segmented/0000755000176200001440000000000013160062250012220 5ustar liggesuserssegmented/inst/0000755000176200001440000000000013157724622013213 5ustar liggesuserssegmented/inst/CITATION0000644000176200001440000000236313157724622014354 0ustar liggesuserscitHeader("To cite segmented in publications use:") citEntry(entry="Article", title = "Estimating regression models with unknown break-points.", author = personList(as.person("Vito M.R. Muggeo")), journal = "Statistics in Medicine", year = "2003", volume = "22", pages = "3055--3071", textVersion = paste("Vito M. R. Muggeo (2003).", "Estimating regression models with unknown break-points.", "Statistics in Medicine, 22, 3055-3071.") ) citEntry(entry="Article", title = "segmented: an R Package to Fit Regression Models with Broken-Line Relationships.", author = personList(as.person("Vito M.R. Muggeo")), journal = "R News", year = "2008", volume = "8", number = "1", pages = "20--25", url = "https://cran.r-project.org/doc/Rnews/", textVersion = paste("Vito M. R. Muggeo (2008).", "segmented: an R Package to Fit Regression Models with Broken-Line Relationships.", "R News, 8/1, 20-25.", "URL https://cran.r-project.org/doc/Rnews/.") ) segmented/NAMESPACE0000644000176200001440000000315113157724622013455 0ustar liggesusersimportFrom("grDevices", "grey") importFrom("graphics", "abline", "axis", "box", "lines", "matlines", "matplot", "par", "plot", "points", "polygon", "segments") importFrom("stats", "approx", "as.formula", "coef", "contrasts", "family", "fitted", "formula", "gaussian", "glm.control", "glm.fit", "is.empty.model", "lm.wfit", "median", "model.frame", "model.matrix", "model.offset", "model.response", "model.weights", "na.omit", "pnorm", "predict", "printCoefmat", "pt", "qnorm", "qt", "quantile", "resid", "residuals", "runif", "summary.glm", "summary.lm", "update", "update.formula", "vcov", "weights") importFrom("utils", "flush.console") importFrom("stats", "complete.cases") importFrom("stats", "add1") export(segmented, segmented.default, segmented.lm, segmented.glm, segmented.Arima, broken.line ,confint.segmented,davies.test,pscore.test,draw.history, intercept,lines.segmented,plot.segmented,print.segmented, seg.control,seg.lm.fit,seg.glm.fit,seg.lm.fit.boot,seg.glm.fit.boot, seg.def.fit,seg.def.fit.boot, seg.Ar.fit,seg.Ar.fit.boot, slope, summary.segmented,print.summary.segmented,vcov.segmented, predict.segmented, points.segmented) S3method(segmented,default) S3method(segmented,lm) S3method(segmented,glm) S3method(segmented,Arima) S3method(plot,segmented) S3method(print,segmented) S3method(summary,segmented) S3method(print, summary.segmented) S3method(lines,segmented) S3method(confint,segmented) S3method(vcov,segmented) S3method(predict,segmented) S3method(points,segmented) segmented/NEWS0000644000176200001440000003616713157724622012752 0ustar liggesusers************************************* * * * Changes in segmented * * * ************************************* =============== version 0.5-2.2 =============== * When there is a single covariate in the starting (g)lm, seg.Z can be missing when calling the segmented methods. * bug fixed: plot.segmented(.., link=FALSE) did not work correctly (sometimes it returned an error) for glm fits with multiple breakpoints. Weights were not handled appropriately by segmented.lm. =============== version 0.5-2.1 =============== * pscore.test() now works also for "glm" fits * plot.segmented() now plots the partial residuals as "component + working residuals" (rather than Pearson residuals, relevant only for glm fits). * segmented.default() now works for fits obtained by MASS::rlm(). =============== version 0.5-2.0 =============== * pscore.test() introduced. The function tests for a breakpoint using a (pseudo) score statistic which is more powerful than davies.test(), especially when the breakpoint is expected to be in the middle of the covariate range and the signal-to-noise ratio is high. * argument 'digits' added in seg.control() to fix the number of digits of the breakpoint estimate during the iterative estimation algorithm. * bug fixed: conf.level>0 in plot.segmented() did not work for objects returned by segmented.default(). =============== version 0.5-1.5 (not on CRAN) =============== * arguments 'gap' and 'show.gap' removed in intercept() and in plot.segmented(). (they are meaningless, as segmented() always returns joined piecewise lines, i.e. with no gaps). * slope() and broken.line() (and then plot.segmented() which uses them) did not work for objects returned by segmented.default() (Thanks to Marcos Krull for reporting). =============== version 0.5-1.4 =============== * segmented.Arima() should be slightly faster, as starting values are passed in arima() (via 'init') throughout the iterative process. * plot.segmented() is expected to work for objects returned by segmented.Arima. * print.summary.segmented() does not print anymore the t-values for the gap coefficients (this information is meaningless as the gap coeffs are always set to zero in the returned model). * Bug fixed: intercept() ignored argument 'rev.sgn'; points.segmented() missed argument 'transf'. =============== version 0.5-1.3 (not on CRAN) =============== * plot.segmented() gains argument 'transf' to plot 'transf(values)' rather 'values' on the current plot. * print.summary.segmented() now uses round() rather than signif() when displaying the breakpoint estimate. * Bug fixed: psi=NA was not working in the segmented.* methods; this bug was incidentally introduced in the last version (thanks to Bertrand Sudre for first reporting that). =============== version 0.5-1.2 =============== * For 1 breakpoint models, 'psi' argument can be missing (default) when calling the segmented methods. * Bug fixed: lines.segmented() did not plot the dots when the fit object included multiple breakpoints and the argument 'shift' was set to FALSE (thanks to Jan Bull for reporting). There were some troubles with variable names including dots (thanks to Melanie Zoelck which first reported this bug). =============== version 0.5-1.1 =============== * segmented.default now accepts 'gee' fits (Thanks to John Boulanger for his input) * Minor change: argument 'col.dens' changed to 'dens.col' in plot.segmented() ('col.dens' made ineffective 'col') * Minor change: error/warning messages introduced in davies.test() if k<10; print.segmented slightly changed in displaying the estimated breakpoints. * Bug fixed: segmented did not terminate appropriately the algorithm with automatic selection of breakpoints concerning more than one variable (thanks to Ali Hashemi for reporting). =============== version 0.5-1.0 =============== * segmented.Arima() introduced. Now it is possible to estimate segmented relationships in "Arima" fits (although the summarizing and plotting methods do not work..) * plot.segmented() gains arguments 'dens.rug' and 'col.dens' to display in the plot (on the x axis) also the smoothed density of the segmented covariate. * Bug fixed: segmented.lm did not work if it.max=0 (but segmented.glm did), thanks to Eric Nussbaumer for reporting. segmented.lm and segmented.glm did work if the starting linear model included weights (this bug was introduced incidentally since version 0.4-0.1; thanks to Michael Rutter for reporting). segmented.lm and segmented.glm did not check appropriately inadmissible breakpoints (thanks to Erica Tennenhouse for reporting). segmented.lm and segmented.glm did not handle correctly variable names equal to function names. davies.test() did not work with 'segmented' objects (to test for and additional breakpoint). points.segmented() missed the argument 'rev.sgn'. =============== version 0.5-0.0 =============== * segmented.default() introduced. Now it is possible to estimate segmented relationships in arbitrary regression models (besides lm and glm) where specific methods do not exist (e.g. cox or quantile regression models). =============== version 0.4-0.1 (not on CRAN) =============== * segmented.lm() and segmented.glm() did not work if the starting model included additional "variables", such as 'threshold' in 'subset=age0. * The breakpoint starting values when automatic selection is performed are now specified as equally spaced values (optionally as quantiles). see argument 'quant' in seg.control() * added 'Authors@R' entry in the DESCRIPTION file =============== version 0.2-9.1 =============== * Some bugs fixed: segmented.lm() and segmented.glm() did not finish correctly when no breakpoint was found; now segmented.lm() and segmented.glm() take care of flat relationships; plot.segmented() did not compute correctly the partial residuals for segmented glm fits. =============== version 0.2-9.0 =============== * Bootstrap restarting implemented to deal with problems coming from flat segmented relationships. segmented now is less sensitive to starting values supplied for 'psi'. * At the convergence segmented now constrains the gap coefficients to be exactly zero. This is the default and it can be altered by the 'gap' argument in seg.control(). * plot.segmented() has been re-written. It gains argument `res' for plotting partial residuals along with the fitted piecewise lines, and now it produces nicer (and typically smaller) plots. * Some bugs fixed: davies.test() did not work correctly for deterministic data (thanks to Glenn Roberts for finding the error). davies.test() also returns the `process', i.e. the different values of the evaluation points and corresponding test statistic. =============== version 0.2-8.4 =============== * Some bugs fixed: segmented.glm() fitted a simple "lm" (and not "glm") (the error was introduced incidentally from 0.2-8.3, thanks to Vronique Storme for finding the error); broken.line() was not working for models without intercept and a null left slope; intercept() was not working correctly with multiple segmented variables. =============== version 0.2-8.3 =============== * Some minor bugs fixed: segmented.lm() and segmented.glm() did not find the offset variable in the dataframe where the initial (g)lm was called for; segmented.lm() and segmented.glm() sometimes returned an error when the automated algorithm was used (thanks to Paul Cohen for finding the error). =============== version 0.2-8.2 =============== * Some minor bugs fixed (segmented.lm() and segmented.glm() *alway* included the left slope in the estimation process, although the number of parameters was correct in the returned final fit. confint.segmented() did not order the estimated breakpoints for the variable having rev.sgn=TRUE; intercept() missed the (currently meaningless) argument var.diff (thanks to Eric Fuchs for pointing out that). ) =============== version 0.2-8.1 =============== * Some minor bugs fixed (segmented.lm() and segmented.glm() were not working correctly with dataframe subset or when the starting linear model included several intercepts (e.g., see the example about data("plant"); thanks to Nicola Ferrari for finding the error). davies.test() did not work when the variable name of its argument `seg.Z' included reserved words, e.g. `seg.Z~dist'; thanks to Thom White for finding the error). =============== version 0.2-8 =============== * intercept() added. It computes the intercepts of the regression lines for each segment of the fitted segmented relationship. * plot.segmented() now accepts a vector `col' argument to draw the fitted piecewise linear relationships with different colors. * Some minor bugs fixed (summary.segmented were not working correctly). =============== version 0.2-7.3 =============== * argument APC added to the slope() function to compute the `annual percent change'. * Some minor bugs fixed (confint and slope were not working correctly when the estimated breakpoints were returned in non-increasing order; offset was ignored in segmented.lm and segmented.glm; broken.line() was not working correctly (and its argument gap was unimplemented), thanks to M. Rennie for pointing out that; summary.segmented() was not working for models with no linear term, i.e. fitted via segmented(lm(y~0),..)). =============== version 0.2-7.2 =============== * segmented.lm and segmented.glm now accept objects with formulas y~., Thanks to G. Ferrara for finding the error. * Some bugs fixed (slope and confint were using the normal (rather than the t-distribution) to compute the CIs in gaussian models). =============== version 0.2-7.1 =============== * segmented.lm and segmented.glm now accept objects without 'explicit' formulas, namely returned by lm(my_fo,..) (and glm(my_fo,..)) where my_fo was defined earlier. Thanks to Y. Iwasaki for finding the error. =============== version 0.2-7 =============== * A sort of automatic procedure for breakpoint estimation is implemented. See argument stop.if.error in seg.control(). * davies.test() now accepts a one-sided formula (~x) rather than character ("x") to mean the segmented variable to be tested. davies.test also gains the arguments `beta0' and `dispersion'. * Some bugs fixed. =============== version 0.2-6 =============== * vcov.segmented() added. * option var.diff for robust covariance matrix has been added in summary.segmented(), print.summary.segmented(), slope(), and confint(). * Some bugs fixed. segmented/data/0000755000176200001440000000000013157724622013147 5ustar liggesuserssegmented/data/down.rda0000644000176200001440000000062113157724622014605 0ustar liggesusersuMKAɵBCDt3K׮Q^7LC}:t.: *,;K7gfvYIjZe9(̡-U)m gX1)8 g, 90" 4hQ(r 9G#'I$r 9Dkwhk}>Zc׀IF?Wi||Sɗs9Juyy%#nnɛcI^ 7yDW쐥s{{{ȳR>i\̛!)J~8`@Iajuı;@wQ/u*f͞|enXKʥG,k7vS+S TzEl|~[psegmented/data/plant.rda0000644000176200001440000000123013157724622014751 0ustar liggesusersV=hA=,(X"Zhނ(j+L+\mP.wMB.DrIN![ iI!Q"/^.w!cBr{oCV{:0͛ޛye77ԣB4MD5)3 z#bu^8ok3JqX_ya] xgiAbBFajPUD4lIKjj-[ɿNWɨA:[I+;B >7v(QJD_x5I_˃fGx?țt`z$^ھ6Is}+$ۣ<:_·cH|)@p ׇ #~GSms Eu0>o>?+דϓqZTO|B:1̧ -tkJa@yλq~ ƙ8s0i9ȟ?埠97WD{~?1|!vY13:OQ鑿@N%'?:zco8 /?Sf9fdx b$3G̸rp2B޴hKH!%-5-wA5b峍XT segmented/data/stagnant.rda0000644000176200001440000000054013157724622015455 0ustar liggesusers r0b```b`fff`b2Y# '(.ILK+a``c`b/g,D_`8oGVpXc4a'h7ۂ׹la40W4V`djC`߯[aV@3O7g@7@O+ 掷P@@-p%H{W k mK o?,/Cl쁊j9H|CAoּb CA M9gQ~L /HKtssaRKҊ xsegmented/R/0000755000176200001440000000000013157724622012437 5ustar liggesuserssegmented/R/predict.segmented.r0000644000176200001440000000724613157724622016237 0ustar liggesuserspredict.segmented<-function(object, newdata, ...){ #rev: 30/10/2013: it seems to work correctly, even with the minus variable (null right slope..) #rev: 14/4/2014 now it works like predict.lm/glm #BUT problems if type="terms" (in realta' funziona, il problema e' che # restituisce una colonna per "x", "U.x", "psi.x".. (Eventualmente si dovrebbero sommare..) #if(!is.null(object$orig.call$offset)) stop("predict.segmented can not handle argument 'offset'. Include it in formula!") dummy.matrix<-function(x.values, x.name, obj.seg, psi.est=TRUE){ #given the segmented fit 'obj.seg' and a segmented variable x.name with corresponding values x.values, #this function simply returns a matrix with columns (x, (x-psi)_+, -b*I(x>psi)) #or ((x-psi)_+, -b*I(x>psi)) if obj.seg does not include the coef for the linear "x" f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } n<-length(x.values) #le seguenti righe selezionavano (ERRONEAMENTE) sia "U1.x" sia "U1.neg.x" (se "x" e "neg.x" erano segmented covariates) #nameU<- grep(paste("\\.",x.name,"$", sep=""), obj.seg$nameUV$U, value = TRUE) #nameV<- grep(paste("\\.",x.name,"$", sep=""), obj.seg$nameUV$V, value = TRUE) nameU<-obj.seg$nameUV$U[f.U(obj.seg$nameUV$U,x.name)] nameV<-obj.seg$nameUV$V[f.U(obj.seg$nameUV$V,x.name)] diffSlope<-coef(obj.seg)[nameU] est.psi<-obj.seg$psi[nameV,2] k<-length(est.psi) PSI <- matrix(rep(est.psi, rep(n, k)), ncol = k) newZ<-matrix(x.values, nrow=n,ncol=k, byrow = FALSE) dummy1<-pmax(newZ-PSI,0) if(psi.est){ V<-ifelse(newZ>PSI,-1,0) dummy2<- if(k==1) V*diffSlope else V%*%diag(diffSlope) #t(diffSlope*t(-I(newZ>PSI))) newd<-cbind(x.values,dummy1,dummy2) colnames(newd)<-c(x.name,nameU, nameV) } else { newd<-cbind(x.values,dummy1) colnames(newd)<-c(x.name,nameU) } if(!x.name%in%names(coef(obj.seg))) newd<-newd[,-1,drop=FALSE] return(newd) } #-------------------------------------------------------------- if(missing(newdata)){ newd.ok<-model.frame(object) } else { #devi trasformare la variabili segmented attraverso dummy.matrix() nameU<-object$nameUV$U nameV<-object$nameUV$V nameZ<-object$nameUV$Z n<-nrow(newdata) r<-NULL for(i in 1:length(nameZ)){ x.values<-newdata[[nameZ[i]]] DM<-dummy.matrix(x.values, nameZ[i], object) r[[i]]<-DM } newd.ok<-data.frame(matrix(unlist(r), nrow=n, byrow = FALSE)) names(newd.ok)<- unlist(sapply(r, colnames)) idZ<-match(nameZ, names( newdata)) newdata<-cbind(newdata[,-idZ, drop=FALSE], newd.ok) # newdata<-subset(newdata, select=-idZ) newdata<-cbind(newdata, newd.ok) } class(object)<-class(object)[-1] f<-predict(object, newdata=newdata, ...) #f<-if(inherits(object, what = "glm", which = FALSE)) predict.glm(object, newdata=newd.ok, ...) else predict.lm(object, newdata=newd.ok, ...) return(f) #sommare se "terms"? } segmented/R/seg.lm.fit.boot.r0000644000176200001440000001322113157724622015531 0ustar liggesusersseg.lm.fit.boot<-function(y, XREG, Z, PSI, w, offs, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE){ #random se TRUE prende valori random quando e' errore: comunque devi modificare qualcosa (magari con it.max) # per fare restituire la dev in corrispondenza del punto psi-random #nonParm. se TRUE implemneta il case resampling. Quello semiparam dipende dal non-errore di #---------------------------------- # sum.of.squares<-function(obj.seg){ # #computes the "correct" SumOfSquares from a segmented" fit # b<-obj.seg$obj$coef # X<-qr.X(obj.seg$obj$qr) #X<-model.matrix(obj.seg) # X<-X[,!is.na(b)] # b<-b[!is.na(b)] # rev.b<-rev(b) # rev.b[1:length(obj.seg$psi)]<-0 # b<-rev(rev.b) # new.fitted<-drop(X%*%b) # new.res<- obj.seg$obj$residuals + obj.seg$obj$fitted - new.fitted # ss<-sum(new.res^2) # ss # } #-------- extract.psi<-function(lista){ #serve per estrarre il miglior psi.. dev.values<-lista[[1]] psi.values<-lista[[2]] dev.ok<-min(dev.values) id.dev.ok<-which.min(dev.values) if(is.list(psi.values)) psi.values<-matrix(unlist(psi.values), nrow=length(dev.values), byrow=TRUE) if(!is.matrix(psi.values)) psi.values<-matrix(psi.values) psi.ok<-psi.values[id.dev.ok,] r<-list(SumSquares.no.gap=dev.ok, psi=psi.ok) r } #------------- visualBoot<-opz$visualBoot