1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2005, 2006, 2007, 2009, 2010, 2011, 2012 Free Software Foundation, Inc
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation, either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <http://www.gnu.org/licenses/>. */
20 #include <gsl/gsl_cdf.h>
22 #include "data/case.h"
23 #include "data/casegrouper.h"
24 #include "data/casereader.h"
25 #include "data/dataset.h"
26 #include "data/dictionary.h"
27 #include "data/format.h"
28 #include "data/variable.h"
29 #include "data/subcase.h"
30 #include "data/casewriter.h"
31 #include "data/short-names.h"
32 #include "language/command.h"
33 #include "language/lexer/lexer.h"
34 #include "language/lexer/variable-parser.h"
35 #include "language/stats/sort-criteria.h"
36 #include "math/sort.h"
37 #include "libpspp/assertion.h"
38 #include "libpspp/message.h"
39 #include "libpspp/misc.h"
40 #include "libpspp/pool.h"
41 #include "libpspp/taint.h"
43 #include "output/tab.h"
46 #define _(msgid) gettext (msgid)
47 #define N_(msgid) (msgid)
51 typedef double (*rank_function_t) (const struct rank*, double c, double cc, double cc_1,
54 static double rank_proportion (const struct rank *, double c, double cc, double cc_1,
57 static double rank_normal (const struct rank *, double c, double cc, double cc_1,
60 static double rank_percent (const struct rank *, double c, double cc, double cc_1,
63 static double rank_rfraction (const struct rank *, double c, double cc, double cc_1,
66 static double rank_rank (const struct rank *, double c, double cc, double cc_1,
69 static double rank_n (const struct rank *, double c, double cc, double cc_1,
72 static double rank_savage (const struct rank *, double c, double cc, double cc_1,
75 static double rank_ntiles (const struct rank *, double c, double cc, double cc_1,
92 static const struct fmt_spec dest_format[n_RANK_FUNCS] = {
93 {FMT_F, 9, 3}, /* rank */
94 {FMT_F, 6, 4}, /* normal */
95 {FMT_F, 6, 2}, /* percent */
96 {FMT_F, 6, 4}, /* rfraction */
97 {FMT_F, 6, 4}, /* proportion */
98 {FMT_F, 6, 0}, /* n */
99 {FMT_F, 3, 0}, /* ntiles */
100 {FMT_F, 8, 4} /* savage */
103 static const char * const function_name[n_RANK_FUNCS] = {
114 static const rank_function_t rank_func[n_RANK_FUNCS] = {
144 enum rank_func rfunc;
145 struct variable **destvars;
149 /* Create and return a new variable in which to store the ranks of SRC_VAR
150 accoring to the rank function F.
151 VNAME is the name of the variable to be created.
152 If VNAME is NULL, then a name will be automatically chosen.
154 static struct variable *
155 create_rank_variable (struct dictionary *dict, enum rank_func f,
156 const struct variable *src_var,
160 struct variable *var = NULL;
161 char name[SHORT_NAME_LEN + 1];
164 var = dict_create_var(dict, vname, 0);
168 snprintf (name, SHORT_NAME_LEN + 1, "%c%s",
169 function_name[f][0], var_get_name (src_var));
171 var = dict_create_var(dict, name, 0);
177 snprintf(func_abb, 4, "%s", function_name[f]);
178 snprintf(name, SHORT_NAME_LEN + 1, "%s%03d", func_abb,
181 var = dict_create_var(dict, name, 0);
187 while ( NULL == var )
190 snprintf(func_abb, 3, "%s", function_name[f]);
192 snprintf(name, SHORT_NAME_LEN + 1,
193 "RNK%s%02d", func_abb, i);
195 var = dict_create_var(dict, name, 0);
202 msg(ME, _("Cannot create new rank variable. All candidates in use."));
206 var_set_both_formats (var, &dest_format[f]);
213 struct dictionary *dict;
