#include "linreg/sweep.h"
#include "math/categoricals.h"
#include "math/covariance.h"
+#include "math/interaction.h"
#include "math/moments.h"
#include "output/tab.h"
size_t n_factor_vars;
const struct variable **factor_vars;
+ size_t n_interactions;
+ struct interaction **interactions;
+
enum mv_class exclude;
/* The weight variable */
const struct variable *wv;
+ const struct dictionary *dict;
+
bool intercept;
+
+ double alpha;
};
struct glm_workspace
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)
{
+ int i;
struct const_var_set *factors = NULL;
- const struct dictionary *dict = dataset_dict (ds);
struct glm_spec glm;
+ bool design = false;
+ glm.dict = dataset_dict (ds);
glm.n_dep_vars = 0;
glm.n_factor_vars = 0;
+ glm.n_interactions = 0;
+ glm.interactions = NULL;
glm.dep_vars = NULL;
glm.factor_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 0
+ 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;
+ }
+
+ 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"))
{
- size_t n_des;
- const struct variable **des;
lex_match (lexer, T_EQUALS);
- parse_const_var_set_vars (lexer, factors, &des, &n_des, 0);
+ if (! parse_design_spec (lexer, &glm))
+ goto error;
+
+ if ( glm.n_interactions == 0)
+ {
+ msg (ME, _("One or more design variables must be given"));
+ goto error;
+ }
+
+ design = true;
}
-#endif
else
{
lex_error (lexer, NULL);
}
}
+ 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);
const_var_set_destroy (factors);
free (glm.factor_vars);
+ for (i = 0 ; i < glm.n_interactions; ++i)
+ interaction_destroy (glm.interactions[i]);
+ free (glm.interactions);
free (glm.dep_vars);
+
return CMD_SUCCESS;
error:
const_var_set_destroy (factors);
free (glm.factor_vars);
+ for (i = 0 ; i < glm.n_interactions; ++i)
+ interaction_destroy (glm.interactions[i]);
+
+ free (glm.interactions);
free (glm.dep_vars);
return CMD_FAILURE;
}
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_factor_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->factor_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))
}
free (dropped);
- free (vars);
gsl_matrix_free (cm);
}
struct glm_workspace ws;
struct covariance *cov;
- ws.cats = categoricals_create (cmd->factor_vars, cmd->n_factor_vars,
+
+ ws.cats = categoricals_create (cmd->interactions, cmd->n_interactions,
cmd->wv, cmd->exclude,
NULL, NULL, NULL, NULL);
}
casereader_destroy (reader);
- categoricals_done (ws.cats);
-
for (reader = input;
(c = casereader_read (reader)) != NULL; case_unref (c))
{
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);
struct tab_table *t;
const int nc = 6;
- int nr = heading_rows + 4 + cmd->n_factor_vars;
+ int nr = heading_rows + 4 + cmd->n_interactions;
if (cmd->intercept)
nr++;
if (cmd->intercept)
df_corr += 1.0;
- for (f = 0; f < cmd->n_factor_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);
r++;
}
- for (f = 0; f < cmd->n_factor_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->factor_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);
tab_double (t, 5, r, 0, gsl_cdf_fdist_Q (F, df, n_total - df_corr),
NULL);
-
-
r++;
}
printf ("\n");
}
#endif
+
+
+\f
+
+/* 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))
+ {
+ // lex_error (lexer, "Interactions are not yet implemented"); return false;
+ 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))
+ return false;
+
+ if (lex_match (lexer, T_LPAREN))
+ {
+ if ( ! parse_nested_variable (lexer, glm))
+ return false;
+
+ if ( ! lex_force_match (lexer, T_RPAREN))
+ return false;
+ }
+
+ lex_error (lexer, "Nested variables are not yet implemented"); return false;
+ return true;
+}
+
+/* A design term is an interaction OR a nested variable */
+static bool
+parse_design_term (struct lexer *lexer, struct glm_spec *glm)
+{
+ 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;
+
+ 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)
+{
+ 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);
+}
+