opz.boot<-opz opz.boot$pow=c(1.1,1.2) opz1<-opz opz1$it.max <-1 n<-length(y) o0<-try(seg.lm.fit(y, XREG, Z, PSI, w, offs, opz, return.all.sol=FALSE), silent=TRUE) rangeZ <- apply(Z, 2, range) #serve sempre if(!is.list(o0)) { o0<- seg.lm.fit(y, XREG, Z, PSI, w, offs, opz, return.all.sol=TRUE) o0<-extract.psi(o0) if(!nonParam) {warning("using nonparametric boot");nonParam<-TRUE} } if(is.list(o0)){ est.psi00<-est.psi0<-o0$psi ss00<-o0$SumSquares.no.gap if(!nonParam) fitted.ok<-fitted(o0) } else { if(!nonParam) stop("the first fit failed and I cannot extract fitted values for the semipar boot") if(random) { est.psi00<-est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0<-try(seg.lm.fit(y, XREG, Z, PSI1, w, offs, opz1), silent=TRUE) ss00<-o0$SumSquares.no.gap } else { est.psi00<-est.psi0<-apply(PSI,2,mean) ss00<-opz$dev0 } } all.est.psi.boot<-all.selected.psi<-all.est.psi<-matrix(, nrow=n.boot, ncol=length(est.psi0)) all.ss<-all.selected.ss<-rep(NA, n.boot) if(is.null(size.boot)) size.boot<-n # na<- ,,apply(...,2,function(x)mean(is.na(x))) Z.orig<-Z if(visualBoot) cat(0, " ", formatC(opz$dev0, 3, format = "f"),"", "(No breakpoint(s))", "\n") count.random<-0 for(k in seq(n.boot)){ PSI <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) if(jt) Z<-apply(Z.orig,2,jitter) if(nonParam){ id<-sample(n, size=size.boot, replace=TRUE) o.boot<-try(seg.lm.fit(y[id], XREG[id,,drop=FALSE], Z[id,,drop=FALSE], PSI[id,,drop=FALSE], w[id], offs[id], opz.boot), silent=TRUE) } else { yy<-fitted.ok+sample(residuals(o0),size=n, replace=TRUE) o.boot<-try(seg.lm.fit(yy, XREG, Z.orig, PSI, weights, offs, opz.boot), silent=TRUE) } if(is.list(o.boot)){ all.est.psi.boot[k,]<-est.psi.boot<-o.boot$psi } else { est.psi.boot<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) } PSI <- matrix(rep(est.psi.boot, rep(nrow(Z), length(est.psi.boot))), ncol = length(est.psi.boot)) opz$h<-max(opz$h*.9, .2) opz$it.max<-opz$it.max+1 o<-try(seg.lm.fit(y, XREG, Z.orig, PSI, w, offs, opz, return.all.sol=TRUE), silent=TRUE) if(!is.list(o) && random){ est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o<-try(seg.lm.fit(y, XREG, Z, PSI1, w, offs, opz1), silent=TRUE) count.random<-count.random+1 } if(is.list(o)){ if(!"coefficients"%in%names(o$obj)) o<-extract.psi(o) all.est.psi[k,]<-o$psi all.ss[k]<-o$SumSquares.no.gap if(o$SumSquares.no.gap<=ifelse(is.list(o0), o0$SumSquares.no.gap, 10^12)) o0<-o est.psi0<-o0$psi all.selected.psi[k,] <- est.psi0 all.selected.ss[k]<-o0$SumSquares.no.gap #min(c(o$SumSquares.no.gap, o0$SumSquares.no.gap)) } if(visualBoot) { flush.console() spp <- if (k < 10) "" else NULL cat(k, spp, "", formatC(o0$SumSquares.no.gap, 3, format = "f"), "\n") } } #end n.boot all.selected.psi<-rbind(est.psi00,all.selected.psi) all.selected.ss<-c(ss00, all.selected.ss) ris<-list(all.selected.psi=drop(all.selected.psi),all.selected.ss=all.selected.ss, all.psi=all.est.psi, all.ss=all.ss) if(is.null(o0$obj)){ PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0<-try(seg.lm.fit(y, XREG, Z, PSI1, w, offs, opz1), silent=TRUE) } if(!is.list(o0)) return(0) o0$boot.restart<-ris return(o0) }segmented/R/summary.segmented.R0000644000176200001440000001177513157724622016244 0ustar liggesusers`summary.segmented` <- function(object, short=FALSE, var.diff=FALSE, ...){ if(is.null(object$psi)) object<-object[[length(object)]] #i seguenti per calcolare aa,bb,cc funzionano per lm e glm, da verificare con arima.... # nome<-rownames(object$psi) # nome<-as.character(parse("",text=nome)) # aa<-grep("U",names(coef(object)[!is.na(coef(object))])) # bb<-unlist(sapply(nome,function(x){grep(x,names(coef(object)[!is.na(coef(object))]))},simplify=FALSE,USE.NAMES=FALSE)) # cc<-intersect(aa,bb) #indices of diff-slope parameters # iV<- -grep("psi.",names(coef(object)[!is.na(coef(object))]))#indices of all but the Vs if(var.diff && length(object$nameUV$Z)>1) { var.diff<-FALSE warning("var.diff set to FALSE with multiple segmented variables", call.=FALSE) } nomiU<-object$nameUV[[1]] nomiV<-object$nameUV[[2]] idU<-match(nomiU,names(coef(object)[!is.na(coef(object))])) idV<-match(nomiV,names(coef(object)[!is.na(coef(object))])) beta.c<- coef(object)[nomiU] #per metodo default.. if( !inherits(object, "segmented")){ summ <- c(summary(object, ...), object["psi"]) summ[c("it","epsilon")]<-object[c("it","epsilon")] coeff<-coef(object) v<-try(vcov(object), silent=TRUE) if(class(v)!="try-error"){ v<-sqrt(diag(v)) summ$gap<-cbind(coeff[idV]*beta.c,abs(v[idV]*beta.c),coeff[idV]/v[idV]) colnames(summ$gap)<-c("Est.","SE","t value") rownames(summ$gap)<-nomiU } else { summ$gap<-cbind(coeff[idV]*beta.c,NA,NA) colnames(summ$gap)<-c("Est.","SE","t value") rownames(summ$gap)<-nomiU } return(summ) } if("lm"%in%class(object) && !"glm"%in%class(object)){ #if(!inherits(object, "glm")){ summ <- c(summary.lm(object, ...), object["psi"]) summ$Ttable<-summ$coefficients if(var.diff){ sigma2.new<-tapply(object$residuals, object$id.group, function(xx){sum(xx^2)}) summ$df.new<-tapply(object$residuals, object$id.group, function(xx){(length(xx)-length(object$coef))}) summ$sigma.new<-sqrt(sigma2.new/summ$df.new) #modifica gli SE Qr <- object$qr p <- object$rank p1 <- 1L:p inv.XtX <- chol2inv(Qr$qr[p1, p1, drop = FALSE]) X <- qr.X(Qr,FALSE) attr(X, "assign") <- NULL sigma.i<-rowSums(model.matrix(~0+factor(object$id.group))%*%diag(summ$sigma.new)) var.b<-inv.XtX%*%crossprod(X*sigma.i)%*%inv.XtX dimnames(var.b)<-dimnames(summ$cov.unscaled) summ$cov.var.diff<-var.b summ$Ttable[,2]<-sqrt(diag(var.b)) summ$Ttable[,3]<-summ$Ttable[,1]/summ$Ttable[,2] summ$Ttable[,4]<- 2 * pnorm(abs(summ$Ttable[,3]), lower.tail = FALSE) dimnames(summ$Ttable) <- list(names(object$coefficients)[Qr$pivot[p1]], c("Estimate", "Std. Error", "z value", "Pr(>|z|)")) } coeff<-summ$Ttable[,1]#summ$coefficients[,1] v<-summ$Ttable[,2] #summ$coefficients[,2] summ$gap<-cbind(coeff[idV]*beta.c,abs(v[idV]*beta.c),coeff[idV]/v[idV]) summ$Ttable[idU,4]<-NA summ$Ttable<-summ$Ttable[-idV,] #dimnames(summ$gap)<-list(rep("",nrow(object$psi)),c("Est.","SE","t value")) colnames(summ$gap)<-c("Est.","SE","t value") rownames(summ$gap)<-nomiU summ[c("it","epsilon")]<-object[c("it","epsilon")] summ$var.diff<-var.diff summ$short<-short class(summ) <- c("summary.segmented", "summary.lm") return(summ) } #if("glm"%in%class(object)){ if(inherits(object, "glm")){ summ <- c(summary.glm(object, ...), object["psi"]) summ$Ttable<-summ$coefficients[-idV,] summ$Ttable[idU,4]<-NA coeff<-summ$coefficients[,1] v<-summ$coefficients[,2] summ$gap<-cbind(coeff[idV]*beta.c,abs(v[idV]*beta.c),coeff[idV]/v[idV]) #dimnames(summ$gap)<-list(rep("",nrow(object$psi)),c("Est.","SE","t value")) colnames(summ$gap)<-c("Est.","SE","t value") rownames(summ$gap)<-nomiU summ[c("it","epsilon")]<-object[c("it","epsilon")] summ$short<-short class(summ) <- c("summary.segmented", "summary.glm") return(summ)} if("Arima"%in%class(object)){ #da controllare coeff<-object$coef v<-sqrt(diag(object$var.coef)) Ttable<-cbind(coeff[-idV],v[-idV],coeff[-idV]/v[-idV]) object$gap<-cbind(coeff[idV]*beta.c,v[idV]*beta.c,coeff[idV]/v[idV]) #dimnames(object$gap)<-list(rep("",nrow(object$psi)),c("Est.","SE","t value")) colnames(object$gap)<-c("Est.","SE","t value") rownames(object$gap)<-nomiU colnames(Ttable)<-c("Estimate","Std. Error","t value") object$Ttable<-Ttable object$short<-short summ<-object class(summ) <- c("summary.segmented", "summary.Arima") return(summ)} } segmented/R/intercept.r0000644000176200001440000000716013157724622014623 0ustar liggesusersintercept<-function (ogg, parm, rev.sgn = FALSE, var.diff = FALSE, digits = max(3, getOption("digits") - 3)){ #corregge in caso di no model intercept -- CHE VOLEVO DIRE?? #forse che adesso funziona se nel modello non c'e' l'interc. #-- f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #-- #if (!"segmented" %in% class(ogg)) stop("A segmented model is needed") if (var.diff && length(ogg$nameUV$Z) > 1) { var.diff <- FALSE warning("var.diff set to FALSE with multiple segmented variables", call. = FALSE) } nomepsi <- rownames(ogg$psi) nomeU <- ogg$nameUV[[1]] nomeZ <- ogg$nameUV[[3]] if (missing(parm)) { nomeZ <- ogg$nameUV[[3]] if (length(rev.sgn) == 1) rev.sgn <- rep(rev.sgn, length(nomeZ)) } else { if (!all(parm %in% ogg$nameUV[[3]])) { stop("invalid parm") } else { nomeZ <- parm } } if (length(rev.sgn) != length(nomeZ)) rev.sgn <- rep(rev.sgn, length.out = length(nomeZ)) nomi <- names(coef(ogg)) nomi <- nomi[-match(nomepsi, nomi)] Allpsi <- index <- vector(mode = "list", length = length(nomeZ)) # gapCoef<-summary.segmented(ogg)$gap ##eliminato 10/11/15 Ris <- list() rev.sgn <- rep(rev.sgn, length.out = length(nomeZ)) if("(Intercept)"%in%names(coef(ogg))){ alpha0 <- alpha00 <- coef(ogg)["(Intercept)"]} else {alpha0 <- alpha00 <-0} #per ogni variabile segmented... for (i in 1:length(nomeZ)) { # id.cof.U <- grep(paste("\\.", nomeZ[i], "$", sep = ""), nomi, value = FALSE) # psii <- ogg$psi[grep(paste("\\.", nomeZ[i], "$", sep = ""), rownames(ogg$psi), value = FALSE), 2] id.cof.U <- f.U(ogg$nameUV$U, nomeZ[i]) + (match(ogg$nameUV$U[1], nomi)-1) psii<- ogg$psi[f.U(ogg$nameUV$V, nomeZ[i]) , "Est."] Allpsi[[i]] <- sort(psii, decreasing = FALSE) id.cof.U <- id.cof.U[order(psii)] index[[i]] <- id.cof.U alpha0<-if("(Intercept)"%in%names(coef(ogg))) coef(ogg)["(Intercept)"] else 0 ind <- as.numeric(na.omit(unlist(index[[i]]))) cof <- coef(ogg)[ind] alpha <- vector(length = length(ind)) #gapCoef.i<-gapCoef[grep(paste("\\.",nomeZ[i],"$",sep=""), rownames(gapCoef), value = FALSE),"Est."] # gapCoef.i<-gapCoef[f.U(rownames(gapCoef), nomeZ[i]) ,"Est."] ###eliminato 10/11/15 for (j in 1:length(cof)) { alpha[j] <- alpha0 - Allpsi[[i]][j] * cof[j] # if(gap) alpha[j] <- alpha[j] - gapCoef.i[j] ###eliminato 10/11/15 alpha0 <- alpha[j] } #if(gap) alpha<-alpha -gapCoef[grep(paste("\\.",nomeZ[i],"$",sep=""), rownames(gapCoef), value = FALSE),"Est."] cof.out <- c(alpha00, alpha) if(rev.sgn[i]) cof.out <- cof.out[length(cof.out):1] ris <- matrix(cof.out) dimnames(ris) <- list(paste("intercept", 1:nrow(ris), sep = ""), "Est.") Ris[[nomeZ[i]]] <- signif(ris, digits) } Ris } segmented/R/seg.control.R0000644000176200001440000000107213157724622015017 0ustar liggesusers`seg.control` <- function(toll=.0001, it.max=10, display=FALSE, stop.if.error=TRUE, K=10, quant=FALSE, last=TRUE, maxit.glm=25, h=1, n.boot=20, size.boot=NULL, gap=FALSE, jt=FALSE, nonParam=TRUE, random=TRUE, powers=c(1,1), seed=NULL, fn.obj=NULL, digits=NULL){ list(toll=toll,it.max=it.max,visual=display,stop.if.error=stop.if.error, K=K,last=last,maxit.glm=maxit.glm,h=h,n.boot=n.boot, size.boot=size.boot, gap=gap, jt=jt, nonParam=nonParam, random=random, pow=powers, seed=seed, quant=quant, fn.obj=fn.obj, digits=digits)} segmented/R/seg.def.fit.boot.r0000644000176200001440000001237313157724622015666 0ustar liggesusersseg.def.fit.boot<-function(obj, Z, PSI, mfExt, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE){ #random se TRUE prende valori random quando e' errore: comunque devi modificare qualcosa (magari con it.max) # per fare restituire la dev in corrispondenza del punto psi-random #nonParm. se TRUE implemneta il case resampling. Quello semiparam dipende dal non-errore di extract.psi<-function(lista){ #serve per estrarre il miglior psi.. dev.values<-lista[[1]] psi.values<-lista[[2]] dev.ok<-min(dev.values) id.dev.ok<-which.min(dev.values) if(is.list(psi.values)) psi.values<-matrix(unlist(psi.values), nrow=length(dev.values), byrow=TRUE) if(!is.matrix(psi.values)) psi.values<-matrix(psi.values) psi.ok<-psi.values[id.dev.ok,] r<-list(SumSquares.no.gap=dev.ok, psi=psi.ok) r } #------------- visualBoot<-opz$visualBoot opz.boot<-opz opz.boot$pow=c(1.1,1.2) opz1<-opz opz1$it.max <-1 n<-nrow(mfExt) o0<-try(seg.def.fit(obj, Z, PSI, mfExt, opz), silent=TRUE) rangeZ <- apply(Z, 2, range) #serve sempre if(!is.list(o0)) { o0<- seg.def.fit(obj, Z, PSI, mfExt, opz, return.all.sol=TRUE) o0<-extract.psi(o0) if(!nonParam) {warning("using nonparametric boot");nonParam<-TRUE} } if(is.list(o0)){ est.psi00<-est.psi0<-o0$psi ss00<-o0$SumSquares.no.gap if(!nonParam) fitted.ok<-fitted(o0) } else { if(!nonParam) stop("the first fit failed and I cannot extract fitted values for the semipar boot") if(random) { est.psi00<-est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0<-try(seg.def.fit(obj, Z, PSI1, mfExt, opz1), silent=TRUE) ss00<-o0$SumSquares.no.gap } else { est.psi00<-est.psi0<-apply(PSI,2,mean) ss00<-opz$dev0 } } all.est.psi.boot<-all.selected.psi<-all.est.psi<-matrix(, nrow=n.boot, ncol=length(est.psi0)) all.ss<-all.selected.ss<-rep(NA, n.boot) if(is.null(size.boot)) size.boot<-n # na<- ,,apply(...,2,function(x)mean(is.na(x))) Z.orig<-Z if(visualBoot) cat(0, " ", formatC(opz$dev0, 3, format = "f"),"", "(No breakpoint(s))", "\n") count.random<-0 for(k in seq(n.boot)){ PSI <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) if(jt) Z<-apply(Z.orig,2,jitter) if(nonParam){ id<-sample(n, size=size.boot, replace=TRUE) o.boot<-try(seg.def.fit(obj, Z[id,,drop=FALSE], PSI[id,,drop=FALSE], mfExt[id,,drop=FALSE], opz.boot), silent=TRUE) } else { yy<-fitted.ok+sample(residuals(o0),size=n, replace=TRUE) ##----> o.boot<-try(seg.lm.fit(yy, XREG, Z.orig, PSI, weights, offs, opz.boot), silent=TRUE) #in realta' la risposta dovrebbe essere "yy" da cambiare in mfExt o.boot<- try(seg.def.fit(obj, Z.orig, PSI, mfExt, opz.boot), silent=TRUE) } if(is.list(o.boot)){ all.est.psi.boot[k,]<-est.psi.boot<-o.boot$psi } else { est.psi.boot<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) } PSI <- matrix(rep(est.psi.boot, rep(nrow(Z), length(est.psi.boot))), ncol = length(est.psi.boot)) opz$h<-max(opz$h*.9, .2) opz$it.max<-opz$it.max+1 o <- try(seg.def.fit(obj, Z.orig, PSI, mfExt, opz, return.all.sol=TRUE), silent=TRUE) if(!is.list(o) && random){ est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o <- try(seg.def.fit(obj, Z, PSI1, mfExt, opz1), silent=TRUE) count.random<-count.random+1 } if(is.list(o)){ if(!"coefficients"%in%names(o$obj)) o<-extract.psi(o) all.est.psi[k,]<-o$psi all.ss[k]<-o$SumSquares.no.gap if(o$SumSquares.no.gap<=ifelse(is.list(o0), o0$SumSquares.no.gap, 10^12)) o0<-o est.psi0<-o0$psi all.selected.psi[k,] <- est.psi0 all.selected.ss[k]<-o0$SumSquares.no.gap #min(c(o$SumSquares.no.gap, o0$SumSquares.no.gap)) } if(visualBoot) { flush.console() spp <- if (k < 10) "" else NULL cat(k, spp, "", formatC(o0$SumSquares.no.gap, 3, format = "f"), "\n") } } #end n.boot all.selected.psi<-rbind(est.psi00,all.selected.psi) all.selected.ss<-c(ss00, all.selected.ss) ris<-list(all.selected.psi=drop(all.selected.psi),all.selected.ss=all.selected.ss, all.psi=all.est.psi, all.ss=all.ss) if(is.null(o0$obj)){ PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0 <- try(seg.def.fit(obj, Z, PSI1, mfExt, opz1), silent=TRUE) } if(!is.list(o0)) return(0) o0$boot.restart<-ris return(o0) }segmented/R/segmented.Arima.r0000644000176200001440000003503313157724622015631 0ustar liggesusers#`segmented.default` <- #o1<-segmented(out.lm, seg.Z=~x +z,psi=list(x=c(30,60),z=.3), control=seg.control(display=FALSE, n.boot=20, seed=1515)) #o2<-ss(out.lm, seg.Z=~x +z,psi=list(x=c(30,60),z=.3), control=seg.control(display=FALSE, n.boot=20, seed=1515)) #o2<-ss(out.lm, seg.Z=~x +z,psi=list(x=c(30,60),z=.3), control=seg.control(display=FALSE, n.boot=0)) #o2<-ss(o, seg.Z=~age, psi=41, control=seg.control(display=FALSE, n.boot=0)) segmented.Arima<- function(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) { #Richiede control$f.obj that should be a string like "sum(x$residuals^2)" or "x$dev" #----------------- dpmax<-function(x,y,pow=1){ #deriv pmax if(pow==1) ifelse(x>y, -1, 0) else -pow*pmax(x-y,0)^(pow-1) } #----------- n.Seg<-1 if(missing(psi)){if(length(all.vars(seg.Z))>1) stop("provide psi") else psi<-Inf} if(length(all.vars(seg.Z))>1 & !is.list(psi)) stop("`psi' should be a list with more than one covariate in `seg.Z'") if(is.list(psi)){ if(length(all.vars(seg.Z))!=length(psi)) stop("A wrong number of terms in `seg.Z' or `psi'") if(any(is.na(match(all.vars(seg.Z),names(psi), nomatch = NA)))) stop("Variables in `seg.Z' and `psi' do not match") n.Seg <- length(psi) } if(length(all.vars(seg.Z))!=n.Seg) stop("A wrong number of terms in `seg.Z' or `psi'") it.max <- old.it.max<- control$it.max digits<-control$digits toll <- control$toll visual <- control$visual stop.if.error<-control$stop.if.error n.boot<-control$n.boot # n.boot<-0 size.boot<-control$size.boot gap<-control$gap random<-control$random pow<-control$pow visualBoot<-FALSE if(n.boot>0){ if(!is.null(control$seed)) { set.seed(control$seed) employed.Random.seed<-control$seed } else { employed.Random.seed<-eval(parse(text=paste(sample(0:9, size=6), collapse=""))) set.seed(employed.Random.seed) } if(visual) {visual<-FALSE; visualBoot<-TRUE}# warning("`display' set to FALSE with bootstrap restart", call.=FALSE)} if(!stop.if.error) stop("Bootstrap restart only with a fixed number of breakpoints") } last <- control$last K<-control$K h<-min(abs(control$h),1) if(h<1) it.max<-it.max+round(it.max/2) name.Z <-all.vars(seg.Z) if(length(name.Z)!=n.Seg) stop("errore strano 1") Z<-sapply(name.Z, function(xx) eval(parse(text=xx))) #e' sempre una matrice if(length(name.Z)!=ncol(Z)) stop("errore strano 2") n<-nrow(Z) n.psi<- length(unlist(psi)) ################# if(ncol(Z)==1 && length(psi)==1 && n.psi==1 && !any(is.na(psi))) { if(psi==Inf) psi<-median(Z)} ################# if(ncol(Z)==1 && is.vector(psi) && (is.numeric(psi)||is.na(psi))){ psi <- list(as.numeric(psi)) names(psi)<-name.Z } if (!is.list(psi) || is.null(names(psi))) stop("psi should be a *named* list") id.nomiZpsi <- match(colnames(Z), names(psi)) if ((ncol(Z)!