X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=src%2Flanguage%2Fstats%2Flogistic.c;h=b8ebb70cf7f452c10dafe81b8f787b03a2974861;hb=d4ff0e074d703dbeb8af5aa3ac470ddda5ebe301;hp=fb9a26521a7273a948641de57fc07a0438d0c020;hpb=9d1bfb34842de4a129140622ee3d800297c0e69d;p=pspp
diff --git a/src/language/stats/logistic.c b/src/language/stats/logistic.c
index fb9a26521a..b8ebb70cf7 100644
--- a/src/language/stats/logistic.c
+++ b/src/language/stats/logistic.c
@@ -15,21 +15,21 @@
along with this program. If not, see . */
-/*
- References:
+/*
+ References:
1. "Coding Logistic Regression with Newton-Raphson", James McCaffrey
http://msdn.microsoft.com/en-us/magazine/jj618304.aspx
2. "SPSS Statistical Algorithms" Chapter LOGISTIC REGRESSION Algorithms
- The Newton Raphson method finds successive approximations to $\bf b$ where
+ The Newton Raphson method finds successive approximations to $\bf b$ where
approximation ${\bf b}_t$ is (hopefully) better than the previous ${\bf b}_{t-1}$.
$ {\bf b}_t = {\bf b}_{t -1} + ({\bf X}^T{\bf W}_{t-1}{\bf X})^{-1}{\bf X}^T({\bf y} - {\bf \pi}_{t-1})$
where:
- $\bf X$ is the $n \times p$ design matrix, $n$ being the number of cases,
+ $\bf X$ is the $n \times p$ design matrix, $n$ being the number of cases,
$p$ the number of parameters, \par
$\bf W$ is the diagonal matrix whose diagonal elements are
$\hat{\pi}_0(1 - \hat{\pi}_0), \, \hat{\pi}_1(1 - \hat{\pi}_2)\dots \hat{\pi}_{n-1}(1 - \hat{\pi}_{n-1})$
@@ -39,7 +39,7 @@
#include
-#include
+#include
#include
#include
@@ -64,6 +64,10 @@
#include "libpspp/message.h"
#include "libpspp/misc.h"
#include "math/categoricals.h"
+#include "math/interaction.h"
+#include "libpspp/hmap.h"
+#include "libpspp/hash-functions.h"
+
#include "output/tab.h"
#include "gettext.h"
@@ -91,8 +95,19 @@ struct lr_spec
/* The dependent variable */
const struct variable *dep_var;
- size_t n_predictor_vars;
+ /* The predictor variables (excluding categorical ones) */
const struct variable **predictor_vars;
+ size_t n_predictor_vars;
+
+ /* The categorical predictors */
+ struct interaction **cat_predictors;
+ size_t n_cat_predictors;
+
+
+ /* The union of the categorical and non-categorical variables */
+ const struct variable **indep_vars;
+ size_t n_indep_vars;
+
/* Which classes of missing vars are to be excluded */
enum mv_class exclude;
@@ -100,6 +115,7 @@ struct lr_spec
/* The weight variable */
const struct variable *wv;
+ /* The dictionary of the dataset */
const struct dictionary *dict;
/* True iff the constant (intercept) is to be included in the model */
@@ -119,18 +135,21 @@ struct lr_spec
/* What results should be presented */
unsigned int print;
- double cut_point;
+ /* Inverse logit of the cut point */
+ double ilogit_cut_point;
};
+
/* The results and intermediate result of the procedure.
These are mutated as the procedure runs. Used for
temporary variables etc.
*/
struct lr_result
{
+ /* Used to indicate if a pass should flag a warning when
+ invalid (ie negative or missing) weight values are encountered */
bool warn_bad_weight;
-
/* The two values of the dependent variable. */
union value y0;
union value y1;
@@ -139,37 +158,58 @@ struct lr_result
/* The sum of caseweights */
double cc;
+ /* The number of missing and nonmissing cases */
casenumber n_missing;
casenumber n_nonmissing;
+
+
+ gsl_matrix *hessian;
+
+ /* The categoricals and their payload. Null if the analysis has no
+ categorical predictors */
+ struct categoricals *cats;
+ struct payload cp;
+
+
+ /* The estimates of the predictor coefficients */
+ gsl_vector *beta_hat;
+
+ /* The predicted classifications:
+ True Negative, True Positive, False Negative, False Positive */
+ double tn, tp, fn, fp;
};
/*
- Convert INPUT into a dichotomous scalar. For simple cases, this is a 1:1 mapping
+ Convert INPUT into a dichotomous scalar, according to how the dependent variable's
+ values are mapped.