217 const struct variable **vars;
220 const struct variable **group_vars;
224 enum mv_class exclude;
226 struct rank_spec *rs;
231 enum fraction fraction;
236 /* Pool on which cell functions may allocate data */
242 destroy_rank (struct rank *rank)
245 free (rank->group_vars);
246 subcase_destroy (&rank->sc);
247 pool_destroy (rank->pool);
251 parse_into (struct lexer *lexer, struct rank *cmd)
254 struct rank_spec *rs = NULL;
256 cmd->rs = pool_realloc (cmd->pool, cmd->rs, sizeof (*cmd->rs) * (cmd->n_rs + 1));
257 rs = &cmd->rs[cmd->n_rs];
259 if (lex_match_id (lexer, "RANK"))
263 else if (lex_match_id (lexer, "NORMAL"))
267 else if (lex_match_id (lexer, "RFRACTION"))
269 rs->rfunc = RFRACTION;
271 else if (lex_match_id (lexer, "N"))
275 else if (lex_match_id (lexer, "SAVAGE"))
279 else if (lex_match_id (lexer, "PERCENT"))
283 else if (lex_match_id (lexer, "PROPORTION"))
285 rs->rfunc = PROPORTION;
287 else if (lex_match_id (lexer, "NTILES"))
289 if ( !lex_force_match (lexer, T_LPAREN))
292 if (! lex_force_int (lexer) )
295 cmd->k_ntiles = lex_integer (lexer);
298 if ( !lex_force_match (lexer, T_RPAREN))
310 rs->destvars = pool_calloc (cmd->pool, cmd->n_vars, sizeof (*rs->destvars));
312 if (lex_match_id (lexer, "INTO"))
314 while( lex_token (lexer) == T_ID )
316 const char *name = lex_tokcstr (lexer);
317 if ( dict_lookup_var (cmd->dict, name) != NULL )
319 msg (SE, _("Variable %s already exists."), name);
323 if ( var_count >= subcase_get_n_fields (&cmd->sc) )
325 msg (SE, _("Too many variables in INTO clause."));
328 rs->destvars[var_count] =
329 create_rank_variable (cmd->dict, rs->rfunc, cmd->vars[var_count], name);
338 /* Hardly a rank function !! */
340 rank_n (const struct rank *cmd UNUSED, double c UNUSED, double cc UNUSED, double cc_1 UNUSED,
341 int i UNUSED, double w)
348 rank_rank (const struct rank *cmd, double c, double cc, double cc_1,
349 int i, double w UNUSED)
364 rank = cc_1 + (c + 1.0)/ 2.0;
384 rank = cc_1 + c / 2.0 ;
399 rank_rfraction (const struct rank *cmd, double c, double cc, double cc_1,
402 return rank_rank (cmd, c, cc, cc_1, i, w) / w ;
407 rank_percent (const struct rank *cmd, double c, double cc, double cc_1,
410 return rank_rank (cmd, c, cc, cc_1, i, w) * 100.0 / w ;
415 rank_proportion (const struct rank *cmd, double c, double cc, double cc_1,
418 const double r = rank_rank (cmd, c, cc, cc_1, i, w) ;
422 switch ( cmd->fraction )
425 f = (r - 3.0/8.0) / (w + 0.25);
431 f = (r - 1.0/3.0) / (w + 1.0/3.0);
441 return (f > 0) ? f : SYSMIS;
445 rank_normal (const struct rank *cmd, double c, double cc, double cc_1,
448 double f = rank_proportion (cmd, c, cc, cc_1, i, w);
450 return gsl_cdf_ugaussian_Pinv (f);
454 rank_ntiles (const struct rank *cmd, double c, double cc, double cc_1,
457 double r = rank_rank (cmd, c, cc, cc_1, i, w);
460 return ( floor (( r * cmd->k_ntiles) / ( w + 1) ) + 1);
463 /* Expected value of the order statistics from an exponential distribution */
465 ee (int j, double w_star)
470 for (k = 1 ; k <= j; k++)
471 sum += 1.0 / ( w_star + 1 - k );
478 rank_savage (const struct rank *cmd UNUSED, double c, double cc, double cc_1,
479 int i UNUSED, double w)
482 const int i_1 = floor (cc_1);
483 const int i_2 = floor (cc);
485 const double w_star = (modf (w, &int_part) == 0 ) ? w : floor (w) + 1;
487 const double g_1 = cc_1 - i_1;
488 const double g_2 = cc - i_2;
490 /* The second factor is infinite, when the first is zero.