=length(psi)) || any(is.na(id.nomiZpsi))) stop("Length or names of Z and psi do not match") nome <- names(psi)[id.nomiZpsi] psi <- psi[nome] initial.psi<-psi for(i in 1:length(psi)) { if(any(is.na(psi[[i]]))) psi[[i]]<-if(control$quant) {quantile(Z[,i], prob= seq(0,1,l=K+2)[-c(1,K+2)], names=FALSE)} else {(min(Z[,i])+ diff(range(Z[,i]))*(1:K)/(K+1))} } a <- sapply(psi, length) #per evitare che durante il processo iterativo i psi non siano ordinati id.psi.group <- rep(1:length(a), times = a) #identificativo di apparteneza alla variabile Z<-matrix(unlist(mapply(function(x,y)rep(x,y),Z,a,SIMPLIFY = TRUE)),nrow=n,byrow = TRUE) #negli altri metodi Z e' una lista per cui la linea di sopra diventa #Z<-matrix(unlist(mapply(function(x,y)rep(x,y),Z,a,SIMPLIFY = TRUE)),nrow=n) colnames(Z) <- nomiZ.vett <- rep(nome, times = a) #SERVE??? si perche' Z e' senza colnames psi <- unlist(psi) #se psi e' numerico, la seguente linea restituisce i valori ordinati all'interno della variabile.. psi<-unlist(tapply(psi,id.psi.group,sort)) k <- ncol(Z) PSI <- matrix(rep(psi, rep(n, k)), ncol = k) #controllo se psi e' ammissibile.. c1 <- apply((Z <= PSI), 2, all) #dovrebbero essere tutti FALSE (prima era solo <) c2 <- apply((Z >= PSI), 2, all) #dovrebbero essere tutti FALSE (prima era solo >) if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2)) ) stop("starting psi out of the admissible range") #ripetizioni <- as.numeric(unlist(sapply(table(nomiZ)[order(unique(nomiZ))], function(xxx) {1:xxx}))) ripetizioni <- as.vector(unlist(tapply(id.psi.group, id.psi.group, function(x) 1:length(x) ))) nomiU <- paste("U", ripetizioni, sep = "") nomiU <- paste(nomiU, nomiZ.vett, sep = ".") nomiV <- paste("V", ripetizioni, sep = "") nomiV <- paste(nomiV, nomiZ.vett, sep = ".") nnomi <- c(nomiU, nomiV) # U <- pmax((Z - PSI), 0)^pow[1]#U <- pmax((Z - PSI), 0) # #V <- dpmax(Z,PSI,pow=pow[2])# ifelse((Z > PSI), -1, 0) # V<-ifelse((Z > PSI), -1, 0) # # for(i in 1:k) { # mfExt[nomiU[i]] <- U[,i] # mfExt[nomiV[i]] <- V[,i] # } # Fo <- update.formula(formula(obj), as.formula(paste(".~.+", paste(nnomi, collapse = "+")))) # Fo.noV <- update.formula(formula(obj), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) # call.ok <- update(obj, formula = Fo, evaluate=FALSE, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) # call.noV <- update(obj, formula = Fo.noV, evaluate=FALSE, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) XREG<-eval(obj$call$xreg) nomiXREG<-setdiff(names(obj$coef),c("intercept", paste("ar",1:100,sep=""), paste("ma",1:100,sep=""))) XREG<-matrix(XREG, ncol=length(nomiXREG)) colnames(XREG)<-nomiXREG mio.init<-mio.init.noV<-NULL call.ok <- update(obj, xreg = cbind(XREG,U,V), init=mio.init, evaluate=FALSE) #, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) call.noV <- update(obj, xreg = cbind(XREG,U), init=mio.init.noV, evaluate=FALSE) #, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) # call.noV <- update(obj, formula = Fo.noV, evaluate=FALSE, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) if (it.max == 0) { obj1 <- eval(call.noV) #, envir=mfExt) return(obj1) } #obj1 <- eval(call.ok, envir=mfExt) initial <- psi obj0 <- obj # browser() dev0<- -obj$loglik if(is.na(dev0)) dev0<-10 list.obj <- list(obj) nomiOK<-nomiU opz<-list(toll=toll,h=h,stop.if.error=stop.if.error,dev0=dev0,visual=visual,it.max=it.max, nomiOK=nomiOK, id.psi.group=id.psi.group, gap=gap, visualBoot=visualBoot, pow=pow, digits=digits) opz$call.ok<-call.ok opz$call.noV<-call.noV opz$nomiU<-nomiU opz$nomiV<-nomiV # opz$fn.obj <- fn.obj if(n.boot<=0){ obj<-seg.Ar.fit(obj, XREG, Z, PSI, opz) } else { obj<-seg.Ar.fit.boot(obj, XREG, Z, PSI, opz, n.boot=n.boot, size.boot=size.boot, random=random) #jt, nonParam } if(!is.list(obj)){ warning("No breakpoint estimated", call. = FALSE) return(obj0) } if(!is.null(obj$obj$df.residual)){ if(obj$obj$df.residual==0) warning("no residual degrees of freedom (other warnings expected)", call.=FALSE) } id.psi.group<-obj$id.psi.group nomiOK<-obj$nomiOK #sarebbe nomiU #-- nomiVxb<-paste("psi",sapply(strsplit(nomiOK,"U"), function(x){x[2]}), sep="") #nomiFINALI<-unique(sapply(strsplit(nomiOK, split="[.]"), function(x)x[2])) #nomi delle variabili con breakpoint stimati! nomiFINALI<-unique(sub("U[1-9]*[0-9].", "", nomiOK)) #se e' stata usata una proc automatica "nomiFINALI" sara' differente da "name.Z" nomiSenzaPSI<-setdiff(name.Z,nomiFINALI) if(length(nomiSenzaPSI)>=1) warning("no breakpoints found for: ", paste(nomiSenzaPSI," "), call. = FALSE) #-- it<-obj$it psi<-obj$psi psi.values<-if(n.boot<=0) obj$psi.values else obj$boot.restart U<-obj$U V<-obj$V # return(obj) #if(any(table(rowSums(V))<=1)) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close") for(jj in colnames(V)) { VV<-V[, which(colnames(V)==jj), drop=FALSE] sumV<-abs(rowSums(VV)) if( #(any(diff(sumV)>=2)|| #se ci sono due breakpoints uguali any(table(sumV)<=1) && stop.if.error) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") } rangeZ<-obj$rangeZ obj<-obj$obj k<-length(psi) # beta.c<-if(k == 1) coef(obj)["U"] else coef(obj)[paste("U", 1:ncol(U), sep = "")] beta.c<- coef(obj)[nomiU] psi.values[[length(psi.values) + 1]] <- psi id.warn <- FALSE if (n.boot<=0 && it > it.max) { #it >= (it.max+1) warning("max number of iterations attained", call. = FALSE) id.warn <- TRUE } Vxb <- V %*% diag(beta.c, ncol = length(beta.c)) # #se usi una procedura automatica devi cambiare ripetizioni, nomiU e nomiV, e quindi: # length.psi<-tapply(as.numeric(as.character(names(psi))), as.numeric(as.character(names(psi))), length) # forma.nomiU<-function(xx,yy)paste("U",1:xx, ".", yy, sep="") # forma.nomiVxb<-function(xx,yy)paste("psi",1:xx, ".", yy, sep="") # nomiU <- unlist(mapply(forma.nomiU, length.psi, name.Z)) #in realta' non serve, c'era gia'! # nomiVxb <- unlist(mapply(forma.nomiVxb, length.psi, name.Z)) # nnomi <- c(nomiU, nomiVxb) # # colnames(U)<-nomiU # colnames(Vxb)<-nomiVxb # ## for(i in 1:ncol(U)) { # mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] # mfExt[nomiVxb[i]]<-mf[nomiVxb[i]]<-Vxb[,i] # } nnomi <- c(nomiU, nomiVxb) # browser() # Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nnomi, collapse = "+")))) # objF <- update(obj0, formula = Fo, evaluate=FALSE, data = mfExt) # objF<- eval(objF, envir=mfExt) #browser() # objF <- update(obj0, xreg = cbind(XREG,U,Vxb), evaluate=TRUE) #, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) XREG.ok<-cbind(XREG, U, Vxb) colnames(XREG.ok)[((ncol(XREG.ok)-length(nnomi)+1):ncol(XREG.ok))]<- nnomi objF <- update(obj0, xreg = XREG.ok, evaluate=TRUE) #, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) #Puo' capitare che psi sia ai margini e ci sono 1 o 2 osservazioni in qualche intervallo. Oppure ce ne #sono di piu' ma hanno gli stessi valori di x #objF$coef puo' avere mancanti.. names(which(is.na(coef(objF)))) if(any(is.na(objF$coef)) && stop.if.error){ stop("at least one coef estimate is NA: breakpoint(s) at the boundary? (possibly with many x-values replicated)", call. = FALSE) } # CONTROLLARE!!!! # objF$offset<- obj0$offset #sostituire i valori: objF include le U e V, obj solo le U if(!gap){ #names.coef <- names(objF$coefficients) #names(obj$coefficients)[match(nomiV, names(coef(obj)))]<-nomiVxb #objF$coefficients[names.coef]<-obj$coefficients[names.coef] names.coef <- names(obj$coef) objF$coef[names.coef]<-obj$coef[names.coef] objF$coef[nomiVxb]<-rep(0, k) #if(!is.null(objF$fitted.values)) objF$fitted.values<-obj$fitted.values if(!is.null(objF$residuals)) objF$residuals<-obj$residuals if(!is.null(objF$weights)) objF$weights<-obj$weights if(!is.null(objF$aic)) objF$aic<-obj$aic + 2*k } if(any(is.na(objF$coef))){ #Se gap==FALSE qui non ci possono essere NA (sono sostituiti dagli 0) stop("some estimate is NA: premature stopping with a large number of breakpoints?", call. = FALSE) } Cov <- try(vcov(objF), silent=TRUE) idd <- match(nomiVxb, names(coef(objF))) if(class(Cov)!="try-error") { vv <- if (length(idd) == 1) Cov[idd, idd] else diag(Cov[idd, idd]) } else { vv<-NA } a<-tapply(id.psi.group, id.psi.group, length) #ho sovrascritto "a" di sopra, ma non dovrebbe servire.. ris.psi<-matrix(,length(psi),3) colnames(ris.psi) <- c("Initial", "Est.", "St.Err") rownames(ris.psi) <- nomiVxb ris.psi[,2]<-psi ris.psi[,3]<-sqrt(vv) a.ok<-NULL for(j in name.Z){ if(j %in% nomiFINALI) { a.ok[length(a.ok)+1]<-a[1] a<-a[-1] } else { a.ok[length(a.ok)+1]<-0 } #ifelse(name.Z %in% nomiFINALI,1,0) } # initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi, a.ok, SIMPLIFY = TRUE)) initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi[nomiFINALI], a.ok[a.ok!=0], SIMPLIFY = TRUE)) ris.psi[,1]<-initial # a<-tapply(id.psi.group, id.psi.group, length) #ho sovrascritto "a" di sopra, ma non dovrebbe servire.. # initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi, a)) # id <- match(nomiVxb, names(coef(objF))) # Cov <- try(vcov(objF), silent=TRUE) # if(class(Cov)!="try-error") { # vv <- if (length(id) == 1) Cov[id, id] else diag(Cov[id, id]) # #if(length(initial)!=length(psi)) initial<-rep(NA,length(psi)) # psi <- cbind(initial, psi, sqrt(vv)) # rownames(psi) <- colnames(Cov)[id] # colnames(psi) <- c("Initial", "Est.", "St.Err") # } else { # psi <- cbind(initial, psi) # rownames(psi) <- nomiVxb # colnames(psi) <- c("Initial", "Est.") # } objF$Z <- Z objF$rangeZ <- rangeZ objF$psi.history <- psi.values objF$psi <- ris.psi objF$it <- (it - 1) objF$epsilon <- obj$epsilon objF$call <- match.call() #objF$nameUV <- list(U = nomiU, V = rownames(psi), Z = name.Z) objF$nameUV <- list(U = drop(nomiU), V = rownames(ris.psi), Z = nomiFINALI) #Z = name.Z objF$id.group <- if(length(name.Z)<=1) -rowSums(as.matrix(V)) objF$id.psi.group <- id.psi.group objF$id.warn <- id.warn # browser() # objF$orig.call<-orig.call # if (model) objF$model <- mf #objF$mframe <- data.frame(as.list(KK)) if(n.boot>0) objF$seed<-employed.Random.seed class(objF) <- c("segmented", class(obj0)) list.obj[[length(list.obj) + 1]] <- objF class(list.obj) <- "segmented" if (last) list.obj <- list.obj[[length(list.obj)]] warning("'segmented.Arima' is at a preliminary stage. Estimates are OK, but the '*.segmented' methods are not expected to work", call.=FALSE) return(list.obj) } #end function segmented/R/plot.segmented.R0000644000176200001440000002657113157724622015525 0ustar liggesusersplot.segmented<-function (x, term, add = FALSE, res = FALSE, conf.level = 0, interc=TRUE, link = TRUE, res.col = 1, rev.sgn = FALSE, const = 0, shade=FALSE, rug=TRUE, dens.rug=FALSE, dens.col = grey(0.8), transf=I, ...){ #funzione plot.segmented che consente di disegnare anche i pointwise CI f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #-------------- enl.range<-function(..., enlarge=TRUE){ #modifica il min dei valori in ... r<-range(...) if(enlarge) r[1]<-if(sign(r[1])>0) r[1]*.9 else r[1]*1.1 r } #-------------- #se l'oggetto e' segmented.Arima il nome dell'eventuale interc va sostituito.. # if((all(class(x)==c("segmented", "Arima")))) names(x$coef)<-gsub("intercept", "(Intercept)", names(coef(x))) if(all(c("segmented", "Arima") %in% class(x))) names(x$coef)<-gsub("intercept", "(Intercept)", names(coef(x))) #-------------- linkinv <- !link if (inherits(x, what = "glm", which = FALSE) && linkinv && !is.null(x$offset) && res) stop("residuals with offset on the response scale?") if(conf.level< 0 || conf.level>.9999) stop("meaningless 'conf.level'") if ((inherits(x, what = "glm", which = FALSE) && linkinv) || res) { if(!(identical(transf, I) || identical(transf, "I"))) {transf<-I; warning("'transf' set to I..")} } show.gap<-FALSE if (missing(term)) { if (length(x$nameUV$Z) > 1) { stop("please, specify `term'") } else { term <- x$nameUV$Z } } else { dterm<- deparse(substitute(term)) if(dterm %in% x$nameUV$Z) term<-dterm if (! isTRUE(term %in% x$nameUV$Z)) stop("invalid `term'") } opz <- list(...) cols <- opz$col if (length(cols) <= 0) cols <- 1 lwds <- opz$lwd if (length(lwds) <= 0) lwds <- 1 ltys <- opz$lty if (length(ltys) <= 0) ltys <- 1 cexs <- opz$cex if (length(cexs) <= 0) cexs <- 1 pchs <- opz$pch if (length(pchs) <= 0) pchs <- 1 xlabs <- opz$xlab if (length(xlabs) <= 0) xlabs <- term ylabs <- opz$ylab if (length(ylabs) <= 0) ylabs <- paste("Effect of ", term, sep = " ") #a <- intercept(x, term, gap = show.gap)[[1]][, "Est."] a <- intercept(x, term)[[1]][, "Est."] #Poiche' intercept() restituisce quantita' che includono sempre l'intercetta del modello, questa va eliminata se interc=FALSE if(!interc && ("(Intercept)" %in% names(coef(x)))) a<- a-coef(x)["(Intercept)"] b <- slope(x, term)[[1]][, "Est."] #id <- grep(paste("\\.", term, "$", sep = ""), rownames(x$psi), value = FALSE) #confondeva "psi1.x","psi1.neg.x" id <- f.U(rownames(x$psi), term) est.psi <- x$psi[id, "Est."] K <- length(est.psi) val <- sort(c(est.psi, x$rangeZ[, term])) #---------aggiunta per gli IC rangeCI<-NULL n<-length(x$residuals) #fitted.values - Arima non ha "fitted.values", ma ha "residuals".. tipo<- if(inherits(x, what = "glm", which = FALSE) && link) "link" else "response" vall<-sort(c(seq(min(val), max(val), l=150), est.psi)) #ciValues<-predict.segmented(x, newdata=vall, se.fit=TRUE, type=tipo, level=conf.level) vall.list<-list(vall) names(vall.list)<-term if(conf.level>0) { #k.alpha<-if(inherits(x, what = c("glm","Arima"), which = FALSE)) abs(qnorm((1-conf.level)/2)) else abs(qt((1-conf.level)/2, x$df.residual)) #cambiato nella 0.5-2.0: k.alpha<-if(inherits(x, what = "lm", which = FALSE)) abs(qt((1-conf.level)/2, x$df.residual)) else abs(qnorm((1-conf.level)/2)) ciValues<-broken.line(x, vall.list, link=link, interc=interc, se.fit=TRUE) ciValues<-cbind(ciValues$fit, ciValues$fit- k.alpha*ciValues$se.fit, ciValues$fit + k.alpha*ciValues$se.fit) #---> transf... ciValues<-apply(ciValues, 2, transf) rangeCI<-range(ciValues) #ciValues e' una matrice di length(val)x3. Le 3 colonne: stime, inf, sup #polygon(c(vall, rev(vall)), c(ciValues[,2],rev(ciValues[,3])), col = "gray", border=NA) } #--------- a.ok <- c(a[1], a) b.ok <- c(b[1], b) y.val <- a.ok + b.ok * val + const a.ok1 <- c(a, a[length(a)]) b.ok1 <- c(b, b[length(b)]) y.val <- y.val1 <- a.ok1 + b.ok1 * val + const s <- 1:(length(val) - 1) # xvalues <- if(all(class(x)==c("segmented", "Arima"))) x$Z[,1] else x$model[, term] xvalues <- if(all(c("segmented", "Arima") %in% class(x))) x$Z[,1] else x$model[, term] if (rev.sgn) { val <- -val xvalues <- -xvalues } m <- cbind(val[s], y.val1[s], val[s + 1], y.val[s + 1]) #values where to compute predictions (useful only if res=TRUE) if(res){ new.d<-data.frame(ifelse(rep(rev.sgn, length(xvalues)),-xvalues, xvalues)) names(new.d)<-term fit0 <- broken.line(x, new.d, link = link, interc=interc, se.fit=FALSE)$fit } #------------------------------------------------------------------------------- if (inherits(x, what = "glm", which = FALSE) && linkinv) { #se GLM con linkinv fit <- if (res) #predict.segmented(x, ifelse(rep(rev.sgn, length(xvalues)),-xvalues,xvalues), type=tipo) + resid(x, "response") + const #broken.line(x, term, gap = show.gap, link = link) + resid(x, "response") + const fit0 + resid(x, "response") + const else x$family$linkinv(c(y.val, y.val1)) # xout <- sort(c(seq(val[1], val[length(val)], l = 150), val[-c(1, length(val))])) xout <- sort(c(seq(val[1], val[length(val)], l = 150), val[-c(1, length(val))],val[-c(1, length(val))]*1.005)) l <- approx(as.vector(m[, c(1, 3)]), as.vector(m[, c(2, 4)]), xout = xout) val[length(val)]<-max(l$x) #aggiunto 11/09/17 id.group <- cut(l$x, val, FALSE, TRUE) yhat <- l$y xhat <- l$x m[, c(2, 4)] <- x$family$linkinv(m[, c(2, 4)]) if (!add) { plot(as.vector(m[, c(1, 3)]), as.vector(m[, c(2, 4)]), type = "n", xlab = xlabs, ylab = ylabs, main = opz$main, sub = opz$sub, xlim = opz$xlim, ylim = if(is.null(opz$ylim)) enl.range(fit, rangeCI, enlarge=dens.rug) else opz$ylim ) if(dens.rug){ density <- density( xvalues ) # the height of the densityity curve max.density <- max(density$y) # Get the boundaries of the plot to # put the density polygon at the x-line plot_coordinates <- par("usr") # get the "length" and range of the y-axis y.scale <- plot_coordinates[4] - plot_coordinates[3] # transform the y-coordinates of the density # to the lower 10% of the plotting panel density$y <- (0.1 * y.scale / max.density) * density$y + plot_coordinates[3] ## plot the polygon polygon( density$x , density$y , border = F , col = dens.col) box() } if(rug) { segments(xvalues, rep(par()$usr[3],length(xvalues)), xvalues, rep(par()$usr[3],length(xvalues))+ abs(diff(par()$usr[3:4]))/40)} } if(conf.level>0){ if(rev.sgn) vall<- -vall if(shade) polygon(c(vall, rev(vall)), c(ciValues[,2],rev(ciValues[,3])), col = "gray", border=NA) else matlines(vall, ciValues[,-1], type="l", lty=2, col=cols) } if (res) points(xvalues, fit, cex = cexs, pch = pchs, col = res.col) yhat <- x$family$linkinv(yhat) if (length(cols) == 1) cols <- rep(cols, max(id.group)) if (length(lwds) == 1) lwds <- rep(lwds, max(id.group)) if (length(ltys) == 1) ltys <- rep(ltys, max(id.group)) for (i in 1:max(id.group)) { lines(xhat[id.group == i], yhat[id.group == i], col = cols[i], lwd = lwds[i], lty = ltys[i]) } #------------------------------------------------------------------------------- } else { #se LM o "GLM con link=TRUE (ovvero linkinv=FALSE)" ##---> transf!!! y.val<- do.call(transf, list(y.val)) y.val1<-do.call(transf, list(y.val1)) r <- cbind(val, y.val) r1 <- cbind(val, y.val1) rr <- rbind(r, r1) fit <- c(y.val, y.val1) if (res) { ress <- if (inherits(x, what = "glm", which = FALSE)) residuals(x, "working") #* sqrt(x$weights) mgcv::gam() usa " ..