X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fglm.c;h=f5feab0e369eaae1541288789b54408bac14b82f;hb=ed38ada34331b3b1e0167c350b375a3fb38099a2;hp=1acf03ffb0a098655fd9a9ea2727a0a57a0cdf32;hpb=4399391aa1efe1cf6cbbcfde90ff52d7f88c9c40;p=pspp diff --git a/src/language/stats/glm.c b/src/language/stats/glm.c index 1acf03ffb0..f5feab0e36 100644 --- a/src/language/stats/glm.c +++ b/src/language/stats/glm.c @@ -1,5 +1,5 @@ /* PSPP - a program for statistical analysis. - Copyright (C) 2010 Free Software Foundation, Inc. + Copyright (C) 2010, 2011 Free Software Foundation, Inc. This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -16,38 +16,31 @@ #include -#include -#include -#include - -#include -#include -#include -#include -#include - -#include - -#include -#include -#include -#include - -#include -#include -#include - -#include -#include -#include - #include +#include #include -#include - -#include -#include +#include "data/case.h" +#include "data/casegrouper.h" +#include "data/casereader.h" +#include "data/dataset.h" +#include "data/dictionary.h" +#include "data/format.h" +#include "data/value.h" +#include "language/command.h" +#include "language/dictionary/split-file.h" +#include "language/lexer/lexer.h" +#include "language/lexer/value-parser.h" +#include "language/lexer/variable-parser.h" +#include "libpspp/ll.h" +#include "libpspp/message.h" +#include "libpspp/misc.h" +#include "libpspp/taint.h" +#include "linreg/sweep.h" +#include "math/categoricals.h" +#include "math/covariance.h" +#include "math/moments.h" +#include "output/tab.h" #include "gettext.h" #define _(msgid) gettext (msgid) @@ -60,67 +53,99 @@ 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; + enum mv_class exclude; /* The weight variable */ const struct variable *wv; + const struct dictionary *dict; + bool intercept; + + double alpha; }; struct glm_workspace { double total_ssq; struct moments *totals; + + struct categoricals *cats; + + /* + Sums of squares due to different variables. Element 0 is the SSE + for the entire model. For i > 0, element i is the SS due to + variable i. + */ + gsl_vector *ssq; }; -static void output_glm (const struct glm_spec *, const struct glm_workspace *ws); -static void run_glm (const struct glm_spec *cmd, struct casereader *input, const struct dataset *ds); +static void output_glm (const struct glm_spec *, + const struct glm_workspace *ws); +static void run_glm (struct glm_spec *cmd, struct casereader *input, + const struct dataset *ds); + + +static bool parse_design_spec (struct lexer *lexer, struct glm_spec *glm); + int cmd_glm (struct lexer *lexer, struct dataset *ds) { - const struct dictionary *dict = dataset_dict (ds); - struct glm_spec glm ; + 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.dep_vars = NULL; glm.factor_vars = NULL; + glm.design_vars = NULL; glm.exclude = MV_ANY; glm.intercept = true; - glm.wv = dict_get_weight (dict); + glm.wv = dict_get_weight (glm.dict); + glm.alpha = 0.05; - - if (!parse_variables_const (lexer, dict, + if (!parse_variables_const (lexer, glm.dict, &glm.dep_vars, &glm.n_dep_vars, PV_NO_DUPLICATE | PV_NUMERIC)) goto error; lex_force_match (lexer, T_BY); - if (!parse_variables_const (lexer, dict, + if (!parse_variables_const (lexer, glm.dict, &glm.factor_vars, &glm.n_factor_vars, PV_NO_DUPLICATE | PV_NUMERIC)) goto error; - if ( glm.n_dep_vars > 1) + if (glm.n_dep_vars > 1) { msg (ME, _("Multivariate analysis is not yet implemented")); return CMD_FAILURE; } - struct const_var_set *factors = const_var_set_create_from_array (glm.factor_vars, glm.n_factor_vars); + factors = + const_var_set_create_from_array (glm.factor_vars, glm.n_factor_vars); - - while (lex_token (lexer) != '.') + while (lex_token (lexer) != T_ENDCMD) { - lex_match (lexer, '/'); + lex_match (lexer, T_SLASH); if (lex_match_id (lexer, "MISSING")) - { - lex_match (lexer, '='); - while (lex_token (lexer) != '.' && lex_token (lexer) != '/') - { + { + lex_match (lexer, T_EQUALS); + while (lex_token (lexer) != T_ENDCMD + && lex_token (lexer) != T_SLASH) + { if (lex_match_id (lexer, "INCLUDE")) { glm.exclude = MV_SYSTEM; @@ -131,16 +156,17 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) } else { - lex_error (lexer, NULL); + lex_error (lexer, NULL); goto error; } } } else if (lex_match_id (lexer, "INTERCEPT")) - { - lex_match (lexer, '='); - while (lex_token (lexer) != '.' && lex_token (lexer) != '/') - { + { + lex_match (lexer, T_EQUALS); + while (lex_token (lexer) != T_ENDCMD + && lex_token (lexer) != T_SLASH) + { if (lex_match_id (lexer, "INCLUDE")) { glm.intercept = true; @@ -151,18 +177,88 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) } else { - lex_error (lexer, NULL); + lex_error (lexer, NULL); goto error; } } } - else if (lex_match_id (lexer, "DESIGN")) - { - size_t n_des; - const struct variable **des; - lex_match (lexer, '='); + else if (lex_match_id (lexer, "CRITERIA")) + { + lex_match (lexer, T_EQUALS); + if (lex_match_id (lexer, "ALPHA")) + { + if (lex_force_match (lexer, T_LPAREN)) + { + if (! lex_force_num (lexer)) + { + lex_error (lexer, NULL); + goto error; + } + + glm.alpha = lex_number (lexer); + lex_get (lexer); + if ( ! lex_force_match (lexer, T_RPAREN)) + { + lex_error (lexer, NULL); + goto error; + } + } + } + else + { + lex_error (lexer, NULL); + goto error; + } + } + else if (lex_match_id (lexer, "METHOD")) + { + lex_match (lexer, T_EQUALS); + if ( !lex_force_match_id (lexer, "SSTYPE")) + { + lex_error (lexer, NULL); + goto error; + } + + if ( ! lex_force_match (lexer, T_LPAREN)) + { + lex_error (lexer, NULL); + goto error; + } + + if ( ! lex_force_int (lexer)) + { + lex_error (lexer, NULL); + goto error; + } - parse_const_var_set_vars (lexer, factors, &des, &n_des, 0); + if (3 != lex_integer (lexer)) + { + msg (ME, _("Only type 3 sum of squares are currently implemented")); + goto error; + } + + lex_get (lexer); + + if ( ! lex_force_match (lexer, T_RPAREN)) + { + lex_error (lexer, NULL); + goto error; + } + } + else if (lex_match_id (lexer, "DESIGN")) + { + lex_match (lexer, T_EQUALS); + + if (! parse_design_spec (lexer, &glm)) + goto error; + + if ( glm.n_design_vars == 0) + { + msg (ME, _("One or more design variables must be given")); + goto error; + } + + design = true; } else { @@ -171,30 +267,127 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) } } + if ( ! design ) + { + lex_error (lexer, _("/DESIGN is mandatory in GLM")); + goto error; + } { struct casegrouper *grouper; struct casereader *group; bool ok; - grouper = casegrouper_create_splits (proc_open (ds), dict); + grouper = casegrouper_create_splits (proc_open (ds), glm.dict); while (casegrouper_get_next_group (grouper, &group)) run_glm (&glm, group, ds); ok = casegrouper_destroy (grouper); ok = proc_commit (ds) && ok; } + const_var_set_destroy (factors); + free (glm.factor_vars); + free (glm.design_vars); + free (glm.dep_vars); + + return CMD_SUCCESS; - error: +error: + + const_var_set_destroy (factors); + free (glm.factor_vars); + free (glm.design_vars); + free (glm.dep_vars); + return CMD_FAILURE; } -static void dump_matrix (const gsl_matrix *m); +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) +{ + size_t i; + + for (i = 0; i < n_dropped; i++) + { + if (j == dropped[i]) + return false; + } + return true; +} static void -run_glm (const struct glm_spec *cmd, struct casereader *input, const struct dataset *ds) +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); + + for (k = 0; k < cmd->n_design_vars; k++) + { + n_dropped = 0; + for (i = 1; i < covariance_dim (cov); i++) + { + if (vars[i] == cmd->design_vars[k]) + { + 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)) + { + m = 0; + for (j = 0; j < cm->size2; j++) + { + if (not_dropped (j, dropped, n_dropped)) + { + gsl_matrix_set (small_cov, n, m, + gsl_matrix_get (cm, i, j)); + m++; + } + } + n++; + } + } + reg_sweep (small_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); + } + + free (dropped); + free (vars); + gsl_matrix_free (cm); +} + +//static void dump_matrix (const gsl_matrix *m); + +static void +run_glm (struct