+ For simple cases, this is a 1:1 mapping
The return value is always either 0 or 1
*/
static double
map_dependent_var (const struct lr_spec *cmd, const struct lr_result *res, const union value *input)
{
- int width = var_get_width (cmd->dep_var);
+ const int width = var_get_width (cmd->dep_var);
if (value_equal (input, &res->y0, width))
return 0;
if (value_equal (input, &res->y1, width))
return 1;
-
+
+ /* This should never happen. If it does, then y0 and/or y1 have probably not been set */
NOT_REACHED ();
return SYSMIS;
}
+static void output_classification_table (const struct lr_spec *cmd, const struct lr_result *res);
+static void output_categories (const struct lr_spec *cmd, const struct lr_result *res);
static void output_depvarmap (const struct lr_spec *cmd, const struct lr_result *);
-static void output_variables (const struct lr_spec *cmd,
- const gsl_vector *,
- const gsl_vector *);
+static void output_variables (const struct lr_spec *cmd,
+ const struct lr_result *);
static void output_model_summary (const struct lr_result *,
double initial_likelihood, double likelihood);
@@ -177,26 +217,55 @@ static void output_model_summary (const struct lr_result *,
static void case_processing_summary (const struct lr_result *);
+/* Return the value of case C corresponding to the INDEX'th entry in the
+ model */
+static double
+predictor_value (const struct ccase *c,
+ const struct variable **x, size_t n_x,
+ const struct categoricals *cats,
+ size_t index)
+{
+ /* Values of the scalar predictor variables */
+ if (index < n_x)
+ return case_data (c, x[index])->f;
+
+ /* Coded values of categorical predictor variables (or interactions) */
+ if (cats && index - n_x < categoricals_df_total (cats))
+ {
+ double x = categoricals_get_dummy_code_for_case (cats, index - n_x, c);
+ return x;
+ }
+
+ /* The constant term */
+ return 1.0;
+}
+
+
/*
- Return the probability estimator (that is the estimator of logit(y) )
- corresponding to the coefficient estimator beta_hat for case C
+ Return the probability beta_hat (that is the estimator logit(y) )
+ corresponding to the coefficient estimator for case C
*/
-static double
-pi_hat (const struct lr_spec *cmd,
- const gsl_vector *beta_hat,
+static double
+pi_hat (const struct lr_spec *cmd,
+ const struct lr_result *res,
const struct variable **x, size_t n_x,
const struct ccase *c)
{
int v0;
double pi = 0;
- for (v0 = 0; v0 < n_x; ++v0)
+ size_t n_coeffs = res->beta_hat->size;
+
+ if (cmd->constant)
{
- pi += gsl_vector_get (beta_hat, v0) *
- case_data (c, x[v0])->f;
+ pi += gsl_vector_get (res->beta_hat, res->beta_hat->size - 1);
+ n_coeffs--;
}
- if (cmd->constant)
- pi += gsl_vector_get (beta_hat, beta_hat->size - 1);
+ for (v0 = 0; v0 < n_coeffs; ++v0)
+ {
+ pi += gsl_vector_get (res->beta_hat, v0) *
+ predictor_value (c, x, n_x, res->cats, v0);
+ }
pi = 1.0 / (1.0 + exp(-pi));
@@ -207,32 +276,31 @@ pi_hat (const struct lr_spec *cmd,
/*
Calculates the Hessian matrix X' V X,
where: X is the n by N_X matrix comprising the n cases in INPUT
- V is a diagonal matrix { (pi_hat_0)(1 - pi_hat_0), (pi_hat_1)(1 - pi_hat_1), ... (pi_hat_{N-1})(1 - pi_hat_{N-1})}
+ V is a diagonal matrix { (pi_hat_0)(1 - pi_hat_0), (pi_hat_1)(1 - pi_hat_1), ... (pi_hat_{N-1})(1 - pi_hat_{N-1})}
(the partial derivative of the predicted values)
If ALL predicted values derivatives are close to zero or one, then CONVERGED
will be set to true.
*/
-static gsl_matrix *
-hessian (const struct lr_spec *cmd,
+static void
+hessian (const struct lr_spec *cmd,
struct lr_result *res,
struct casereader *input,
const struct variable **x, size_t n_x,
- const gsl_vector *beta_hat,
- bool *converged
- )
+ bool *converged)
{
struct casereader *reader;
struct ccase *c;
- gsl_matrix *output = gsl_matrix_calloc (beta_hat->size, beta_hat->size);
double max_w = -DBL_MAX;
+ gsl_matrix_set_zero (res->hessian);
+
for (reader = casereader_clone (input);
(c = casereader_read (reader)) != NULL; case_unref (c))
{
int v0, v1;
- double pi = pi_hat (cmd, beta_hat, x, n_x, c);
+ double pi = pi_hat (cmd, res, x, n_x, c);
double weight = dict_get_case_weight (cmd->dict, c, &res->warn_bad_weight);
double w = pi * (1 - pi);
@@ -240,36 +308,35 @@ hessian (const struct lr_spec *cmd,
max_w = w;
w *= weight;
- for (v0 = 0; v0 < beta_hat->size; ++v0)
+ for (v0 = 0; v0 < res->beta_hat->size; ++v0)
{
- double in0 = v0 < n_x ? case_data (c, x[v0])->f : 1.0;
- for (v1 = 0; v1 < beta_hat->size; ++v1)
+ double in0 = predictor_value (c, x, n_x, res->cats, v0);
+ for (v1 = 0; v1 < res->beta_hat->size; ++v1)
{
- double in1 = v1 < n_x ? case_data (c, x[v1])->f : 1.0 ;
- double *o = gsl_matrix_ptr (output, v0, v1);
+ double in1 = predictor_value (c, x, n_x, res->cats, v1);
+ double *o = gsl_matrix_ptr (res->hessian, v0, v1);
*o += in0 * w * in1;
}
}
}
casereader_destroy (reader);
-
if ( max_w < cmd->min_epsilon)
{
*converged = true;
msg (MN, _("All predicted values are either 1 or 0"));
}
-
- return output;
}
/* Calculates the value X' (y - pi)
- where X is the design model,
+ where X is the design model,
y is the vector of observed independent variables
pi is the vector of estimates for y
- As a side effect, the likelihood is stored in LIKELIHOOD
+ Side effects:
+ the likelihood is stored in LIKELIHOOD;
+ the predicted values are placed in the respective tn, fn, tp fp values in RES
*/
static gsl_vector *
xt_times_y_pi (const struct lr_spec *cmd,
@@ -277,31 +344,50 @@ xt_times_y_pi (const struct lr_spec *cmd,
struct casereader *input,
const struct variable **x, size_t n_x,
const struct variable *y_var,
- const gsl_vector *beta_hat,
- double *likelihood)
+ double *llikelihood)
{
struct casereader *reader;
struct ccase *c;
- gsl_vector *output = gsl_vector_calloc (beta_hat->size);
+ gsl_vector *output = gsl_vector_calloc (res->beta_hat->size);
- *likelihood = 1.0;
+ *llikelihood = 0.0;
+ res->tn = res->tp = res->fn = res->fp = 0;
for (reader = casereader_clone (input);
(c = casereader_read (reader)) != NULL; case_unref (c))
{
+ double pred_y = 0;
int v0;
- double pi = pi_hat (cmd, beta_hat, x, n_x, c);
+ double pi = pi_hat (cmd, res, x, n_x, c);
double weight = dict_get_case_weight (cmd->dict, c, &res->warn_bad_weight);
double y = map_dependent_var (cmd, res, case_data (c, y_var));
- *likelihood *= pow (pi, weight * y) * pow (1 - pi, weight * (1 - y));
+ *llikelihood += (weight * y) * log (pi) + log (1 - pi) * weight * (1 - y);
- for (v0 = 0; v0 < beta_hat->size; ++v0)
+ for (v0 = 0; v0 < res->beta_hat->size; ++v0)
{
- double in0 = v0 < n_x ? case_data (c, x[v0])->f : 1.0;
+ double in0 = predictor_value (c, x, n_x, res->cats, v0);
double *o = gsl_vector_ptr (output, v0);
*o += in0 * (y - pi) * weight;
+ pred_y += gsl_vector_get (res->beta_hat, v0) * in0;
+ }
+
+ /* Count the number of cases which would be correctly/incorrectly classified by this
+ estimated model */
+ if (pred_y <= cmd->ilogit_cut_point)
+ {
+ if (y == 0)
+ res->tn += weight;
+ else
+ res->fn += weight;
+ }
+ else
+ {
+ if (y == 0)
+ res->fp += weight;
+ else
+ res->tp += weight;
}
}
@@ -310,21 +396,54 @@ xt_times_y_pi (const struct lr_spec *cmd,
return output;
}
+
-/*
- Makes an initial pass though the data, checks that the dependent variable is
- dichotomous, and calculates necessary initial values.
