#include <config.h>
-#include "roc.h"
-#include <data/procedure.h>
-#include <language/lexer/variable-parser.h>
-#include <language/lexer/value-parser.h>
-#include <language/lexer/lexer.h>
+#include <language/stats/roc.h>
#include <data/casegrouper.h>
#include <data/casereader.h>
#include <data/casewriter.h>
#include <data/dictionary.h>
#include <data/format.h>
-#include <math/sort.h>
+#include <data/procedure.h>
#include <data/subcase.h>
-
-
+#include <language/command.h>
+#include <language/lexer/lexer.h>
+#include <language/lexer/value-parser.h>
+#include <language/lexer/variable-parser.h>
#include <libpspp/misc.h>
+#include <math/sort.h>
+#include <output/chart-item.h>
+#include <output/charts/roc-chart.h>
+#include <output/tab.h>
#include <gsl/gsl_cdf.h>
-#include <output/table.h>
-
-#include <output/charts/plot-chart.h>
-#include <output/charts/cartesian.h>
#include "gettext.h"
#define _(msgid) gettext (msgid)
const struct variable **vars;
const struct dictionary *dict;
- struct variable *state_var ;
+ const struct variable *state_var;
union value state_value;
/* Plot the roc curve */
roc.pos = roc.pos_weighted = 0;
roc.neg = roc.neg_weighted = 0;
roc.dict = dataset_dict (ds);
+ roc.state_var = NULL;
+ lex_match (lexer, '/');
if (!parse_variables_const (lexer, dict, &roc.vars, &roc.n_vars,
PV_APPEND | PV_NO_DUPLICATE | PV_NUMERIC))
- return 2;
+ goto error;
if ( ! lex_force_match (lexer, T_BY))
{
- return 2;
+ goto error;
}
roc.state_var = parse_variable (lexer, dict);
if ( !lex_force_match (lexer, '('))
{
- return 2;
+ goto error;
}
+ value_init (&roc.state_value, var_get_width (roc.state_var));
parse_value (lexer, &roc.state_value, var_get_width (roc.state_var));
if ( !lex_force_match (lexer, ')'))
{
- return 2;
+ goto error;
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
}
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
}
else if (lex_match_id (lexer, "PRINT"))
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
}
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
lex_force_match (lexer, ')');
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
lex_force_match (lexer, ')');
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
lex_force_match (lexer, ')');
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;
}
}
}
}
}
- run_roc (ds, &roc);
+ if ( ! run_roc (ds, &roc))
+ goto error;
+
+ value_destroy (&roc.state_value, var_get_width (roc.state_var));
+ free (roc.vars);
+ return CMD_SUCCESS;
- return 1;
+ error:
+ if ( roc.state_var)
+ value_destroy (&roc.state_value, var_get_width (roc.state_var));
+ free (roc.vars);
+ return CMD_FAILURE;
}
}
#endif
+
+/*
+ Return true iff the state variable indicates that C has positive actual state.
+
+ As a side effect, this function also accumulates the roc->{pos,neg} and
+ roc->{pos,neg}_weighted counts.
+ */
static bool
match_positives (const struct ccase *c, void *aux)
{
const struct variable *wv = dict_get_weight (roc->dict);
const double weight = wv ? case_data (c, wv)->f : 1.0;
- bool positive = ( 0 == value_compare_3way (case_data (c, roc->state_var),
- &roc->state_value,
- var_get_width (roc->state_var)));
+ const bool positive =
+ ( 0 == value_compare_3way (case_data (c, roc->state_var), &roc->state_value,
+ var_get_width (roc->state_var)));
if ( positive )
{
#define N_EQ 1
#define N_PRED 2
+/* Some intermediate state for calculating the cutpoints and the
+ standard error values */
struct roc_state
{
- double auc;
+ double auc; /* Area under the curve */
- double n1;
- double n2;
+ double n1; /* total weight of positives */
+ double n2; /* total weight of negatives */
- double q1hat;
+ /* intermediates for standard error */
+ double q1hat;
double q2hat;
+ /* intermediates for cutpoints */
struct casewriter *cutpoint_wtr;
struct casereader *cutpoint_rdr;
double prev_result;
double max;
};
-
-
-#define CUTPOINT 0
-#define TP 1
-#define FN 2
-#define TN 3
-#define FP 4
-
-
+/*
+ Return a new casereader based upon CUTPOINT_RDR.
+ The number of "positive" cases are placed into
+ the position TRUE_INDEX, and the number of "negative" cases
+ into FALSE_INDEX.
