; k++) { - size_t n = 0; - size_t m = 0; - gsl_matrix *small_cov = NULL; - size_t n_dropped = 0; + gsl_matrix *model_cov = NULL; + gsl_matrix *submodel_cov = NULL; + + n_dropped_submodel = n_dropped_model; + for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) + { + submodel_dropped[i] = model_dropped[i]; + } + for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) { - if (categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars) - == cmd->interactions[k]) + const struct interaction * x = + categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars); + + if ( x == cmd->interactions [k]) { - assert (n_dropped < covariance_dim (cov)); - dropped[n_dropped++] = i; + model_dropped[i] = false; + n_dropped_model--; } } - small_cov = - gsl_matrix_alloc (cm->size1 - n_dropped, cm->size2 - n_dropped); - gsl_matrix_set (small_cov, 0, 0, gsl_matrix_get (cm, 0, 0)); - for (i = 0; i < cm->size1; i++) + + model_cov = gsl_matrix_alloc (cm->size1 - n_dropped_model, cm->size2 - n_dropped_model); + submodel_cov = gsl_matrix_alloc (cm->size1 - n_dropped_submodel, cm->size2 - n_dropped_submodel); + + fill_submatrix (cm, model_cov, model_dropped); + fill_submatrix (cm, submodel_cov, submodel_dropped); + + reg_sweep (model_cov, 0); + reg_sweep (submodel_cov, 0); + + gsl_vector_set (ssq, k + 1, + gsl_matrix_get (submodel_cov, 0, 0) - gsl_matrix_get (model_cov, 0, 0) + ); + + gsl_matrix_free (model_cov); + gsl_matrix_free (submodel_cov); + } + + free (model_dropped); + free (submodel_dropped); + gsl_matrix_free (cm); +} + +/* + Type 2 sums of squares. + Populate SSQ with the Type 2 sums of squares according to COV + */ +static void +ssq_type2 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) +{ + gsl_matrix *cm = covariance_calculate_unnormalized (cov); + size_t i; + size_t k; + bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped)); + bool *submodel_dropped = xcalloc (covariance_dim (cov), sizeof (*submodel_dropped)); + const struct categoricals *cats = covariance_get_categoricals (cov); + + for (k = 0; k < cmd->n_interactions; k++) + { + gsl_matrix *model_cov = NULL; + gsl_matrix *submodel_cov = NULL; + size_t n_dropped_model = 0; + size_t n_dropped_submodel = 0; + for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) { - if (not_dropped (i, dropped, n_dropped)) + const struct interaction * x = + categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars); + + model_dropped[i] = false; + submodel_dropped[i] = false; + if (interaction_is_subset (cmd->interactions [k], x)) { - m = 0; - for (j = 0; j < cm->size2; j++) + assert (n_dropped_submodel < covariance_dim (cov)); + n_dropped_submodel++; + submodel_dropped[i] = true; + + if ( cmd->interactions [k]->n_vars < x->n_vars) { - if (not_dropped (j, dropped, n_dropped)) - { - gsl_matrix_set (small_cov, n, m, - gsl_matrix_get (cm, i, j)); - m++; - } + assert (n_dropped_model < covariance_dim (cov)); + n_dropped_model++; + model_dropped[i] = true; } - n++; } } - reg_sweep (small_cov, 0); + + model_cov = gsl_matrix_alloc (cm->size1 - n_dropped_model, cm->size2 - n_dropped_model); + submodel_cov = gsl_matrix_alloc (cm->size1 - n_dropped_submodel, cm->size2 - n_dropped_submodel); + + fill_submatrix (cm, model_cov, model_dropped); + fill_submatrix (cm, submodel_cov, submodel_dropped); + + reg_sweep (model_cov, 0); + reg_sweep (submodel_cov, 0); + + gsl_vector_set (ssq, k + 1, + gsl_matrix_get (submodel_cov, 0, 0) - gsl_matrix_get (model_cov, 0, 0) + ); + + gsl_matrix_free (model_cov); + gsl_matrix_free (submodel_cov); + } + + free (model_dropped); + free (submodel_dropped); + gsl_matrix_free (cm); +} + +/* + Type 3 sums of squares. + Populate SSQ with the Type 2 sums of squares according to COV + */ +static void +ssq_type3 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) +{ + gsl_matrix *cm = covariance_calculate_unnormalized (cov); + size_t i; + size_t k; + bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped)); + bool *submodel_dropped = xcalloc (covariance_dim (cov), sizeof (*submodel_dropped)); + const struct categoricals *cats = covariance_get_categoricals (cov); + + double ss0; + gsl_matrix *submodel_cov = gsl_matrix_alloc (cm->size1, cm->size2); + fill_submatrix (cm, submodel_cov, submodel_dropped); + reg_sweep (submodel_cov, 0); + ss0 = gsl_matrix_get (submodel_cov, 0, 0); + gsl_matrix_free (submodel_cov); + free (submodel_dropped); + + for (k = 0; k < cmd->n_interactions; k++) + { + gsl_matrix *model_cov = NULL; + size_t n_dropped_model = 0; + + for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) + { + const struct interaction * x = + categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars); + + model_dropped[i] = false; + + if ( cmd->interactions [k] == x) + { + assert (n_dropped_model < covariance_dim (cov)); + n_dropped_model++; + model_dropped[i] = true; + } + } + + model_cov = gsl_matrix_alloc (cm->size1 - n_dropped_model, cm->size2 - n_dropped_model); + + fill_submatrix (cm, model_cov, model_dropped); + + reg_sweep (model_cov, 0); + gsl_vector_set (ssq, k + 1, - gsl_matrix_get (small_cov, 0, 0) - - gsl_vector_get (ssq, 0)); - gsl_matrix_free (small_cov); + gsl_matrix_get (model_cov, 0, 0) - ss0); + + gsl_matrix_free (model_cov); } + free (model_dropped); - free (dropped); gsl_matrix_free (cm); } + + //static void dump_matrix (const gsl_matrix *m); static void @@ -427,7 +623,12 @@ run_glm (struct glm_spec *cmd, struct casereader *input, } casereader_destroy (reader); - for (reader = input; + if (cmd->dump_coding) + reader = casereader_clone (input); + else + reader = input; + + for (; (c = casereader_read (reader)) != NULL; case_unref (c)) { double weight = dict_get_case_weight (dict, c, &warn_bad_weight); @@ -440,6 +641,21 @@ run_glm (struct glm_spec *cmd, struct casereader *input, } casereader_destroy (reader); + + if (cmd->dump_coding) + { + struct tab_table *t = + covariance_dump_enc_header (cov, + 1 + casereader_count_cases (input)); + for (reader = input; + (c = casereader_read (reader)) != NULL; case_unref (c)) + { + covariance_dump_enc (cov, c, t); + } + casereader_destroy (reader); + tab_submit (t); + } + { gsl_matrix *cm = covariance_calculate_unnormalized (cov); @@ -454,7 +670,21 @@ run_glm (struct glm_spec *cmd, struct casereader *input, */ ws.ssq = gsl_vector_alloc (cm->size1); gsl_vector_set (ws.ssq, 0, gsl_matrix_get (cm, 0, 0)); - get_ssq (cov, ws.ssq, cmd); + switch (cmd->ss_type) + { + case 1: + ssq_type1 (cov, ws.ssq, cmd); + break; + case 2: + ssq_type2 (cov, ws.ssq, cmd); + break; + case 3: + ssq_type3 (cov, ws.ssq, cmd); + break; + default: + NOT_REACHED (); + break; + } // dump_matrix (cm); gsl_matrix_free (cm); @@ -471,14 +701,25 @@ run_glm (struct glm_spec *cmd, struct casereader *input, taint_destroy (taint); } +static const char *roman[] = + { + "", /* The Romans had no concept of zero */ + "I", + "II", + "III", + "IV" + }; + static void output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) { const struct fmt_spec *wfmt = cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0; + double intercept_ssq; + double ssq_effects; double n_total, mean; - double df_corr = 0.0; + double df_corr = 1.0; double mse = 0; int f; @@ -488,10 +729,11 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) struct tab_table *t; const int nc = 6; - int nr = heading_rows + 4 + cmd->n_interactions; + int nr = heading_rows + 3 + cmd->n_interactions; if (cmd->intercept) - nr++; + nr += 2; + msg (MW, "GLM is experimental. Do not rely on these results."); t = tab_create (nc, nr); tab_title (t, _("Tests of Between-Subjects Effects")); @@ -506,7 +748,8 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) /* TRANSLATORS: The parameter is a roman numeral */ tab_text_format (t, 1, 0, TAB_CENTER | TAT_TITLE, - _("Type %s Sum of Squares"), "III"); + _("Type %s Sum of