+/* "payload" functions for the categoricals.
+ The only function is to accumulate the frequency of each
+ category.
+ */
- Returns an initial value for \hat\beta the vector of estimators of \beta
+static void *
+frq_create (const void *aux1 UNUSED, void *aux2 UNUSED)
+{
+ return xzalloc (sizeof (double));
+}
+
+static void
+frq_update (const void *aux1 UNUSED, void *aux2 UNUSED,
+ void *ud, const struct ccase *c UNUSED , double weight)
+{
+ double *freq = ud;
+ *freq += weight;
+}
+
+static void
+frq_destroy (const void *aux1 UNUSED, void *aux2 UNUSED, void *user_data UNUSED)
+{
+ free (user_data);
+}
+
+
+
+/*
+ Makes an initial pass though the data, doing the following:
+
+ * Checks that the dependent variable is dichotomous,
+ * Creates and initialises the categoricals,
+ * Accumulates summary results,
+ * Calculates necessary initial values.
+ * Creates an initial value for \hat\beta the vector of beta_hats of \beta
+
+ Returns true if successful
*/
-static gsl_vector *
-beta_hat_initial (const struct lr_spec *cmd, struct lr_result *res, struct casereader *input)
+static bool
+initial_pass (const struct lr_spec *cmd, struct lr_result *res, struct casereader *input)
{
const int width = var_get_width (cmd->dep_var);
struct ccase *c;
struct casereader *reader;
- gsl_vector *b0 ;
+
double sum;
double sumA = 0.0;
double sumB = 0.0;
@@ -336,7 +455,19 @@ beta_hat_initial (const struct lr_spec *cmd, struct lr_result *res, struct caser
if (cmd->constant)
n_coefficients++;
- b0 = gsl_vector_calloc (n_coefficients);
+ /* Create categoricals if appropriate */
+ if (cmd->n_cat_predictors > 0)
+ {
+ res->cp.create = frq_create;
+ res->cp.update = frq_update;
+ res->cp.calculate = NULL;
+ res->cp.destroy = frq_destroy;
+
+ res->cats = categoricals_create (cmd->cat_predictors, cmd->n_cat_predictors,
+ cmd->wv, cmd->exclude, MV_ANY);
+
+ categoricals_set_payload (res->cats, &res->cp, cmd, res);
+ }
res->cc = 0;
for (reader = casereader_clone (input);
@@ -347,24 +478,30 @@ beta_hat_initial (const struct lr_spec *cmd, struct lr_result *res, struct caser
double weight = dict_get_case_weight (cmd->dict, c, &res->warn_bad_weight);
const union value *depval = case_data (c, cmd->dep_var);
- for (v = 0; v < cmd->n_predictor_vars; ++v)
+ if (var_is_value_missing (cmd->dep_var, depval, cmd->exclude))
{
- const union value *val = case_data (c, cmd->predictor_vars[v]);
- if (var_is_value_missing (cmd->predictor_vars[v], val, cmd->exclude))
+ missing = true;
+ }
+ else
+ for (v = 0; v < cmd->n_indep_vars; ++v)
+ {
+ const union value *val = case_data (c, cmd->indep_vars[v]);
+ if (var_is_value_missing (cmd->indep_vars[v], val, cmd->exclude))
{
missing = true;
break;
}
}
+ /* Accumulate the missing and non-missing counts */
if (missing)
{
res->n_missing++;
continue;
}
-
res->n_nonmissing++;
+ /* Find the values of the dependent variable */
if (!v0set)
{
value_clone (&res->y0, depval, width);
@@ -386,21 +523,26 @@ beta_hat_initial (const struct lr_spec *cmd, struct lr_result *res, struct caser
)
{
msg (ME, _("Dependent variable's values are not dichotomous."));
+ case_unref (c);
goto error;
}
}
- if (value_equal (&res->y0, depval, width))
+ if (v0set && value_equal (&res->y0, depval, width))
sumA += weight;
- if (value_equal (&res->y1, depval, width))
+ if (v1set && value_equal (&res->y1, depval, width))
sumB += weight;
res->cc += weight;
+
+ categoricals_update (res->cats, c);
}
casereader_destroy (reader);
+ categoricals_done (res->cats);
+
sum = sumB;
/* Ensure that Y0 is less than Y1. Otherwise the mapping gets
@@ -415,97 +557,115 @@ beta_hat_initial (const struct lr_spec *cmd, struct lr_result *res, struct caser
sum = sumA;
}
- if ( cmd->constant)
+ n_coefficients += categoricals_df_total (res->cats);
+ res->beta_hat = gsl_vector_calloc (n_coefficients);
+
+ if (cmd->constant)
{
double mean = sum / res->cc;
- gsl_vector_set (b0, b0->size - 1, log (mean / (1 - mean)));
+ gsl_vector_set (res->beta_hat, res->beta_hat->size - 1, log (mean / (1 - mean)));
}
- return b0;
+ return true;
error:
casereader_destroy (reader);
- return NULL;
+ return false;
}
+/* Start of the logistic regression routine proper */
static bool
run_lr (const struct lr_spec *cmd, struct casereader *input,
const struct dataset *ds UNUSED)
{
- int i,j;
-
- gsl_vector *beta_hat;
- gsl_vector *se ;
+ int i;
bool converged = false;
- double likelihood = -1;
- double prev_likelihood = -1;
- double initial_likelihood = -1;
+
+ /* Set the log likelihoods to a sentinel value */
+ double log_likelihood = SYSMIS;
+ double prev_log_likelihood = SYSMIS;
+ double initial_log_likelihood = SYSMIS;
struct lr_result work;
work.n_missing = 0;
work.n_nonmissing = 0;
work.warn_bad_weight = true;
+ work.cats = NULL;
+ work.beta_hat = NULL;
+ work.hessian = NULL;
+ /* Get the initial estimates of \beta and their standard errors.