491 Therefore, evaluate the second, only when the first is non-zero */
492 const double expr1 = (1 - g_1) ? (1 - g_1) * ee(i_1+1, w_star) : ( 1 - g_1);
493 const double expr2 = g_2 ? g_2 * ee (i_2+1, w_star) : g_2 ;
496 return ee (i_1 + 1, w_star) - 1;
498 if ( i_1 + 1 == i_2 )
499 return ( ( expr1 + expr2 )/c ) - 1;
501 if ( i_1 + 2 <= i_2 )
505 for (j = i_1 + 2 ; j <= i_2; ++j )
506 sigma += ee (j, w_star);
507 return ( (expr1 + expr2 + sigma) / c) -1;
515 rank_sorted_file (struct casereader *input,
516 struct casewriter *output,
517 const struct dictionary *dict,
519 const struct rank *cmd
522 struct casereader *pass1, *pass2, *pass2_1;
523 struct casegrouper *tie_grouper;
529 input = casereader_create_filter_missing (input, &cmd->vars[dest_idx], 1,
530 cmd->exclude, NULL, output);
531 input = casereader_create_filter_weight (input, dict, NULL, output);
533 casereader_split (input, &pass1, &pass2);
535 /* Pass 1: Get total group weight. */
536 for (; (c = casereader_read (pass1)) != NULL; case_unref (c))
537 w += dict_get_case_weight (dict, c, NULL);
538 casereader_destroy (pass1);
540 /* Pass 2: Do ranking. */
541 tie_grouper = casegrouper_create_vars (pass2, &cmd->vars[dest_idx], 1);
542 while (casegrouper_get_next_group (tie_grouper, &pass2_1))
544 struct casereader *pass2_2;
549 pass2_2 = casereader_clone (pass2_1);
550 taint_propagate (casereader_get_taint (pass2_2),
551 casewriter_get_taint (output));
553 /* Pass 2.1: Sum up weight for tied cases. */
554 for (; (c = casereader_read (pass2_1)) != NULL; case_unref (c))
555 tw += dict_get_case_weight (dict, c, NULL);
557 casereader_destroy (pass2_1);
559 /* Pass 2.2: Rank tied cases. */
560 while ((c = casereader_read (pass2_2)) != NULL)
562 c = case_unshare (c);
563 for (i = 0; i < cmd->n_rs; ++i)
565 const struct variable *dst_var = cmd->rs[i].destvars[dest_idx];
566 double *dst_value = &case_data_rw (c, dst_var)->f;
567 *dst_value = rank_func[cmd->rs[i].rfunc] (cmd, tw, cc, cc_1, tie_group, w);
569 casewriter_write (output, c);
571 casereader_destroy (pass2_2);
575 casegrouper_destroy (tie_grouper);
579 /* Transformation function to enumerate all the cases */
581 create_resort_key (void *key_var_, struct ccase **cc, casenumber case_num)
583 struct variable *key_var = key_var_;
585 *cc = case_unshare (*cc);
586 case_data_rw (*cc, key_var)->f = case_num;
588 return TRNS_CONTINUE;
592 rank_cmd (struct dataset *ds, const struct rank *cmd);
596 fraction_name (const struct rank *cmd)
598 static char name[10];
599 switch (cmd->fraction )
602 strcpy (name, "BLOM");
605 strcpy (name, "RANKIT");
608 strcpy (name, "TUKEY");
619 /* Create a label on DEST_VAR, describing its derivation from SRC_VAR and F */
621 create_var_label (struct rank *cmd, struct variable *dest_var,
622 const struct variable *src_var, enum rank_func f)
625 ds_init_empty (&label);
627 if ( cmd->n_group_vars > 0 )