+ POS_COND and RESULT determine the semantics of what is
+ "positive".
+ WEIGHT is the value of a single count.
+ */
static struct casereader *
accumulate_counts (struct casereader *cutpoint_rdr,
double result, double weight,
struct ccase *cpc;
double prev_cp = SYSMIS;
-
for ( ; (cpc = casereader_read (r) ); case_unref (cpc))
{
struct ccase *new_case;
- const double cp = case_data_idx (cpc, CUTPOINT)->f;
+ const double cp = case_data_idx (cpc, ROC_CUTPOINT)->f;
+
+ assert (cp != SYSMIS);
/* We don't want duplicates here */
if ( cp == prev_cp )
new_case = case_clone (cpc);
if ( pos_cond (result, cp))
- {
- case_data_rw_idx (new_case, true_index)->f += weight;
- }
+ case_data_rw_idx (new_case, true_index)->f += weight;
else
- {
- case_data_rw_idx (new_case, false_index)->f += weight;
- }
+ case_data_rw_idx (new_case, false_index)->f += weight;
prev_cp = cp;
static void output_roc (struct roc_state *rs, const struct cmd_roc *roc);
+/*
+ This function does 3 things:
+
+ 1. Counts the number of cases which are equal to every other case in READER,
+ and those cases for which the relationship between it and every other case
+ satifies PRED (normally either > or <). VAR is variable defining a case's value
+ for this purpose.
+ 2. Counts the number of true and false cases in reader, and populates
+ CUTPOINT_RDR accordingly. TRUE_INDEX and FALSE_INDEX are the indices
+ which receive these values. POS_COND is the condition defining true
+ and false.
+
+ 3. CC is filled with the cumulative weight of all cases of READER.
+*/
static struct casereader *
process_group (const struct variable *var, struct casereader *reader,
bool (*pred) (double, double),
return casewriter_make_reader (wtr);
}
+/* Some more indeces into case data */
+#define N_POS_EQ 1 /* number of positive cases with values equal to n */
+#define N_POS_GT 2 /* number of postive cases with values greater than n */
+#define N_NEG_EQ 3 /* number of negative cases with values equal to n */
+#define N_NEG_LT 4 /* number of negative cases with values less than n */
+
static bool
gt (double d1, double d2)
{
return d1 < d2;
}
+
+/*
+ Return a casereader with width 3,
+ populated with cases based upon READER.
+ The cases will have the values:
+ (N, number of cases equal to N, number of cases greater than N)
+ As a side effect, update RS->n1 with the number of positive cases.
+*/
static struct casereader *
process_positive_group (const struct variable *var, struct casereader *reader,
const struct dictionary *dict,
return process_group (var, reader, gt, dict, &rs->n1,
&rs->cutpoint_rdr,
ge,
- TP, FN);
+ ROC_TP, ROC_FN);
}
-
+/*
+ Return a casereader with width 3,
+ populated with cases based upon READER.
+ The cases will have the values:
+ (N, number of cases equal to N, number of cases less than N)
+ As a side effect, update RS->n2 with the number of negative cases.
+*/
static struct casereader *
process_negative_group (const struct variable *var, struct casereader *reader,
const struct dictionary *dict,
return process_group (var, reader, lt, dict, &rs->n2,
&rs->cutpoint_rdr,
lt,
- TN, FP);
+ ROC_TN, ROC_FP);
}
{
struct ccase *cc = case_create (casewriter_get_proto (writer));
- case_data_rw_idx (cc, CUTPOINT)->f = cutpoint;
- case_data_rw_idx (cc, TP)->f = 0;
- case_data_rw_idx (cc, FN)->f = 0;
- case_data_rw_idx (cc, TN)->f = 0;
- case_data_rw_idx (cc, FP)->f = 0;
-
+ case_data_rw_idx (cc, ROC_CUTPOINT)->f = cutpoint;
+ case_data_rw_idx (cc, ROC_TP)->f = 0;
+ case_data_rw_idx (cc, ROC_FN)->f = 0;
+ case_data_rw_idx (cc, ROC_TN)->f = 0;
+ case_data_rw_idx (cc, ROC_FP)->f = 0;
casewriter_write (writer, cc);
}
-/* Prepare the cutpoints */
+/*
+ Create and initialise the rs[x].cutpoint_rdr casereaders. That is, the readers will
+ be created with width 5, ready to take the values (cutpoint, ROC_TP, ROC_FN, ROC_TN, ROC_FP), and the
+ reader will be populated with its final number of cases.
