#include <gsl/gsl_matrix.h>
#include "alloc.h"
#include "case.h"
+#include "casefile.h"
+#include "cat.h"
+#include "cat-routines.h"
+#include "command.h"
+#include "design-matrix.h"
#include "dictionary.h"
+#include "error.h"
#include "file-handle.h"
-#include "command.h"
+#include "gettext.h"
#include "lexer.h"
+#include <linreg/pspp_linreg.h>
+#include "missing-values.h"
+#include "regression_export.h"
#include "tab.h"
#include "var.h"
#include "vfm.h"
-#include "casefile.h"
-#include <linreg/pspp_linreg.h>
-#include "cat.h"
-/* (headers) */
+#define REG_LARGE_DATA 1000
+
+/* (headers) */
/* (specification)
"REGRESSION" (regression_):
f,
defaults,
all;
+ export=custom;
^dependent=varlist;
^method=enter.
*/
*/
size_t *indep_vars;
+/*
+ File where the model will be saved if the EXPORT subcommand
+ is given.
+ */
+struct file_handle *model_file;
+
+/*
+ Return value for the procedure.
+ */
+int pspp_reg_rc = CMD_SUCCESS;
+
static void run_regression (const struct casefile *, void *);
/*
STATISTICS subcommand output functions.
static void
reg_stats_r (pspp_linreg_cache * c)
{
+ struct tab_table *t;
+ int n_rows = 2;
+ int n_cols = 5;
+ double rsq;
+ double adjrsq;
+ double std_error;
+
assert (c != NULL);
+ rsq = c->ssm / c->sst;
+ adjrsq = 1.0 - (1.0 - rsq) * (c->n_obs - 1.0) / (c->n_obs - c->n_indeps);
+ std_error = sqrt ((c->n_indeps - 1.0) / (c->n_obs - 1.0));
+ t = tab_create (n_cols, n_rows, 0);
+ tab_dim (t, tab_natural_dimensions);
+ tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
+ tab_hline (t, TAL_2, 0, n_cols - 1, 1);
+ tab_vline (t, TAL_2, 2, 0, n_rows - 1);
+ tab_vline (t, TAL_0, 1, 0, 0);
+
+ tab_text (t, 1, 0, TAB_CENTER | TAT_TITLE, _("R"));
+ tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("R Square"));
+ tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Adjusted R Square"));
+ tab_text (t, 4, 0, TAB_CENTER | TAT_TITLE, _("Std. Error of the Estimate"));
+ tab_float (t, 1, 1, TAB_RIGHT, sqrt (rsq), 10, 2);
+ tab_float (t, 2, 1, TAB_RIGHT, rsq, 10, 2);
+ tab_float (t, 3, 1, TAB_RIGHT, adjrsq, 10, 2);
+ tab_float (t, 4, 1, TAB_RIGHT, std_error, 10, 2);
+ tab_title (t, 0, _("Model Summary"));
+ tab_submit (t);
}
/*
struct tab_table *t;
assert (c != NULL);
- n_rows = 2 + c->param_estimates->size;
+ n_rows = c->n_coeffs + 2;
+
t = tab_create (n_cols, n_rows, 0);
tab_headers (t, 2, 0, 1, 0);
tab_dim (t, tab_natural_dimensions);
tab_text (t, 5, 0, TAB_CENTER | TAT_TITLE, _("t"));
tab_text (t, 6, 0, TAB_CENTER | TAT_TITLE, _("Significance"));
tab_text (t, 1, 1, TAB_LEFT | TAT_TITLE, _("(Constant)"));
- coeff = gsl_vector_get (c->param_estimates, 0);
+ coeff = c->coeff[0].estimate;
tab_float (t, 2, 1, 0, coeff, 10, 2);
std_err = sqrt (gsl_matrix_get (c->cov, 0, 0));
tab_float (t, 3, 1, 0, std_err, 10, 2);
tab_float (t, 5, 1, 0, t_stat, 10, 2);
pval = 2 * gsl_cdf_tdist_Q (fabs (t_stat), 1.0);
tab_float (t, 6, 1, 0, pval, 10, 2);
- for (j = 0; j < c->n_indeps; j++)
+ for (j = 1; j <= c->n_indeps; j++)
{
i = indep_vars[j];
- struct variable *v = cmd.v_variables[i];
- label = var_to_string (v);
- tab_text (t, 1, j + 2, TAB_CENTER, label);
+ label = var_to_string (c->coeff[j].v);
+ tab_text (t, 1, j + 1, TAB_CENTER, label);
/*
Regression coefficients.