+ And perform other auxilliary initialisation. */
+ if (! initial_pass (cmd, &work, input))
+ goto error;
- /* Get the initial estimates of \beta and their standard errors */
- beta_hat = beta_hat_initial (cmd, &work, input);
- if (NULL == beta_hat)
- return false;
+ for (i = 0; i < cmd->n_cat_predictors; ++i)
+ {
+ if (1 >= categoricals_n_count (work.cats, i))
+ {
+ struct string str;
+ ds_init_empty (&str);
- output_depvarmap (cmd, &work);
+ interaction_to_string (cmd->cat_predictors[i], &str);
- se = gsl_vector_alloc (beta_hat->size);
+ msg (ME, _("Category %s does not have at least two distinct values. Logistic regression will not be run."),
+ ds_cstr(&str));
+ ds_destroy (&str);
+ goto error;
+ }
+ }
+
+ output_depvarmap (cmd, &work);
case_processing_summary (&work);
input = casereader_create_filter_missing (input,
- cmd->predictor_vars,
- cmd->n_predictor_vars,
+ cmd->indep_vars,
+ cmd->n_indep_vars,
+ cmd->exclude,
+ NULL,
+ NULL);
+
+ input = casereader_create_filter_missing (input,
+ &cmd->dep_var,
+ 1,
cmd->exclude,
NULL,
NULL);
+ work.hessian = gsl_matrix_calloc (work.beta_hat->size, work.beta_hat->size);
/* Start the Newton Raphson iteration process... */
for( i = 0 ; i < cmd->max_iter ; ++i)
{
double min, max;
- gsl_matrix *m ;
gsl_vector *v ;
- m = hessian (cmd, &work, input,
- cmd->predictor_vars, cmd->n_predictor_vars,
- beta_hat,
- &converged);
- gsl_linalg_cholesky_decomp (m);
- gsl_linalg_cholesky_invert (m);
+ hessian (cmd, &work, input,
+ cmd->predictor_vars, cmd->n_predictor_vars,
+ &converged);
+
+ gsl_linalg_cholesky_decomp (work.hessian);
+ gsl_linalg_cholesky_invert (work.hessian);
v = xt_times_y_pi (cmd, &work, input,
cmd->predictor_vars, cmd->n_predictor_vars,
cmd->dep_var,
- beta_hat,
- &likelihood);
+ &log_likelihood);
{
/* delta = M.v */
gsl_vector *delta = gsl_vector_alloc (v->size);
- gsl_blas_dgemv (CblasNoTrans, 1.0, m, v, 0, delta);
+ gsl_blas_dgemv (CblasNoTrans, 1.0, work.hessian, v, 0, delta);
gsl_vector_free (v);
- for (j = 0; j < se->size; ++j)
- {
- double *ptr = gsl_vector_ptr (se, j);
- *ptr = gsl_matrix_get (m, j, j);
- }
- gsl_matrix_free (m);
-
- gsl_vector_add (beta_hat, delta);
+ gsl_vector_add (work.beta_hat, delta);
gsl_vector_minmax (delta, &min, &max);
@@ -519,43 +679,86 @@ run_lr (const struct lr_spec *cmd, struct casereader *input,
gsl_vector_free (delta);
}
- if ( prev_likelihood >= 0)
+ if (i > 0)
{
- if (-log (likelihood) > -(1.0 - cmd->lcon) * log (prev_likelihood))
+ if (-log_likelihood > -(1.0 - cmd->lcon) * prev_log_likelihood)
{
msg (MN, _("Estimation terminated at iteration number %d because Log Likelihood decreased by less than %g%%"), i + 1, 100 * cmd->lcon);
converged = true;
}
}
if (i == 0)
- initial_likelihood = likelihood;
- prev_likelihood = likelihood;
+ initial_log_likelihood = log_likelihood;
+ prev_log_likelihood = log_likelihood;
if (converged)
break;
}
- casereader_destroy (input);
- assert (initial_likelihood >= 0);
- for (i = 0; i < se->size; ++i)
- {
- double *ptr = gsl_vector_ptr (se, i);
- *ptr = sqrt (*ptr);
- }
- output_model_summary (&work, initial_likelihood, likelihood);
- output_variables (cmd, beta_hat, se);
- gsl_vector_free (beta_hat);
- gsl_vector_free (se);
+ if ( ! converged)
+ msg (MW, _("Estimation terminated at iteration number %d because maximum iterations has been reached"), i );
+
+
+ output_model_summary (&work, initial_log_likelihood, log_likelihood);
+
+ if (work.cats)
+ output_categories (cmd, &work);
+
+ output_classification_table (cmd, &work);
+ output_variables (cmd, &work);
+
+ casereader_destroy (input);
+ gsl_matrix_free (work.hessian);
+ gsl_vector_free (work.beta_hat);
+ categoricals_destroy (work.cats);
return true;
+
+ error:
+ casereader_destroy (input);
+ gsl_matrix_free (work.hessian);
+ gsl_vector_free (work.beta_hat);
+ categoricals_destroy (work.cats);
+
+ return false;
+}
+
+struct variable_node
+{
+ struct hmap_node node; /* Node in hash map. */
+ const struct variable *var; /* The variable */
+};
+
+static struct variable_node *
+lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
+{
+ struct variable_node *vn = NULL;
+ HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
+ {
+ if (vn->var == var)
+ break;
+
+ fprintf (stderr, "Warning: Hash table collision\n");
+ }
+
+ return vn;
}
+
/* Parse the LOGISTIC REGRESSION command syntax */
int
cmd_logistic (struct lexer *lexer, struct dataset *ds)
{
+ int i;
+ /* Temporary location for the predictor variables.