+ However on exit from this function, only ROC_CUTPOINT entries will be set to their final
+ value. The other entries will be initialised to zero.
+*/
static void
prepare_cutpoints (struct cmd_roc *roc, struct roc_state *rs, struct casereader *input)
{
struct caseproto *proto = caseproto_create ();
struct subcase ordering;
- subcase_init (&ordering, CUTPOINT, 0, SC_ASCEND);
+ subcase_init (&ordering, ROC_CUTPOINT, 0, SC_ASCEND);
proto = caseproto_add_width (proto, 0); /* cutpoint */
- proto = caseproto_add_width (proto, 0); /* TP */
- proto = caseproto_add_width (proto, 0); /* FN */
- proto = caseproto_add_width (proto, 0); /* TN */
- proto = caseproto_add_width (proto, 0); /* FP */
+ proto = caseproto_add_width (proto, 0); /* ROC_TP */
+ proto = caseproto_add_width (proto, 0); /* ROC_FN */
+ proto = caseproto_add_width (proto, 0); /* ROC_TN */
+ proto = caseproto_add_width (proto, 0); /* ROC_FP */
for (i = 0 ; i < roc->n_vars; ++i)
{
struct subcase up_ordering;
struct subcase down_ordering;
+ struct casewriter *neg_wtr = NULL;
+
struct casereader *input = casereader_create_filter_missing (reader,
roc->vars, roc->n_vars,
roc->exclude,
NULL,
NULL);
+ input = casereader_create_filter_missing (input,
+ &roc->state_var, 1,
+ roc->exclude,
+ NULL,
+ NULL);
- struct casewriter *neg_wtr = autopaging_writer_create (casereader_get_proto (input));
+ neg_wtr = autopaging_writer_create (casereader_get_proto (input));
prepare_cutpoints (roc, rs, input);
+
+ /* Separate the positive actual state cases from the negative ones */
positives =
casereader_create_filter_func (input,
match_positives,
struct casereader *neg ;
struct casereader *pos = casereader_clone (positives);
+
struct casereader *n_pos =
process_positive_group (var, pos, dict, &rs[i]);
n_neg = process_negative_group (var, neg, dict, &rs[i]);
+
+ /* Merge the n_pos and n_neg casereaders */
w = sort_create_writer (&up_ordering, n_proto);
for ( ; (cpos = casereader_read (n_pos) ); case_unref (cpos))
{
const double jneg = case_data_idx (cneg, VALUE)->f;
case_data_rw_idx (nc, VALUE)->f = jneg;
- case_data_rw_idx (nc, N_EQ)->f = 0;
+ case_data_rw_idx (nc, N_POS_EQ)->f = 0;
- case_data_rw_idx (nc, N_PRED)->f = SYSMIS;
+ case_data_rw_idx (nc, N_POS_GT)->f = SYSMIS;
- *case_data_rw_idx (nc, 3) = *case_data_idx (cneg, N_EQ);
- *case_data_rw_idx (nc, 4) = *case_data_idx (cneg, N_PRED);
+ *case_data_rw_idx (nc, N_NEG_EQ) = *case_data_idx (cneg, N_EQ);
+ *case_data_rw_idx (nc, N_NEG_LT) = *case_data_idx (cneg, N_PRED);
casewriter_write (w, nc);
}
case_data_rw_idx (pos_case, VALUE)->f = jpos;
- *case_data_rw_idx (pos_case, N_EQ) = *case_data_idx (cpos, N_EQ);
- *case_data_rw_idx (pos_case, N_PRED) = *case_data_idx (cpos, N_PRED);
- case_data_rw_idx (pos_case, 3)->f = 0;
- case_data_rw_idx (pos_case, 4)->f = SYSMIS;
+ *case_data_rw_idx (pos_case, N_POS_EQ) = *case_data_idx (cpos, N_EQ);
+ *case_data_rw_idx (pos_case, N_POS_GT) = *case_data_idx (cpos, N_PRED);
+ case_data_rw_idx (pos_case, N_NEG_EQ)->f = 0;
+ case_data_rw_idx (pos_case, N_NEG_LT)->f = SYSMIS;
casewriter_write (w, pos_case);
}
+/* These aren't used anymore */