*/
- coeff = gsl_vector_get (c->param_estimates, j + 1);
- tab_float (t, 2, j + 2, 0, coeff, 10, 2);
+ coeff = c->coeff[j].estimate;
+ tab_float (t, 2, j + 1, 0, coeff, 10, 2);
/*
Standard error of the coefficients.
*/
- std_err = sqrt (gsl_matrix_get (c->cov, j + 1, j + 1));
- tab_float (t, 3, j + 2, 0, std_err, 10, 2);
+ std_err = sqrt (gsl_matrix_get (c->cov, j, j));
+ tab_float (t, 3, j + 1, 0, std_err, 10, 2);
/*
'Standardized' coefficient, i.e., regression coefficient
if all variables had unit variance.
*/
- beta = gsl_vector_get (c->indep_std, j + 1);
+ beta = gsl_vector_get (c->indep_std, j);
beta *= coeff / c->depvar_std;
- tab_float (t, 4, j + 2, 0, beta, 10, 2);
+ tab_float (t, 4, j + 1, 0, beta, 10, 2);
/*
Test statistic for H0: coefficient is 0.
*/
t_stat = coeff / std_err;
- tab_float (t, 5, j + 2, 0, t_stat, 10, 2);
+ tab_float (t, 5, j + 1, 0, t_stat, 10, 2);
/*
P values for the test statistic above.
*/
pval = 2 * gsl_cdf_tdist_Q (fabs (t_stat), 1.0);
- tab_float (t, 6, j + 2, 0, pval, 10, 2);
+ tab_float (t, 6, j + 1, 0, pval, 10, 2);
}
tab_title (t, 0, _("Coefficients"));
tab_submit (t);
static void
reg_stats_bcov (pspp_linreg_cache * c)
{
+ int n_cols;
+ int n_rows;
+ int i;
+ int j;
+ int k;
+ int row;
+ int col;
+ const char *label;
+ struct tab_table *t;
+
assert (c != NULL);
+ n_cols = c->n_indeps + 1 + 2;
+ n_rows = 2 * (c->n_indeps + 1);
+ t = tab_create (n_cols, n_rows, 0);
+ tab_headers (t, 2, 0, 1, 0);
+ tab_dim (t, tab_natural_dimensions);
+ tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, n_cols - 1, n_rows - 1);
+ tab_hline (t, TAL_2, 0, n_cols - 1, 1);
+ tab_vline (t, TAL_2, 2, 0, n_rows - 1);
+ tab_vline (t, TAL_0, 1, 0, 0);
+ tab_text (t, 0, 0, TAB_CENTER | TAT_TITLE, _("Model"));
+ tab_text (t, 1, 1, TAB_CENTER | TAT_TITLE, _("Covariances"));
+ for (i = 1; i < c->n_indeps + 1; i++)
+ {
+ j = indep_vars[(i - 1)];
+ struct variable *v = cmd.v_variables[j];
+ label = var_to_string (v);
+ tab_text (t, 2, i, TAB_CENTER, label);
+ tab_text (t, i + 2, 0, TAB_CENTER, label);
+ for (k = 1; k < c->n_indeps + 1; k++)
+ {
+ col = (i <= k) ? k : i;
+ row = (i <= k) ? i : k;
+ tab_float (t, k + 2, i, TAB_CENTER,
+ gsl_matrix_get (c->cov, row, col), 8, 3);
+ }
+ }
+ tab_title (t, 0, _("Coefficient Correlations"));
+ tab_submit (t);
}
static void
reg_stats_ses (pspp_linreg_cache * c)
*/
for (i = 0; i < f; i++)
{
- *(keywords + i) = 1;
+ keywords[i] = 1;
}
}
else
*/
if (keywords[defaults] | d)
{
- *(keywords + anova) = 1;
- *(keywords + outs) = 1;
- *(keywords + coeff) = 1;
- *(keywords + r) = 1;
+ keywords[anova] = 1;
+ keywords[outs] = 1;
+ keywords[coeff] = 1;
+ keywords[r] = 1;
}
}