+ These may or may not include the categorical predictors */
+ const struct variable **pred_vars;
+ size_t n_pred_vars;
+ double cp = 0.5;
+
+ int v, x;
struct lr_spec lr;
lr.dict = dataset_dict (ds);
lr.n_predictor_vars = 0;
@@ -566,10 +769,12 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
lr.lcon = 0.0000;
lr.bcon = 0.001;
lr.min_epsilon = 0.00000001;
- lr.cut_point = 0.5;
lr.constant = true;
lr.confidence = 95;
lr.print = PRINT_DEFAULT;
+ lr.cat_predictors = NULL;
+ lr.n_cat_predictors = 0;
+ lr.indep_vars = NULL;
if (lex_match_id (lexer, "VARIABLES"))
@@ -578,11 +783,12 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
if (! (lr.dep_var = parse_variable_const (lexer, lr.dict)))
goto error;
- lex_force_match (lexer, T_WITH);
+ if (! lex_force_match (lexer, T_WITH))
+ goto error;
if (!parse_variables_const (lexer, lr.dict,
- &lr.predictor_vars, &lr.n_predictor_vars,
- PV_NO_DUPLICATE | PV_NUMERIC))
+ &pred_vars, &n_pred_vars,
+ PV_NO_DUPLICATE))
goto error;
@@ -627,6 +833,19 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
{
/* This is for compatibility. It does nothing */
}
+ else if (lex_match_id (lexer, "CATEGORICAL"))
+ {
+ lex_match (lexer, T_EQUALS);
+ do
+ {
+ lr.cat_predictors = xrealloc (lr.cat_predictors,
+ sizeof (*lr.cat_predictors) * ++lr.n_cat_predictors);
+ lr.cat_predictors[lr.n_cat_predictors - 1] = 0;
+ }
+ while (parse_design_interaction (lexer, lr.dict,
+ lr.cat_predictors + lr.n_cat_predictors - 1));
+ lr.n_cat_predictors--;
+ }
else if (lex_match_id (lexer, "PRINT"))
{
lex_match (lexer, T_EQUALS);
@@ -659,12 +878,12 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
lr.print |= PRINT_CI;
if (lex_force_match (lexer, T_LPAREN))
{
- if (! lex_force_int (lexer))
+ if (! lex_force_num (lexer))
{
lex_error (lexer, NULL);
goto error;
}
- lr.confidence = lex_integer (lexer);
+ lr.confidence = lex_number (lexer);
lex_get (lexer);
if ( ! lex_force_match (lexer, T_RPAREN))
{
@@ -761,6 +980,30 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
}
}
}
+ else if (lex_match_id (lexer, "CUT"))
+ {
+ if (lex_force_match (lexer, T_LPAREN))
+ {
+ if (! lex_force_num (lexer))
+ {
+ lex_error (lexer, NULL);
+ goto error;
+ }
+ cp = lex_number (lexer);
+
+ if (cp < 0 || cp > 1.0)
+ {
+ msg (ME, _("Cut point value must be in the range [0,1]"));
+ goto error;
+ }
+ lex_get (lexer);
+ if ( ! lex_force_match (lexer, T_RPAREN))
+ {
+ lex_error (lexer, NULL);
+ goto error;
+ }
+ }
+ }
else
{
lex_error (lexer, NULL);
@@ -775,8 +1018,80 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
}
}
+ lr.ilogit_cut_point = - log (1/cp - 1);
+ /* Copy the predictor variables from the temporary location into the
+ final one, dropping any categorical variables which appear there.
+ FIXME: This is O(NxM).
+ */
+ {
+ struct variable_node *vn, *next;
+ struct hmap allvars;
+ hmap_init (&allvars);
+ for (v = x = 0; v < n_pred_vars; ++v)
+ {
+ bool drop = false;
+ const struct variable *var = pred_vars[v];
+ int cv = 0;
+
+ unsigned int hash = hash_pointer (var, 0);
+ struct variable_node *vn = lookup_variable (&allvars, var, hash);
+ if (vn == NULL)
+ {
+ vn = xmalloc (sizeof *vn);
+ vn->var = var;
+ hmap_insert (&allvars, &vn->node, hash);
+ }
+
+ for (cv = 0; cv < lr.n_cat_predictors ; ++cv)
+ {
+ int iv;
+ const struct interaction *iact = lr.cat_predictors[cv];
+ for (iv = 0 ; iv < iact->n_vars ; ++iv)
+ {
+ const struct variable *ivar = iact->vars[iv];
+ unsigned int hash = hash_pointer (ivar, 0);
+ struct variable_node *vn = lookup_variable (&allvars, ivar, hash);
+ if (vn == NULL)
+ {
+ vn = xmalloc (sizeof *vn);
+ vn->var = ivar;
+
+ hmap_insert (&allvars, &vn->node, hash);
+ }
+
+ if (var == ivar)
+ {
+ drop = true;
+ }
+ }
+ }
+
+ if (drop)
+ continue;
+
+ lr.predictor_vars = xrealloc (lr.predictor_vars, sizeof *lr.predictor_vars * (x + 1));
+ lr.predictor_vars[x++] = var;
+ lr.n_predictor_vars++;
+ }
+ free (pred_vars);
+
+ lr.n_indep_vars = hmap_count (&allvars);
+ lr.indep_vars = xmalloc (lr.n_indep_vars * sizeof *lr.indep_vars);
