X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fglm.c;h=6401defdeac8774daaf43d906819fa0d27e1650c;hb=6f1fc2655c5e4aea0b39e5d309a1cc4826f16d4e;hp=fa418583200781af8fb186b4a160f4dba5602049;hpb=df4eeb7d759173fa12eddd067d799e9db9b6bb48;p=pspp diff --git a/src/language/stats/glm.c b/src/language/stats/glm.c index fa41858320..6401defdea 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" @@ -39,6 +40,7 @@ #include "linreg/sweep.h" #include "math/categoricals.h" #include "math/covariance.h" +#include "math/interaction.h" #include "math/moments.h" #include "output/tab.h" @@ -53,13 +55,8 @@ struct glm_spec size_t n_factor_vars; const struct variable **factor_vars; - /* In the current implementation, design_vars will - normally be the same as factor_vars. - This will change once interactions, nested variables - and repeated measures become involved. - */ - size_t n_design_vars; - const struct variable **design_vars; + size_t n_interactions; + struct interaction **interactions; enum mv_class exclude; @@ -88,6 +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, @@ -96,20 +124,23 @@ static void run_glm (struct glm_spec *cmd, struct casereader *input, static bool parse_design_spec (struct lexer *lexer, struct glm_spec *glm); +/* Define to 1 if the /DESIGN subcommand should not be optional */ +#define DESIGN_MANDATORY 1 int cmd_glm (struct lexer *lexer, struct dataset *ds) { + int i; struct const_var_set *factors = NULL; struct glm_spec glm; bool design = false; glm.dict = dataset_dict (ds); glm.n_dep_vars = 0; glm.n_factor_vars = 0; - glm.n_design_vars = 0; + glm.n_interactions = 0; + glm.interactions = NULL; glm.dep_vars = NULL; glm.factor_vars = NULL; - glm.design_vars = NULL; glm.exclude = MV_ANY; glm.intercept = true; glm.wv = dict_get_weight (glm.dict); @@ -251,14 +282,19 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) if (! parse_design_spec (lexer, &glm)) goto error; - - if ( glm.n_design_vars == 0) + +#if DESIGN_MANDATORY + if ( glm.n_interactions == 0) { msg (ME, _("One or more design variables must be given")); goto error; } design = true; +#else + if (glm.n_interactions > 0) + design = true; +#endif } else { @@ -269,8 +305,12 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) if ( ! design ) { +#if DESIGN_MANDATORY lex_error (lexer, _("/DESIGN is mandatory in GLM")); goto error; +#endif + + design_full (&glm); } { @@ -287,7 +327,9 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) const_var_set_destroy (factors); free (glm.factor_vars); - free (glm.design_vars); + for (i = 0 ; i < glm.n_interactions; ++i) + interaction_destroy (glm.interactions[i]); + free (glm.interactions); free (glm.dep_vars); @@ -297,7 +339,10 @@ error: const_var_set_destroy (factors); free (glm.factor_vars); - free (glm.design_vars); + for (i = 0 ; i < glm.n_interactions; ++i) + interaction_destroy (glm.interactions[i]); + + free (glm.interactions); free (glm.dep_vars); return CMD_FAILURE; @@ -307,7 +352,7 @@ 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) +not_dropped (size_t j, const size_t *dropped, size_t n_dropped) { size_t i; @@ -320,38 +365,33 @@ not_dropped (size_t j, size_t * dropped, size_t n_dropped) } 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) { - const struct variable **vars; - gsl_matrix *small_cov = NULL; gsl_matrix *cm = covariance_calculate_unnormalized (cov); size_t i; size_t j; size_t k; - size_t n; - size_t m; - size_t *dropped; - size_t n_dropped; - - dropped = xcalloc (covariance_dim (cov), sizeof (*dropped)); - vars = xcalloc (covariance_dim (cov), sizeof (*vars)); - covariance_get_var_indices (cov, vars); + size_t *dropped = xcalloc (covariance_dim (cov), sizeof (*dropped)); + const struct categoricals *cats = covariance_get_categoricals (cov); - for (k = 0; k < cmd->n_design_vars; k++) + for (k = 0; k < cmd->n_interactions; k++) { - n_dropped = 0; - for (i = 1; i < covariance_dim (cov); i++) + size_t n = 0; + size_t m = 0; + gsl_matrix *small_cov = NULL; + size_t n_dropped = 0; + for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) { - if (vars[i] == cmd->design_vars[k]) + 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)); - n = 0; - m = 0; for (i = 0; i < cm->size1; i++) { if (not_dropped (i, dropped, n_dropped)) @@ -377,7 +417,6 @@ get_ssq (struct covariance *cov, gsl_vector * ssq, const struct glm_spec *cmd) } free (dropped); - free (vars); gsl_matrix_free (cm); } @@ -396,7 +435,8 @@ run_glm (struct glm_spec *cmd, struct casereader *input, struct glm_workspace ws; struct covariance *cov; - ws.cats = categoricals_create (cmd->design_vars, cmd->n_design_vars, + + ws.cats = categoricals_create (cmd->interactions, cmd->n_interactions, cmd->wv, cmd->exclude, NULL, NULL, NULL, NULL); @@ -430,8 +470,6 @@ 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)) { @@ -455,8 +493,8 @@ run_glm (struct glm_spec *cmd, struct casereader *input, reg_sweep (cm, 0); /* - Store the overall SSE. - */ + Store the overall SSE. + */ 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); @@ -493,7 +531,7 @@ 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_design_vars; + int nr = heading_rows + 4 + cmd->n_interactions; if (cmd->intercept) nr++; @@ -522,8 +560,7 @@ 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_design_vars; ++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); @@ -547,13 +584,15 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) r++; } - for (f = 0; f < cmd->n_design_vars; ++f) + for (f = 0; f < cmd->n_interactions; ++f) { - const double df = categoricals_n_count (ws->cats, f) - 1.0; + 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; - tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, - var_to_string (cmd->design_vars[f])); + interaction_to_string (cmd->interactions[f], &str); + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, ds_cstr (&str)); + ds_destroy (&str); tab_double (t, 1, r, 0, ssq, NULL); tab_double (t, 2, r, 0, df, wfmt); @@ -562,8 +601,6 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL); - - r++; } @@ -651,9 +688,11 @@ lex_match_variable (struct lexer *lexer, const struct glm_spec *glm, const struc /* An interaction is a variable followed by {*, BY} followed by an interaction */ static bool -parse_design_interaction (struct lexer *lexer, struct glm_spec *glm) +parse_design_interaction (struct lexer *lexer, struct glm_spec *glm, struct interaction **iact) { const struct variable *v = NULL; + assert (iact); + switch (lex_next_token (lexer, 1)) { case T_ENDCMD: @@ -669,18 +708,25 @@ parse_design_interaction (struct lexer *lexer, struct glm_spec *glm) } if (! lex_match_variable (lexer, glm, &v)) - return false; + { + interaction_destroy (*iact); + *iact = NULL; + return false; + } + + assert (v); + + if ( *iact == NULL) + *iact = interaction_create (v); + else + interaction_add_variable (*iact, v); 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); + return parse_design_interaction (lexer, glm, iact); } - glm->n_design_vars++; - glm->design_vars = xrealloc (glm->design_vars, sizeof (*glm->design_vars) * glm->n_design_vars); - glm->design_vars[glm->n_design_vars - 1] = v; - return true; } @@ -708,8 +754,14 @@ parse_nested_variable (struct lexer *lexer, struct glm_spec *glm) static bool parse_design_term (struct lexer *lexer, struct glm_spec *glm) { - if (parse_design_interaction (lexer, glm)) - return true; + struct interaction *iact = NULL; + if (parse_design_interaction (lexer, glm, &iact)) + { + /* Interaction parsing successful. Add to list of interactions */ + glm->interactions = xrealloc (glm->interactions, sizeof *glm->interactions * ++glm->n_interactions); + glm->interactions[glm->n_interactions - 1] = iact; + return true; + } if ( parse_nested_variable (lexer, glm)) return true;