statistics_keyword_output (reg_stats_r, keywords[r], c);
statistics_keyword_output (reg_stats_selection, keywords[selection], c);
}
+static void
+reg_print_categorical_encoding (FILE *fp, pspp_linreg_cache *c)
+{
+ int i;
+ size_t j;
+ struct pspp_linreg_coeff coeff;
+ union value *val;
+
+ fprintf (fp, "%s", reg_export_categorical_encode_1);
+
+ for (i = 1; i < c->n_indeps; i++) /* c->coeff[0] is the intercept. */
+ {
+ coeff = c->coeff[i];
+ if (coeff.v->type == ALPHA)
+ {
+ fprintf (fp, "struct pspp_reg_categorical_variable %s;\n\t", coeff.v->name);
+ }
+ }
+ for (i = 1; i < c->n_indeps; i++)
+ {
+ coeff = c->coeff[i];
+ if (coeff.v->type == ALPHA)
+ {
+ fprintf (fp, "%s.name = \"%s\";\n\t", coeff.v->name, coeff.v->name);
+ fprintf (fp, "%s.n_vals = %d;\n\t", coeff.v->name, coeff.v->obs_vals->n_categories);
+ fprintf (fp, "%s.values = {", coeff.v->name);
+ for (j = 0; j < coeff.v->obs_vals->n_categories - 1; j++)
+ {
+ val = cat_subscript_to_value ( (const size_t) j, coeff.v);
+ fprintf (fp, "\"%s\",\n\t\t", val->s);
+ }
+ val = cat_subscript_to_value ( (const size_t) j, coeff.v);
+ fprintf (fp, "\"%s\"};\n\n\t", val->s);
+ }
+ }
+}
+
+static void
+reg_print_depvars (FILE *fp, pspp_linreg_cache *c)
+{
+ int i;
+ struct pspp_linreg_coeff coeff;
+
+ fprintf (fp, "char *model_depvars[%d] = {", c->n_indeps);
+ for (i = 1; i < c->n_indeps; i++)
+ {
+ coeff = c->coeff[i];
+ fprintf (fp, "\"%s\",\n\t\t", coeff.v->name);
+ }
+ coeff = c->coeff[i];
+ fprintf (fp, "\"%s\"};\n\t", coeff.v->name);
+}
+static void
+reg_print_getvar (FILE *fp, pspp_linreg_cache *c)
+{
+ fprintf (fp, "static int\npspp_reg_getvar (char *v_name)\n{\n\t");
+ fprintf (fp, "int i;\n\tint n_vars = %d;\n\t",c->n_indeps);
+ reg_print_depvars (fp, c);
+ fprintf (fp, "for (i = 0; i < n_vars; i++)\n\t{\n\t\t");
+ fprintf (fp, "if (strcmp (v_name, model_depvars[i]) == 0)\n\t\t{\n\t\t\t");
+ fprintf (fp, "return i;\n\t\t}\n\t}\n}\n");
+}
+static void
+subcommand_export (int export, pspp_linreg_cache *c)
+{
+ FILE *fp;
+ size_t i;
+ size_t j;
+ int n_quantiles = 100;
+ double increment;
+ double tmp;
+ struct pspp_linreg_coeff coeff;
+
+ if (export)
+ {
+ assert (c != NULL);
+ assert (model_file != NULL);
+ assert (fp != NULL);
+ fp = fopen (handle_get_filename (model_file), "w");
+ fprintf (fp, "%s", reg_preamble);
+ fprintf (fp, "#include <string.h>\n#include <math.h>\n\n");
+ reg_print_getvar (fp, c);
+ reg_print_categorical_encoding (fp, c);
+ fprintf (fp, "%s", reg_export_t_quantiles_1);
+ increment = 0.5 / (double) increment;
+ for (i = 0; i < n_quantiles - 1; i++)
+ {
+ tmp = 0.5 + 0.005 * (double) i;
+ fprintf (fp, "%.15e,\n\t\t", gsl_cdf_tdist_Pinv (tmp, c->n_obs - c->n_indeps));