+#undef N_EQ
+#undef N_PRED
+
r = casewriter_make_reader (w);
+ /* Propagate the N_POS_GT values from the positive cases
+ to the negative ones */
{
double prev_pos_gt = rs[i].n1;
w = sort_create_writer (&down_ordering, n_proto);
for ( ; (c = casereader_read (r) ); case_unref (c))
{
- double n_pos_gt = case_data_idx (c, N_PRED)->f;
+ double n_pos_gt = case_data_idx (c, N_POS_GT)->f;
struct ccase *nc = case_clone (c);
if ( n_pos_gt == SYSMIS)
{
n_pos_gt = prev_pos_gt;
- case_data_rw_idx (nc, N_PRED)->f = n_pos_gt;
+ case_data_rw_idx (nc, N_POS_GT)->f = n_pos_gt;
}
casewriter_write (w, nc);
r = casewriter_make_reader (w);
}
+ /* Propagate the N_NEG_LT values from the negative cases
+ to the positive ones */
{
double prev_neg_lt = rs[i].n2;
w = sort_create_writer (&up_ordering, n_proto);
for ( ; (c = casereader_read (r) ); case_unref (c))
{
- double n_neg_lt = case_data_idx (c, 4)->f;
+ double n_neg_lt = case_data_idx (c, N_NEG_LT)->f;
struct ccase *nc = case_clone (c);
if ( n_neg_lt == SYSMIS)
{
n_neg_lt = prev_neg_lt;
- case_data_rw_idx (nc, 4)->f = n_neg_lt;
+ case_data_rw_idx (nc, N_NEG_LT)->f = n_neg_lt;
}
casewriter_write (w, nc);
const struct ccase *next_case = casereader_peek (r, 0);
const double j = case_data_idx (c, VALUE)->f;
- double n_pos_eq = case_data_idx (c, N_EQ)->f;
- double n_pos_gt = case_data_idx (c, N_PRED)->f;
- double n_neg_eq = case_data_idx (c, 3)->f;
- double n_neg_lt = case_data_idx (c, 4)->f;
+ double n_pos_eq = case_data_idx (c, N_POS_EQ)->f;
+ double n_pos_gt = case_data_idx (c, N_POS_GT)->f;
+ double n_neg_eq = case_data_idx (c, N_NEG_EQ)->f;
+ double n_neg_lt = case_data_idx (c, N_NEG_LT)->f;
if ( prev_case && j == case_data_idx (prev_case, VALUE)->f)
{
- if ( 0 == case_data_idx (c, N_EQ)->f)
+ if ( 0 == case_data_idx (c, N_POS_EQ)->f)
{
- n_pos_eq = case_data_idx (prev_case, N_EQ)->f;
- n_pos_gt = case_data_idx (prev_case, N_PRED)->f;
+ n_pos_eq = case_data_idx (prev_case, N_POS_EQ)->f;
+ n_pos_gt = case_data_idx (prev_case, N_POS_GT)->f;
}
- if ( 0 == case_data_idx (c, 3)->f)
+ if ( 0 == case_data_idx (c, N_NEG_EQ)->f)
{
- n_neg_eq = case_data_idx (prev_case, 3)->f;
- n_neg_lt = case_data_idx (prev_case, 4)->f;
+ n_neg_eq = case_data_idx (prev_case, N_NEG_EQ)->f;
+ n_neg_lt = case_data_idx (prev_case, N_NEG_LT)->f;
}
}
const int n_fields = roc->print_se ? 5 : 1;
const int n_cols = roc->n_vars > 1 ? n_fields + 1: n_fields;
const int n_rows = 2 + roc->n_vars;
- struct tab_table *tbl = tab_create (n_cols, n_rows, 0);
+ struct tab_table *tbl = tab_create (n_cols, n_rows);
if ( roc->n_vars > 1)
tab_title (tbl, _("Area Under the Curve"));
tab_headers (tbl, n_cols - n_fields, 0, 1, 0);
- tab_dim (tbl, tab_natural_dimensions, NULL);
tab_text (tbl, n_cols - n_fields, 1, TAT_TITLE, _("Area"));
tab_text (tbl, n_cols - 2, 1, TAT_TITLE, _("Lower Bound"));
tab_text (tbl, n_cols - 1, 1, TAT_TITLE, _("Upper Bound"));
- tab_joint_text (tbl, n_cols - 2, 0, 4, 0,
- TAT_TITLE | TAB_CENTER | TAT_PRINTF,
- _("Asymp. %g%% Confidence Interval"), roc->ci);
+ tab_joint_text_format (tbl, n_cols - 2, 0, 4, 0,
+ TAT_TITLE | TAB_CENTER,
+ _("Asymp. %g%% Confidence Interval"), roc->ci);