629 struct string group_var_str;
632 ds_init_empty (&group_var_str);
634 for (g = 0 ; g < cmd->n_group_vars ; ++g )
636 if ( g > 0 ) ds_put_cstr (&group_var_str, " ");
637 ds_put_cstr (&group_var_str, var_get_name (cmd->group_vars[g]));
640 ds_put_format (&label, _("%s of %s by %s"), function_name[f],
641 var_get_name (src_var), ds_cstr (&group_var_str));
642 ds_destroy (&group_var_str);
645 ds_put_format (&label, _("%s of %s"),
646 function_name[f], var_get_name (src_var));
648 var_set_label (dest_var, ds_cstr (&label), false);
654 cmd_rank (struct lexer *lexer, struct dataset *ds)
657 struct variable *order;
661 subcase_init_empty (&rank.sc);
665 rank.exclude = MV_ANY;
666 rank.n_group_vars = 0;
667 rank.group_vars = NULL;
668 rank.dict = dataset_dict (ds);
669 rank.ties = TIES_MEAN;
670 rank.fraction = FRAC_BLOM;
672 rank.pool = pool_create ();
674 if (lex_match_id (lexer, "VARIABLES"))
675 lex_force_match (lexer, T_EQUALS);
677 if (!parse_sort_criteria (lexer, rank.dict,
682 rank.n_vars = rank.sc.n_fields;
684 if (lex_match (lexer, T_BY) )
686 if ( ! parse_variables_const (lexer, rank.dict,
687 &rank.group_vars, &rank.n_group_vars,
688 PV_NO_DUPLICATE | PV_NO_SCRATCH))
693 while (lex_token (lexer) != T_ENDCMD )
695 lex_force_match (lexer, T_SLASH);
696 if (lex_match_id (lexer, "TIES"))
698 lex_force_match (lexer, T_EQUALS);
699 if (lex_match_id (lexer, "MEAN"))
701 rank.ties = TIES_MEAN;
703 else if (lex_match_id (lexer, "LOW"))
705 rank.ties = TIES_LOW;
707 else if (lex_match_id (lexer, "HIGH"))
709 rank.ties = TIES_HIGH;
711 else if (lex_match_id (lexer, "CONDENSE"))
713 rank.ties = TIES_CONDENSE;
717 lex_error (lexer, NULL);
721 else if (lex_match_id (lexer, "FRACTION"))
723 lex_force_match (lexer, T_EQUALS);
724 if (lex_match_id (lexer, "BLOM"))
726 rank.fraction = FRAC_BLOM;
728 else if (lex_match_id (lexer, "TUKEY"))
730 rank.fraction = FRAC_TUKEY;
732 else if (lex_match_id (lexer, "VW"))
734 rank.fraction = FRAC_VW;
736 else if (lex_match_id (lexer, "RANKIT"))
738 rank.fraction = FRAC_RANKIT;
742 lex_error (lexer, NULL);
746 else if (lex_match_id (lexer, "PRINT"))
748 lex_force_match (lexer, T_EQUALS);
749 if (lex_match_id (lexer, "YES"))
753 else if (lex_match_id (lexer, "NO"))
759 lex_error (lexer, NULL);
763 else if (lex_match_id (lexer, "MISSING"))
765 lex_force_match (lexer, T_EQUALS);
766 if (lex_match_id (lexer, "INCLUDE"))
768 rank.exclude = MV_SYSTEM;
770 else if (lex_match_id (lexer, "EXCLUDE"))
772 rank.exclude = MV_ANY;
776 lex_error (lexer, NULL);
780 else if (! parse_into (lexer, &rank))
785 /* If no rank specs are given, then apply a default */
788 rank.rs = pool_calloc (rank.pool, 1, sizeof (*rank.rs));
790 rank.rs[0].rfunc = RANK;
791 rank.rs[0].destvars = pool_calloc (rank.pool, rank.n_vars, sizeof (*rank.rs[0].destvars));
794 /* Create variables for all rank destinations which haven't
795 already been created with INTO.