+ }
+ fprintf (fp, "%.15e};\n\t", gsl_cdf_tdist_Pinv (.9995, c->n_obs - c->n_indeps));
+ fprintf (fp, "%s", reg_export_t_quantiles_2);
+ fprintf (fp, "%s", reg_mean_cmt);
+ fprintf (fp, "double\npspp_reg_estimate (const double *var_vals,");
+ fprintf (fp, "const char *var_names[])\n{\n\t");
+ fprintf (fp, "double model_coeffs[%d] = {", c->n_indeps);
+ for (i = 1; i < c->n_indeps; i++)
+ {
+ coeff = c->coeff[i];
+ fprintf (fp, "%.15e,\n\t\t", coeff.estimate);
+ }
+ coeff = c->coeff[i];
+ fprintf (fp, "%.15e};\n\t", coeff.estimate);
+ coeff = c->coeff[0];
+ fprintf (fp, "double estimate = %.15e;\n\t", coeff.estimate);
+ fprintf (fp, "int i;\n\tint j;\n\n\t");
+ fprintf (fp, "for (i = 0; i < %d; i++)\n\t", c->n_indeps);
+ fprintf (fp, "%s", reg_getvar);
+ fprintf (fp, "const double cov[%d][%d] = {\n\t", c->n_coeffs, c->n_coeffs);
+ for (i = 0; i < c->cov->size1 - 1; i++)
+ {
+ fprintf (fp, "{");
+ for (j = 0; j < c->cov->size2 - 1; j++)
+ {
+ fprintf (fp, "%.15e, ", gsl_matrix_get (c->cov, i, j));
+ }
+ fprintf (fp, "%.15e},\n\t", gsl_matrix_get (c->cov, i, j));
+ }
+ fprintf (fp, "{");
+ for (j = 0; j < c->cov->size2 - 1; j++)
+ {
+ fprintf (fp, "%.15e, ", gsl_matrix_get (c->cov, c->cov->size1 - 1, j));
+ }
+ fprintf (fp, "%.15e}\n\t", gsl_matrix_get (c->cov, c->cov->size1 - 1, c->cov->size2 - 1));
+ fprintf (fp, "};\n\tint n_vars = %d;\n\tint i;\n\tint j;\n\t", c->n_indeps);
+ fprintf (fp, "double unshuffled_vals[%d];\n\t",c->n_indeps);
+ fprintf (fp, "%s", reg_variance);
+ fprintf (fp, "%s", reg_export_confidence_interval);
+ tmp = c->mse * c->mse;
+ fprintf (fp, "%s %.15e", reg_export_prediction_interval_1, tmp);
+ fprintf (fp, "%s %.15e", reg_export_prediction_interval_2, tmp);
+ fprintf (fp, "%s", reg_export_prediction_interval_3);
+ fclose (fp);
+ fp = fopen ("pspp_model_reg.h", "w");
+ fprintf (fp, "%s", reg_header);
+ fclose (fp);
+ }
+}
+static int
+regression_custom_export (struct cmd_regression *cmd)
+{
+ /* 0 on failure, 1 on success, 2 on failure that should result in syntax error */
+ if (!lex_force_match ('('))
+ return 0;
+
+ if (lex_match ('*'))
+ model_file = NULL;
+ else
+ {
+ model_file = fh_parse ();
+ if (model_file == NULL)
+ return 0;
+ }
+
+ if (!lex_force_match (')'))
+ return 0;
+
+ return 1;
+}
+
int
cmd_regression (void)
{
}
multipass_procedure_with_splits (run_regression, &cmd);
- return CMD_SUCCESS;
+ return pspp_reg_rc;
}
/*
}
static void
-run_regression (const struct casefile *cf, void *cmd_)
+run_regression (const struct casefile *cf, void *cmd_ UNUSED)
{
size_t i;
- size_t k;
size_t n_data = 0;
size_t row;
+ size_t case_num;
int n_indep;
+ int j = 0;
+ /*
+ Keep track of the missing cases.