*sqrt(x$weights)/mean(sqrt(x$weights))" else resid(x) #if(!is.null(x$offset)) ress<- ress - x$offset #fit <- broken.line(x, term, gap = show.gap, link = link, interc = TRUE) + ress + const #fit <- predict.segmented(x, ifelse(rep(rev.sgn, length(xvalues)),-xvalues,xvalues), type=tipo) + ress + const fit <- fit0 + ress + const } if (!add) plot(rr, type = "n", xlab = xlabs, ylab = ylabs, main = opz$main, sub = opz$sub, xlim = opz$xlim, #ylim = if(is.null(opz$ylim)) enl.range(fit, rangeCI, enlarge=dens.rug) else opz$ylim) ylim = if(is.null(opz$ylim)) enl.range(fit, rangeCI, do.call(transf, list(m[, c(2,4)])), enlarge=dens.rug) else opz$ylim) if(dens.rug){ density <- density( xvalues ) # the height of the densityity curve max.density <- max(density$y) # Get the boundaries of the plot to # put the density polygon at the x-line plot_coordinates <- par("usr") # get the "length" and range of the y-axis y.scale <- plot_coordinates[4] - plot_coordinates[3] # transform the y-coordinates of the density # to the lower 10% of the plotting panel density$y <- (0.1 * y.scale / max.density) * density$y + plot_coordinates[3] ## plot the polygon polygon( density$x , density$y , border = F , col = dens.col) box() } if(rug) {segments(xvalues, rep(par()$usr[3],length(xvalues)), xvalues, rep(par()$usr[3],length(xvalues))+ abs(diff(par()$usr[3:4]))/40)} if(conf.level>0) { if(rev.sgn) vall<- -vall if(shade) polygon(c(vall, rev(vall)), c(ciValues[,2],rev(ciValues[,3])), col = "gray", border=NA) else matlines(vall, ciValues[,-1], type="l", lty=2, col=cols) } if (res) points(xvalues, fit, cex = cexs, pch = pchs, col = res.col) segments(m[, 1], do.call(transf, list(m[, 2])), m[, 3], do.call(transf, list(m[, 4])), col = cols, lwd = lwds, lty = ltys) } invisible(NULL) } segmented/R/seg.Ar.fit.r0000644000176200001440000001333713157724622014531 0ustar liggesusersseg.Ar.fit<-function(obj, XREG, Z, PSI, opz, return.all.sol=FALSE){ #----------------- dpmax<-function(x,y,pow=1){ #deriv pmax if(pow==1) ifelse(x>y, -1, 0) else -pow*pmax(x-y,0)^(pow-1) } #----------- c1 <- apply((Z <= PSI), 2, all) c2 <- apply((Z >= PSI), 2, all) if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2))) stop("psi out of the range") # digits<-opz$digits pow<-opz$pow nomiOK<-opz$nomiOK toll<-opz$toll h<-opz$h gap<-opz$gap stop.if.error<-opz$stop.if.error dev.new<-opz$dev0 visual<-opz$visual id.psi.group<-opz$id.psi.group it.max<-old.it.max<-opz$it.max rangeZ <- apply(Z, 2, range) psi<-PSI[1,] names(psi)<-id.psi.group #H<-1 it <- 1 epsilon <- 10 dev.values<-psi.values <- NULL id.psi.ok<-rep(TRUE, length(psi)) nomiU<- opz$nomiU nomiV<- opz$nomiV call.ok <- opz$call.ok call.noV <- opz$call.noV toll<-opz$toll k<-ncol(Z) mio.init<-NULL mio.init.noV<-NULL while (abs(epsilon) > toll) { U <- pmax((Z - PSI), 0)^pow[1]#U <- pmax((Z - PSI), 0) colnames(U)<-nomiU V <- dpmax(Z,PSI,pow=pow[2])# ifelse((Z > PSI), -1, 0) colnames(V)<-nomiV # for(i in 1:k) { # mfExt[nomiU[i]] <- U[,i] # mfExt[nomiV[i]] <- V[,i] # } # obj <- suppressWarnings(eval(call.ok, envir=mfExt)) obj <- suppressWarnings(eval(call.ok)) call.ok$init=quote(coef(obj)) #mio.init<- c(0,obj$coef[-1]) dev.old<-dev.new dev.new <- dev.new1 <- -obj$loglik #control$f.obj should be something like "sum(x$residuals^2)" or "x$dev" if(return.all.sol) { obj.noV <- suppressWarnings(eval(call.noV)) #, envir=mfExt #mio.init.noV<-obj.noV$coef #mio.init.noV<- c(0,obj.noV$coef[-1]) dev.new1 <- -obj.noV$loglik #dev.new1 <- sum(mylm(x = cbind(XREG, U), y = y, w = w, offs = offs)$residuals^2) } dev.values[[length(dev.values) + 1]] <- dev.new1 if (visual) { flush.console() if (it == 1) cat(0, " ", formatC(dev.old, 3, format = "f"), "", "(No breakpoint(s))", "\n") spp <- if (it < 10) "" else NULL cat(it, spp, "", formatC(dev.new, 3, format = "f"), "",length(psi),"\n") #cat(paste("iter = ", it, spp," dev = ",formatC(dev.new,digits=3,format="f"), " n.psi = ",formatC(length(psi),digits=0,format="f"), sep=""), "\n") } epsilon <- (dev.new - dev.old)/(dev.old + .001) obj$epsilon <- epsilon it <- it + 1 obj$it <- it beta.c<-coef(obj)[nomiU] gamma.c<-coef(obj)[nomiV] if (it > it.max) break psi.values[[length(psi.values) + 1]] <- psi.old <- psi # if(it>=old.it.max && h<1) H<-h psi <- psi.old + h*gamma.c/beta.c if(!is.null(digits)) psi<-round(psi, digits) PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #check if psi is admissible.. a <- apply((Z <= PSI), 2, all) #prima era solo < b <- apply((Z >= PSI), 2, all) #prima era solo > if(stop.if.error) { isErr<- (sum(a + b) != 0 || is.na(sum(a + b))) if(isErr) { if(return.all.sol) return(list(dev.values, psi.values)) else stop("(Some) estimated psi out of its range") } } else { id.psi.ok<-!is.na((a+b)<=0)&(a+b)<=0 Z <- Z[,id.psi.ok,drop=FALSE] psi <- psi[id.psi.ok] PSI <- PSI[,id.psi.ok,drop=FALSE] nomiOK<-nomiOK[id.psi.ok] #salva i nomi delle U per i psi ammissibili id.psi.group<-id.psi.group[id.psi.ok] names(psi)<-id.psi.group if(ncol(PSI)<=0) return(0) k<-ncol(Z) } #end else #obj$psi <- psi } #end while psi<-unlist(tapply(psi, id.psi.group, sort)) names(psi)<-id.psi.group PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #aggiunto da qua.. U <- pmax((Z - PSI), 0) colnames(U)<-nomiU V <- ifelse((Z > PSI), -1, 0) colnames(V)<-nomiV # for(i in 1:k) { # mfExt[nomiU[i]] <- U[,i] # mfExt[nomiV[i]] <- V[,i] # } ##LA DOMANDA E': PERCHE' QUI STIMA UN MODELLO SENZA V SE POI VIENE RISTIMATO in segmented.default (o segmented.lm o segmented.glm?) ##RE: il valore di SS.new serve per il boot restart. #Invece la domanda e': non si puo' restituire direttamente obj.new senza bisogno di sostituire i valori in obj ? obj.new <- suppressWarnings(eval(call.noV)) #, envir=mfExt)) SS.new <- -obj.new$loglik #sum(obj.new$residuals^2) if(!gap){ obj<-obj.new #names.coef<-names(obj$coefficients) #obj$coefficients<-c(obj.new$coefficients, rep(0,ncol(V))) #names(obj$coefficients)<-names.coef #obj$residuals<-obj.new$residuals #obj$fitted.values<-obj.new$fitted.values #obj$linear.predictors<-obj.new$linear.predictors #obj$deviance<-obj.new$deviance #obj$weights<-obj.new$weights #obj$aic<-obj.new$aic #+ 2*ncol(V) #ho fatto la modifica in segmented.glm(): "objF$aic<-obj$aic + 2*k" } else { obj <- suppressWarnings(eval(call.ok)) #, envir=mfExt)) } obj$epsilon <- epsilon obj$it <- it #fino a qua.. obj<-list(obj=obj,it=it,psi=psi,psi.values=psi.values,U=U,V=V,rangeZ=rangeZ, epsilon=epsilon,nomiOK=nomiOK, SumSquares.no.gap=SS.new, id.psi.group=id.psi.group) #inserire id.psi.ok? return(obj) } segmented/R/confint.segmented.R0000644000176200001440000000700613157724622016177 0ustar liggesusers`confint.segmented` <- function(object, parm, level=0.95, rev.sgn=FALSE, var.diff=FALSE, digits=max(3, getOption("digits") - 3), ...){ #-- f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #-- # if(!"segmented"%in%class(object)) stop("A segmented model is needed") if(var.diff && length(object$nameUV$Z)>1) { var.diff<-FALSE warning("var.diff set to FALSE with multiple segmented variables", call.=FALSE) } #nomi delle variabili segmented: if(missing(parm)) { nomeZ<- object$nameUV[[3]] if(length(rev.sgn)==1) rev.sgn<-rep(rev.sgn,length(nomeZ)) } else { if(! parm %in% object$nameUV[[3]]) {stop("invalid parm")} else {nomeZ<-parm} } if(length(rev.sgn)!=length(nomeZ)) rev.sgn<-rep(rev.sgn, length.out=length(nomeZ)) rr<-list() z<-if("lm"%in%class(object)) abs(qt((1-level)/2,df=object$df.residual)) else abs(qnorm((1-level)/2)) for(i in 1:length(nomeZ)){ #per ogni variabile segmented `parm' (tutte o selezionata).. #nomi.U<-grep(paste("\\.",nomeZ[i],"$",sep=""),object$nameUV$U,value=TRUE) #nomi.V<-grep(paste("\\.",nomeZ[i],"$",sep=""),object$nameUV$V,value=TRUE) nomi.U<- object$nameUV$U[f.U(object$nameUV$U, nomeZ[i])] nomi.V<- object$nameUV$V[f.U(object$nameUV$V, nomeZ[i])] m<-matrix(,length(nomi.U),3) rownames(m)<-nomi.V colnames(m)<-c("Est.",paste("CI","(",level*100,"%",")",c(".l",".u"),sep="")) for(j in 1:length(nomi.U)){ #per ogni psi della stessa variabile segmented.. sel<-c(nomi.V[j],nomi.U[j]) V<-vcov(object,var.diff=var.diff)[sel,sel] #questa e' vcov di (psi,U) b<-coef(object)[sel[2]] #diff-Slope th<-c(b,1) orig.coef<-drop(diag(th)%*%coef(object)[sel]) #sono i (gamma,beta) th*coef(ogg)[sel] gammma<-orig.coef[1] est.psi<-object$psi[sel[1],2] V<-diag(th)%*%V%*%diag(th) #2x2 vcov() di gamma e beta se.psi<-sqrt((V[1,1]+V[2,2]*(gammma/b)^2-2*V[1,2]*(gammma/b))/b^2) r<-c(est.psi, est.psi-z*se.psi, est.psi+z*se.psi) if(rev.sgn[i]) r<-c(-r[1],rev(-r[2:3])) m[j,]<-r } #end loop j (ogni psi della stessa variabile segmented) #CONTROLLA QUESTO:..sarebbe piu' bello if(nrow(m)==1) rownames(m)<-"" else m<-m[order(m[,1]),] if(rev.sgn[i]) { #m<-m[nrow(m):1,] rownames(m)<-rev(rownames(m)) } rr[[length(rr)+1]]<-signif(m,digits) } #end loop i (ogni variabile segmented) names(rr)<-nomeZ return(rr) } #end_function segmented/R/seg.glm.fit.r0000644000176200001440000001441013157724622014737 0ustar liggesusersseg.glm.fit<-function(y,XREG,Z,PSI,w,offs,opz,return.all.sol=FALSE){ #------------------------- dpmax<-function(x,y,pow=1){ #deriv pmax if(pow==1) ifelse(x>y, -1, 0) else -pow*pmax(x-y,0)^(pow-1) } #-------------------- c1 <- apply((Z <= PSI), 2, all) #prima era solo < c2 <- apply((Z >= PSI), 2, all) #prima era solo > if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2))) stop("psi out of the range") digits<-opz$digits pow<-opz$pow eta0<-opz$eta0 fam<-opz$fam maxit.glm<-opz$maxit.glm #-------------- nomiOK<-opz$nomiOK toll<-opz$toll h<-opz$h stop.if.error<-opz$stop.if.error dev.new<-opz$dev0 visual<-opz$visual id.psi.group<-opz$id.psi.group it.max<-old.it.max<-opz$it.max id.psi.group<-opz$id.psi.group gap<-opz$gap rangeZ <- apply(Z, 2, range) #k<-ncol(Z) psi<-PSI[1,] names(psi)<-id.psi.group it <- 1 epsilon <- 10 dev.values<-psi.values <- NULL id.psi.ok<-rep(TRUE, length(psi)) sel.col.XREG<-unique(sapply(colnames(XREG), function(x)match(x,colnames(XREG)))) if(is.numeric(sel.col.XREG))XREG<-XREG[,sel.col.XREG,drop=FALSE] #elimina le ripetizioni, ad es. le due intercette.. while (abs(epsilon) > toll) { k<-ncol(Z) U <- pmax((Z - PSI), 0)^pow[1] V <- dpmax(Z,PSI,pow=pow[2])# ifelse((Z > PSI), -1, 0) X <- cbind(XREG, U, V) rownames(X) <- NULL if (ncol(V) == 1) colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c("U", "V") else colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c(paste("U", 1:ncol(U), sep = ""), paste("V", 1:ncol(V), sep = "")) #obj <- lm.wfit(x = X, y = y, w = w, offset = o) #controlla****************** obj <- suppressWarnings(glm.fit(x = X, y = y, offset = offs, weights = w, family = fam, control = glm.control(maxit = maxit.glm), etastart = eta0)) eta0 <- obj$linear.predictors dev.old<-dev.new dev.new <- dev.new1<- obj$dev if(return.all.sol) dev.new1 <- glm.fit(x=cbind(XREG, U),y=y, family=fam, weights=w, offset=offs, etastart=eta0)$dev dev.values[[length(dev.values) + 1]] <- dev.new1 if (visual) { flush.console() if (it == 1) cat(0, " ", formatC(dev.old, 3, format = "f"), "", "(No breakpoint(s))", "\n") spp <- if (it < 10) "" else NULL cat(it, spp, "", formatC(dev.new1, 3, format = "f"), "",length(psi),"\n") #cat(paste("iter = ", it, spp," dev = ",formatC(dev.new,digits=3,format="f"), " n.psi = ",formatC(length(psi),digits=0,format="f"), sep=""), "\n") } epsilon <- (dev.new - dev.old)/(dev.old+.001) # epsilon <- (dev.new1 - dev.old)/dev.old #se vuoi usare la *vera* (e non la working che tiene conto dei gap) deviance obj$epsilon <- epsilon it <- it + 1 obj$it <- it #class(obj) <- c("segmented", class(obj)) #list.obj[[length(list.obj) + ifelse(last == TRUE, 0, 1)]] <- obj if (k == 1) { beta.c <- coef(obj)["U"] gamma.c <- coef(obj)["V"] } else { beta.c <- coef(obj)[paste("U", 1:ncol(U), sep = "")] gamma.c <- coef(obj)[paste("V", 1:ncol(V), sep = "")] } if (it > it.max) break psi.values[[length(psi.values) + 1]] <- psi.old <- psi #if(it>=old.it.max && h<1) H<-h psi <- psi.old + h*gamma.c/beta.c if(!is.null(digits)) psi<-round(psi, digits) PSI <- matrix(rep(psi, rep(nrow(Z), ncol(Z))), ncol = ncol(Z)) #check if psi is admissible.. a <- apply((Z <= PSI), 2, all) #prima era solo < b <- apply((Z >= PSI), 2, all) #prima era solo > if(stop.if.error) { isErr <- sum(a + b) != 0 || is.na(sum(a + b)) if(isErr) { if(return.all.sol) return(list(dev.values, psi.values)) else stop("(Some) estimated psi out of its range") } } else { id.psi.ok<-!is.na((a+b)<=0)&(a+b)<=0 Z <- Z[,id.psi.ok,drop=FALSE] psi <- psi[id.psi.ok] PSI <- PSI[,id.psi.ok,drop=FALSE] nomiOK<-nomiOK[id.psi.ok] #salva i nomi delle U per i psi ammissibili id.psi.group<-id.psi.group[id.psi.ok] names(psi)<-id.psi.group if(ncol(PSI)<=0) return(0) } #end else #obj$psi <- psi } #end while #queste due righe aggiunte nella versione 0.2.9-3 (adesso i breakpoints sono sempre ordinati) psi<-unlist(tapply(psi, id.psi.group, sort)) names(psi)<-id.psi.group PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #aggiunto da qua.. U <- pmax((Z - PSI), 0) V <- ifelse((Z > PSI), -1, 0) X <- cbind(XREG, U, V) rownames(X) <- NULL if (ncol(V) == 1) colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c("U", "V") else colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c(paste("U", 1:ncol(U), sep = ""), paste("V", 1:ncol(V), sep = "")) #tolto il suppressWarnings( obj <- glm.fit(x = X, y = y, offset = offs, weights = w, family = fam, control = glm.control(maxit = maxit.glm), etastart = eta0) obj$epsilon <- epsilon obj$it <- it obj.new <- glm.fit(x = cbind(XREG, U), y = y, offset = offs, weights = w, family = fam, control = glm.control(maxit = maxit.glm), etastart = eta0) SS.new<- obj.new$dev if(!gap){ names.coef<-names(obj$coefficients) obj$coefficients<-c(obj.new$coefficients, rep(0,ncol(V))) names(obj$coefficients)<-names.coef obj$residuals<-obj.new$residuals obj$fitted.values<-obj.new$fitted.values obj$linear.predictors<-obj.new$linear.predictors obj$deviance<-obj.new$deviance obj$weights<-obj.new$weights obj$aic<-obj.new$aic #+ 2*ncol(V) #ho fatto la modifica in segmented.glm(): "objF$aic<-obj$aic + 2*k" } #fino a qua.. obj<-list(obj=obj,it=it,psi=psi,psi.values=psi.values,U=U,V=V,rangeZ=rangeZ, epsilon=epsilon,nomiOK=nomiOK, dev.no.gap=SS.new, id.psi.group=id.psi.group) return(obj) } segmented/R/pscore.test.r0000644000176200001440000003262513157724622015103 0ustar liggesusers`pscore.test` <- function(obj, seg.Z, k = 10, alternative = c("two.sided", "less", "greater"), values=NULL, dispersion=NULL, df.t=NULL, more.break=FALSE) { #------------------------------------------------------------------------------- test.Sc2<-function(y, z, xreg, sigma=NULL, values=NULL, fn="pmax(x-p,0)", df.t="Inf", alternative, w=NULL){ #xreg: la matrice del disegno del modello nullo. Se mancante viene assunta solo l'intercetta. #Attenzione che se invXtX e xx vengono entrambe fornite, non viene fatto alcun controllo #invXtX: {X'X}^{-1}. if missing it is computed from xreg #sigma: the sd. If missing it is computed from data (under the *null* model) #values: the values with respect to ones to compute the average term. If NULL 10 values from min(z) to max(z) are taken. n<-length(y) if(missing(xreg)) xreg<-cbind(rep(1,n)) id.ok<-complete.cases(cbind(y,z,xreg)) y<-y[id.ok] z<-z[id.ok] xreg<-xreg[id.ok,,drop=FALSE] n<-length(y) k=ncol(xreg) #per un modello ~1+x if(is.null(values)) values<-seq(min(z), max(z), length=10) n1<-length(values) PSI<-matrix(values, nrow=n, ncol=n1, byrow=TRUE) #(era X2) matrice di valori di psi if(is.matrix(z)) { X1<-matrix(z[,1], nrow=n, ncol=n1, byrow=FALSE) X2<-matrix(z[,2], nrow=n, ncol=n1, byrow=FALSE) X<-eval(parse(text=fn), list(x=X1, y=X2, p=PSI)) #X<-pmax(X1-X2,0) pmaxMedio<-rowMeans(X) } else { X1<-matrix(z, nrow=n, ncol=n1, byrow=FALSE) #matrice della variabile Z if(length(fn)<=1){ X<-eval(parse(text=fn), list(x=X1, p=PSI)) #X<-pmax(X1-X2,0) pmaxMedio<-rowMeans(X) } else { pmaxMedio<-matrix(,n,length(fn)) #list.X<-vector("list", length=length(fn)) for(j in 1:length(fn)){ #list.X[[j]]<-eval(parse(text=fn[j]), list(x=X1, p=PSI)) X<-eval(parse(text=fn[[j]]), list(x=X1, p=PSI)) pmaxMedio[,j]<-rowMeans(X) } } } if(is.null(w)){ invXtX<-solve(crossprod(xreg)) IA<-diag(n) - xreg%*%tcrossprod(invXtX, xreg) #I- hat matrix r<-IA%*%y sc<- sum(pmaxMedio*r) v.s<-crossprod(pmaxMedio,IA)%*% pmaxMedio } else { invXtX<-solve(crossprod(sqrt(w)*xreg)) IA<-diag(n) - xreg%*%tcrossprod(invXtX, xreg*w) #I-hat matrix sc<-t(pmaxMedio*w) %*% IA %*% y v.s<- t(pmaxMedio*w) %*% crossprod(t(IA)/sqrt(w))%*%(w*pmaxMedio) } # ris<-if(length(fn)<=1) sc/(sigma*sqrt(v.s)) else drop(crossprod(sc,solve(v.s,sc)))/(sigma^2) # if(length(fn)<=1 && cadj) ris<- sign(ris)*sqrt((ris^2)*(1-(3-(ris^2))/(2*n))) ris<- drop(sc/(sigma*sqrt(v.s))) df.t<-eval(parse(text=df.t)) pvalue<- switch(alternative, less = pt(ris, df=df.t, lower.tail =TRUE) , greater = pt(abs(ris), df=df.t, lower.tail =FALSE) , two.sided = 2*pt(abs(ris), df=df.t, lower.tail =FALSE) ) #pvalue<- 2*pt(abs(ris), df=df.t, lower.tail =FALSE) r<-c(ris, pvalue, pmaxMedio) r } #------------------------------------------------------------------------------- scGLM<-function(y, z, xreg, family, values = NULL, size=1, weights.var, fn="pmax(x-p,0)", alternative=alternative){ #score test for GLM #size (only if family=binomial()) #weights.var: weights to be used for variance computations. If missing the weights come from the null fit output<-match.arg(output) n<-length(y) if(missing(xreg)) xreg<-cbind(rep(1,n)) id.ok<-complete.cases(cbind(y,z,xreg)) y<-y[id.ok] z<-z[id.ok] xreg<-xreg[id.ok,,drop=FALSE] n<-length(y) if(family$family=="poisson") size=1 if(length(size)==1) size<-rep(size,n) yN<-y/size k=ncol(xreg) #per un modello ~1+x if(is.null(values)) values<-seq(min(z), max(z), length=10) n1<-length(values) PSI<-matrix(values, nrow=n, ncol=n1, byrow=TRUE) #(era X2) matrice di valori di psi X1<-matrix(z, nrow=n, ncol=n1, byrow=FALSE) #matrice della variabile Z X<-eval(parse(text=fn), list(x=X1, p=PSI)) #X<-pmax(X1-X2,0) pmaxMedio<-rowMeans(X) o<-glm.fit(yN, x=xreg, weights=size, family=family) r<-y-(o$fitted*size) sc<-drop(crossprod(r, pmaxMedio)) # if(output=="unst.score") return(drop(sc)) p <- o$rank Qr <- o$qr COV <- chol2inv(Qr$qr[1:p, 1:p, drop = FALSE]) #vcov(glm(y~x, family=poisson)) A<-xreg%*%COV%*%crossprod(xreg, diag(o$weights)) h<- drop(tcrossprod(pmaxMedio, diag(n)- A)) if(missing(weights.var)) weights.var<-o$weights v.s<- drop(crossprod(h*sqrt(weights.var))) #t(h)%*%diag(exp(lp))%*%h ris<-if(length(fn)<=1) sc/sqrt(v.s) else drop(crossprod(sc,solve(v.s,sc))) # if(output=="score") return(drop(ris)) pvalue<- switch(alternative, less = pnorm(ris, lower.tail =TRUE) , greater = pnorm(abs(ris), lower.tail =FALSE) , two.sided = 2*pnorm(abs(ris), lower.tail =FALSE) ) # pvalue<- if(length(fn)<=1) 2*pnorm(abs(ris), lower.tail =FALSE) else pchisq(ris,df=length(fn), lower.tail =FALSE) # NB: se calcoli ris<-drop(t(sc)%*%solve(v.s,sc))/(length(fn)*sigma^2) devi usare pf(ris,df1=length(fn),df2=df.t, lower.tail =FALSE) return(c(ris, pvalue)) } #---------------------------------------------------- fn="pmax(x-p,0)" # if(inherits(obj, "glm")) stop("Currently only 'lm', or 'segmented-lm' models are allowed") if(!inherits(obj, "lm")) stop("A 'lm', or 'segmented-lm' model is requested") if(inherits(obj, "segmented") && length(obj$nameUV$Z)==1) seg.Z<- as.formula(paste("~", obj$nameUV$Z )) if(!inherits(obj, "segmented") && length(all.vars(formula(obj)))==2) seg.Z<- as.formula(paste("~", all.vars(formula(obj))[2])) if(class(seg.Z)!="formula") stop("'seg.Z' should be an one-sided formula") name.Z <- all.vars(seg.Z) if(length(name.Z)>1) stop("Only a single segmented variable can be specified in 'seg.Z' ") if(k<=1) stop("k>1 requested! k>=10 is recommended") if(k<10) warnings("k>=10 is recommended") alternative <- match.arg(alternative) #if(length(all.vars(seg.Z))>1) warning("multiple segmented variables ignored in 'seg.Z'",call.=FALSE) isGLM<-"glm"%in%class(obj) if(isGLM){ if(is.null(dispersion)) dispersion<- summary.glm(obj)$dispersion if(inherits(obj, "segmented")){ mf<-model.frame(obj) X0<-model.matrix(obj) Z<-X0[,name.Z] n<-length(Z) if(is.null(values)) values<-seq(min(Z), max(Z), length=k) #values<-seq(sort(Z)[2], sort(Z)[(n - 1)], length = k) n1<-length(values) #escludi *tutte* le variabili psi (sia da X0 che dalla formula) #X0<-X0[, -match(obj$nameUV$V, colnames(X0))] #formulaNull <-update.formula(formula(obj),paste("~.