796 Add labels to all the destination variables.
798 for (i = 0 ; i < rank.n_rs ; ++i )
801 struct rank_spec *rs = &rank.rs[i];
803 for ( v = 0 ; v < rank.n_vars ; v ++ )
805 if ( rs->destvars[v] == NULL )
808 create_rank_variable (rank.dict, rs->rfunc, rank.vars[v], NULL);
811 create_var_label (&rank, rs->destvars[v],
821 tab_output_text (0, _("Variables Created By RANK"));
822 tab_output_text (0, "");
824 for (i = 0 ; i < rank.n_rs ; ++i )
826 for ( v = 0 ; v < rank.n_vars ; v ++ )
828 if ( rank.n_group_vars > 0 )
830 struct string varlist;
833 ds_init_empty (&varlist);
834 for ( g = 0 ; g < rank.n_group_vars ; ++g )
836 ds_put_cstr (&varlist, var_get_name (rank.group_vars[g]));
838 if ( g < rank.n_group_vars - 1)
839 ds_put_cstr (&varlist, " ");
842 if ( rank.rs[i].rfunc == NORMAL ||
843 rank.rs[i].rfunc == PROPORTION )
844 tab_output_text_format (0,
845 _("%s into %s(%s of %s using %s BY %s)"),
846 var_get_name (rank.vars[v]),
847 var_get_name (rank.rs[i].destvars[v]),
848 function_name[rank.rs[i].rfunc],
849 var_get_name (rank.vars[v]),
850 fraction_name (&rank),
854 tab_output_text_format (0,
855 _("%s into %s(%s of %s BY %s)"),
856 var_get_name (rank.vars[v]),
857 var_get_name (rank.rs[i].destvars[v]),
858 function_name[rank.rs[i].rfunc],
859 var_get_name (rank.vars[v]),
861 ds_destroy (&varlist);
865 if ( rank.rs[i].rfunc == NORMAL ||
866 rank.rs[i].rfunc == PROPORTION )
867 tab_output_text_format (0,
868 _("%s into %s(%s of %s using %s)"),
869 var_get_name (rank.vars[v]),
870 var_get_name (rank.rs[i].destvars[v]),
871 function_name[rank.rs[i].rfunc],
872 var_get_name (rank.vars[v]),
873 fraction_name (&rank));
876 tab_output_text_format (0,
877 _("%s into %s(%s of %s)"),
878 var_get_name (rank.vars[v]),
879 var_get_name (rank.rs[i].destvars[v]),
880 function_name[rank.rs[i].rfunc],
881 var_get_name (rank.vars[v]));
887 /* Add a variable which we can sort by to get back the original
889 order = dict_create_var_assert (dataset_dict (ds), "$ORDER_", 0);
891 add_transformation (ds, create_resort_key, 0, order);
894 result = rank_cmd (ds, &rank);
896 /* Put the active dataset back in its original order. Delete
897 our sort key, which we don't need anymore. */
899 struct casereader *sorted;
902 /* FIXME: loses error conditions. */
904 proc_discard_output (ds);
905 sorted = sort_execute_1var (proc_open (ds), order);
906 result = proc_commit (ds) && result;
908 dict_delete_var (dataset_dict (ds), order);
909 result = dataset_set_source (ds, sorted) && result;
914 destroy_rank (&rank);
919 destroy_rank (&rank);
926 rank_cmd (struct dataset *ds, const struct rank *cmd)
928 struct dictionary *d = dataset_dict (ds);
932 for (i = 0 ; i < subcase_get_n_fields (&cmd->sc) ; ++i )
934 /* Rank variable at index I in SC. */
935 struct casegrouper *split_grouper;
936 struct casereader *split_group;
937 struct casewriter *output;
939 proc_discard_output (ds);
940 split_grouper = casegrouper_create_splits (proc_open (ds), d);
941 output = autopaging_writer_create (dict_get_proto (d));
943 while (casegrouper_get_next_group (split_grouper, &split_group))
945 struct subcase ordering;
946 struct casereader *ordered;
947 struct casegrouper *by_grouper;
948 struct casereader *by_group;
950 /* Sort this split group by the BY variables as primary
951 keys and the rank variable as secondary key. */
952 subcase_init_vars (&ordering, cmd->group_vars, cmd->n_group_vars);
953 subcase_add_var (&ordering, cmd->vars[i],
954 subcase_get_direction (&cmd->sc, i));
955 ordered = sort_execute (split_group, &ordering);
956 subcase_destroy (&ordering);
958 /* Rank the rank variable within this split group. */
959 by_grouper = casegrouper_create_vars (ordered,
960 cmd->group_vars, cmd->n_group_vars);
961 while (casegrouper_get_next_group (by_grouper, &by_group))
963 /* Rank the rank variable within this BY group
964 within the split group. */
966 rank_sorted_file (by_group, output, d, i, cmd);
969 ok = casegrouper_destroy (by_grouper) && ok;
971 ok = casegrouper_destroy (split_grouper);
972 ok = proc_commit (ds) && ok;
973 ok = (dataset_set_source (ds, casewriter_make_reader (output))