+ */
+ int *is_missing_case;
const union value *val;
struct casereader *r;
struct casereader *r2;
struct ccase c;
- const struct variable *v;
- struct recoded_categorical_array *ca;
- struct recoded_categorical *rc;
+ struct variable *v;
+ struct variable **indep_vars;
struct design_matrix *X;
gsl_vector *Y;
pspp_linreg_cache *lcache;
pspp_linreg_opts lopts;
n_data = casefile_get_case_cnt (cf);
+
+ is_missing_case = xnmalloc (n_data, sizeof (*is_missing_case));
+ for (i = 0; i < n_data; i++)
+ is_missing_case[i] = 0;
+
n_indep = cmd.n_variables - cmd.n_dependent;
- indep_vars = (size_t *) malloc (n_indep * sizeof (*indep_vars));
+ indep_vars = xnmalloc (n_indep, sizeof *indep_vars);
- Y = gsl_vector_alloc (n_data);
lopts.get_depvar_mean_std = 1;
- lopts.get_indep_mean_std = (int *) malloc (n_indep * sizeof (int));
+ lopts.get_indep_mean_std = xnmalloc (n_indep, sizeof (int));
- lcache = pspp_linreg_cache_alloc (n_data, n_indep);
- lcache->indep_means = gsl_vector_alloc (n_indep);
- lcache->indep_std = gsl_vector_alloc (n_indep);
/*
Read from the active file. The first pass encodes categorical
- variables.
+ variables and drops cases with missing values.
*/
- ca = cr_recoded_cat_ar_create (cmd.n_variables, cmd.v_variables);
- for (r = casefile_get_reader (cf);
- casereader_read (r, &c); case_destroy (&c))
+ j = 0;
+ for (i = 0; i < cmd.n_variables; i++)
{
- for (i = 0; i < ca->n_vars; i++)
+ if (!is_depvar (i))
{
- v = (*(ca->a + i))->v;
- val = case_data (&c, v->fv);
- cr_value_update (*(ca->a + i), val);
+ v = cmd.v_variables[i];
+ indep_vars[j] = v;
+ j++;
+ if (v->type == ALPHA)
+ {
+ /* Make a place to hold the binary vectors
+ corresponding to this variable's values. */
+ cat_stored_values_create (v);
+ }
+ for (r = casefile_get_reader (cf);
+ casereader_read (r, &c); case_destroy (&c))
+ {
+ row = casereader_cnum (r) - 1;
+
+ val = case_data (&c, v->fv);
+ cat_value_update (v, val);
+ if (mv_is_value_missing (&v->miss, val))
+ {
+ if (!is_missing_case[row])
+ {
+ /* Now it is missing. */
+ n_data--;
+ is_missing_case[row] = 1;
+ }
+ }
+ }
}
- n_data++;
}
- cr_create_value_matrices (ca);
+
+ Y = gsl_vector_alloc (n_data);
X =
- design_matrix_create (n_indep, (const struct variable **) cmd.v_variables,
- ca, n_data);
+ design_matrix_create (n_indep, (const struct variable **) indep_vars,
+ n_data);
+ lcache = pspp_linreg_cache_alloc (X->m->size1, X->m->size2);
+ lcache->indep_means = gsl_vector_alloc (X->m->size2);
+ lcache->indep_std = gsl_vector_alloc (X->m->size2);
/*
The second pass creates the design matrix.
