IndependenceTest.c

Go to the documentation of this file.
00001 
00009 #include "party.h"
00010 
00011 
00021 void C_TeststatPvalue(const SEXP linexpcov, const SEXP varctrl, 
00022                       double *ans_teststat, double *ans_pvalue) {
00023     
00024     double releps, abseps, tol;
00025     int maxpts;
00026     
00027     maxpts = get_maxpts(varctrl);
00028     tol = get_tol(varctrl);
00029     abseps = get_abseps(varctrl);
00030     releps = get_releps(varctrl);
00031     
00032     /* compute the test statistic */
00033     ans_teststat[0] = C_TestStatistic(linexpcov, get_teststat(varctrl), 
00034                                   get_tol(varctrl));
00035 
00036     /* compute the p-value if requested */                                  
00037     if (get_pvalue(varctrl))
00038         ans_pvalue[0] =  C_ConditionalPvalue(ans_teststat[0], linexpcov, 
00039                                          get_teststat(varctrl),
00040                                          tol, &maxpts, &releps, &abseps);
00041     else
00042         ans_pvalue[0] = 1.0;
00043 }
00044 
00053 void C_TeststatCriterion(const SEXP linexpcov, const SEXP varctrl, 
00054                          double *ans_teststat, double *ans_criterion) {
00055     
00056     C_TeststatPvalue(linexpcov, varctrl, ans_teststat, ans_criterion);
00057     
00058     /* the node criterion is to be MAXIMISED, 
00059        i.e. 1-pvalue or test statistic \in \[0, \infty\] */
00060     if (get_pvalue(varctrl))
00061         ans_criterion[0] = 1 - ans_criterion[0];
00062     else
00063         ans_criterion[0] = ans_teststat[0];
00064     
00065 }
00066 
00067 
00078 void C_IndependenceTest(const SEXP x, const SEXP y, const SEXP weights, 
00079                         SEXP linexpcov, SEXP varctrl, 
00080                         SEXP ans) {
00081     
00082     /* compute linear statistic and its conditional expectation and
00083        covariance
00084     */
00085     C_LinStatExpCov(REAL(x), ncol(x), REAL(y), ncol(y), 
00086                     REAL(weights), nrow(x), 1, 
00087                     GET_SLOT(linexpcov, PL2_expcovinfSym), linexpcov);
00088 
00089     /* compute test statistic */
00090     if (get_teststat(varctrl) == 2) 
00091         C_LinStatExpCovMPinv(linexpcov, get_tol(varctrl));
00092     C_TeststatPvalue(linexpcov, varctrl, &REAL(ans)[0], &REAL(ans)[1]);
00093 }
00094 
00095 
00105 SEXP R_IndependenceTest(SEXP x, SEXP y, SEXP weights, SEXP linexpcov, SEXP varctrl) {
00106                         
00107     SEXP ans;
00108     
00109     PROTECT(ans = allocVector(REALSXP, 2));
00110     C_IndependenceTest(x, y, weights, linexpcov, varctrl, ans);
00111     UNPROTECT(1);
00112     return(ans);
00113 }
00114 
00115 
00129 void C_GlobalTest(const SEXP learnsample, const SEXP weights, 
00130                   SEXP fitmem, const SEXP varctrl, 
00131                   const SEXP gtctrl, const double minsplit, 
00132                   double *ans_teststat, double *ans_criterion) {
00133 
00134     int ninputs, nobs, j, i, k, RECALC = 1, type;
00135     SEXP responses, inputs, y, x, xmem, expcovinf;
00136     SEXP thiswhichNA;
00137     double *thisweights, *dweights, *pvaltmp, stweights = 0.0;
00138     int *ithiswhichNA, RANDOM, mtry, *randomvar, *index;
00139     int *dontuse, *dontusetmp;
00140     
00141     ninputs = get_ninputs(learnsample);
00142     nobs = get_nobs(learnsample);
00143     responses = GET_SLOT(learnsample, PL2_responsesSym);
00144     inputs = GET_SLOT(learnsample, PL2_inputsSym);
00145     dweights = REAL(weights);
00146     
00147     /* y = get_transformation(responses, 1); */
00148     y = get_test_trafo(responses);
00149     
00150     expcovinf = GET_SLOT(fitmem, PL2_expcovinfSym);
00151     C_ExpectCovarInfluence(REAL(y), ncol(y), REAL(weights), 
00152                            nobs, expcovinf);
00153     
00154     if (REAL(GET_SLOT(expcovinf, PL2_sumweightsSym))[0] < minsplit) {
00155         for (j = 0; j < ninputs; j++) {
00156             ans_teststat[j] = 0.0;
00157             ans_criterion[j] = 0.0;
00158         }
00159     } else {
00160 
00161         dontuse = INTEGER(get_dontuse(fitmem));
00162         dontusetmp = INTEGER(get_dontusetmp(fitmem));
00163     
00164         for (j = 0; j < ninputs; j++) dontusetmp[j] = !dontuse[j];
00165     
00166         /* random forest */
00167         RANDOM = get_randomsplits(gtctrl);
00168         mtry = get_mtry(gtctrl);
00169         if (RANDOM & (mtry > ninputs)) {
00170             warning("mtry is larger than ninputs, using mtry = inputs");
00171             mtry = ninputs;
00172             RANDOM = 0;
00173         }
00174         if (RANDOM) {
00175             index = Calloc(ninputs, int);
00176             randomvar = Calloc(mtry, int);
00177             C_SampleNoReplace(index, ninputs, mtry, randomvar);