+
+ /* Interate over each variable and push it into the array */
+ x = 0;
+ HMAP_FOR_EACH_SAFE (vn, next, struct variable_node, node, &allvars)
+ {
+ lr.indep_vars[x++] = vn->var;
+ free (vn);
+ }
+ hmap_destroy (&allvars);
+ }
+
+
+ /* logistical regression for each split group */
{
struct casegrouper *grouper;
struct casereader *group;
@@ -789,12 +1104,26 @@ cmd_logistic (struct lexer *lexer, struct dataset *ds)
ok = proc_commit (ds) && ok;
}
+ for (i = 0 ; i < lr.n_cat_predictors; ++i)
+ {
+ interaction_destroy (lr.cat_predictors[i]);
+ }
free (lr.predictor_vars);
+ free (lr.cat_predictors);
+ free (lr.indep_vars);
+
return CMD_SUCCESS;
error:
+ for (i = 0 ; i < lr.n_cat_predictors; ++i)
+ {
+ interaction_destroy (lr.cat_predictors[i]);
+ }
free (lr.predictor_vars);
+ free (lr.cat_predictors);
+ free (lr.indep_vars);
+
return CMD_FAILURE;
}
@@ -840,8 +1169,8 @@ output_depvarmap (const struct lr_spec *cmd, const struct lr_result *res)
tab_text (t, 0, 1 + heading_rows, 0, ds_cstr (&str));
- tab_double (t, 1, 0 + heading_rows, 0, map_dependent_var (cmd, res, &res->y0), &F_8_0);
- tab_double (t, 1, 1 + heading_rows, 0, map_dependent_var (cmd, res, &res->y1), &F_8_0);
+ tab_double (t, 1, 0 + heading_rows, 0, map_dependent_var (cmd, res, &res->y0), NULL, RC_INTEGER);
+ tab_double (t, 1, 1 + heading_rows, 0, map_dependent_var (cmd, res, &res->y1), NULL, RC_INTEGER);
ds_destroy (&str);
tab_submit (t);
@@ -850,20 +1179,20 @@ output_depvarmap (const struct lr_spec *cmd, const struct lr_result *res)
/* Show the Variables in the Equation box */
static void
-output_variables (const struct lr_spec *cmd,
- const gsl_vector *beta,
- const gsl_vector *se)
+output_variables (const struct lr_spec *cmd,
+ const struct lr_result *res)
{
int row = 0;
const int heading_columns = 1;
int heading_rows = 1;
struct tab_table *t;
- int idx;
- int n_rows = cmd->n_predictor_vars;
-
int nc = 8;
int nr ;
+ int i = 0;
+ int ivar = 0;
+ int idx_correction = 0;
+
if (cmd->print & PRINT_CI)
{
nc += 2;
@@ -874,6 +1203,9 @@ output_variables (const struct lr_spec *cmd,
if (cmd->constant)
nr++;
+ if (res->cats)
+ nr += categoricals_df_total (res->cats) + cmd->n_cat_predictors;
+
t = tab_create (nc, nr);
tab_title (t, _("Variables in the Equation"));
@@ -901,46 +1233,108 @@ output_variables (const struct lr_spec *cmd,
tab_text (t, 8, row, TAB_CENTER | TAT_TITLE, _("Lower"));
tab_text (t, 9, row, TAB_CENTER | TAT_TITLE, _("Upper"));
}
-
- if (cmd->constant)
- n_rows++;
- for (idx = 0 ; idx < n_rows; ++idx)
+ for (row = heading_rows ; row < nr; ++row)
{
- const int r = idx + heading_rows;
+ const int idx = row - heading_rows - idx_correction;
- const double b = gsl_vector_get (beta, idx);
- const double sigma = gsl_vector_get (se, idx);
- const double wald = pow2 (b / sigma);
+ const double b = gsl_vector_get (res->beta_hat, idx);
+ const double sigma2 = gsl_matrix_get (res->hessian, idx, idx);
+ const double wald = pow2 (b) / sigma2;
const double df = 1;
if (idx < cmd->n_predictor_vars)
- tab_text (t, 1, r, TAB_LEFT | TAT_TITLE,
- var_to_string (cmd->predictor_vars[idx]));
+ {
+ tab_text (t, 1, row, TAB_LEFT | TAT_TITLE,
+ var_to_string (cmd->predictor_vars[idx]));
+ }
+ else if (i < cmd->n_cat_predictors)
+ {
+ double wald;
+ bool summary = false;
+ struct string str;
+ const struct interaction *cat_predictors = cmd->cat_predictors[i];
+ const int df = categoricals_df (res->cats, i);
+
+ ds_init_empty (&str);
+ interaction_to_string (cat_predictors, &str);
- tab_double (t, 2, r, 0, b, 0);
- tab_double (t, 3, r, 0, sigma, 0);
- tab_double (t, 4, r, 0, wald, 0);
- tab_double (t, 5, r, 0, df, &F_8_0);
- tab_double (t, 6, r, 0, gsl_cdf_chisq_Q (wald, df), 0);
- tab_double (t, 7, r, 0, exp (b), 0);
+ if (ivar == 0)
+ {
+ /* Calculate the Wald statistic,
+ which is \beta' C^-1 \beta .
+ where \beta is the vector of the coefficient estimates comprising this
+ categorial variable. and C is the corresponding submatrix of the
+ hessian matrix.