-",paste(obj$nameUV$V, collapse="-"))) #togli tutti i termini "V" X1<-matrix(Z, nrow=n, ncol=n1, byrow=FALSE) #matrice della variabile Z fn="pmax(x-p,0)" PSI<-matrix(values, nrow=n, ncol=n1, byrow=TRUE) X<-eval(parse(text=fn), list(x=X1, p=PSI)) # fn t.c. length(fn)<=1 mf$pmaxMedio<-rowMeans(X) nc<-obj$orig.call #se c'e' un break e vuoi saggiare uno in piu' devi aggiustare la call if(more.break){ nc$formula<-update.formula(nc$formula, paste("~.+",paste(obj$nameUV$U, collapse="+"))) } formulaNull<-nc$formula nc$data=quote(mf) a<-eval(nc) #assign("mf", mf, envir=sys.frame()) #funziona ma R CMD check da problemi.. #ne <- new.env(parent = baseenv()) pos<-1 assign("mf", mf, envir=as.environment(pos)) #r<-as.numeric(add1(a, ~.+pmaxMedio, scale=dispersion, test="Rao")[c("scaled Rao sc.", "Pr(>Chi)")][2,]) r<- as.numeric(add1(a, ~.+pmaxMedio, scale=dispersion, test="Rao")[2,4:5]) } else { Call<-mf<-obj$call mf$formula<-formula(obj) m <- match(c("formula", "data", "subset", "weights", "na.action","offset"), names(mf), 0L) mf <- mf[c(1, m)] mf$drop.unused.levels <- TRUE mf[[1L]] <- as.name("model.frame") mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) formulaNull <- formula(obj) mf <- eval(mf, parent.frame()) mt <- attr(mf, "terms") XREG <- if (!is.empty.model(mt)) model.matrix(mt, mf, contrasts) n <- nrow(XREG) Z<- XREG[,match(name.Z, colnames(XREG))] if(!name.Z %in% names(coef(obj))) XREG<-XREG[,-match(name.Z, colnames(XREG)),drop=FALSE] if(is.null(values)) values<-seq(min(Z), max(Z), length=k) #values<-seq(sort(Z)[2], sort(Z)[(n - 1)], length = k) n1<-length(values) PSI<-matrix(values, nrow=n, ncol=n1, byrow=TRUE) #(era X2) matrice di valori di psi X1<-matrix(Z, nrow=n, ncol=n1, byrow=FALSE) #matrice della variabile Z fn="pmax(x-p,0)" X<-eval(parse(text=fn), list(x=X1, p=PSI)) # fn t.c. length(fn)<=1 pmaxMedio<-rowMeans(X) #r<-as.numeric(add1(update(obj, data=mf), ~.+pmaxMedio, scale=dispersion, test="Rao")[c("scaled Rao sc.", "Pr(>Chi)")][2,]) r<-as.numeric(add1(update(obj, data=mf), ~.+pmaxMedio, scale=dispersion, test="Rao")[2,4:5]) } } else { #SE E' un LM if(is.null(dispersion)) dispersion<- summary(obj)$sigma^2 if(is.null(df.t)) df.t <- obj$df.residual #df.ok<- if(!is.null(df.t)) df.t else obj$df.residual # se e' LM-segmented if(inherits(obj, "segmented")){ if(!is.null(eval(obj$call$obj)$call$data)) mf$data <- eval(obj$call$obj)$call$data y<- obj$res+obj$fitted if(!is.null(obj$offset)) y<- y-obj$offset weights<- obj$weights X0<-model.matrix(obj) Z<-X0[,name.Z] #escludi *tutte* le variabili psi (sia da X0 che dalla formula) X0<-X0[, -match(obj$nameUV$V, colnames(X0))] formulaNull <-update.formula(formula(obj),paste("~.-",paste(obj$nameUV$V, collapse="-"))) #togli tutti i termini "V" #se vuoi saggiare per un additional breakpoint if(!more.break){ idU<-grep(name.Z, obj$nameUV$U) X0<-X0[, -match(obj$nameUV$U[idU], colnames(X0))] formulaNull <-update.formula(formulaNull, paste("~.-",paste(obj$nameUV$U[idU], collapse="-"))) } #browser() if(is.null(values)) values<-seq(min(Z), max(Z), length=k) #values<-seq(sort(Z)[2], sort(Z)[(n - 1)], length = k) # formulaOrig<-mf$formula<-update.formula(mf$formula,paste("~.-",paste(obj$nameUV$V, collapse="-"))) #togli tutti i termini "V" # update.formula(formulaOrig, paste("~.-",paste(obj$nameUV$U[idU], collapse="-"))) # #for(i in 1:length(obj$nameUV$U)) assign(obj$nameUV$U[i], obj$model[,obj$nameUV$U[i]], envir=parent.frame()) # formulaOrig<-update.formula(formulaOrig, paste("~.-",paste(obj$nameUV$U, collapse="-"))) r<-test.Sc2(y=y, z=Z, xreg=X0, sigma=sqrt(dispersion), values=values, fn=fn, df.t=df.t, alternative=alternative, w=weights) } else { #SE l'oggetto obj NON e' "segmented".. # ci possono essere mancanti nella variabile seg.Z, quindi devi costruire un nuovo dataframe.... Call<-mf<-obj$call mf$formula<-formula(obj) m <- match(c("formula", "data", "subset", "weights", "na.action","offset"), names(mf), 0L) mf <- mf[c(1, m)] mf$drop.unused.levels <- TRUE mf[[1L]] <- as.name("model.frame") mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) formulaNull <- formula(obj) mf <- eval(mf, parent.frame()) weights <- as.vector(model.weights(mf)) offs <- as.vector(model.offset(mf)) mt <- attr(mf, "terms") interc<-attr(mt,"intercept") y <- model.response(mf, "any") XREG <- if (!is.empty.model(mt)) model.matrix(mt, mf, contrasts) n <- nrow(XREG) if (is.null(weights)) weights <- rep(1, n) if (!is.null(offs)) y<-y-offs Z<- XREG[,match(name.Z, colnames(XREG))] if(!name.Z %in% names(coef(obj))) XREG<-XREG[,-match(name.Z, colnames(XREG)),drop=FALSE] if(is.null(values)) values<-seq(min(Z), max(Z), length=k) #values<-seq(sort(Z)[2], sort(Z)[(n - 1)], length = k) r<-test.Sc2(y=y, z=Z, xreg=XREG, sigma=sqrt(dispersion), values=values, fn=fn, df.t=df.t, alternative=alternative, w=weights) } } if(is.null(obj$family$family)) { famiglia<-"gaussian" legame<-"identity" } else { famiglia<-obj$family$family legame<-obj$family$link } out <- list(method = "Score test for one change in the slope", # data.name=paste("Model = ",famiglia,", link =", legame, # "\nformula =", as.expression(formulaOrig), # "\nsegmented variable =", name.Z), data.name=paste("formula =", as.expression(formulaNull), ", method =", obj$call[[1]] , "\nmodel =",famiglia,", link =", legame, NULL , "\nsegmented variable =", name.Z), statistic = c(`observed value` = r[1]), parameter = c(n.points = length(values)), p.value = r[2], alternative = alternative) class(out) <- "htest" return(out) } segmented/R/segmented.lm.R0000644000176200001440000005220013157724622015143 0ustar liggesusers`segmented.lm` <- function(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) { n.Seg<-1 if(missing(seg.Z) && length(all.vars(formula(obj)))==2) seg.Z<- as.formula(paste("~", all.vars(formula(obj))[2])) if(missing(psi)){if(length(all.vars(seg.Z))>1) stop("provide psi") else psi<-Inf} if(length(all.vars(seg.Z))>1 & !is.list(psi)) stop("`psi' should be a list with more than one covariate in `seg.Z'") if(is.list(psi)){ if(length(all.vars(seg.Z))!=length(psi)) stop("A wrong number of terms in `seg.Z' or `psi'") if(any(is.na(match(all.vars(seg.Z),names(psi), nomatch = NA)))) stop("Variables in `seg.Z' and `psi' do not match") n.Seg <- length(psi) } if(length(all.vars(seg.Z))!=n.Seg) stop("A wrong number of terms in `seg.Z' or `psi'") it.max <- old.it.max<- control$it.max digits<-control$digits toll <- control$toll visual <- control$visual stop.if.error<-control$stop.if.error n.boot<-control$n.boot size.boot<-control$size.boot gap<-control$gap random<-control$random pow<-control$pow visualBoot<-FALSE if(n.boot>0){ if(!is.null(control$seed)) { set.seed(control$seed) employed.Random.seed<-control$seed } else { employed.Random.seed<-eval(parse(text=paste(sample(0:9, size=6), collapse=""))) set.seed(employed.Random.seed) } if(visual) {visual<-FALSE; visualBoot<-TRUE}# warning("`display' set to FALSE with bootstrap restart", call.=FALSE)} # if(!stop.if.error) stop("Bootstrap restart only with a fixed number of breakpoints") } last <- control$last K<-control$K h<-min(abs(control$h),1) if(h<1) it.max<-it.max+round(it.max/2) # if(!stop.if.error) objInitial<-obj #------------------------------- # #una migliore soluzione......... # objframe <- update(obj, model = TRUE, x = TRUE, y = TRUE) # y <- objframe$y # a <- model.matrix(seg.Z, data = eval(obj$call$data)) # a <- subset(a, select = colnames(a)[-1]) orig.call<-Call<-mf<-obj$call orig.call$formula<- mf$formula<-formula(obj) #per consentire lm(y~.) m <- match(c("formula", "data", "subset", "weights", "na.action","offset"), names(mf), 0L) mf <- mf[c(1, m)] mf$drop.unused.levels <- TRUE mf[[1L]] <- as.name("model.frame") if(class(mf$formula)=="name" && !"~"%in%paste(mf$formula)) mf$formula<-eval(mf$formula) #orig.call$formula<-update.formula(orig.call$formula, paste("~.-",all.vars(seg.Z))) # #genn 2013. dalla versione 0.2.9-4 ho tolto if(length(.. Tra l'altro non capisco perche' lo avevo fatto # if(length(all.vars(formula(obj)))>1){ # mf$formula<-update.formula(mf$formula,paste(paste(seg.Z,collapse=".+"),"+",paste(all.vars(formula(obj))[-1],collapse="+"))) # } else { # mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) # } #nov 2013 dalla versione 0.3-0.0 (che dovrebbe essere successiva alla 0.2-9.5) viene creato anche il modelframe esteso che comprende # termini "originali", prima che fossero trasformati (Ad es., x prima che ns(x) costruisca le basi). Questo permette di avere termini # ns(), poly(), bs() nel modello di partenza mfExt<- mf mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) #mfExt$formula<- update.formula(mfExt$formula,paste(paste(seg.Z,collapse=".+"),"+",paste(all.vars(formula(obj)),collapse="+"))) # mfExt$formula<- if(!is.null(obj$call$data)) # update.formula(mf$formula,paste(".~",paste(all.vars(obj$call), collapse="+"),"-",obj$call$data,sep="")) # else update.formula(mf$formula,paste(".~",paste(all.vars(obj$call), collapse="+"),sep="")) #----------- # browser() if(!is.null(obj$call$offset) || !is.null(obj$call$weights) || !is.null(obj$call$subset)){ mfExt$formula <- update.formula(mf$formula, paste(".~.+", paste( c(all.vars(obj$call$offset), all.vars(obj$call$weights), all.vars(obj$call$subset)), collapse = "+") )) } mf <- eval(mf, parent.frame()) n<-nrow(mf) #questo serve per inserire in mfExt le eventuali variabili contenute nella formula con offset(..) nomiOff<-setdiff(all.vars(formula(obj)), names(mf)) if(length(nomiOff)>=1) mfExt$formula<-update.formula(mfExt$formula,paste(".~.+", paste( nomiOff, collapse="+"), sep="")) #---------------------------------------------------- # browser() #ago 2014 c'e' la questione di variabili aggiuntive... nomiTUTTI<-all.vars(mfExt$formula) #comprende anche altri nomi (ad es., threshold) "variabili" nomiNO<-NULL for(i in nomiTUTTI){ r<-try(eval(parse(text=i), parent.frame()), silent=TRUE) if(class(r)!="try-error" && length(r)==1 && !is.function(r)) nomiNO[[length(nomiNO)+1]]<-i } #nomiNO dovrebbe contenere i nomi delle "altre variabili" (come th in subset=x0 if(any(id.duplic)) { #new.mf<-mf[,id.duplic,drop=FALSE] new.mf<-mf[,all.vars(formula(obj))[id.duplic],drop=FALSE] new.XREGseg<-data.matrix(new.mf) XREG<-cbind(XREG,new.XREGseg) } n.psi<- length(unlist(psi)) id.n.Seg<-(ncol(XREG)-n.Seg+1):ncol(XREG) XREGseg<-XREG[,id.n.Seg,drop=FALSE] #XREG<-XREG[,-id.n.Seg,drop=FALSE] #XREG<-model.matrix(obj0) non va bene perche' non elimina gli eventuali mancanti in seg.Z.. #Due soluzioni #XREG<-XREG[,colnames(model.matrix(obj)),drop=FALSE] #XREG<-XREG[,match(c("(Intercept)",all.vars(formula(obj))[-1]),colnames(XREG),nomatch =0),drop=FALSE] XREG <- XREG[, match(c("(Intercept)", namesXREG0),colnames(XREG), nomatch = 0), drop = FALSE] XREG<-XREG[,unique(colnames(XREG)), drop=FALSE] ################# if(ncol(XREGseg)==1 && length(psi)==1 && n.psi==1 && !any(is.na(psi))) { if(psi==Inf) psi<-median(XREGseg)} ################# n <- nrow(XREG) #Z <- list(); for (i in colnames(XREGseg)) Z[[length(Z) + 1]] <- XREGseg[, i] Z<-lapply(apply(XREGseg,2,list),unlist) #prende anche i nomi! name.Z <- names(Z) <- colnames(XREGseg) if(length(Z)==1 && is.vector(psi) && (is.numeric(psi)||is.na(psi))){ psi <- list(as.numeric(psi)) names(psi)<-name.Z } if (!is.list(Z) || !is.list(psi) || is.null(names(Z)) || is.null(names(psi))) stop("Z and psi have to be *named* list") id.nomiZpsi <- match(names(Z), names(psi)) if ((length(Z)!=length(psi)) || any(is.na(id.nomiZpsi))) stop("Length or names of Z and psi do not match") #dd <- match(names(Z), names(psi)) nome <- names(psi)[id.nomiZpsi] psi <- psi[nome] initial.psi<-psi for(i in 1:length(psi)) { if(any(is.na(psi[[i]]))) psi[[i]]<-if(control$quant) {quantile(Z[[i]], prob= seq(0,1,l=K+2)[-c(1,K+2)], names=FALSE)} else {(min(Z[[i]])+ diff(range(Z[[i]]))*(1:K)/(K+1))} } a <- sapply(psi, length) #per evitare che durante il processo iterativo i psi non siano ordinati id.psi.group <- rep(1:length(a), times = a) #identificativo di apparteneza alla variabile # #Znew <- list() #for (i in 1:length(psi)) Znew[[length(Znew) + 1]] <- rep(Z[i], a[i]) #Z <- matrix(unlist(Znew), nrow = n) Z<-matrix(unlist(mapply(function(x,y)rep(x,y),Z,a,SIMPLIFY = TRUE)),nrow=n) psi <- unlist(psi) #se psi e' numerico, la seguente linea restituisce i valori ordinati all'interno della variabile.. psi<-unlist(tapply(psi,id.psi.group,sort)) k <- ncol(Z) PSI <- matrix(rep(psi, rep(n, k)), ncol = k) #controllo se psi e' ammissibile.. c1 <- apply((Z <= PSI), 2, all) #dovrebbero essere tutti FALSE (prima era solo <) c2 <- apply((Z >= PSI), 2, all) #dovrebbero essere tutti FALSE (prima era solo >) if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2)) ) stop("starting psi out of the admissible range") #questo dovrebbe eliminare i psi non-ammissib. ma non sono sicuro cosa succede se ci sono piu' variabili # if(sum(c1 + c2) != 0){ # id.val.psi<-!((c1+c1)>0) #individua i psi ammissibili (i.e. interni) # psi<-psi[id.val.psi] # Z<-Z[,id.val.psi] # PSI<-PSI[,id.val.psi] # } # if(is.na(sum(c1 + c2))) stop("psi out of the range") colnames(Z) <- nomiZ <- rep(nome, times = a) ripetizioni <- as.numeric(unlist(sapply(table(nomiZ)[order(unique(nomiZ))], function(xxx) {1:xxx}))) nomiU <- paste("U", ripetizioni, sep = "") nomiU <- paste(nomiU, nomiZ, sep = ".") nomiV <- paste("V", ripetizioni, sep = "") nomiV <- paste(nomiV, nomiZ, sep = ".") #forse non serve crearsi l'ambiente KK, usa mf.. #obj <- update(obj, formula = Fo, data = mf) #if (model.frame) obj$model <- mf #controlla che model.frame() funzioni sull'oggetto restituito # KK <- new.env() # for (i in 1:ncol(objframe$model)) assign(names(objframe$model[i]), objframe$model[[i]], envir = KK) if (it.max == 0) { #mf<-cbind(mf, mfExt) U <- pmax((Z - PSI), 0) colnames(U) <- paste(ripetizioni, nomiZ, sep = ".") nomiU <- paste("U", colnames(U), sep = "") #for (i in 1:ncol(U)) assign(nomiU[i], U[, i], envir = KK) #e' necessario il for? puoi usare colnames(U)<-nomiU;mf[nomiU]<-U for(i in 1:ncol(U)) mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] Fo <- update.formula(formula(obj), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) obj <- update(obj, formula = Fo, evaluate=FALSE, data=mfExt) #data = mf, if(!is.null(obj[["subset"]])) obj[["subset"]]<-NULL obj<-eval(obj, envir=mfExt) if (model) obj$model <-mf #obj$model <- data.frame(as.list(KK)) psi <- cbind(psi, psi, 0) rownames(psi) <- paste(paste("psi", ripetizioni, sep = ""), nomiZ, sep=".") colnames(psi) <- c("Initial", "Est.", "St.Err") #names(psi)<-paste(paste("psi", ripetizioni, sep = ""), nomiZ, sep=".") obj$psi <- psi return(obj) } #XREG <- model.matrix(obj) creata sopra #o <- model.offset(objframe) #w <- model.weights(objframe) if (is.null(weights)) weights <- rep(1, n) if (is.null(offs)) offs <- rep(0, n) initial <- psi obj0 <- obj dev0<-sum(obj$residuals^2) list.obj <- list(obj) # psi.values <- NULL nomiOK<-nomiU opz<-list(toll=toll,h=h,stop.if.error=stop.if.error,dev0=dev0,visual=visual,it.max=it.max, nomiOK=nomiOK,id.psi.group=id.psi.group,gap=gap,visualBoot=visualBoot,pow=pow,digits=digits) if(n.boot<=0){ obj<-seg.lm.fit(y,XREG,Z,PSI,weights,offs,opz) } else { obj<-seg.lm.fit.boot(y, XREG, Z, PSI, weights, offs, opz, n.boot=n.boot, size.boot=size.boot, random=random) #jt, nonParam } if(!is.list(obj)){ warning("No breakpoint estimated", call. = FALSE) return(obj0) } if(obj$obj$df.residual==0) warning("no residual degrees of freedom (other warnings expected)", call.=FALSE) id.psi.group<-obj$id.psi.group nomiOK<-obj$nomiOK #nomiFINALI<-unique(sapply(strsplit(nomiOK, split="[.]"), function(x)x[2])) #nomi delle variabili con breakpoint stimati! #nomiFINALI<-sub("U[1-9].", "", nomiOK) #nomi originali delle variabili con breakpoint stimati! nomiFINALI<- unique(sub("U[1-9]*[0-9].", "", nomiOK)) #se e' stata usata una proc automatica "nomiFINALI" sara' differente da "name.Z" nomiSenzaPSI<-setdiff(name.Z,nomiFINALI) if(length(nomiSenzaPSI)>=1) warning("no breakpoints found for: ", paste(nomiSenzaPSI," "), call. = FALSE) it<-obj$it psi<-obj$psi psi.values<-if(n.boot<=0) obj$psi.values else obj$boot.restart U<-obj$U V<-obj$V rangeZ<-obj$rangeZ obj<-obj$obj k<-length(psi) beta.c<-if(k == 1) coef(obj)["U"] else coef(obj)[paste("U", 1:ncol(U), sep = "")] psi.values[[length(psi.values) + 1]] <- psi id.warn <- FALSE if (n.boot<=0 && it > it.max) { #it >= (it.max+1) warning("max number of iterations attained", call. = FALSE) id.warn <- TRUE } Vxb <- V %*% diag(beta.c, ncol = length(beta.c)) #se usi una procedura automatica devi cambiare ripetizioni, nomiU e nomiV, e quindi: length.psi<-tapply(as.numeric(as.character(names(psi))), as.numeric(as.character(names(psi))), length) forma.nomiU<-function(xx,yy)paste("U",1:xx, ".", yy, sep="") forma.nomiVxb<-function(xx,yy)paste("psi",1:xx, ".", yy, sep="") #nomiU <- unlist(mapply(forma.nomiU, length.psi, name.Z)) #invece di un ciclo #paste("U",1:length.psi[i], ".", name.Z[i]) #nomiVxb <- unlist(mapply(forma.nomiVxb, length.psi, name.Z)) nomiU <- unlist(mapply(forma.nomiU, length.psi, nomiFINALI)) #invece di un ciclo #paste("U",1:length.psi[i], ".", name.Z[i]) nomiVxb <- unlist(mapply(forma.nomiVxb, length.psi, nomiFINALI)) # nomiVxb <-paste("psi",ripetizioni, ".",nomiZ ,sep="") #Dalla 0.2.9-5 eliminati i seguenti. La linea sopra sembra sufficiente # colnames(U) <- colnames(Vxb) <-sapply(strsplit(nomiOK,"U"),function(x)x[2]) # #colnames(U) <- paste(ripetizioni, nomiZ, sep = ".") # #colnames(Vxb) <- paste(ripetizioni, nomiZ, sep = ".") # nomiU <- paste("U", colnames(U), sep = "") # nomiVxb <- paste("psi", colnames(Vxb), sep = "") # for (i in 1:ncol(U)) { # assign(nomiU[i], U[, i], envir = KK) # assign(nomiVxb[i], Vxb[, i], envir = KK) # } #mf<-cbind(mf, mfExt) #questo creava ripetizioni.. for(i in 1:ncol(U)) { mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] mfExt[nomiVxb[i]]<-mf[nomiVxb[i]]<-Vxb[,i] } nnomi <- c(nomiU, nomiVxb) # browser() Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nnomi, collapse = "+")))) #objF <- update(obj0, formula = Fo, data = KK) objF <- update(obj0, formula = Fo, evaluate=FALSE, data = mfExt) #eliminiamo subset, perche' se e' del tipo subset=x>min(x) allora continuerebbe a togliere 1 osservazione if(!is.null(objF[["subset"]])) objF[["subset"]]<-NULL objF<-eval(objF, envir=mfExt) # #11/10/16 il controllo e' stato commentato in modo tale da restituire anche un oggetto lm in cui psi viene considerato fisso.. # for(jj in colnames(V)) { # VV<-V[, which(colnames(V)==jj), drop=FALSE] # sumV<-abs(rowSums(VV)) ## if( (any(diff(sumV)>=2) #se ci sono due breakpoints uguali ## || any(table(sumV)<=1)) && stop.if.error) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") ## Tolto perche' se la variabile segmented non e' ordinata non ha senso.. ##magari potresti fare un abs(diff(psi))<=.0001? ma clusterizzato.. # if(any(table(sumV)<=1) && stop.if.error) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") # } #Puo' capitare che psi sia ai margini o molto vicini (e ci sono solo 1 o 2 osservazioni in qualche intervallo. Oppure ce ne #sono di piu' ma hanno gli stessi valori di x. In questo caso objF$coef puo' avere mancanti.. names(which(is.na(coef(objF)))) objF$offset<- obj0$offset isNAcoef<-any(is.na(objF$coefficients)) if(isNAcoef){ if(stop.if.error) {stop("at least one coef is NA: breakpoint(s) at the boundary? (possibly with many x-values replicated)", call. = FALSE) } else { warning("some estimate is NA: too many breakpoints? 