glm_spec *cmd, struct casereader *input, + const struct dataset *ds) +{ + bool warn_bad_weight = true; int v; struct taint *taint; struct dictionary *dict = dataset_dict (ds); @@ -202,15 +395,13 @@ run_glm (const struct glm_spec *cmd, struct casereader *input, const struct data struct ccase *c; struct glm_workspace ws; + struct covariance *cov; + ws.cats = categoricals_create (cmd->design_vars, cmd->n_design_vars, + cmd->wv, cmd->exclude, + NULL, NULL, NULL, NULL); - struct categoricals *cats = categoricals_create (cmd->factor_vars, cmd->n_factor_vars, - cmd->wv, cmd->exclude, - NULL, NULL, - NULL, NULL); - - struct covariance *cov = covariance_2pass_create (cmd->n_dep_vars, cmd->dep_vars, - cats, - cmd->wv, cmd->exclude); + cov = covariance_2pass_create (cmd->n_dep_vars, cmd->dep_vars, + ws.cats, cmd->wv, cmd->exclude); c = casereader_peek (input, 0); @@ -226,28 +417,29 @@ run_glm (const struct glm_spec *cmd, struct casereader *input, const struct data ws.totals = moments_create (MOMENT_VARIANCE); - bool warn_bad_weight = true; for (reader = casereader_clone (input); (c = casereader_read (reader)) != NULL; case_unref (c)) { double weight = dict_get_case_weight (dict, c, &warn_bad_weight); - for ( v = 0; v < cmd->n_dep_vars; ++v) - moments_pass_one (ws.totals, case_data (c, cmd->dep_vars[v])->f, weight); + for (v = 0; v < cmd->n_dep_vars; ++v) + moments_pass_one (ws.totals, case_data (c, cmd->dep_vars[v])->f, + weight); covariance_accumulate_pass1 (cov, c); } casereader_destroy (reader); - categoricals_done (cats); + categoricals_done (ws.cats); - for (reader = casereader_clone (input); + for (reader = input; (c = casereader_read (reader)) != NULL; case_unref (c)) { double weight = dict_get_case_weight (dict, c, &warn_bad_weight); - for ( v = 0; v < cmd->n_dep_vars; ++v) - moments_pass_two (ws.totals, case_data (c, cmd->dep_vars[v])->f, weight); + for (v = 0; v < cmd->n_dep_vars; ++v) + moments_pass_two (ws.totals, case_data (c, cmd->dep_vars[v])->f, + weight); covariance_accumulate_pass2 (cov, c); } @@ -256,34 +448,52 @@ run_glm (const struct glm_spec *cmd, struct casereader *input, const struct data { gsl_matrix *cm = covariance_calculate_unnormalized (cov); - dump_matrix (cm); + // dump_matrix (cm); ws.total_ssq = gsl_matrix_get (cm, 0, 0); reg_sweep (cm, 0); - dump_matrix (cm); + /* + 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); + // dump_matrix (cm); + + gsl_matrix_free (cm); } if (!taint_has_tainted_successor (taint)) output_glm (cmd, &ws); + gsl_vector_free (ws.ssq); + + covariance_destroy (cov); + moments_destroy (ws.totals); + taint_destroy (taint); } 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; + const struct fmt_spec *wfmt = + cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0; + + double n_total, mean; + double df_corr = 0.0; + double mse = 0; int f; int r; const int heading_columns = 1; const int heading_rows = 1; - struct tab_table *t ; + struct tab_table *t; const int nc = 6; - int nr = heading_rows + 4 + cmd->n_factor_vars; + int nr = heading_rows + 4 + cmd->n_design_vars; if (cmd->intercept) nr++; @@ -292,11 +502,7 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) tab_headers (t, heading_columns, 0, heading_rows, 0); - tab_box (t, - TAL_2, TAL_2, - -1, TAL_1, - 0, 0, - nc - 1, nr - 1); + tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, nc - 1, nr - 1); tab_hline (t, TAL_2, 0, nc - 1, heading_rows); tab_vline (t, TAL_2, heading_columns, 0, nr - 1); @@ -304,42 +510,92 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Source")); /* TRANSLATORS: The parameter is a roman numeral */ - tab_text_format (t, 1, 0, TAB_CENTER | TAT_TITLE, _("Type %s Sum of Squares"), "III"); + tab_text_format (t, 1, 0, TAB_CENTER | TAT_TITLE, + _("Type %s Sum of Squares"), "III"); 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")); tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("Sig.")); + moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL); + + 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; + + 