+ */
+ gsl_matrix_const_view mv =
+ gsl_matrix_const_submatrix (res->hessian, idx, idx, df, df);
+ gsl_matrix *subhessian = gsl_matrix_alloc (mv.matrix.size1, mv.matrix.size2);
+ gsl_vector_const_view vv = gsl_vector_const_subvector (res->beta_hat, idx, df);
+ gsl_vector *temp = gsl_vector_alloc (df);
+
+ gsl_matrix_memcpy (subhessian, &mv.matrix);
+ gsl_linalg_cholesky_decomp (subhessian);
+ gsl_linalg_cholesky_invert (subhessian);
+
+ gsl_blas_dgemv (CblasTrans, 1.0, subhessian, &vv.vector, 0, temp);
+ gsl_blas_ddot (temp, &vv.vector, &wald);
+
+ tab_double (t, 4, row, 0, wald, NULL, RC_OTHER);
+ tab_double (t, 5, row, 0, df, NULL, RC_INTEGER);
+ tab_double (t, 6, row, 0, gsl_cdf_chisq_Q (wald, df), NULL, RC_PVALUE);
+
+ idx_correction ++;
+ summary = true;
+ gsl_matrix_free (subhessian);
+ gsl_vector_free (temp);
+ }
+ else
+ {
+ ds_put_format (&str, "(%d)", ivar);
+ }
+
+ tab_text (t, 1, row, TAB_LEFT | TAT_TITLE, ds_cstr (&str));
+ if (ivar++ == df)
+ {
+ ++i; /* next interaction */
+ ivar = 0;
+ }
+
+ ds_destroy (&str);
+
+ if (summary)
+ continue;
+ }
+ else
+ {
+ tab_text (t, 1, row, TAB_LEFT | TAT_TITLE, _("Constant"));
+ }
+
+ tab_double (t, 2, row, 0, b, NULL, RC_OTHER);
+ tab_double (t, 3, row, 0, sqrt (sigma2), NULL, RC_OTHER);
+ tab_double (t, 4, row, 0, wald, NULL, RC_OTHER);
+ tab_double (t, 5, row, 0, df, NULL, RC_INTEGER);
+ tab_double (t, 6, row, 0, gsl_cdf_chisq_Q (wald, df), NULL, RC_PVALUE);
+ tab_double (t, 7, row, 0, exp (b), NULL, RC_OTHER);
if (cmd->print & PRINT_CI)
{
+ int last_ci = nr;
double wc = gsl_cdf_ugaussian_Pinv (0.5 + cmd->confidence / 200.0);
- wc *= sigma;
+ wc *= sqrt (sigma2);
+
+ if (cmd->constant)
+ last_ci--;
- if (idx < cmd->n_predictor_vars)
+ if (row < last_ci)
{
- tab_double (t, 8, r, 0, exp (b - wc), 0);
- tab_double (t, 9, r, 0, exp (b + wc), 0);
+ tab_double (t, 8, row, 0, exp (b - wc), NULL, RC_OTHER);
+ tab_double (t, 9, row, 0, exp (b + wc), NULL, RC_OTHER);
}
}
}
- if ( cmd->constant)
- tab_text (t, 1, nr - 1, TAB_LEFT | TAT_TITLE, _("Constant"));
-
tab_submit (t);
}
@@ -948,7 +1342,7 @@ output_variables (const struct lr_spec *cmd,
/* Show the model summary box */
static void
output_model_summary (const struct lr_result *res,
- double initial_likelihood, double likelihood)
+ double initial_log_likelihood, double log_likelihood)
{
const int heading_columns = 0;
const int heading_rows = 1;
@@ -970,15 +1364,15 @@ output_model_summary (const struct lr_result *res,
tab_text (t, 0, 0, TAB_LEFT | TAT_TITLE, _("Step 1"));
tab_text (t, 1, 0, TAB_CENTER | TAT_TITLE, _("-2 Log likelihood"));
- tab_double (t, 1, 1, 0, -2 * log (likelihood), 0);
+ tab_double (t, 1, 1, 0, -2 * log_likelihood, NULL, RC_OTHER);
tab_text (t, 2, 0, TAB_CENTER | TAT_TITLE, _("Cox & Snell R Square"));
- cox = 1.0 - pow (initial_likelihood /likelihood, 2 / res->cc);
- tab_double (t, 2, 1, 0, cox, 0);
+ cox = 1.0 - exp((initial_log_likelihood - log_likelihood) * (2 / res->cc));
+ tab_double (t, 2, 1, 0, cox, NULL, RC_OTHER);
tab_text (t, 3, 0, TAB_CENTER | TAT_TITLE, _("Nagelkerke R Square"));
- tab_double (t, 3, 1, 0, cox / ( 1.0 - pow (initial_likelihood, 2 / res->cc)), 0);
+ tab_double (t, 3, 1, 0, cox / ( 1.0 - exp(initial_log_likelihood * (2 / res->cc))), NULL, RC_OTHER);
tab_submit (t);
@@ -1015,16 +1409,206 @@ case_processing_summary (const struct lr_result *res)
tab_text (t, 0, 2, TAB_LEFT | TAT_TITLE, _("Missing Cases"));
tab_text (t, 0, 3, TAB_LEFT | TAT_TITLE, _("Total"));
- tab_double (t, 1, 1, 0, res->n_nonmissing, &F_8_0);
- tab_double (t, 1, 2, 0, res->n_missing, &F_8_0);
+ tab_double (t, 1, 1, 0, res->n_nonmissing, NULL, RC_INTEGER);
+ tab_double (t, 1, 2, 0, res->n_missing, NULL, RC_INTEGER);
total = res->n_nonmissing + res->n_missing;
- tab_double (t, 1, 3, 0, total , &F_8_0);
+ tab_double (t, 1, 3, 0, total , NULL, RC_INTEGER);
+
+ tab_double (t, 2, 1, 0, 100 * res->n_nonmissing / (double) total, NULL, RC_OTHER);
+ tab_double (t, 2, 2, 0, 100 * res->n_missing / (double) total, NULL, RC_OTHER);
+ tab_double (t, 2, 3, 0, 100 * total / (double) total, NULL, RC_OTHER);
+
+ tab_submit (t);
+}
+
+static void
+output_categories (const struct lr_spec *cmd, const struct lr_result *res)
+{
+ const struct fmt_spec *wfmt =
+ cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0;
+
+ int cumulative_df;
+ int i = 0;
+ const int heading_columns = 2;
+ const int heading_rows = 2;
+ struct tab_table *t;
+
+ int nc ;
+ int nr ;
+
+ int v;
+ int r = 0;
+
+ int max_df = 0;
+ int total_cats = 0;
+ for (i = 0; i < cmd->n_cat_predictors; ++i)
+ {
+ size_t n = categoricals_n_count (res->cats, i);
+ size_t df = categoricals_df (res->cats, i);
+ if (max_df < df)
+ max_df = df;
+ total_cats += n;
+ }
+
+ nc = heading_columns + 1 + max_df;