'var(hat.psi)' cannot be computed \n ..returning a 'lm' model", call. = FALSE) Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) objF <- update(obj0, formula = Fo, evaluate=TRUE, data = mfExt) names(psi)<-nomiVxb objF$psi<-psi return(objF) } } if(!gap){ names.coef<-names(objF$coefficients) #questi codici funzionano e si basano sull'assunzioni che le U e le V siano ordinate.. if(k==1){ names(obj$coefficients)[match(c("U","V"), names(coef(obj)))]<- nnomi } else { names(obj$coefficients)[match(c(paste("U",1:k, sep=""), paste("V",1:k, sep="")), names(coef(obj)))]<- nnomi } objF$coefficients[names.coef]<-obj$coefficients[names.coef] #sostituisce gli 0 #objF$coefficients<-obj$coefficients #names(objF$coefficients)<-names.coef objF$fitted.values<-obj$fitted.values objF$residuals<-obj$residuals } Cov <- vcov(objF) id <- match(nomiVxb, names(coef(objF))) vv <- if (length(id) == 1) Cov[id, id] else diag(Cov[id, id]) #if(length(initial)!=length(psi)) initial<-rep(NA,length(psi)) #browser() a<-tapply(id.psi.group, id.psi.group, length) #ho sovrascritto "a" di sopra, ma non dovrebbe servire.. ris.psi<-matrix(,length(psi),3) colnames(ris.psi) <- c("Initial", "Est.", "St.Err") rownames(ris.psi) <- nomiVxb ris.psi[,2]<-psi ris.psi[,3]<-sqrt(vv) #NB "a" deve essere un vettore che si appatta con "initial.psi" per ottnetere "initial" sotto... Se una variabile alla fine risulta # senza breakpoint questo non avviene e ci sono problemi nella formazione di "initial". Allora costruisco a.ok a.ok<-NULL for(j in name.Z){ if(j %in% nomiFINALI) { a.ok[length(a.ok)+1]<-a[1] a<-a[-1] } else { a.ok[length(a.ok)+1]<-0 } #ifelse(name.Z %in% nomiFINALI,1,0) } #initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi, a.ok, SIMPLIFY = TRUE)) initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi[nomiFINALI], a.ok[a.ok!=0], SIMPLIFY = TRUE)) ris.psi[,1]<-initial #psi <- cbind(initial, psi, sqrt(vv)) #rownames(psi) <- colnames(Cov)[id] objF$rangeZ <- rangeZ objF$psi.history <- psi.values objF$psi <- ris.psi objF$it <- (it - 1) objF$epsilon <- obj$epsilon objF$call <- match.call() objF$nameUV <- list(U = drop(nomiU), V = rownames(ris.psi), Z = nomiFINALI) #Z = name.Z objF$id.group <- if(length(name.Z)<=1) -rowSums(as.matrix(V)) objF$id.psi.group <- id.psi.group objF$id.warn <- id.warn objF$orig.call<-orig.call if (model) objF$model <- mf #objF$mframe <- data.frame(as.list(KK)) # PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) # SE.PSI <- matrix(rep(sqrt(vv), rep(nrow(Z), length(psi))), ncol = length(psi)) # X.is<-model.matrix(Fo, data=objF$model) # X.is[,nomiVxb]<-pnorm((Z-PSI)/SE.PSI)%*% diag(-beta.c, ncol = length(beta.c)) # objF$cov.unscaled.is<-crossprod(X.is) #browser() if(n.boot>0) objF$seed<-employed.Random.seed class(objF) <- c("segmented", class(obj0)) list.obj[[length(list.obj) + 1]] <- objF class(list.obj) <- "segmented" if (last) list.obj <- list.obj[[length(list.obj)]] return(list.obj) } segmented/R/lines.segmented.R0000644000176200001440000000172413157724622015652 0ustar liggesuserslines.segmented<-function(x, term, bottom=TRUE, shift=TRUE, conf.level=0.95, k=50, pch=18, rev.sgn=FALSE,...){ if(missing(term)){ if(length(x$nameUV$Z)>1 ) {stop("please, specify `term'")} else {term<-x$nameUV$Z} } ss<-list(...) colore<- if(is.null(ss$col)) 1 else ss$col usr <- par("usr") h<-(usr[4]-usr[3])/abs(k) y<- if(bottom) usr[3]+h else usr[4]-h r<- confint.segmented(object=x,parm=term,level=conf.level,rev.sgn=rev.sgn,digits=15) m<-r[[term]] #FORSE non e' necessaria #if(rev.sgn) m<- -m #ma invece serve il seguente (se length(psi)=1 e rev.sgn=T): m<-matrix(m,ncol=3) if(nrow(m)>1) m<-m[order(m[,1]),] est.psi<-m[,1] lower.psi<-m[,2] upper.psi<-m[,3] if(length(est.psi)>1) { y<- if(shift) y+seq(-h/2,h/2,length=length(est.psi)) else rep(y,length(est.psi)) } segments(lower.psi, y, upper.psi, y, ...) points(est.psi,y,type="p",pch=pch,col=colore) } segmented/R/seg.lm.fit.r0000644000176200001440000001337013157724622014574 0ustar liggesusersseg.lm.fit<-function(y,XREG,Z,PSI,w,offs,opz,return.all.sol=FALSE){ #aggiunge la SS.ok (che esclude i gap) #argomento return.all.sol #opz$pow (passata da seg.control) che necessita di dpmax() #step halving more straightforward (deleted H) #----------------- dpmax<-function(x,y,pow=1){ #deriv pmax if(pow==1) ifelse(x>y, -1, 0) else -pow*pmax(x-y,0)^(pow-1) } #----------- mylm<-function(x,y,w,offs=rep(0,length(y))){ x1<-x*sqrt(w) y<-y-offs y1<-y*sqrt(w) b<-drop(solve(crossprod(x1),crossprod(x1,y1))) fit<-drop(tcrossprod(x,t(b))) r<-y-fit o<-list(coefficients=b,fitted.values=fit,residuals=r) o } #----------- c1 <- apply((Z <= PSI), 2, all) c2 <- apply((Z >= PSI), 2, all) if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2))) stop("psi out of the range") # digits<-opz$digits pow<-opz$pow nomiOK<-opz$nomiOK toll<-opz$toll h<-opz$h gap<-opz$gap stop.if.error<-opz$stop.if.error dev.new<-opz$dev0 visual<-opz$visual id.psi.group<-opz$id.psi.group it.max<-old.it.max<-opz$it.max rangeZ <- apply(Z, 2, range) psi<-PSI[1,] names(psi)<-id.psi.group #H<-1 it <- 1 epsilon <- 10 dev.values<-psi.values <- NULL id.psi.ok<-rep(TRUE, length(psi)) sel.col.XREG<-unique(sapply(colnames(XREG), function(x)match(x,colnames(XREG)))) if(is.numeric(sel.col.XREG))XREG<-XREG[,sel.col.XREG,drop=FALSE] #elimina le ripetizioni, ad es. le due intercette.. while (abs(epsilon) > toll) { k<-ncol(Z) U <- pmax((Z - PSI), 0)^pow[1]#U <- pmax((Z - PSI), 0) V <- dpmax(Z,PSI,pow=pow[2])# ifelse((Z > PSI), -1, 0) X <- cbind(XREG, U, V) rownames(X) <- NULL if (ncol(V) == 1) colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c("U", "V") else colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c(paste("U", 1:ncol(U), sep = ""), paste("V", 1:ncol(V), sep = "")) obj <- lm.wfit(x = X, y = y, w = w, offset = offs) dev.old<-dev.new dev.new <- dev.new1 <-sum(obj$residuals^2) if(return.all.sol) dev.new1 <- sum(mylm(x = cbind(XREG, U), y = y, w = w, offs = offs)$residuals^2*w/sum(w)) dev.values[[length(dev.values) + 1]] <- dev.new1 if (visual) { flush.console() if (it == 1) cat(0, " ", formatC(dev.old, 3, format = "f"), "", "(No breakpoint(s))", "\n") spp <- if (it < 10) "" else NULL cat(it, spp, "", formatC(dev.new, 3, format = "f"), "",length(psi),"\n") #cat(paste("iter = ", it, spp," dev = ",formatC(dev.new,digits=3,format="f"), " n.psi = ",formatC(length(psi),digits=0,format="f"), sep=""), "\n") } epsilon <- (dev.new - dev.old)/(dev.old + .001) obj$epsilon <- epsilon it <- it + 1 obj$it <- it #class(obj) <- c("segmented", class(obj)) #list.obj[[length(list.obj) + ifelse(last == TRUE, 0, 1)]] <- obj if (k == 1) { beta.c <- coef(obj)["U"] gamma.c <- coef(obj)["V"] } else { beta.c <- coef(obj)[paste("U", 1:ncol(U), sep = "")] gamma.c <- coef(obj)[paste("V", 1:ncol(V), sep = "")] } if (it > it.max) break psi.values[[length(psi.values) + 1]] <- psi.old <- psi # if(it>=old.it.max && h<1) H<-h psi <- psi.old + h*gamma.c/beta.c if(!is.null(digits)) psi<-round(psi, digits) PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #check if psi is admissible.. a <- apply((Z <= PSI), 2, all) #prima era solo < b <- apply((Z >= PSI), 2, all) #prima era solo > if(stop.if.error) { isErr<- (sum(a + b) != 0 || is.na(sum(a + b))) if(isErr) { if(return.all.sol) return(list(dev.values, psi.values)) else stop("(Some) estimated psi out of its range") } } else { id.psi.ok<-!is.na((a+b)<=0)&(a+b)<=0 Z <- Z[,id.psi.ok,drop=FALSE] psi <- psi[id.psi.ok] PSI <- PSI[,id.psi.ok,drop=FALSE] nomiOK<-nomiOK[id.psi.ok] #salva i nomi delle U per i psi ammissibili id.psi.group<-id.psi.group[id.psi.ok] names(psi)<-id.psi.group if(ncol(PSI)<=0) return(0) } #end else #obj$psi <- psi } #end while #queste due righe aggiunte nella versione 0.2.9-3 (adesso i breakpoints sono sempre ordinati) psi<-unlist(tapply(psi, id.psi.group, sort)) names(psi)<-id.psi.group PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #aggiunto da qua.. U <- pmax((Z - PSI), 0) V <- ifelse((Z > PSI), -1, 0) X <- cbind(XREG, U, V) rownames(X) <- NULL if (ncol(V) == 1) colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c("U", "V") else colnames(X)[(ncol(XREG) + 1):ncol(X)] <- c(paste("U", 1:ncol(U), sep = ""), paste("V", 1:ncol(V), sep = "")) obj <- lm.wfit(x = X, y = y, w = w, offset = offs) obj$epsilon <- epsilon obj$it <- it obj.new <- lm.wfit(x = cbind(XREG, U), y = y, w = w, offset = offs) SS.new<-sum(obj.new$residuals^2*w/sum(w)) if(!gap){ names.coef<-names(obj$coefficients) obj$coefficients<-c(obj.new$coefficients, rep(0,ncol(V))) names(obj$coefficients)<-names.coef obj$residuals<-obj.new$residuals obj$fitted.values<-obj.new$fitted.values } #fino a qua.. obj<-list(obj=obj,it=it,psi=psi,psi.values=psi.values,U=U,V=V,rangeZ=rangeZ, epsilon=epsilon,nomiOK=nomiOK, SumSquares.no.gap=SS.new, id.psi.group=id.psi.group) #inserire id.psi.ok? return(obj) } segmented/R/print.summary.segmented.R0000644000176200001440000000636713157724622017400 0ustar liggesusers`print.summary.segmented` <- function(x, short = x$short, var.diff = x$var.diff, digits = max(3, getOption("digits") - 3), signif.stars = getOption("show.signif.stars"),...){ cat("\n\t***Regression Model with Segmented Relationship(s)***\n\n") cat( "Call: \n" ) print( x$call ) cat("\nEstimated Break-Point(s):\n ") print(round(x$psi[,-1],3)) #era "signif(,4)" # cat("\nt value for the gap-variable(s) V: ",x$gap[,3],"\n") #if(any(abs(x$gap[,3])>1.96)) cat(" Warning:", sum(abs(x$gap[,3])>1.96),"gap coefficient(s) significant at 0.05 level\n") if(short){ cat("\nDifference-in-slopes parameter(s):\n") #print(x$Ttable[(nrow(x$Ttable)-nrow(x$psi)+1):nrow(x$Ttable),])} nome<-rownames(x$psi) #nome<-as.character(parse("",text=nome)) #aa<-grep("U",rownames(x$Ttable)) #bb<-unlist(sapply(nome,function(xx){grep(xx,rownames(x$Ttable))},simplify=FALSE,USE.NAMES=FALSE)) #cc<-intersect(aa,bb) #indices of diff-slope parameters nomiU<-rownames(x$gap) #idU<-match(nomiU,rownames(x$Ttable)) print(x$Ttable[nomiU,]) } else {cat("\nMeaningful coefficients of the linear terms:\n") if(is.null(dim(x$Ttable))){ print(x$Ttable) #printCoefmat(matrix(x$Ttable,nrow=1,ncol=4,dimnames=list(" ",names(x$Ttable))),has.Pvalue=FALSE) } else { printCoefmat(x$Ttable, digits = digits, signif.stars = signif.stars,na.print = "NA", ...) } } if("summary.lm"%in%class(x)){ #for lm if(var.diff){ for(i in 1:length(x$sigma.new)){ cat("\nResidual standard error ",i,":", format(signif(x$sigma.new[i], digits)), "on", x$df.new[i], "degrees of freedom")} cat("\n") } else { cat("\nResidual standard error:", format(signif(x$sigma, digits)), "on", x$df[2], "degrees of freedom\n")} if (!is.null(x$fstatistic)) { cat("Multiple R-Squared:", formatC(x$r.squared, digits = digits)) cat(", Adjusted R-squared:", formatC(x$adj.r.squared, digits = digits), "\n")} } if("summary.glm"%in%class(x)){ #for glm cat("(Dispersion parameter for ", x$family$family, " family taken to be ", format(x$dispersion), ")\n\n", apply(cbind(paste(format.default(c("Null", "Residual"), width = 8, flag = ""), "deviance:"), format(unlist(x[c("null.deviance", "deviance")]), digits = max(5, digits + 1)), " on", format(unlist(x[c("df.null", "df.residual")])), " degrees of freedom\n"), 1, paste, collapse = " "), "AIC: ", format(x$aic, digits = max(4, digits + 1)), "\n", sep = "") } if("summary.Arima"%in%class(x)){#for Arima cm <- x$call$method if (is.null(cm) || cm != "CSS") cat("\nsigma^2 estimated as ", format(x$sigma2, digits = digits), ", log likelihood = ", format(round(x$loglik, 2)), ", aic = ", format(round(x$aic, 2)), "\n", sep = "") else cat("\nsigma^2 estimated as ", format(x$sigma2, digits = digits), ", part log likelihood = ", format(round(x$loglik, 2)), "\n", sep = "") } invisible(x) cat("\nConvergence attained in",x$it,"iterations with relative change",x$epsilon,"\n") } segmented/R/slope.R0000644000176200001440000001373413157724622013714 0ustar liggesusers`slope` <- function(ogg, parm, conf.level=0.95, rev.sgn=FALSE, var.diff=FALSE, APC=FALSE, digits = max(3, getOption("digits") - 3)){ #-- f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") #riscostruisce il nome con il "." (che era stato scomposto da strsplit()) nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #-- # if(!"segmented"%in%class(ogg)) stop("A 'segmented' model is requested") if(var.diff && length(ogg$nameUV$Z)>1) { var.diff<-FALSE warning("var.diff set to FALSE with multiple segmented variables", call.=FALSE) } #se e' un "newsegmented" # if(!is.null(ogg$R.slope)) { # covv<-old.coef.var(ogg) # ogg$coefficients<- covv$b # covv<- covv$cov # ogg$psi<-old.psi(ogg) # ogg$nameUV<-old.nomi(ogg) # } else { covv<-try(vcov(ogg,var.diff=var.diff), silent=TRUE) # } nomepsi<-rownames(ogg$psi) #OK nomeU<-ogg$nameUV$U nomeZ<-ogg$nameUV$Z if(missing(parm)) { nomeZ<- ogg$nameUV[[3]] if(length(rev.sgn)==1) rev.sgn<-rep(rev.sgn,length(nomeZ)) } else { if(! all(parm %in% ogg$nameUV$Z)) {stop("invalid parm")} else {nomeZ<-parm} } if(length(rev.sgn)!=length(nomeZ)) rev.sgn<-rep(rev.sgn, length.out=length(nomeZ)) nomi<-names(coef(ogg)) nomi<-nomi[-match(nomepsi,nomi)] #escludi i coef delle V index<-vector(mode = "list", length = length(nomeZ)) for(i in 1:length(nomeZ)) { #---> DA RIMUOVERE E SOSTITUIRE CON QUELLI DI SUBITO DOPO? # #id.cof.U<-grep(paste("\\.",nomeZ[i],"$",sep=""), nomi, value=FALSE) # #psii<-ogg$psi[grep(paste("\\.",nomeZ[i],"$",sep=""), rownames(ogg$psi), value=FALSE),2] # #id.cof.U<- match(grep(nomeZ[i], ogg$nameUV$U, value=TRUE), nomi) # #psii<-ogg$psi[grep(nomeZ[i], ogg$nameUV$V, value=TRUE),2] # #il paste con "$" (paste("\\.",nomeZ[i],"$",sep="")) e' utile perche' serve a distinguere variabili con nomi simili (ad es., "x" e "xx") # #Comunque nella versione dopo la 0.3-1.0 ho (FINALMENTE) risolto mettendo f.U # id.cof.U<- f.U(ogg$nameUV$U, nomeZ[i]) # #id.cof.U e' la posizione nel vettore ogg$nameUV$U; la seguente corregge per eventuali variabili che ci sono prima (ad es., interc) # id.cof.U<- id.cof.U + (match(ogg$nameUV$U[1], nomi)-1) # psii<- ogg$psi[f.U(ogg$nameUV$V, nomeZ[i]) , "Est."] # id.cof.U <- id.cof.U[order(psii)] #---> #questi funzionano anche con oggetti da segreg nomiPsi<-grep(paste(".", nomeZ[i], sep="") , ogg$nameUV$V, value=TRUE) psii<- ogg$psi[nomiPsi , "Est."] nomiU<-grep(paste(".", nomeZ[i], sep="") , ogg$nameUV$U, value=TRUE) #cof<-coef(ogg)[nomiU] id.cof.U<- match(nomiU, names(ogg$coefficients)) index[[i]]<-c(match(nomeZ[i],nomi), id.cof.U) } Ris<-list() digits <- max(3, getOption("digits") - 3) rev.sgn<-rep(rev.sgn, length.out=length(nomeZ)) # transf=c("x","1") # if( (length(transf)!=2) || !(length(transf)==1 && transf=="APC")) stop("'error in transf'") # if(transf=="APC") transf<-c("100*(exp(x)-1)", "100*exp(x)") # my.f<-function(x)eval(parse(text=transf[1])) # my.f.deriv<-function(x)eval(parse(text=transf[2])) for(i in 1:length(index)){ ind<-as.numeric(na.omit(unlist(index[[i]]))) M<-matrix(1,length(ind),length(ind)) M[row(M)1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #------------- if(missing(term)){ if(length(x$nameUV$Z)>1 ) {stop("please, specify `term'")} else {term<-x$nameUV$Z} } opz<-list(...) nameV<- x$nameUV$V[f.U(x$nameUV$V, term)] psii<- x$psi[nameV, "Est."] d<-data.frame(a=psii) names(d)<-term opz$y<-broken.line(x,d, se.fit=FALSE, interc=interc, link=link)[[1]] if(rev.sgn) psii<- -psii opz$x<- psii if(is.null(opz$cex)) opz$cex<-1.5 if(is.null(opz$lwd)) opz$lwd<-2 opz$y<-do.call(transf, list(opz$y)) do.call(points, opz) invisible(NULL) } segmented/R/segmented.default.r0000644000176200001440000004455313157724622016233 0ustar liggesusers#`segmented.default` <- #o1<-segmented(out.lm, seg.Z=~x +z,psi=list(x=c(30,60),z=.3), control=seg.control(display=FALSE, n.boot=20, seed=1515)) #o2<-ss(out.lm, seg.Z=~x +z,psi=list(x=c(30,60),z=.3), control=seg.control(display=FALSE, n.boot=20, seed=1515)) #o2<-ss(out.lm, seg.Z=~x +z,psi=list(x=c(30,60),z=.3), control=seg.control(display=FALSE, n.boot=0)) #o2<-ss(o, seg.Z=~age, psi=41, control=seg.control(display=FALSE, n.boot=0)) segmented.default<- function(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) { #Richiede control$f.obj that should be a string like "sum(x$residuals^2)" or "x$dev" #----------------- dpmax<-function(x,y,pow=1){ #deriv pmax if(pow==1) ifelse(x>y, -1, 0) else -pow*pmax(x-y,0)^(pow-1) } #----------- # control$fn.obj<-"sum(x$residuals^2)" # control$fn.obj<-"x$dev" # control$fn.obj<-"-x$loglik[2]" # control$fn.obj<-"x$rho" # if(is.null(control$fn.obj)) stop("'segmented.default' needs 'fn.obj' specified in seg.control") else fn.obj<-control$fn.obj if(is.null(control$fn.obj)) fn.obj<-"-as.numeric(logLik(x))" else fn.obj<-control$fn.obj #----------- n.Seg<-1 if(missing(psi)){if(length(all.vars(seg.Z))>1) stop("provide psi") else psi<-Inf} if(length(all.vars(seg.Z))>1 & !is.list(psi)) stop("`psi' should be a list with more than one covariate in `seg.Z'") if(is.list(psi)){ if(length(all.vars(seg.Z))!=length(psi)) stop("A wrong number of terms in `seg.Z' or `psi'") if(any(is.na(match(all.vars(seg.Z),names(psi), nomatch = NA)))) stop("Variables in `seg.Z' and `psi' do not match") n.Seg <- length(psi) } if(length(all.vars(seg.Z))!=n.Seg) stop("A wrong number of terms in `seg.Z' or `psi'") it.max <- old.it.max<- control$it.max digits<-control$digits toll <- control$toll visual <- control$visual stop.if.error<-control$stop.if.error n.boot<-control$n.boot # n.boot<-0 size.boot<-control$size.boot gap<-control$gap random<-control$random pow<-control$pow visualBoot<-FALSE if(n.boot>0){ if(!is.null(control$seed)) { set.seed(control$seed) employed.Random.seed<-control$seed } else { employed.Random.seed<-eval(parse(text=paste(sample(0:9, size=6), collapse=""))) set.seed(employed.Random.seed) } if(visual) {visual<-FALSE; visualBoot<-TRUE}# warning("`display' set to FALSE with bootstrap restart", call.