X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fregression.c;h=72649ba4ac633667d678f188927cbf0cb339dc6e;hb=f790dbda9d498eef9c9c0a49078adbeecf768d56;hp=9379ce70ef5d1591ad3fa855edb49ef54a957ba6;hpb=a016af53bdf60d1ab95e680e6b31311b21518b75;p=pspp diff --git a/src/language/stats/regression.c b/src/language/stats/regression.c index 9379ce70ef..72649ba4ac 100644 --- a/src/language/stats/regression.c +++ b/src/language/stats/regression.c @@ -1,5 +1,6 @@ /* PSPP - a program for statistical analysis. - Copyright (C) 2005, 2009, 2010, 2011, 2012, 2013, 2014, 2016 Free Software Foundation, Inc. + Copyright (C) 2005, 2009, 2010, 2011, 2012, 2013, 2014, + 2016, 2017 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 @@ -19,6 +20,7 @@ #include #include +#include #include #include @@ -51,8 +53,6 @@ #include -#define REG_LARGE_DATA 1000 - #define STATS_R 1 #define STATS_COEFF 2 #define STATS_ANOVA 4 @@ -79,15 +79,17 @@ struct regression bool resid; bool pred; + + bool origin; }; struct regression_workspace { /* The new variables which will be introduced by /SAVE */ - const struct variable **predvars; + const struct variable **predvars; const struct variable **residvars; - /* A reader/writer pair to temporarily hold the + /* A reader/writer pair to temporarily hold the values of the new variables */ struct casewriter *writer; struct casereader *reader; @@ -135,7 +137,7 @@ create_aux_var (struct dataset *ds, const char *prefix) return var; } -/* Auxilliary data for transformation when /SAVE is entered */ +/* Auxiliary data for transformation when /SAVE is entered */ struct save_trans_data { int n_dep_vars; @@ -155,7 +157,7 @@ save_trans_free (void *aux) return true; } -static int +static int save_trans_func (void *aux, struct ccase **c, casenumber x UNUSED) { struct save_trans_data *save_trans_data = aux; @@ -174,7 +176,7 @@ save_trans_func (void *aux, struct ccase **c, casenumber x UNUSED) double pred = case_data_idx (in, ws->extras * k + ws->pred_idx)->f; case_data_rw (*c, ws->predvars[k])->f = pred; } - + if (ws->res_idx != -1) { double resid = case_data_idx (in, ws->extras * k + ws->res_idx)->f; @@ -204,46 +206,74 @@ cmd_regression (struct lexer *lexer, struct dataset *ds) regression.resid = false; regression.ds = ds; + regression.origin = false; - /* Accept an optional, completely pointless "/VARIABLES=" */ - lex_match (lexer, T_SLASH); - if (lex_match_id (lexer, "VARIABLES")) - { - if (!lex_force_match (lexer, T_EQUALS)) - goto error; - } - - if (!parse_variables_const (lexer, dict, - ®ression.vars, ®ression.n_vars, - PV_NO_DUPLICATE | PV_NUMERIC)) - goto error; - - + bool variables_seen = false; + bool method_seen = false; + bool dependent_seen = false; while (lex_token (lexer) != T_ENDCMD) { lex_match (lexer, T_SLASH); - if (lex_match_id (lexer, "DEPENDENT")) + if (lex_match_id (lexer, "VARIABLES")) { + if (method_seen) + { + msg (SE, _("VARIABLES may not appear after %s"), "METHOD"); + goto error; + } + if (dependent_seen) + { + msg (SE, _("VARIABLES may not appear after %s"), "DEPENDENT"); + goto error; + } + variables_seen = true; + lex_match (lexer, T_EQUALS); + + if (!parse_variables_const (lexer, dict, + ®ression.vars, ®ression.n_vars, + PV_NO_DUPLICATE | PV_NUMERIC)) + goto error; + } + else if (lex_match_id (lexer, "DEPENDENT")) + { + dependent_seen = true; lex_match (lexer, T_EQUALS); free (regression.dep_vars); regression.n_dep_vars = 0; - + if (!parse_variables_const (lexer, dict, ®ression.dep_vars, ®ression.n_dep_vars, PV_NO_DUPLICATE | PV_NUMERIC)) goto error; } + else if (lex_match_id (lexer, "ORIGIN")) + { + regression.origin = true; + } + else if (lex_match_id (lexer, "NOORIGIN")) + { + regression.origin = false; + } else if (lex_match_id (lexer, "METHOD")) { + method_seen = true; lex_match (lexer, T_EQUALS); if (!lex_force_match_id (lexer, "ENTER")) { goto error; } + + if (! variables_seen) + { + if (!parse_variables_const (lexer, dict, + ®ression.vars, ®ression.n_vars, + PV_NO_DUPLICATE | PV_NUMERIC)) + goto error; + } } else if (lex_match_id (lexer, "STATISTICS")) { @@ -281,11 +311,13 @@ cmd_regression (struct lexer *lexer, struct dataset *ds) { statistics |= STATS_CI; - if (lex_match (lexer, T_LPAREN)) + if (lex_match (lexer, T_LPAREN) && + lex_force_num (lexer)) { regression.ci = lex_number (lexer) / 100.0; lex_get (lexer); - lex_force_match (lexer, T_RPAREN); + if (! lex_force_match (lexer, T_RPAREN)) + goto error; } } else @@ -337,7 +369,7 @@ cmd_regression (struct lexer *lexer, struct dataset *ds) workspace.extras = 0; workspace.res_idx = -1; workspace.pred_idx = -1; - workspace.writer = NULL; + workspace.writer = NULL; workspace.reader = NULL; workspace.residvars = NULL; workspace.predvars = NULL; @@ -348,8 +380,7 @@ cmd_regression (struct lexer *lexer, struct dataset *ds) if (regression.resid) { - workspace.extras ++; - workspace.res_idx = 0; + workspace.res_idx = workspace.extras ++; workspace.residvars = xcalloc (regression.n_dep_vars, sizeof (*workspace.residvars)); for (i = 0; i < regression.n_dep_vars; ++i) @@ -361,8 +392,7 @@ cmd_regression (struct lexer *lexer, struct dataset *ds) if (regression.pred) { - workspace.extras ++; - workspace.pred_idx = 1; + workspace.pred_idx = workspace.extras ++; workspace.predvars = xcalloc (regression.n_dep_vars, sizeof (*workspace.predvars)); for (i = 0; i < regression.n_dep_vars; ++i) @@ -413,7 +443,7 @@ cmd_regression (struct lexer *lexer, struct dataset *ds) save_trans_data->ws = xmalloc (sizeof (workspace)); memcpy (save_trans_data->ws, &workspace, sizeof (workspace)); save_trans_data->n_dep_vars = regression.n_dep_vars; - + add_transformation (ds, save_trans_func, save_trans_free, save_trans_data); } @@ -521,7 +551,6 @@ identify_indep_vars (const struct regression *cmd, return n_indep_vars; } - static double fill_covariance (gsl_matrix * cov, struct covariance *all_cov, const struct variable **vars, @@ -597,20 +626,20 @@ fill_covariance (gsl_matrix * cov, struct covariance *all_cov, /* STATISTICS subcommand output functions. */ -static void reg_stats_r (const linreg *, const struct variable *); -static void reg_stats_coeff (const linreg *, const gsl_matrix *, const struct variable *, const struct regression *); -static void reg_stats_anova (const linreg *, const struct variable *); -static void reg_stats_bcov (const linreg *, const struct variable *); +static void reg_stats_r (const struct linreg *, const struct variable *); +static void reg_stats_coeff (const struct linreg *, const gsl_matrix *, const struct variable *, const struct regression *); +static void reg_stats_anova (const struct linreg *, const struct variable *); +static void reg_stats_bcov (const struct linreg *, const struct variable *); static void -subcommand_statistics (const struct regression *cmd, const linreg * c, const gsl_matrix * cm, +subcommand_statistics (const struct regression *cmd, const struct linreg * c, const gsl_matrix * cm, const struct variable *var) { - if (cmd->stats & STATS_R) + if (cmd->stats & STATS_R) reg_stats_r (c, var); - if (cmd->stats & STATS_ANOVA) + if (cmd->stats & STATS_ANOVA) reg_stats_anova (c, var); if (cmd->stats & STATS_COEFF) @@ -622,12 +651,12 @@ subcommand_statistics (const struct regression *cmd, const linreg * c, const gsl static void -run_regression (const struct regression *cmd, +run_regression (const struct regression *cmd, struct regression_workspace *ws, struct casereader *input) { size_t i; - linreg **models; + struct linreg **models; int k; struct ccase *c; @@ -641,7 +670,7 @@ run_regression (const struct regression *cmd, fill_all_vars (all_vars, cmd); cov = covariance_1pass_create (n_all_vars, all_vars, dict_get_weight (dataset_dict (cmd->ds)), - MV_ANY); + MV_ANY, cmd->origin == false); reader = casereader_clone (input); reader = casereader_create_filter_missing (reader, all_vars, n_all_vars, @@ -667,21 +696,12 @@ run_regression (const struct regression *cmd, gsl_matrix *this_cm = gsl_matrix_alloc (n_indep + 1, n_indep + 1); double n_data = fill_covariance (this_cm, cov, vars, n_indep, dep_var, all_vars, n_all_vars, means); - models[k] = linreg_alloc (dep_var, vars, n_data, n_indep); - models[k]->depvar = dep_var; + models[k] = linreg_alloc (dep_var, vars, n_data, n_indep, cmd->origin); for (i = 0; i < n_indep; i++) { linreg_set_indep_variable_mean (models[k], i, means[i]); } linreg_set_depvar_mean (models[k], means[i]); - /* - For large data sets, use QR decomposition. - */ - if (n_data > sqrt (n_indep) && n_data > REG_LARGE_DATA) - { - models[k]->method = LINREG_QR; - } - if (n_data > 0) { /* @@ -706,10 +726,10 @@ run_regression (const struct regression *cmd, if (ws->extras > 0) { struct casereader *r = casereader_clone (reader); - + for (; (c = casereader_read (r)) != NULL; case_unref (c)) { - struct ccase *outc = case_clone (c); + struct ccase *outc = case_create (casewriter_get_proto (ws->writer)); for (k = 0; k < cmd->n_dep_vars; k++) { const struct variable **vars = xnmalloc (cmd->n_vars, sizeof (*vars)); @@ -730,13 +750,13 @@ run_regression (const struct regression *cmd, if (cmd->resid) { - double obs = case_data (c, models[k]->depvar)->f; + double obs = case_data (c, linreg_dep_var (models[k]))->f; double res = linreg_residual (models[k], obs, vals, n_indep); case_data_rw_idx (outc, k * ws->extras + ws->res_idx)->f = res; } free (vals); free (vars); - } + } casewriter_write (ws->writer, outc); } casereader_destroy (r); @@ -760,7 +780,7 @@ run_regression (const struct regression *cmd, static void -reg_stats_r (const linreg * c, const struct variable *var) +reg_stats_r (const struct linreg * c, const struct variable *var) { struct tab_table *t; int n_rows = 2; @@ -797,14 +817,13 @@ reg_stats_r (const linreg * c, const struct variable *var) Table showing estimated regression coefficients. */ static void -reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable *var, const struct regression *cmd) +reg_stats_coeff (const struct linreg * c, const gsl_matrix *cov, const struct variable *var, const struct regression *cmd) { size_t j; int n_cols = 7; const int heading_rows = 2; int n_rows; - int this_row; - double t_stat; + int this_row = heading_rows; double pval; double std_err; double beta; @@ -831,7 +850,7 @@ reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable tab_vline (t, TAL_0, 1, 0, 0); - tab_hline (t, TAL_1, 2, 4, 1); + tab_hline (t, TAL_1, 2, 4, 1); tab_joint_text (t, 2, 0, 3, 0, TAB_CENTER | TAT_TITLE, _("Unstandardized Coefficients")); tab_text (t, 2, 1, TAB_CENTER | TAT_TITLE, _("B")); tab_text (t, 3, 1, TAB_CENTER | TAT_TITLE, _("Std. Error")); @@ -839,8 +858,7 @@ reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable tab_text (t, 4, 1, TAB_CENTER | TAT_TITLE, _("Beta")); tab_text (t, 5, 1, TAB_CENTER | TAT_TITLE, _("t")); tab_text (t, 6, 1, TAB_CENTER | TAT_TITLE, _("Sig.")); - tab_text (t, 1, heading_rows, TAB_LEFT | TAT_TITLE, _("(Constant)")); - tab_double (t, 2, heading_rows, 0, linreg_intercept (c), NULL, RC_OTHER); + std_err = sqrt (gsl_matrix_get (linreg_cov (c), 0, 0)); if (cmd->stats & STATS_CI) @@ -851,24 +869,31 @@ reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable tab_double (t, 8, heading_rows, 0, upper, NULL, RC_OTHER); tab_joint_text_format (t, 7, 0, 8, 0, TAB_CENTER | TAT_TITLE, _("%g%% Confidence Interval for B"), cmd->ci * 100); - tab_hline (t, TAL_1, 7, 8, 1); + tab_hline (t, TAL_1, 7, 8, 1); tab_text (t, 7, 1, TAB_CENTER | TAT_TITLE, _("Lower Bound")); tab_text (t, 8, 1, TAB_CENTER | TAT_TITLE, _("Upper Bound")); } - tab_double (t, 3, heading_rows, 0, std_err, NULL, RC_OTHER); - tab_double (t, 4, heading_rows, 0, 0.0, NULL, RC_OTHER); - t_stat = linreg_intercept (c) / std_err; - tab_double (t, 5, heading_rows, 0, t_stat, NULL, RC_OTHER); - pval = - 2 * gsl_cdf_tdist_Q (fabs (t_stat), - (double) (linreg_n_obs (c) - linreg_n_coeffs (c))); - tab_double (t, 6, heading_rows, 0, pval, NULL, RC_PVALUE); - - for (j = 0; j < linreg_n_coeffs (c); j++) + + if (!cmd->origin) + { + tab_text (t, 1, this_row, TAB_LEFT | TAT_TITLE, _("(Constant)")); + tab_double (t, 2, this_row, 0, linreg_intercept (c), NULL, RC_OTHER); + tab_double (t, 3, this_row, 0, std_err, NULL, RC_OTHER); + tab_double (t, 4, this_row, 0, 0.0, NULL, RC_OTHER); + double t_stat = linreg_intercept (c) / std_err; + tab_double (t, 5, this_row, 0, t_stat, NULL, RC_OTHER); + + double pval = + 2 * gsl_cdf_tdist_Q (fabs (t_stat), + (double) (linreg_n_obs (c) - linreg_n_coeffs (c))); + tab_double (t, 6, this_row, 0, pval, NULL, RC_PVALUE); + this_row++; + } + + for (j = 0; j < linreg_n_coeffs (c); j++, this_row++) { struct string tstr; ds_init_empty (&tstr); - this_row = j + heading_rows + 1; v = linreg_indep_var (c, j); label = var_to_string (v); @@ -896,7 +921,7 @@ reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable /* Test statistic for H0: coefficient is 0. */ - t_stat = linreg_coeff (c, j) / std_err; + double t_stat = linreg_coeff (c, j) / std_err; tab_double (t, 5, this_row, 0, t_stat, NULL, RC_OTHER); /* P values for the test statistic above. @@ -909,7 +934,7 @@ reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable { double lower = linreg_coeff (c, j) - tval * std_err ; double upper = linreg_coeff (c, j) + tval * std_err ; - + tab_double (t, 7, this_row, 0, lower, NULL, RC_OTHER); tab_double (t, 8, this_row, 0, upper, NULL, RC_OTHER); } @@ -922,14 +947,15 @@ reg_stats_coeff (const linreg * c, const gsl_matrix *cov, const struct variable Display the ANOVA table. */ static void -reg_stats_anova (const linreg * c, const struct variable *var) +reg_stats_anova (const struct linreg * c, const struct variable *var) { int n_cols = 7; int n_rows = 4; const double msm = linreg_ssreg (c) / linreg_dfmodel (c); const double mse = linreg_mse (c); const double F = msm / mse; - const double pval = gsl_cdf_fdist_Q (F, c->dfm, c->dfe); + const double pval = gsl_cdf_fdist_Q (F, linreg_dfmodel (c), + linreg_dferror (c)); struct tab_table *t; @@ -960,9 +986,9 @@ reg_stats_anova (const linreg * c, const struct variable *var) /* Degrees of freedom */ - tab_text_format (t, 3, 1, TAB_RIGHT, "%.*g", DBL_DIG + 1, c->dfm); - tab_text_format (t, 3, 2, TAB_RIGHT, "%.*g", DBL_DIG + 1, c->dfe); - tab_text_format (t, 3, 3, TAB_RIGHT, "%.*g", DBL_DIG + 1, c->dft); + tab_text_format (t, 3, 1, TAB_RIGHT, "%.*g", DBL_DIG + 1, linreg_dfmodel (c)); + tab_text_format (t, 3, 2, TAB_RIGHT, "%.*g", DBL_DIG + 1, linreg_dferror (c)); + tab_text_format (t, 3, 3, TAB_RIGHT, "%.*g", DBL_DIG + 1, linreg_dftotal (c)); /* Mean Squares */ tab_double (t, 4, 1, TAB_RIGHT, msm, NULL, RC_OTHER); @@ -978,7 +1004,7 @@ reg_stats_anova (const linreg * c, const struct variable *var) static void -reg_stats_bcov (const linreg * c, const struct variable *var) +reg_stats_bcov (const struct linreg * c, const struct variable *var) { int n_cols; int n_rows; @@ -990,8 +1016,8 @@ reg_stats_bcov (const linreg * c, const struct variable *var) struct tab_table *t; assert (c != NULL); - n_cols = c->n_indeps + 1 + 2; - n_rows = 2 * (c->n_indeps + 1); + n_cols = linreg_n_indeps (c) + 1 + 2; + n_rows = 2 * (linreg_n_indeps (c) + 1); t = tab_create (n_cols, n_rows); tab_headers (t, 2, 0, 1, 0); tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1); @@ -1011,7 +1037,7 @@ reg_stats_bcov (const linreg * c, const struct variable *var) col = (i <= k) ? k : i; row = (i <= k) ? i : k; tab_double (t, k + 2, i, TAB_CENTER, - gsl_matrix_get (c->cov, row, col), NULL, RC_OTHER); + gsl_matrix_get (linreg_cov (c), row, col), NULL, RC_OTHER); } } tab_title (t, _("Coefficient Correlations (%s)"), var_to_string (var));