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")); + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Corrected Model")); + + r++; - double intercept, n_total; if (cmd->intercept) { - double mean; - moments_calculate (ws->totals, &n_total, &mean, NULL, NULL, NULL); - intercept = pow2 (mean * n_total) / n_total; - + const double intercept = pow2 (mean * n_total) / n_total; + const double df = 1.0; + const double F = intercept / df / mse; tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept")); tab_double (t, 1, r, 0, intercept, NULL); tab_double (t, 2, r, 0, 1.00, wfmt); - - tab_double (t, 3, r, 0, intercept / 1.0 , NULL); + tab_double (t, 3, r, 0, intercept / 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); r++; } - for (f = 0; f < cmd->n_factor_vars; ++f) + for (f = 0; f < cmd->n_design_vars; ++f) { - tab_text (t, 0, r++, TAB_LEFT | TAT_TITLE, - var_to_string (cmd->factor_vars[f])); + const double df = categoricals_n_count (ws->cats, f) - 1.0; + 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])); + + tab_double (t, 1, r, 0, ssq, NULL); + tab_double (t, 2, r, 0, df, wfmt); + tab_double (t, 3, r, 0, 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); + + + r++; } - tab_text (t, 0, r++, TAB_LEFT | TAT_TITLE, _("Error")); + { + /* 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; + 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); + tab_double (t, 4, heading_rows, 0, F, NULL); + + tab_double (t, 5, heading_rows, 0, + gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL); + } + + { + const double df = n_total - df_corr; + const double ssq = gsl_vector_get (ws->ssq, 0); + const double mse = ssq / df; + tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Error")); + tab_double (t, 1, r, 0, ssq, NULL); + tab_double (t, 2, r, 0, df, wfmt); + tab_double (t, 3, r++, 0, mse, NULL); + } if (cmd->intercept) { - double ssq = intercept + ws->total_ssq; - double mse = ssq / n_total; + 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); @@ -349,14 +605,16 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) 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); } -static -void dump_matrix (const gsl_matrix *m) +#if 0 +static void +dump_matrix (const gsl_matrix * m) { size_t i, j; for (i = 0; i < m->size1; ++i) @@ -370,3 +628,83 @@ void dump_matrix (const gsl_matrix *m) } printf ("\n"); } +#endif + + + + +/* Match a variable. + If the match succeeds, the variable will be placed in VAR. + Returns true if successful */ +static bool +lex_match_variable (struct lexer *lexer, const struct glm_spec *glm, const struct variable **var) +{ + if (lex_token (lexer) != T_ID) + return false; + + *var = parse_variable_const (lexer, glm->dict); + + if ( *var == NULL) + return false; + return true; +} + +/* An interaction is a variable followed by {*, BY} followed by an interaction */ +static bool +parse_design_interaction (struct lexer *lexer, struct glm_spec *glm) +{ + const struct variable *v = NULL; + if (! lex_match_variable (lexer, glm, &v)) + return false; + + 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); + } + + 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; +} + +/* A design term is a varible OR an interaction */ +static bool +parse_design_term (struct lexer *lexer, struct glm_spec *glm) +{ + const struct variable *v = NULL; + if (parse_design_interaction (lexer, glm)) + return true; + + /* FIXME: This should accept nexted variables */ + if ( lex_match_variable (lexer, glm, &v)) + { + return true; + } + + return false; +} + + + +/* Parse a complete DESIGN specification. + A design spec is a design term, optionally followed by a comma, + and another design spec. +*/ +static bool +parse_design_spec (struct lexer *lexer, struct glm_spec *glm) +{ + /* Kludge: Return success if end of design spec */ + if (lex_token (lexer) == T_ENDCMD || lex_token (lexer) == T_SLASH) + return true; + + if ( ! parse_design_term (lexer, glm)) + return false; + + lex_match (lexer, T_COMMA); + + return parse_design_spec (lexer, glm); +} +