X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Fglm.c;h=74e918b886b21f26d05b1fadf2d245faa60096bb;hb=edd5c738dfef01c90d02e06a33b93fc9d38320b8;hp=92e35ccd27e1f5f4c1f8082d63d402e90f05ca4d;hpb=72d873a1af4914c2bfe1cdda1cb0108da243f534;p=pspp diff --git a/src/language/stats/glm.c b/src/language/stats/glm.c index 92e35ccd27..74e918b886 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, 2011 Free Software Foundation, Inc. + Copyright (C) 2010, 2011, 2012 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 @@ -81,7 +81,7 @@ struct glm_workspace 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. @@ -155,7 +155,8 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) PV_NO_DUPLICATE | PV_NUMERIC)) goto error; - lex_force_match (lexer, T_BY); + if (! lex_force_match (lexer, T_BY)) + goto error; if (!parse_variables_const (lexer, glm.dict, &glm.factor_vars, &glm.n_factor_vars, @@ -229,7 +230,7 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) lex_error (lexer, NULL); goto error; } - + glm.alpha = lex_number (lexer); lex_get (lexer); if ( ! lex_force_match (lexer, T_RPAREN)) @@ -267,7 +268,7 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) } glm.ss_type = lex_integer (lexer); - if (1 > glm.ss_type && 3 < glm.ss_type ) + if (1 > glm.ss_type || 3 < glm.ss_type ) { msg (ME, _("Only types 1, 2 & 3 sums of squares are currently implemented")); goto error; @@ -326,6 +327,7 @@ cmd_glm (struct lexer *lexer, struct dataset *ds) free (glm.factor_vars); for (i = 0 ; i < glm.n_interactions; ++i) interaction_destroy (glm.interactions[i]); + free (glm.interactions); free (glm.dep_vars); @@ -358,11 +360,11 @@ fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f) size_t j; size_t n = 0; size_t m = 0; - + for (i = 0; i < cov->size1; i++) { if (not_dropped (i, dropped_f)) - { + { m = 0; for (j = 0; j < cov->size2; j++) { @@ -371,7 +373,7 @@ fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f) gsl_matrix_set (submatrix, n, m, gsl_matrix_get (cov, i, j)); m++; - } + } } n++; } @@ -379,14 +381,14 @@ fill_submatrix (const gsl_matrix * cov, gsl_matrix * submatrix, bool *dropped_f) } -/* +/* Type 1 sums of squares. Populate SSQ with the Type 1 sums of squares according to COV */ static void ssq_type1 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) { - gsl_matrix *cm = covariance_calculate_unnormalized (cov); + const gsl_matrix *cm = covariance_calculate_unnormalized (cov); size_t i; size_t k; bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped)); @@ -408,7 +410,7 @@ ssq_type1 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) { 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++) { @@ -417,7 +419,7 @@ ssq_type1 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) { - const struct interaction * x = + const struct interaction * x = categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars); if ( x == cmd->interactions [k]) @@ -446,17 +448,16 @@ ssq_type1 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) 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); + const gsl_matrix *cm = covariance_calculate_unnormalized (cov); size_t i; size_t k; bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped)); @@ -471,7 +472,7 @@ ssq_type2 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) size_t n_dropped_submodel = 0; for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) { - const struct interaction * x = + const struct interaction * x = categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars); model_dropped[i] = false; @@ -510,17 +511,16 @@ ssq_type2 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) 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); + const gsl_matrix *cm = covariance_calculate_unnormalized (cov); size_t i; size_t k; bool *model_dropped = xcalloc (covariance_dim (cov), sizeof (*model_dropped)); @@ -542,7 +542,7 @@ ssq_type3 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) for (i = cmd->n_dep_vars; i < covariance_dim (cov); i++) { - const struct interaction * x = + const struct interaction * x = categoricals_get_interaction_by_subscript (cats, i - cmd->n_dep_vars); model_dropped[i] = false; @@ -567,8 +567,6 @@ ssq_type3 (struct covariance *cov, gsl_vector *ssq, const struct glm_spec *cmd) gsl_matrix_free (model_cov); } free (model_dropped); - - gsl_matrix_free (cm); } @@ -589,12 +587,21 @@ run_glm (struct glm_spec *cmd, struct casereader *input, struct glm_workspace ws; struct covariance *cov; + input = casereader_create_filter_missing (input, + cmd->dep_vars, cmd->n_dep_vars, + cmd->exclude, + NULL, NULL); + + input = casereader_create_filter_missing (input, + cmd->factor_vars, cmd->n_factor_vars, + cmd->exclude, + NULL, NULL); + ws.cats = categoricals_create (cmd->interactions, cmd->n_interactions, - cmd->wv, cmd->exclude, - NULL, NULL, NULL, NULL); + cmd->wv, cmd->exclude, MV_ANY); cov = covariance_2pass_create (cmd->n_dep_vars, cmd->dep_vars, - ws.cats, cmd->wv, cmd->exclude); + ws.cats, cmd->wv, cmd->exclude, true); c = casereader_peek (input, 0); @@ -657,7 +664,9 @@ run_glm (struct glm_spec *cmd, struct casereader *input, } { - gsl_matrix *cm = covariance_calculate_unnormalized (cov); + const gsl_matrix *ucm = covariance_calculate_unnormalized (cov); + gsl_matrix *cm = gsl_matrix_alloc (ucm->size1, ucm->size2); + gsl_matrix_memcpy (cm, ucm); // dump_matrix (cm); @@ -686,7 +695,6 @@ run_glm (struct glm_spec *cmd, struct casereader *input, break; } // dump_matrix (cm); - gsl_matrix_free (cm); } @@ -701,7 +709,7 @@ run_glm (struct glm_spec *cmd, struct casereader *input, taint_destroy (taint); } -static const char *roman[] = +static const char *roman[] = { "", /* The Romans had no concept of zero */ "I", @@ -716,6 +724,8 @@ 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 = 1.0; double mse = 0; @@ -731,8 +741,8 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) if (cmd->intercept) nr += 2; - msg (MW, "GLM is experimental. Do not rely on these results."); t = tab_create (nc, nr); + tab_set_format (t, RC_WEIGHT, wfmt); tab_title (t, _("Tests of Between-Subjects Effects")); tab_headers (t, heading_columns, 0, heading_rows, 0); @@ -746,7 +756,7 @@ 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"), + _("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")); @@ -767,20 +777,25 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) mse = gsl_vector_get (ws->ssq, 0) / (n_total - df_corr); - const double intercept_ssq = pow2 (mean * n_total) / n_total; + intercept_ssq = pow2 (mean * n_total) / n_total; - double ssq_effects = 0.0; + ssq_effects = 0.0; if (cmd->intercept) { const double df = 1.0; const double F = intercept_ssq / df / mse; tab_text (t, 0, r, TAB_LEFT | TAT_TITLE, _("Intercept")); - 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_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); + /* The intercept for unbalanced models is of limited use and + nobody knows how to calculate it properly */ + if (categoricals_isbalanced (ws->cats)) + { + tab_double (t, 1, r, 0, intercept_ssq, NULL, RC_OTHER); + tab_double (t, 2, r, 0, 1.00, NULL, RC_WEIGHT); + tab_double (t, 3, r, 0, intercept_ssq / df, NULL, RC_OTHER); + tab_double (t, 4, r, 0, F, NULL, RC_OTHER); + tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr), + NULL, RC_PVALUE); + } r++; } @@ -790,26 +805,28 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) double df = categoricals_df (ws->cats, f); double ssq = gsl_vector_get (ws->ssq, f + 1); + double F; + ssq_effects += ssq; - if (! cmd->intercept) + if (! cmd->intercept) { df++; ssq += intercept_ssq; } - const double F = ssq / df / mse; + 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); - 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, 1, r, 0, ssq, NULL, RC_OTHER); + tab_double (t, 2, r, 0, df, NULL, RC_WEIGHT); + tab_double (t, 3, r, 0, ssq / df, NULL, RC_OTHER); + tab_double (t, 4, r, 0, F, NULL, RC_OTHER); tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr), - NULL); + NULL, RC_PVALUE); r++; } @@ -817,19 +834,21 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) /* 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; - 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); + F = ssq / df / mse; + tab_double (t, 1, heading_rows, 0, ssq, NULL, RC_OTHER); + tab_double (t, 2, heading_rows, 0, df, NULL, RC_WEIGHT); + tab_double (t, 3, heading_rows, 0, ssq / df, NULL, RC_OTHER); + tab_double (t, 4, heading_rows, 0, F, NULL, RC_OTHER); tab_double (t, 5, heading_rows, 0, - gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL); + gsl_cdf_fdist_Q (F, df, n_total - df_corr), NULL, RC_PVALUE); } { @@ -837,24 +856,24 @@ output_glm (const struct glm_spec *cmd, const struct glm_workspace *ws) 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); + tab_double (t, 1, r, 0, ssq, NULL, RC_OTHER); + tab_double (t, 2, r, 0, df, NULL, RC_WEIGHT); + tab_double (t, 3, r++, 0, mse, NULL, RC_OTHER); } { 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); - + tab_double (t, 1, r, 0, ws->total_ssq + intercept_ssq, NULL, RC_OTHER); + tab_double (t, 2, r, 0, n_total, NULL, RC_WEIGHT); + r++; } if (cmd->intercept) { 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_double (t, 1, r, 0, ws->total_ssq, NULL, RC_OTHER); + tab_double (t, 2, r, 0, n_total - 1.0, NULL, RC_WEIGHT); } tab_submit (t); @@ -880,71 +899,11 @@ dump_matrix (const gsl_matrix * m) - -/* 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, struct interaction **iact) -{ - const struct variable *v = NULL; - assert (iact); - - switch (lex_next_token (lexer, 1)) - { - case T_ENDCMD: - case T_SLASH: - case T_COMMA: - case T_ID: - case T_BY: - case T_ASTERISK: - break; - default: - return false; - break; - } - - if (! lex_match_variable (lexer, glm, &v)) - { - 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)) - { - return parse_design_interaction (lexer, glm, iact); - } - - return true; -} - static bool parse_nested_variable (struct lexer *lexer, struct glm_spec *glm) { const struct variable *v = NULL; - if ( ! lex_match_variable (lexer, glm, &v)) + if ( ! lex_match_variable (lexer, glm->dict, &v)) return false; if (lex_match (lexer, T_LPAREN)) @@ -956,7 +915,7 @@ parse_nested_variable (struct lexer *lexer, struct glm_spec *glm) return false; } - lex_error (lexer, "Nested variables are not yet implemented"); return false; + lex_error (lexer, "Nested variables are not yet implemented"); return false; return true; } @@ -965,7 +924,7 @@ static bool parse_design_term (struct lexer *lexer, struct glm_spec *glm) { struct interaction *iact = NULL; - if (parse_design_interaction (lexer, glm, &iact)) + if (parse_design_interaction (lexer, glm->dict, &iact)) { /* Interaction parsing successful. Add to list of interactions */ glm->interactions = xrealloc (glm->interactions, sizeof *glm->interactions * ++glm->n_interactions);