X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fglm.c;h=2536c5f52de069c9e326f3891e18600d61394c1e;hb=7235f7f42b61c2b111174c3ee5ca72aac8815cd5;hp=3194fd8178df9f30dcd944e49e92c5aab0754e89;hpb=a3dd22c0ed3bcf0d3cf5cf3e009c215eeb35b567;p=pspp diff --git a/src/language/stats/glm.c b/src/language/stats/glm.c index 3194fd8178..2536c5f52d 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" @@ -84,6 +85,37 @@ 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, @@ -248,14 +280,9 @@ 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 { @@ -266,8 +293,7 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) if ( ! design ) { - lex_error (lexer, _("/DESIGN is mandatory in GLM")); - goto error; + design_full (&glm); } { @@ -308,72 +334,97 @@ error: static void get_ssq (struct covariance *, gsl_vector *, const struct glm_spec *); -static bool -not_dropped (size_t j, size_t * dropped, size_t n_dropped) +static inline bool +not_dropped (size_t j, const bool *ff) { - size_t i; + return ! ff[j]; +} - for (i = 0; i < n_dropped; i++) +static void +fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f) +{ + size_t 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; } - + static void -get_ssq (struct covariance *cov, gsl_vector * ssq, const struct glm_spec *cmd) +get_ssq (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); 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; + size_t n_dropped_model = 0; + size_t n_dropped_submodel = 0; 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]) - { - assert (n_dropped < covariance_dim (cov)); - dropped[n_dropped++] = i; - } - } - 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++) - { - 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); + gsl_matrix_set (model_cov, 0, 0, gsl_matrix_get (cm, 0, 0)); + submodel_cov = gsl_matrix_calloc (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 (small_cov, 0, 0) - - gsl_vector_get (ssq, 0)); - gsl_matrix_free (small_cov); + 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 (dropped); + free (model_dropped); + free (submodel_dropped); gsl_matrix_free (cm); } @@ -427,8 +478,6 @@ run_glm (struct glm_spec *cmd, struct casereader *input, } casereader_destroy (reader); - categoricals_done (ws.cats); - for (reader = input; (c = casereader_read (reader)) != NULL; case_unref (c)) { @@ -494,6 +543,7 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) if (cmd->intercept) nr++; + msg (MW, "GLM is experimental. Do not rely on these results."); t = tab_create (nc, nr); tab_title (t, _("Tests of Between-Subjects Effects")); @@ -519,8 +569,7 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) if (cmd->intercept) df_corr += 1.0; - for (f = 0; f < cmd->n_interactions; ++f) - df_corr += categoricals_n_count (ws->cats, f) - 1.0; + df_corr += categoricals_df_total (ws->cats); mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr); @@ -547,7 +596,7 @@ 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_n_count (ws->cats, f) - 1.0; + const double df = categoricals_df (ws->cats, f); const double ssq = gsl_vector_get (ws->ssq, f + 1); const double F = ssq / df / mse; interaction_to_string (cmd->interactions[f], &str); @@ -683,7 +732,6 @@ 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); }