*/
+ row = 0;
for (r2 = casefile_get_reader (cf); casereader_read (r2, &c);
case_destroy (&c))
/* Iterate over the cases. */
{
- k = 0;
- row = casereader_cnum (r2) - 1;
- for (i = 0; i < cmd.n_variables; ++i) /* Iterate over the variables
+ case_num = casereader_cnum (r2) - 1;
+ if (!is_missing_case[case_num])
+ {
+ for (i = 0; i < cmd.n_variables; ++i) /* Iterate over the variables
for the current case.
*/
- {
- v = cmd.v_variables[i];
- val = case_data (&c, v->fv);
- /*
- Independent/dependent variable separation. The
- 'variables' subcommand specifies a varlist which contains
- both dependent and independent variables. The dependent
- variables are specified with the 'dependent'
- subcommand. We need to separate the two.
- */
- if (is_depvar (i))
{
- if (v->type == NUMERIC)
+ v = cmd.v_variables[i];
+ val = case_data (&c, v->fv);
+ /*
+ Independent/dependent variable separation. The
+ 'variables' subcommand specifies a varlist which contains
+ both dependent and independent variables. The dependent
+ variables are specified with the 'dependent'
+ subcommand. We need to separate the two.
+ */
+ if (is_depvar (i))
{
+ if (v->type != NUMERIC)
+ {
+ msg (SE,
+ gettext ("Dependent variable must be numeric."));
+ pspp_reg_rc = CMD_FAILURE;
+ return;
+ }
+ lcache->depvar = (const struct variable *) v;
gsl_vector_set (Y, row, val->f);
}
else
{
- errno = EINVAL;
- fprintf (stderr,
- "%s:%d: Dependent variable should be numeric: %s\n",
- __FILE__, __LINE__, strerror (errno));
- err_cond_fail ();
- }
- }
- else
- {
- if (v->type == ALPHA)
- {
- rc = cr_var_to_recoded_categorical (v, ca);
- design_matrix_set_categorical (X, row, v, val, rc);
- }
- else if (v->type == NUMERIC)
- {
- design_matrix_set_numeric (X, row, v, val);
+ if (v->type == ALPHA)
+ {
+ design_matrix_set_categorical (X, row, v, val);
+ }
+ else if (v->type == NUMERIC)
+ {
+ design_matrix_set_numeric (X, row, v, val);
+ }
+
+ lopts.get_indep_mean_std[i] = 1;
}
-
- indep_vars[k] = i;
- k++;
- lopts.get_indep_mean_std[i] = 1;
}
+ row++;
}
}
+ /*
+ Now that we know the number of coefficients, allocate space
+ and store pointers to the variables that correspond to the
+ coefficients.
+ */
+ lcache->coeff = xnmalloc (X->m->size2 + 1, sizeof (*lcache->coeff));
+ for (i = 0; i < X->m->size2; i++)
+ {
+ j = i + 1; /* The first coeff is the intercept. */
+ lcache->coeff[j].v =
+ (const struct variable *) design_matrix_col_to_var (X, i);
+ assert (lcache->coeff[j].v != NULL);
+ }
+ /*
+ For large data sets, use QR decomposition.
+ */
+ if (n_data > sqrt (n_indep) && n_data > REG_LARGE_DATA)
+ {
+ lcache->method = PSPP_LINREG_SVD;
+ }
/*
Find the least-squares estimates and other statistics.
*/
pspp_linreg ((const gsl_vector *) Y, X->m, &lopts, lcache);
subcommand_statistics (cmd.a_statistics, lcache);
+ subcommand_export (cmd.sbc_export, lcache);
gsl_vector_free (Y);
design_matrix_destroy (X);
pspp_linreg_cache_free (lcache);
free (lopts.get_indep_mean_std);
free (indep_vars);
+ free (is_missing_case);
casereader_destroy (r);
+ return;
}
/*