+ nr = heading_rows + total_cats;
+
+ t = tab_create (nc, nr);
+ tab_set_format (t, RC_WEIGHT, wfmt);
+
+ tab_title (t, _("Categorical Variables' Codings"));
+
+ tab_headers (t, heading_columns, 0, heading_rows, 0);
+
+ tab_box (t, TAL_2, TAL_2, -1, TAL_1, 0, 0, nc - 1, nr - 1);
+
+ tab_hline (t, TAL_2, 0, nc - 1, heading_rows);
+ tab_vline (t, TAL_2, heading_columns, 0, nr - 1);
+
+
+ tab_text (t, heading_columns, 1, TAB_CENTER | TAT_TITLE, _("Frequency"));
+
+ tab_joint_text_format (t, heading_columns + 1, 0, nc - 1, 0,
+ TAB_CENTER | TAT_TITLE, _("Parameter coding"));
+
+
+ for (i = 0; i < max_df; ++i)
+ {
+ int c = heading_columns + 1 + i;
+ tab_text_format (t, c, 1, TAB_CENTER | TAT_TITLE, _("(%d)"), i + 1);
+ }
+
+ cumulative_df = 0;
+ for (v = 0; v < cmd->n_cat_predictors; ++v)
+ {
+ int cat;
+ const struct interaction *cat_predictors = cmd->cat_predictors[v];
+ int df = categoricals_df (res->cats, v);
+ struct string str;
+ ds_init_empty (&str);
+
+ interaction_to_string (cat_predictors, &str);
- tab_double (t, 2, 1, 0, 100 * res->n_nonmissing / (double) total, 0);
- tab_double (t, 2, 2, 0, 100 * res->n_missing / (double) total, 0);
- tab_double (t, 2, 3, 0, 100 * total / (double) total, 0);
+ tab_text (t, 0, heading_rows + r, TAB_LEFT | TAT_TITLE, ds_cstr (&str) );
+
+ ds_destroy (&str);
+
+ for (cat = 0; cat < categoricals_n_count (res->cats, v) ; ++cat)
+ {
+ struct string str;
+ const struct ccase *c = categoricals_get_case_by_category_real (res->cats, v, cat);
+ const double *freq = categoricals_get_user_data_by_category_real (res->cats, v, cat);
+
+ int x;
+ ds_init_empty (&str);
+
+ for (x = 0; x < cat_predictors->n_vars; ++x)
+ {
+ const union value *val = case_data (c, cat_predictors->vars[x]);
+ var_append_value_name (cat_predictors->vars[x], val, &str);
+
+ if (x < cat_predictors->n_vars - 1)
+ ds_put_cstr (&str, " ");
+ }
+
+ tab_text (t, 1, heading_rows + r, 0, ds_cstr (&str));
+ ds_destroy (&str);
+ tab_double (t, 2, heading_rows + r, 0, *freq, NULL, RC_WEIGHT);
+
+ for (x = 0; x < df; ++x)
+ {
+ tab_double (t, heading_columns + 1 + x, heading_rows + r, 0, (cat == x), NULL, RC_INTEGER);
+ }
+ ++r;
+ }
+ cumulative_df += df;
+ }
tab_submit (t);
+
}
+
+static void
+output_classification_table (const struct lr_spec *cmd, const struct lr_result *res)
+{
+ const struct fmt_spec *wfmt =
+ cmd->wv ? var_get_print_format (cmd->wv) : &F_8_0;
+
+ const int heading_columns = 3;
+ const int heading_rows = 3;
+
+ struct string sv0, sv1;
+
+ const int nc = heading_columns + 3;
+ const int nr = heading_rows + 3;
+
+ struct tab_table *t = tab_create (nc, nr);
+ tab_set_format (t, RC_WEIGHT, wfmt);
+
+ ds_init_empty (&sv0);
+ ds_init_empty (&sv1);
+
+ tab_title (t, _("Classification Table"));
+
+ tab_headers (t, heading_columns, 0, heading_rows, 0);
+
+ tab_box (t, TAL_2, TAL_2, -1, -1, 0, 0, nc - 1, nr - 1);
+ tab_box (t, -1, -1, -1, TAL_1, heading_columns, 0, nc - 1, nr - 1);
+
+ tab_hline (t, TAL_2, 0, nc - 1, heading_rows);
+ tab_vline (t, TAL_2, heading_columns, 0, nr - 1);
+
+ tab_text (t, 0, heading_rows, TAB_CENTER | TAT_TITLE, _("Step 1"));
+
+
+ tab_joint_text (t, heading_columns, 0, nc - 1, 0,
+ TAB_CENTER | TAT_TITLE, _("Predicted"));
+
+ tab_joint_text (t, heading_columns, 1, heading_columns + 1, 1,
+ 0, var_to_string (cmd->dep_var) );
+
+ tab_joint_text (t, 1, 2, 2, 2,
+ TAB_LEFT | TAT_TITLE, _("Observed"));
+
+ tab_text (t, 1, 3, TAB_LEFT, var_to_string (cmd->dep_var) );
+
+
+ tab_joint_text (t, nc - 1, 1, nc - 1, 2,
+ TAB_CENTER | TAT_TITLE, _("Percentage\nCorrect"));
+
+
+ tab_joint_text (t, 1, nr - 1, 2, nr - 1,
+ TAB_LEFT | TAT_TITLE, _("Overall Percentage"));
+
+
+ tab_hline (t, TAL_1, 1, nc - 1, nr - 1);
+
+ var_append_value_name (cmd->dep_var, &res->y0, &sv0);
+ var_append_value_name (cmd->dep_var, &res->y1, &sv1);
+
+ tab_text (t, 2, heading_rows, TAB_LEFT, ds_cstr (&sv0));
+ tab_text (t, 2, heading_rows + 1, TAB_LEFT, ds_cstr (&sv1));
+
+ tab_text (t, heading_columns, 2, 0, ds_cstr (&sv0));
+ tab_text (t, heading_columns + 1, 2, 0, ds_cstr (&sv1));
+
+ ds_destroy (&sv0);
+ ds_destroy (&sv1);
+
+ tab_double (t, heading_columns, 3, 0, res->tn, NULL, RC_WEIGHT);
+ tab_double (t, heading_columns + 1, 4, 0, res->tp, NULL, RC_WEIGHT);
+
+ tab_double (t, heading_columns + 1, 3, 0, res->fp, NULL, RC_WEIGHT);
+ tab_double (t, heading_columns, 4, 0, res->fn, NULL, RC_WEIGHT);
+
+ tab_double (t, heading_columns + 2, 3, 0, 100 * res->tn / (res->tn + res->fp), NULL, RC_OTHER);
+ tab_double (t, heading_columns + 2, 4, 0, 100 * res->tp / (res->tp + res->fn), NULL, RC_OTHER);
+
+ tab_double (t, heading_columns + 2, 5, 0,
+ 100 * (res->tp + res->tn) / (res->tp + res->tn + res->fp + res->fn), NULL, RC_OTHER);
+
+
+ tab_submit (t);
+}