Squares"), + roman[cmd->ss_type]); tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("df")); tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Mean Square")); tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("F")); @@ -514,27 +757,29 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL); - if (cmd->intercept) - df_corr += 1.0; - df_corr += categoricals_df_total (ws->cats); - mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr); - r = heading_rows; - tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model")); + if (cmd->intercept) + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model")); + else + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Model")); r++; + mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr); + + intercept_ssq = pow2 (mean * n_total) / n_total; + + ssq_effects = 0.0; if (cmd->intercept) { - const double intercept = pow2 (mean * n_total) / n_total; const double df = 1.0; - const double F = intercept / df / mse; + const double F = intercept_ssq / df / mse; tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept")); - tab_double (t, 1, r, 0, intercept, NULL); + tab_double (t, 1, r, 0, intercept_ssq, NULL); tab_double (t, 2, r, 0, 1.00, wfmt); - tab_double (t, 3, r, 0, intercept / df, NULL); + tab_double (t, 3, r, 0, intercept_ssq / df, NULL); tab_double (t, 4, r, 0, F, NULL); tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL); @@ -544,9 +789,20 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) for (f = 0; f < cmd->n_interactions; ++f) { struct string str = DS_EMPTY_INITIALIZER; - const double df = categoricals_df (ws->cats, f); - const double ssq = gsl_vector_get (ws->ssq, f + 1); - const double F = ssq / df / mse; + double df = categoricals_df (ws->cats, f); + + double ssq = gsl_vector_get (ws->ssq, f + 1); + double F; + + ssq_effects += ssq; + + if (! cmd->intercept) + { + df++; + ssq += intercept_ssq; + } + + F = ssq / df / mse; interaction_to_string (cmd->interactions[f], &str); tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, ds_cstr (&str)); ds_destroy (&str); @@ -562,10 +818,17 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) } { - /* Corrected Model */ - const double df = df_corr - 1.0; - const double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0); - const double F = ssq / df / mse; + /* Model / Corrected Model */ + double df = df_corr; + double ssq = ws->total_ssq - gsl_vector_get (ws->ssq, 0); + double F; + + if ( cmd->intercept ) + df --; + else + ssq += intercept_ssq; + + F = ssq / df / mse; tab_double (t, 1, heading_rows, 0, ssq, NULL); tab_double (t, 2, heading_rows, 0, df, wfmt); tab_double (t, 3, heading_rows, 0, ssq / df, NULL); @@ -585,24 +848,21 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) tab_double (t, 3, r++, 0, mse, NULL); } + { + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total")); + tab_double (t, 1, r, 0, ws->total_ssq + intercept_ssq, NULL); + tab_double (t, 2, r, 0, n_total, wfmt); + + r++; + } + if (cmd->intercept) { - const double intercept = pow2 (mean * n_total) / n_total; - const double ssq = intercept + ws->total_ssq; - - tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Total")); - tab_double (t, 1, r, 0, ssq, NULL); - tab_double (t, 2, r, 0, n_total, wfmt); - - r++; + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total")); + tab_double (t, 1, r, 0, ws->total_ssq, NULL); + tab_double (t, 2, r, 0, n_total - 1.0, wfmt); } - tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Total")); - - - tab_double (t, 1, r, 0, ws->total_ssq, NULL); - tab_double (t, 2, r, 0, n_total - 1.0, wfmt); - tab_submit (t); } @@ -680,7 +940,6 @@ parse_design_interaction (struct lexer *lexer, struct glm_spec *glm, struct inte if ( lex_match (lexer, T_ASTERISK) || lex_match (lexer, T_BY)) { - // lex_error (lexer, "Interactions are not yet implemented"); return false; return parse_design_interaction (lexer, glm, iact); }