=FALSE)} if(!stop.if.error) stop("Bootstrap restart only with a fixed number of breakpoints") } last <- control$last K<-control$K h<-min(abs(control$h),1) if(h<1) it.max<-it.max+round(it.max/2) orig.call<-Call<-mf<-obj$call orig.call$formula<- mf$formula<-formula(obj) #per consentire lm(y~.) m <- match(c("formula", "data", "subset", "weights", "na.action","offset"), names(mf), 0L) mf <- mf[c(1, m)] mf$drop.unused.levels <- TRUE mf[[1L]] <- as.name("model.frame") if(class(mf$formula)=="name" && !"~"%in%paste(mf$formula)) mf$formula<-eval(mf$formula) mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) mfExt<- mf if(!is.null(obj$call$offset) || !is.null(obj$call$weights) || !is.null(obj$call$subset) || !is.null(obj$call$id)){ mfExt$formula <- update.formula(mf$formula, paste(".~.+", paste( c(all.vars(obj$call$offset), all.vars(obj$call$weights), all.vars(obj$call$subset), all.vars(obj$call$id)), collapse = "+") )) } mf <- eval(mf, parent.frame()) n<-nrow(mf) #questo serve per inserire in mfExt le eventuali variabili contenute nella formula con offset(..) nomiOff<-setdiff(all.vars(formula(obj)), names(mf)) if(length(nomiOff)>=1) mfExt$formula<-update.formula(mfExt$formula,paste(".~.+", paste( nomiOff, collapse="+"), sep="")) nomiTUTTI<-all.vars(mfExt$formula) #comprende anche altri nomi (ad es., threshold) "variabili" nomiNO<-NULL #dovrebbe contenere for(i in nomiTUTTI){ r<-try(eval(parse(text=i), parent.frame()), silent=TRUE) if(class(r)!="try-error" && length(r)==1 && !is.function(r)) nomiNO[[length(nomiNO)+1]]<-i } if(!is.null(nomiNO)) mfExt$formula<-update.formula(mfExt$formula,paste(".~.-", paste( nomiNO, collapse="-"), sep="")) mfExt<-eval(mfExt, parent.frame()) #apply(mfExt,2,function(x) {if(is.Surv(x)) x[,1:ncol(x)] else x}) if(inherits(obj, "coxph")){ is.Surv<-NA rm(is.Surv) for(i in 1:ncol(mfExt)){ if(is.Surv(mfExt[,i])) aa<-mfExt[,i][,1:ncol(mfExt[,i])] } mfExt<-cbind(aa,mfExt) } id.seg<-match(all.vars(seg.Z), names(mfExt)) name.Z<-names(mfExt)[id.seg] Z<-mfExt[,id.seg,drop=FALSE] # name.Z <- names(Z) n.psi<-length(unlist(psi)) ################# if(ncol(Z)==1 && length(psi)==1 && n.psi==1 && !any(is.na(psi))) { if(psi==Inf) psi<-median(Z[,1])} #devi selezionare la colonna perche' Z e' un dataframe! ################# if(ncol(Z)==1 && is.vector(psi) && (is.numeric(psi)||is.na(psi))){ psi <- list(as.numeric(psi)) names(psi)<-name.Z } if (!is.list(psi) || is.null(names(psi))) stop("psi should be a *named* list") id.nomiZpsi <- match(colnames(Z), names(psi)) if ((ncol(Z)!=length(psi)) || any(is.na(id.nomiZpsi))) stop("Length or names of Z and psi do not match") nome <- names(psi)[id.nomiZpsi] psi <- psi[nome] initial.psi<-psi for(i in 1:length(psi)) { if(any(is.na(psi[[i]]))) psi[[i]]<-if(control$quant) {quantile(Z[,i], prob= seq(0,1,l=K+2)[-c(1,K+2)], names=FALSE)} else {(min(Z[,i])+ diff(range(Z[,i]))*(1:K)/(K+1))} } a <- sapply(psi, length) #per evitare che durante il processo iterativo i psi non siano ordinati id.psi.group <- rep(1:length(a), times = a) #identificativo di apparteneza alla variabile Z<-matrix(unlist(mapply(function(x,y)rep(x,y),Z,a,SIMPLIFY = TRUE)),nrow=n) colnames(Z) <- nomiZ.vett <- rep(nome, times = a) #SERVE??? si perche' Z e' senza colnames psi <- unlist(psi) #se psi e' numerico, la seguente linea restituisce i valori ordinati all'interno della variabile.. psi<-unlist(tapply(psi,id.psi.group,sort)) k <- ncol(Z) PSI <- matrix(rep(psi, rep(n, k)), ncol = k) #controllo se psi e' ammissibile.. c1 <- apply((Z <= PSI), 2, all) #dovrebbero essere tutti FALSE (prima era solo <) c2 <- apply((Z >= PSI), 2, all) #dovrebbero essere tutti FALSE (prima era solo >) if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2)) ) stop("starting psi out of the admissible range") #ripetizioni <- as.numeric(unlist(sapply(table(nomiZ)[order(unique(nomiZ))], function(xxx) {1:xxx}))) ripetizioni <- as.vector(unlist(tapply(id.psi.group, id.psi.group, function(x) 1:length(x) ))) nomiU <- paste("U", ripetizioni, sep = "") nomiU <- paste(nomiU, nomiZ.vett, sep = ".") nomiV <- paste("V", ripetizioni, sep = "") nomiV <- paste(nomiV, nomiZ.vett, sep = ".") nnomi <- c(nomiU, nomiV) #CONTROLLARE se e' necessario aggiungere a mfExt le nuove variabili U e V. Non dovrebbe perche' # update(, evaluate=FALSE) non "le vuole" e poi comunque vengono calcolate in seg.def.fit() # Servono solo se it.max=0.. #---------------------------------------------------------- U <- pmax((Z - PSI), 0)^pow[1]#U <- pmax((Z - PSI), 0) #V <- dpmax(Z,PSI,pow=pow[2])# ifelse((Z > PSI), -1, 0) V<-ifelse((Z > PSI), -1, 0) for(i in 1:k) { mfExt[nomiU[i]] <- U[,i] mfExt[nomiV[i]] <- V[,i] } Fo <- update.formula(formula(obj), as.formula(paste(".~.+", paste(nnomi, collapse = "+")))) Fo.noV <- update.formula(formula(obj), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) #ho tolto "formula = Fo" call.ok <- update(obj, Fo, evaluate=FALSE, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) call.noV <- update(obj, Fo.noV, evaluate=FALSE, data = mfExt) #objF <- update(obj0, formula = Fo, data = KK) if (it.max == 0) { if(!is.null(call.noV[["subset"]])) call.noV[["subset"]]<-NULL obj1 <- eval(call.noV, envir=mfExt) return(obj1) } #obj1 <- eval(call.ok, envir=mfExt) initial <- psi obj0 <- obj #browser() dev0<- eval(parse(text=fn.obj), list(x=obj)) if(length(dev0)<=0) stop("error in the objective to be minimized, see 'fn.obj' in ?seg.control") #per fit su cui logLik() does not work.. if(length(dev0)>1) stop("the objective to be minimized is not scalar, see 'fn.obj' in ?seg.control") #per fit su cui logLik() does not work.. if(is.na(dev0)) dev0<-10 list.obj <- list(obj) nomiOK<-nomiU opz<-list(toll=toll,h=h,stop.if.error=stop.if.error,dev0=dev0,visual=visual,it.max=it.max, nomiOK=nomiOK, id.psi.group=id.psi.group, gap=gap, visualBoot=visualBoot, pow=pow, digits=digits) opz$call.ok<-call.ok opz$call.noV<-call.noV opz$formula.orig<-formula(obj) opz$nomiU<-nomiU opz$nomiV<-nomiV opz$fn.obj <- fn.obj opz<-c(opz,...) #8/10/16... #browser() if(n.boot<=0){ obj<-seg.def.fit(obj, Z, PSI, mfExt, opz) } else { obj<-seg.def.fit.boot(obj, Z, PSI, mfExt, opz, n.boot=n.boot, size.boot=size.boot, random=random) #jt, nonParam } #browser() if(!is.list(obj)){ warning("No breakpoint estimated", call. = FALSE) return(obj0) } if(!is.null(obj$obj$df.residual) && !is.na(obj$obj$df.residual)){ if(obj$obj$df.residual==0) warning("no residual degrees of freedom (other warnings expected)", call.=FALSE) } id.psi.group<-obj$id.psi.group nomiOK<-obj$nomiOK #sarebbe nomiU nomiVxb<-paste("psi",sapply(strsplit(nomiOK,"U"), function(x){x[2]}), sep="") #nomiFINALI<-unique(sapply(strsplit(nomiOK, split="[.]"), function(x)x[2])) #nomi delle variabili con breakpoint stimati! nomiFINALI<-unique(sub("U[1-9]*[0-9].", "", nomiOK)) #se e' stata usata una proc automatica "nomiFINALI" sara differente da "name.Z" nomiSenzaPSI<-setdiff(name.Z,nomiFINALI) if(length(nomiSenzaPSI)>=1) warning("no breakpoints found for: ", paste(nomiSenzaPSI," "), call. = FALSE) it<-obj$it psi<-obj$psi psi.values<-if(n.boot<=0) obj$psi.values else obj$boot.restart U<-obj$U V<-obj$V # return(obj) #browser() #if(any(table(rowSums(V))<=1)) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close") for(jj in colnames(V)) { VV<-V[, which(colnames(V)==jj), drop=FALSE] sumV<-abs(rowSums(VV)) # if( (any(diff(sumV)>=2) #se ci sono due breakpoints uguali # || any(table(sumV)<=1)) && stop.if.error) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") if(any(table(sumV)<=1) && stop.if.error) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") } rangeZ<-obj$rangeZ mfExt<-obj$mfExt names(mfExt)[match(obj$nomiV, names(mfExt))]<-nomiVxb R<-obj$R R.noV<-obj$R.noV r<-obj$r obj<-obj$obj k<-length(psi) # beta.c<-if(k == 1) coef(obj)["U"] else coef(obj)[paste("U", 1:ncol(U), sep = "")] beta.c<- coef(obj)[nomiOK] #nomiOK e' stato estratto da obj e contiene tutti i nomi delle variabili U inserite nel modello psi.values[[length(psi.values) + 1]] <- psi id.warn <- FALSE if (n.boot<=0 && it > it.max) { #it >= (it.max+1) warning("max number of iterations attained", call. = FALSE) id.warn <- TRUE } Vxb <- V %*% diag(beta.c, ncol = length(beta.c)) # #se usi una procedura automatica devi cambiare ripetizioni, nomiU e nomiV, e quindi: # length.psi<-tapply(as.numeric(as.character(names(psi))), as.numeric(as.character(names(psi))), length) # forma.nomiU<-function(xx,yy)paste("U",1:xx, ".", yy, sep="") # forma.nomiVxb<-function(xx,yy)paste("psi",1:xx, ".", yy, sep="") # nomiU <- unlist(mapply(forma.nomiU, length.psi, nomiFINALI)) #invece di un ciclo #paste("U",1:length.psi[i], ".", name.Z[i]) # nomiVxb <- unlist(mapply(forma.nomiVxb, length.psi, nomiFINALI)) # for(i in 1:ncol(U)) { # mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] # mfExt[nomiVxb[i]]<-mf[nomiVxb[i]]<-Vxb[,i] # } # ToDeletenomiU<-nomiU[!id.psi.ok] #salva i nomi delle U per i psi ammissibili # ToDeletenomiV<-nomiV[!id.psi.ok] #salva i nomi delle V per i psi ammissibili # if(length(ToDeletenomiU)>0 || length(ToDeletenomiV)>0) for(nn in c(ToDeletenomiU, ToDeletenomiV)) mfExt[[nn]]<-NULL #E' inutile lavorare sul mf, utilizzo quello restituito da seg.def.fit.r # for(i in 1:k) { # mfExt[nomiOK[i]]<-mf[nomiOK[i]]<-U[,i] # mfExt[nomiVxb[i]]<-mf[nomiVxb[i]]<-Vxb[,i] # } #ottobre 2016: ci vuole altrimenti Vxb non viene inserita nel dataframe for(i in 1:ncol(U)) { mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] mfExt[nomiVxb[i]]<-mf[nomiVxb[i]]<-Vxb[,i] } nnomi <- c(nomiOK, nomiVxb) Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nnomi, collapse = "+")))) objF <- update(obj0, Fo, evaluate=FALSE, data = mfExt) #tolto "formula =Fo" if(!is.null(objF[["subset"]])) objF[["subset"]]<-NULL if(is.null(opz$constr)) opz$constr<-0 if((opz$constr %in% 1:2) && class(obj0)=="rq"){ objF$method<-"fnc" objF$R<-quote(R) objF$r<-quote(r) } objF<- eval(objF, envir=mfExt) #Puo' capitare che psi sia ai margini e ci sono 1 o 2 osservazioni in qualche intervallo. Oppure ce ne #sono di piu' ma hanno gli stessi valori di x #objF$coef puo' avere mancanti.. names(which(is.na(coef(objF)))) objF$offset<- obj0$offset isNAcoef<-any(is.na(objF$coefficients)) if(isNAcoef){ if(stop.if.error) {stop("at least one coef is NA: breakpoint(s) at the boundary? (possibly with many x-values replicated)", call. = FALSE)} else { warning("some estimate is NA: too many breakpoints? 'var(hat.psi)' cannot be computed \n ..returning a 'lm' model", call. = FALSE) Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) objF <- if((opz$constr %in% 1:2) && class(obj0)=="rq") {update(obj0, formula = Fo, R=R.noV, r=r, method="fnc", evaluate=TRUE, data = mfExt) } else { update(obj0, Fo, evaluate=TRUE, data = mfExt) #ho tolto "formula = Fo" per farlo funzionare con gls() } names(psi)<-nomiVxb objF$psi<-psi return(objF) } } # CONTROLLARE!!!! # objF$offset<- obj0$offset #sostituire i valori: objF include le U e V, obj solo le U if(!gap){ #names.coef <- names(objF$coefficients) #names(obj$coefficients)[match(nomiV, names(coef(obj)))]<-nomiVxb #objF$coefficients[names.coef]<-obj$coefficients[names.coef] names.coef <- names(obj$coefficients) objF$coefficients[names.coef]<-obj$coefficients[names.coef] objF$coefficients[nomiVxb]<-rep(0, k) if(!is.null(objF$geese$beta)) objF$geese$beta <- objF$coefficients#obj$geese$beta if(!is.null(objF$geese$gamma)) objF$geese$gamma <- obj$geese$gamma if(!is.null(objF$geese$alpha)) objF$geese$alpha <- obj$geese$alpha if(!is.null(objF$fitted.values)) objF$fitted.values<-obj$fitted.values if(!is.null(objF$residuals)) objF$residuals<-obj$residuals if(!is.null(objF$linear.predictors)) objF$linear.predictors<-obj$linear.predictors if(!is.null(objF$deviance)) objF$deviance<-obj$deviance if(!is.null(objF$weights)) objF$weights<-obj$weights if(!is.null(objF$aic)) objF$aic<-obj$aic + 2*k if(!is.null(objF$loglik)) objF$loglik<-obj$loglik #per coxph if(!is.null(objF$rho)) objF$rho<-obj$rho #per rq if(!is.null(objF$dual)) objF$dual<-obj$dual #per rq } Cov <- try(vcov(objF), silent=TRUE) idd <- match(nomiVxb, names(coef(objF))) if(class(Cov)!="try-error") { vv <- if (length(idd) == 1) Cov[idd, idd] else diag(Cov[idd, idd]) #if(length(initial)!=length(psi)) initial<-rep(NA,length(psi)) # psi <- cbind(initial, psi, sqrt(vv)) # rownames(psi) <- colnames(Cov)[idd] # colnames(psi) <- c("Initial", "Est.", "St.Err") } else { # psi <- cbind(initial, psi) # rownames(psi) <- nomiVxb # colnames(psi) <- c("Initial", "Est.") vv<-NA } #browser() a<-tapply(id.psi.group, id.psi.group, length) #ho sovrascritto "a" di sopra, ma non dovrebbe servire.. ris.psi<-matrix(,length(psi),3) colnames(ris.psi) <- c("Initial", "Est.", "St.Err") rownames(ris.psi) <- nomiVxb ris.psi[,2]<-psi ris.psi[,3]<-sqrt(vv) #NB "a" deve essere un vettore che si appatta con "initial.psi" per ottnetere "initial" sotto... Se una variabile alla fine risulta # senza breakpoint questo non avviene e ci sono problemi nella formazione di "initial". Allora costruisco a.ok a.ok<-NULL for(j in name.Z){ if(j %in% nomiFINALI) { a.ok[length(a.ok)+1]<-a[1] a<-a[-1] } else { a.ok[length(a.ok)+1]<-0 } #ifelse(name.Z %in% nomiFINALI,1,0) } # initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi, a.ok, SIMPLIFY = TRUE)) initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi[nomiFINALI], a.ok[a.ok!=0], SIMPLIFY = TRUE)) ris.psi[,1]<-initial objF$rangeZ <- rangeZ objF$psi.history <- psi.values objF$psi <- ris.psi objF$it <- (it - 1) objF$epsilon <- obj$epsilon objF$call <- match.call() objF$nameUV <- list(U = drop(nomiU), V = rownames(ris.psi), Z = nomiFINALI) #Z = name.Z objF$id.group <- if(length(name.Z)<=1) -rowSums(as.matrix(V)) objF$id.psi.group <- id.psi.group objF$id.warn <- id.warn objF$orig.call<-orig.call if (model) objF$model <- mf #objF$mframe <- data.frame(as.list(KK)) if(n.boot>0) objF$seed<-employed.Random.seed # class(objF) <- c("segmented", class(obj0)) list.obj[[length(list.obj) + 1]] <- objF class(list.obj) <- "segmented" if (last) list.obj <- list.obj[[length(list.obj)]] return(list.obj) } #end function segmented/R/segmented.glm.R0000644000176200001440000004660613157724622015327 0ustar liggesusers`segmented.glm` <- function(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) { n.Seg<-1 if(missing(seg.Z) && length(all.vars(formula(obj)))==2) seg.Z<- as.formula(paste("~", all.vars(formula(obj))[2])) if(missing(psi)){if(length(all.vars(seg.Z))>1) stop("provide psi") else psi<-Inf} if(length(all.vars(seg.Z))>1 & !is.list(psi)) stop("`psi' should be a list with more than one covariate in `seg.Z'") if(is.list(psi)){ if(length(all.vars(seg.Z))!=length(psi)) stop("A wrong number of terms in `seg.Z' or `psi'") if(any(is.na(match(all.vars(seg.Z),names(psi), nomatch = NA)))) stop("Variables in `seg.Z' and `psi' do not match") n.Seg <- length(psi) } if(length(all.vars(seg.Z))!=n.Seg) stop("A wrong number of terms in `seg.Z' or `psi'") maxit.glm <- control$maxit.glm it.max <- old.it.max<- control$it.max digits<-control$digits toll <- control$toll visual <- control$visual stop.if.error<-control$stop.if.error n.boot<-control$n.boot size.boot<-control$size.boot gap<-control$gap random<-control$random pow<-control$pow visualBoot<-FALSE if(n.boot>0){ if(!is.null(control$seed)) { set.seed(control$seed) employed.Random.seed<-control$seed } else { employed.Random.seed<-eval(parse(text=paste(sample(0:9, size=6), collapse=""))) set.seed(employed.Random.seed) } if(visual) {visual<-FALSE; visualBoot<-TRUE}#warning("`display' set to FALSE with bootstrap restart", call.=FALSE)} if(!stop.if.error) stop("Bootstrap restart only with a fixed number of breakpoints") } last <- control$last K<-control$K h<-min(abs(control$h),1) if(h<1) it.max<-it.max+round(it.max/2) # if(!stop.if.error) objInitial<-obj #------------------------------- # #una migliore soluzione......... # objframe <- update(obj, model = TRUE, x = TRUE, y = TRUE) # y <- objframe$y # a <- model.matrix(seg.Z, data = eval(obj$call$data)) # a <- subset(a, select = colnames(a)[-1]) orig.call<-Call<-mf<-obj$call orig.call$formula<-mf$formula<-formula(obj) #per consentire lm(y~.) m <- match(c("formula", "data", "subset", "weights", "na.action","etastart","mustart","offset"), names(mf), 0L) mf <- mf[c(1, m)] mf$drop.unused.levels <- TRUE mf[[1L]] <- as.name("model.frame") #non so a che serva la seguente linea.. if(class(mf$formula)=="name" && !"~"%in%paste(mf$formula)) mf$formula<-eval(mf$formula) #orig.call$formula<-update.formula(orig.call$formula, paste("~.-",all.vars(seg.Z))) #utile per plotting # nomeRispo<-strsplit(paste(formula(obj))[2],"/")[[1]] #eventuali doppi nomi separati da "/" (tipo "y/n" per GLM binom) #la linea sotto aggiunge nel mf anche la variabile offs.. # if(length(all.vars(formula(obj)))>1){ # id.rispo<-1 # if(length(nomeRispo)>=2) id.rispo<-1:2 # #questo serve quando formula(obj) ha solo l'intercept # agg<-if(length(all.vars(formula(obj))[-id.rispo])==0) "" else "+" # mf$formula<-update.formula(mf$formula,paste(paste(seg.Z,collapse=".+"),agg,paste(all.vars(formula(obj))[-id.rispo],collapse="+"))) # } else { # mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) # } mfExt<- mf mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) # mfExt$formula<- update.formula(mfExt$formula,paste(paste(seg.Z,collapse=".+"),"+",paste(all.vars(formula(obj)),collapse="+"))) # mfExt$formula<- if(!is.null(obj$call$data)) # update.formula(mf$formula,paste(".~",paste(all.vars(obj$call), collapse="+"),"-",obj$call$data,sep="")) # else update.formula(mf$formula,paste(".~",paste(all.vars(obj$call), collapse="+"),sep="")) #----------- if(!is.null(obj$call$offset) || !is.null(obj$call$weights) || !is.null(obj$call$subset)){ mfExt$formula <- update.formula(mf$formula, paste(".~.+", paste( c(all.vars(obj$call$offset), all.vars(obj$call$weights), all.vars(obj$call$subset)), collapse = "+") )) } mf <- eval(mf, parent.frame()) n<-nrow(mf) #La linea sotto serve per inserire in mfExt le eventuali variabili contenute nella formula con offset(..) # o anche variabili che rientrano in espressioni (ad es., y/n o I(y*n)) nomiOff<-setdiff(all.vars(formula(obj)), names(mf)) if(length(nomiOff)>=1) mfExt$formula<-update.formula(mfExt$formula,paste(".