00178             j = 0;
00179             for (k = 0; k < mtry; k++) {
00180                 j = randomvar[k];
00181                 while(dontuse[j] && j < ninputs) j++;
00182                 if (j == ninputs) 
00183                     error("not enough variables to sample from");
00184                 dontusetmp[j] = 0;
00185             }
00186             Free(index);
00187             Free(randomvar);
00188         }
00189 
00190         for (j = 1; j <= ninputs; j++) {
00191 
00192             if ((RANDOM && dontusetmp[j - 1]) || dontuse[j - 1]) {
00193                 ans_teststat[j - 1] = 0.0;
00194                 ans_criterion[j - 1] = 0.0;
00195                 continue; 
00196             }
00197         
00198             x = get_transformation(inputs, j);
00199 
00200             xmem = get_varmemory(fitmem, j);
00201             if (!has_missings(inputs, j)) {
00202                 C_LinStatExpCov(REAL(x), ncol(x), REAL(y), ncol(y),
00203                                 REAL(weights), nrow(x), !RECALC, expcovinf,
00204                                 xmem);
00205             } else {
00206                 thisweights = REAL(get_weights(fitmem, j));
00207                 thiswhichNA = get_missings(inputs, j);
00208                 ithiswhichNA = INTEGER(thiswhichNA);
00209                 for (i = 0; i < nobs; i++) thisweights[i] = dweights[i];
00210                 for (k = 0; k < LENGTH(thiswhichNA); k++)
00211                     thisweights[ithiswhichNA[k] - 1] = 0.0;
00212 
00213                 /* check if minsplit criterion is still met 
00214                    in the presence of missing values
00215                    bug spotted by Han Lee <Han.Lee@geodecapital.com>
00216                        fixed 2006-08-31
00217                 */
00218                 stweights = 0.0;
00219                 for (i = 0; i < nobs; i++) stweights += thisweights[i];
00220                 if (stweights < minsplit) {
00221                     ans_teststat[j - 1] = 0.0;
00222                     ans_criterion[j - 1] = 0.0;
00223                     continue; 
00224                 }
00225 
00226                 C_LinStatExpCov(REAL(x), ncol(x), REAL(y), ncol(y),
00227                                 thisweights, nrow(x), RECALC, 
00228                                 GET_SLOT(xmem, PL2_expcovinfSym),
00229                                 xmem);
00230             }
00231 
00232             if (get_teststat(varctrl) == 2)
00233                 C_LinStatExpCovMPinv(xmem, get_tol(varctrl));
00234             C_TeststatCriterion(xmem, varctrl, &ans_teststat[j - 1], 
00235                                 &ans_criterion[j - 1]);
00236         }                
00237 
00238         type = get_testtype(gtctrl);
00239         switch(type) {
00240             /* Bonferroni: p_adj = 1 - (1 - p)^k */
00241             case BONFERRONI: 
00242                     for (j = 0; j < ninputs; j++)
00243                         ans_criterion[j] = R_pow_di(ans_criterion[j], ninputs);
00244                     break;
00245             /* Monte-Carlo */
00246             case MONTECARLO: 
00247                     pvaltmp = Calloc(ninputs, double);
00248                     C_MonteCarlo(ans_criterion, learnsample, weights, fitmem, 
00249                                  varctrl, gtctrl, pvaltmp);
00250                     for (j = 0; j < ninputs; j++)
00251                         ans_criterion[j] = 1 - pvaltmp[j];
00252                     Free(pvaltmp);
00253                     break;
00254             /* aggregated */
00255             case AGGREGATED: 
00256                     error("C_GlobalTest: aggregated global test not yet implemented");
00257                     break;
00258             /* raw */
00259             case UNIVARIATE: break;
00260             case TESTSTATISTIC: break;
00261             default: error("C_GlobalTest: undefined value for type argument");
00262                      break;
00263         }
00264     }
00265 }
00266 
00267 
00277 SEXP R_GlobalTest(SEXP learnsample, SEXP weights, SEXP fitmem, 
00278                   SEXP varctrl, SEXP gtctrl) {
00279 
00280     SEXP ans, teststat, criterion;
00281 
00282     GetRNGstate();
00283 
00284     PROTECT(ans = allocVector(VECSXP, 2));
00285     SET_VECTOR_ELT(ans, 0, 
00286         teststat = allocVector(REALSXP, get_ninputs(learnsample)));
00287     SET_VECTOR_ELT(ans, 1, 
00288         criterion = allocVector(REALSXP, get_ninputs(learnsample)));
00289 
00290     C_GlobalTest(learnsample, weights, fitmem, varctrl, gtctrl, 0, 
00291                  REAL(teststat), REAL(criterion));
00292                  
00293     PutRNGstate();
00294     
00295     UNPROTECT(1);
00296     return(ans);
00297 }

Generated on Fri Feb 2 11:29:32 2007 for party by  doxygen 1.4.6