tab_vline (tbl, 0, n_cols - 1, 0, 0);
tab_hline (tbl, TAL_1, n_cols - 2, n_cols - 1, 1);
}
{
const int n_cols = 3;
const int n_rows = 4;
- struct tab_table *tbl = tab_create (n_cols, n_rows, 0);
+ struct tab_table *tbl = tab_create (n_cols, n_rows);
tab_title (tbl, _("Case Summary"));
tab_headers (tbl, 1, 0, 2, 0);
- tab_dim (tbl, tab_natural_dimensions, NULL);
-
tab_box (tbl,
TAL_2, TAL_2,
-1, -1,
for (i = 0; i < roc->n_vars; ++i)
n_rows += casereader_count_cases (rs[i].cutpoint_rdr);
- tbl = tab_create (n_cols, n_rows, 0);
+ tbl = tab_create (n_cols, n_rows);
if ( roc->n_vars > 1)
tab_title (tbl, _("Coordinates of the Curve"));
tab_headers (tbl, 1, 0, 1, 0);
- tab_dim (tbl, tab_natural_dimensions, NULL);
-
tab_hline (tbl, TAL_2, 0, n_cols - 1, 1);
if ( roc->n_vars > 1)
for (; (cc = casereader_read (r)) != NULL;
case_unref (cc), x++)
{
- const double se = case_data_idx (cc, TP)->f /
+ const double se = case_data_idx (cc, ROC_TP)->f /
(
- case_data_idx (cc, TP)->f
+ case_data_idx (cc, ROC_TP)->f
+
- case_data_idx (cc, FN)->f
+ case_data_idx (cc, ROC_FN)->f
);
- const double sp = case_data_idx (cc, TN)->f /
+ const double sp = case_data_idx (cc, ROC_TN)->f /
(
- case_data_idx (cc, TN)->f
+ case_data_idx (cc, ROC_TN)->f
+
- case_data_idx (cc, FP)->f
+ case_data_idx (cc, ROC_FP)->f
);
- tab_double (tbl, n_cols - 3, x, 0, case_data_idx (cc, CUTPOINT)->f,
+ tab_double (tbl, n_cols - 3, x, 0, case_data_idx (cc, ROC_CUTPOINT)->f,
var_get_print_format (roc->vars[i]));
tab_double (tbl, n_cols - 2, x, 0, se, NULL);
}
-static void
-draw_roc (struct roc_state *rs, const struct cmd_roc *roc)
-{
- int i;
-
- struct chart *roc_chart = chart_create ();
-
- chart_write_title (roc_chart, _("ROC Curve"));
- chart_write_xlabel (roc_chart, _("1 - Specificity"));
- chart_write_ylabel (roc_chart, _("Sensitivity"));
-
- chart_write_xscale (roc_chart, 0, 1, 5);
- chart_write_yscale (roc_chart, 0, 1, 5);
-
- if ( roc->reference )
- {
- chart_line (roc_chart, 1.0, 0,
- 0.0, 1.0,
- CHART_DIM_X);
- }
-
- for (i = 0; i < roc->n_vars; ++i)
- {
- struct ccase *cc;
- struct casereader *r = casereader_clone (rs[i].cutpoint_rdr);
-
- chart_vector_start (roc_chart, var_get_name (roc->vars[i]));
- for (; (cc = casereader_read (r)) != NULL;
- case_unref (cc))
- {
- double se = case_data_idx (cc, TP)->f;
- double sp = case_data_idx (cc, TN)->f;
-
- se /= case_data_idx (cc, FN)->f +
- case_data_idx (cc, TP)->f ;
-
- sp /= case_data_idx (cc, TN)->f +
- case_data_idx (cc, FP)->f ;
-
- chart_vector (roc_chart, 1 - sp, se);
- }
- chart_vector_end (roc_chart);
- casereader_destroy (r);
- }
-
- chart_write_legend (roc_chart);
-
- chart_submit (roc_chart);
-}
-
-
static void
output_roc (struct roc_state *rs, const struct cmd_roc *roc)
{
show_summary (roc);
if ( roc->curve )
- draw_roc (rs, roc);
+ {
+ struct roc_chart *rc;
+ size_t i;
+
+ rc = roc_chart_create (roc->reference);
+ for (i = 0; i < roc->n_vars; i++)
+ roc_chart_add_var (rc, var_get_name (roc->vars[i]),
+ rs[i].cutpoint_rdr);
+ roc_chart_submit (rc);
+ }
show_auc (rs, roc);
-
if ( roc->print_coords )
show_coords (rs, roc);
}