~.+", paste( nomiOff, collapse="+"), sep="")) #ago 2014 c'e' la questione di variabili aggiuntive... nomiTUTTI<-all.vars(mfExt$formula) #comprende anche altri nomi (ad es., threshold) "variabili" nomiNO<-NULL #dovrebbe contenere for(i in nomiTUTTI){ r<-try(eval(parse(text=i), parent.frame()), silent=TRUE) if(class(r)!="try-error" && length(r)==1 && !is.function(r)) nomiNO[[length(nomiNO)+1]]<-i } #nomiNO dovrebbe contenere i nomi delle "altre variabili" (come th in subset=x=2) mf[nomeRispo[1]]<-weights*y id.duplic<-match(all.vars(formula(obj)),all.vars(seg.Z),nomatch=0)>0 if(any(id.duplic)) { #new.mf<-mf[,id.duplic,drop=FALSE] new.mf<-mf[,all.vars(formula(obj))[id.duplic],drop=FALSE] new.XREGseg<-data.matrix(new.mf) XREG<-cbind(XREG,new.XREGseg) } n.psi<- length(unlist(psi)) id.n.Seg<-(ncol(XREG)-n.Seg+1):ncol(XREG) XREGseg<-XREG[,id.n.Seg,drop=FALSE] #XREG<-XREG[,-id.n.Seg,drop=FALSE] #XREG<-model.matrix(obj0) non va bene perche' non elimina gli eventuali mancanti in seg.Z.. #Due soluzioni #XREG<-XREG[,colnames(model.matrix(obj)),drop=FALSE] #XREG<-XREG[,match(c("(Intercept)",all.vars(formula(obj))[-1]),colnames(XREG),nomatch =0),drop=FALSE] XREG <- XREG[, match(c("(Intercept)", namesXREG0),colnames(XREG), nomatch = 0), drop = FALSE] XREG<-XREG[,unique(colnames(XREG)), drop=FALSE] ################# if(ncol(XREGseg)==1 && length(psi)==1 && n.psi==1 && !any(is.na(psi))) { if(psi==Inf) psi<-median(XREGseg)} ################# n <- nrow(XREG) #Z <- list(); for (i in colnames(XREGseg)) Z[[length(Z) + 1]] <- XREGseg[, i] Z<-lapply(apply(XREGseg,2,list),unlist) #prende anche i nomi! name.Z <- names(Z) <- colnames(XREGseg) if(length(Z)==1 && is.vector(psi) && (is.numeric(psi)||is.na(psi))){ psi <- list(as.numeric(psi)) names(psi)<-name.Z } if (!is.list(Z) || !is.list(psi) || is.null(names(Z)) || is.null(names(psi))) stop("Z and psi have to be *named* list") id.nomiZpsi <- match(names(Z), names(psi)) if ((length(Z)!=length(psi)) || any(is.na(id.nomiZpsi))) stop("Length or names of Z and psi do not match") #dd <- match(names(Z), names(psi)) nome <- names(psi)[id.nomiZpsi] psi <- psi[nome] initial.psi<-psi for(i in 1:length(psi)) { if(any(is.na(psi[[i]]))) psi[[i]]<-quantile(Z[[i]], prob= seq(0,1,l=K+2)[-c(1,K+2)], names=FALSE) } a <- sapply(psi, length)#b <- rep(1:length(a), times = a) id.psi.group <- rep(1:length(a), times = a) #identificativo di apparteneza alla variabile #Znew <- list() #for (i in 1:length(psi)) Znew[[length(Znew) + 1]] <- rep(Z[i], a[i]) #Z <- matrix(unlist(Znew), nrow = n) Z<-matrix(unlist(mapply(function(x,y)rep(x,y),Z,a,SIMPLIFY = TRUE)),nrow=n) psi <- unlist(psi) psi<-unlist(tapply(psi,id.psi.group,sort)) k <- ncol(Z) PSI <- matrix(rep(psi, rep(n, k)), ncol = k) colnames(Z) <- nomiZ <- rep(nome, times = a) ripetizioni <- as.numeric(unlist(sapply(table(nomiZ)[order(unique(nomiZ))], function(xxx) {1:xxx}))) nomiU <- paste("U", ripetizioni, sep = "") nomiU <- paste(nomiU, nomiZ, sep = ".") nomiV <- paste("V", ripetizioni, sep = "") nomiV <- paste(nomiV, nomiZ, sep = ".") #forse non serve crearsi l'ambiente KK, usa mf.. #obj <- update(obj, formula = Fo, data = mf) #if (model.frame) obj$model <- mf #controlla che model.frame() funzioni sull'oggetto restituito # KK <- new.env() # for (i in 1:ncol(objframe$model)) assign(names(objframe$model[i]), objframe$model[[i]], envir = KK) if (it.max == 0) { #mf<-cbind(mf, mfExt) U <- pmax((Z - PSI), 0) colnames(U) <- paste(ripetizioni, nomiZ, sep = ".") nomiU <- paste("U", colnames(U), sep = "") #for (i in 1:ncol(U)) assign(nomiU[i], U[, i], envir = KK) #e' necessario il for? puoi usare colnames(U)<-nomiU;mf[nomiU]<-U for(i in 1:ncol(U)) mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] Fo <- update.formula(formula(obj), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) #obj <- update(obj, formula = Fo, data = KK) obj <- update(obj, formula = Fo, data = mfExt, evaluate=FALSE) if(!is.null(obj[["subset"]])) obj[["subset"]]<-NULL obj<-eval(obj, envir=mfExt) if (model) obj$model <-mf #obj$model <- data.frame(as.list(KK)) names(psi)<-paste(paste("psi", ripetizioni, sep = ""), nomiZ, sep=".") obj$psi <- psi return(obj) } #XREG <- model.matrix(obj) creata sopra #o <- model.offset(objframe) #w <- model.weights(objframe) if (is.null(weights)) weights <- rep(1, n) if (is.null(offs)) offs <- rep(0, n) fam <- family(obj) initial <- psi obj0 <- obj dev0<-obj$dev list.obj <- list(obj) # psi.values <- NULL nomiOK<-nomiU opz<-list(toll=toll,h=h,stop.if.error=stop.if.error,dev0=dev0,visual=visual,it.max=it.max,nomiOK=nomiOK, fam=fam, eta0=obj$linear.predictors, maxit.glm=maxit.glm, id.psi.group=id.psi.group, gap=gap, pow=pow, visualBoot=visualBoot, digits=digits) if(n.boot<=0){ obj<-seg.glm.fit(y,XREG,Z,PSI,weights,offs,opz) } else { obj<-seg.glm.fit.boot(y, XREG, Z, PSI, weights, offs, opz, n.boot=n.boot, size.boot=size.boot, random=random) #jt, nonParam } if(!is.list(obj)){ warning("No breakpoint estimated", call. = FALSE) return(obj0) } id.psi.group<-obj$id.psi.group nomiOK<-obj$nomiOK #nomiFINALI<-unique(sapply(strsplit(nomiOK, split="[.]"), function(x)x[2])) #nomi delle variabili con breakpoint stimati! nomiFINALI<-unique(sub("U[1-9]*[0-9].", "", nomiOK)) #nomi originali delle variabili con breakpoint stimati! #se e' stata usata una proc automatica "nomiFINALI" sara' differente da "name.Z" nomiSenzaPSI<-setdiff(name.Z,nomiFINALI) if(length(nomiSenzaPSI)>=1) warning("no breakpoints found for: ", paste(nomiSenzaPSI," "), call. = FALSE) it<-obj$it psi<-obj$psi k<-length(psi) psi.values<-if(n.boot<=0) obj$psi.values else obj$boot.restart U<-obj$U V<-obj$V #if(any(table(rowSums(V))<=1)) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close") for(jj in colnames(V)) { VV<-V[, which(colnames(V)==jj),drop=FALSE] sumV<-abs(rowSums(VV)) # if( (any(diff(sumV)>=2) #se ci sono due breakpoints equivalenti # || any(table(sumV)<=1))) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") if(any(table(sumV)<=1) && stop.if.error) stop("only 1 datum in an interval: breakpoint(s) at the boundary or too close each other") } rangeZ<-obj$rangeZ obj<-obj$obj beta.c<-if(k == 1) coef(obj)["U"] else coef(obj)[paste("U", 1:ncol(U), sep = "")] psi.values[[length(psi.values) + 1]] <- psi id.warn <- FALSE if (n.boot<=0 && it > it.max) { #it >= (it.max+1) warning("max number of iterations attained", call. = FALSE) id.warn <- TRUE } Vxb <- V %*% diag(beta.c, ncol = length(beta.c)) #se usi una procedura automatica devi cambiare ripetizioni, nomiU e nomiV, e quindi: length.psi<-tapply(as.numeric(as.character(names(psi))), as.numeric(as.character(names(psi))), length) forma.nomiU<-function(xx,yy)paste("U",1:xx, ".", yy, sep="") forma.nomiVxb<-function(xx,yy)paste("psi",1:xx, ".", yy, sep="") nomiU <- unlist(mapply(forma.nomiU, length.psi, nomiFINALI)) #invece di un ciclo #paste("U",1:length.psi[i], ".", name.Z[i]) nomiVxb <- unlist(mapply(forma.nomiVxb, length.psi, nomiFINALI)) #mf<-cbind(mf, mfExt) for(i in 1:ncol(U)) { mfExt[nomiU[i]]<-mf[nomiU[i]]<-U[,i] mfExt[nomiVxb[i]]<-mf[nomiVxb[i]]<-Vxb[,i] } # for (i in 1:ncol(U)) { # assign(nomiU[i], U[, i], envir = KK) # assign(nomiVxb[i], Vxb[, i], envir = KK) # } nnomi <- c(nomiU, nomiVxb) Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nnomi, collapse = "+")))) #la seguente linea si potrebbe rimuovere perche' in mfExt c'e' gia' tutto.. if(is.matrix(y)&& (fam$family=="binomial" || fam$family=="quasibinomial")){ mfExt<-cbind(mfExt[[1]], mfExt[,-1]) } objF <- update(obj0, formula = Fo, data = mfExt, evaluate=FALSE) if(!is.null(objF[["subset"]])) objF[["subset"]]<-NULL objF<-eval(objF, envir=mfExt) #C'e' un problema..controlla obj (ha due "(Intercepts)" - bhu.. al 27/03/14 non mi sembra! #Puo' capitare che psi sia ai margini e ci sono 1 o 2 osservazioni in qualche intervallo. Oppure ce ne # sono di piu' ma hanno gli stessi valori di x objF$offset<- obj0$offset isNAcoef<-any(is.na(objF$coefficients)) if(isNAcoef){ if(stop.if.error) {stop("at least one coef is NA: breakpoint(s) at the boundary? (possibly with many x-values replicated)", call. = FALSE)} else { warning("some estimate is NA: too many breakpoints? 'var(hat.psi)' cannot be computed \n ..returning a 'lm' model", call. = FALSE) Fo <- update.formula(formula(obj0), as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) objF <- update(obj0, formula = Fo, evaluate=TRUE, data = mfExt) names(psi)<-nomiVxb objF$psi<-psi return(objF) } } #aggiornare qui i weights???? (piuttosto che sotto) #------>>> #------>>> #------>>> if(!gap){ names.coef<-names(objF$coefficients) if(k==1) {names(obj$coefficients)[match(c("U","V"), names(coef(obj)))]<- nnomi } else { names(obj$coefficients)[match(c(paste("U",1:k, sep=""), paste("V",1:k, sep="")), names(coef(obj)))]<- nnomi } objF$coefficients[names.coef]<-obj$coefficients[names.coef] # objF$coefficients<- if(sum("(Intercept)"==names(obj$coef))==2) obj$coefficients[-2] else obj$coefficients objF$fitted.values<-obj$fitted.values objF$linear.predictors<-obj$linear.predictors objF$residuals<-obj$residuals objF$deviance<-obj$deviance objF$aic<-obj$aic + 2*k objF$weights<-obj$weights } Cov <- vcov(objF) id <- match(nomiVxb, names(coef(objF))) #cat(id,"\n") #return(objF) vv <- if (length(id) == 1) Cov[id, id] else diag(Cov[id, id]) #if(length(initial)!=length(psi)) initial<-rep(NA,length(psi)) a<-tapply(id.psi.group, id.psi.group, length) #ho sovrascritto "a" di sopra, ma non dovrebbe servire.. ris.psi<-matrix(,length(psi),3) colnames(ris.psi) <- c("Initial", "Est.", "St.Err") rownames(ris.psi) <- nomiVxb ris.psi[,2]<-psi ris.psi[,3]<-sqrt(vv) #NB "a" deve essere un vettore che si appatta con "initial.psi" per ottnetere "initial" sotto... Se una variabile alla fine risulta # senza breakpoint questo non avviene e ci sono problemi nella formazione di "initial". Allora costruisco a.ok a.ok<-NULL for(j in name.Z){ if(j %in% nomiFINALI) { a.ok[length(a.ok)+1]<-a[1] a<-a[-1] } else { a.ok[length(a.ok)+1]<-0 } #ifelse(name.Z %in% nomiFINALI,1,0) } # initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi, a.ok, SIMPLIFY = TRUE)) initial<-unlist(mapply(function(x,y){if(is.na(x)[1])rep(x,y) else x }, initial.psi[nomiFINALI], a.ok[a.ok!=0], SIMPLIFY = TRUE)) ris.psi[,1]<-initial objF$rangeZ <- rangeZ objF$psi.history <- psi.values objF$psi <- ris.psi objF$it <- (it - 1) objF$epsilon <- obj$epsilon objF$call <- match.call() objF$nameUV <- list(U = drop(nomiU), V = rownames(ris.psi), Z = nomiFINALI) #Z = name.Z objF$id.group <- if(length(name.Z)<=1) -rowSums(as.matrix(V)) objF$id.psi.group <- id.psi.group objF$id.warn <- id.warn objF$orig.call<-orig.call if (model) objF$model <- mf #objF$mframe <- data.frame(as.list(KK)) if(n.boot>0) objF$seed<-employed.Random.seed class(objF) <- c("segmented", class(obj0)) list.obj[[length(list.obj) + 1]] <- objF class(list.obj) <- "segmented" if (last) list.obj <- list.obj[[length(list.obj)]] return(list.obj) } segmented/R/seg.glm.fit.boot.r0000644000176200001440000001263713157724622015712 0ustar liggesusersseg.glm.fit.boot<-function(y, XREG, Z, PSI, w, offs, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE){ #random: if TRUE, when the algorithm fails in minimizing f(y), random numbers are used as final estimates. # If the algorithm fails in minimizing f(y*), the final estimates (to be used as starting values with # the original responses y) *always* are replaced by random numbers (regardless of the random argument) #nonParm. se TRUE implemneta il case resampling. Quello semiparam dipende dal non-errore del primo tentativo #show.history() se c'e' stato boot restart potrebbe produrre un grafico 2x1 di "dev vs it" and "no.of distinct vs it" #-------- extract.psi<-function(lista){ #serve per estrarre il miglior psi.. dev.values<-lista[[1]] psi.values<-lista[[2]] dev.ok<-min(dev.values) id.dev.ok<-which.min(dev.values) if(is.list(psi.values)) psi.values<-matrix(unlist(psi.values), nrow=length(dev.values), byrow=TRUE) if(!is.matrix(psi.values)) psi.values<-matrix(psi.values) psi.ok<-psi.values[id.dev.ok,] r<-list(dev.no.gap=dev.ok, psi=psi.ok) r } #------------- if(!nonParam){ nonParam<-TRUE warning("`nonParam' set to TRUE for segmented glm..", call.=FALSE) } visualBoot<-opz$visualBoot opz.boot<-opz opz.boot$pow=c(1.1,1.2) opz1<-opz opz1$it.max <-1 n<-length(y) o0<-try(seg.glm.fit(y, XREG, Z, PSI, w, offs, opz), silent=TRUE) rangeZ <- apply(Z, 2, range) #serve sempre if(!is.list(o0)) { o0<- seg.glm.fit(y, XREG, Z, PSI, w, offs, opz, return.all.sol=TRUE) o0<-extract.psi(o0) if(!nonParam) {warning("using nonparametric boot");nonParam<-TRUE} } if(is.list(o0)){ est.psi00<-est.psi0<-o0$psi ss00<-o0$dev.no.gap if(!nonParam) fitted.ok<-fitted(o0) } else { if(!nonParam) stop("semiparametric boot requires reasonable fitted values. try a different psi or use nonparam boot") if(random) { est.psi00<-est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0<-try(seg.glm.fit(y, XREG, Z, PSI1, w, offs, opz1), silent=TRUE) ss00<-o0$dev.no.gap } else { est.psi00<-est.psi0<-apply(PSI,2,mean) ss00<-opz$dev0 } } all.est.psi.boot<-all.selected.psi<-all.est.psi<-matrix(, nrow=n.boot, ncol=length(est.psi0)) all.ss<-all.selected.ss<-rep(NA, n.boot) if(is.null(size.boot)) size.boot<-n Z.orig<-Z if(visualBoot) cat(0, " ", formatC(opz$dev0, 3, format = "f"),"", "(No breakpoint(s))", "\n") count.random<-0 for(k in seq(n.boot)){ PSI <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) if(jt) Z<-apply(Z.orig,2,jitter) if(nonParam){ id<-sample(n, size=size.boot, replace=TRUE) o.boot<-try(seg.glm.fit(y[id], XREG[id,,drop=FALSE], Z[id,,drop=FALSE], PSI[id,,drop=FALSE], w[id], offs[id], opz), silent=TRUE) } else { yy<-fitted.ok+sample(residuals(o0),size=n, replace=TRUE) o.boot<-try(seg.glm.fit(yy, XREG, Z.orig, PSI, weights, offs, opz), silent=TRUE) } if(is.list(o.boot)){ all.est.psi.boot[k,]<-est.psi.boot<-o.boot$psi } else { est.psi.boot<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) } PSI <- matrix(rep(est.psi.boot, rep(nrow(Z), length(est.psi.boot))), ncol = length(est.psi.boot)) opz$h<-max(opz$h*.9, .2) opz$it.max<-opz$it.max+1 o<-try(seg.glm.fit(y, XREG, Z.orig, PSI, w, offs, opz), silent=TRUE) if(!is.list(o) && random){ est.psi00<-est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o<-try(seg.glm.fit(y, XREG, Z, PSI1, w, offs, opz1), silent=TRUE) count.random<-count.random+1 } if(is.list(o)){ if(!"coefficients"%in%names(o$obj)) o<-extract.psi(o) all.est.psi[k,]<-o$psi all.ss[k]<-o$dev.no.gap if(o$dev.no.gap<=ifelse(is.list(o0), o0$dev.no.gap, 10^12)) o0<-o est.psi0<-o0$psi all.selected.psi[k,] <- est.psi0 all.selected.ss[k]<-o0$dev.no.gap #min(c(o$SumSquares.no.gap, o0$SumSquares.no.gap)) } if(visualBoot) { flush.console() spp <- if (k < 10) "" else NULL cat(k, spp, "", formatC(o0$dev.no.gap, 3, format = "f"), "\n") } } #end n.boot all.selected.psi<-rbind(est.psi00,all.selected.psi) all.selected.ss<-c(ss00, all.selected.ss) ris<-list(all.selected.psi=drop(all.selected.psi),all.selected.ss=all.selected.ss, all.psi=all.est.psi, all.ss=all.ss) if(is.null(o0$obj)){ PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0<-try(seg.glm.fit(y, XREG, Z, PSI1, w, offs, opz1), silent=TRUE) } if(!is.list(o0)) return(0) o0$boot.restart<-ris return(o0) }segmented/R/print.segmented.R0000644000176200001440000000330113157724622015665 0ustar liggesusers`print.segmented` <- function(x,digits = max(3, getOption("digits") - 3),...){ #revisione 15/05/03; 24/02/04 if(is.null(x$psi)) x<-x[[length(x)]] if(!"segmented"%in%class(x)) stop("a `segmented' object is requested") cat( "Call: " ) print( x$call ) cat("\nMeaningful coefficients of the linear terms:\n") #print(x$coef[(1:(length(x$coef)-length(x$psi[,2])))]) iV<- -match(x$nameUV[[2]],names(coef(x)))#iV<- -grep("psi.",names(coef(x)))#indices all but V #print(x$coef[iV]) print.default(format(x$coef[iV], digits = digits), print.gap = 2, quote = FALSE) cat("\n") cat("Estimated Break-Point(s):\n") a<-as.vector(x$psi[,"Est."]) names(a)<-rownames(x$psi) print.default(a, digits, print.gap=2) #print(x$psi[,"Est."], digits) #ng<-max(x$id.psi.group) #for(i in 1:ng){ # nomi<-x$nameUV$V[x$id.psi.group==i] # cat("\t",nomi,":", format(signif(x$psi[nomi,"Est."],digits)),"\n") # } if("glm"%in%class(x)){ cat("\nDegrees of Freedom:", x$df.null, "Total (i.e. Null); ", x$df.residual, "Residual\n") cat("Null Deviance: ", format(signif(x$null.deviance, digits)), "\nResidual Deviance:", format(signif(x$deviance, digits)), " AIC:", format(signif(x$aic, digits)), "\n") } if("Arima"%in%class(x)){ cm <- x$call$method if (is.null(cm) || cm != "CSS") cat("\nsigma^2 estimated as ", format(x$sigma2, digits = digits), ": log likelihood = ", format(round(x$loglik, 2)), ", aic = ", format(round(x$aic, 2)), "\n", sep = "") else cat("\nsigma^2 estimated as ", format(x$sigma2, digits = digits), ": part log likelihood = ", format(round(x$loglik, 2)), "\n", sep = "") } invisible(x) } segmented/R/vcov.segmented.R0000644000176200001440000000163313157724622015514 0ustar liggesusersvcov.segmented<-function (object, var.diff=FALSE, ...){ if(inherits(object, "Arima")){ v<-object$var.coef return(v) } if(inherits(object, "glm")){ if(var.diff) warning("option var.diff=TRUE ignored with `glm' objects", call.=FALSE) so <- summary.glm(object, correlation = FALSE, ...) v<-so$dispersion * so$cov.unscaled } else { if(var.diff){ if(length(object$nameUV$Z)>1) { var.diff<-FALSE warning("var.diff set to FALSE with multiple segmented variables", call.=FALSE) } v<-summary.segmented(object, var.diff=TRUE, correlation = FALSE, ...)$cov.var.diff } else { so<-summary.segmented(object, var.diff=FALSE, correlation = FALSE, ...) v<-so$sigma^2 * so$cov.unscaled #object$cov.unscaled.is } } return(v) } segmented/R/broken.line.r0000644000176200001440000001300413157724622015026 0ustar liggesusersbroken.line<-function(ogg, term=NULL, link=TRUE, interc=TRUE, se.fit=TRUE){ #ogg: l'oggetto segmented #term: una lista *nominata* con i valori rispetto a cui calcolare i fitted # OPPURE una stringa per indicare la variabile segmented OPPURE NULL (se c'e' solo una variabile) dummy.matrix<-NULL dummy.matrix<-function(x.values, x.name, obj.seg, psi.est=TRUE){ #given the segmented fit 'obj.seg' and a segmented variable x.name with corresponding values x.values, #this function simply returns a matrix with columns (x, (x-psi)_+, -b*I(x>psi)) #or ((x-psi)_+, -b*I(x>psi)) if obj.seg does not include the coef for the linear "x" f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } n<-length(x.values) #le seguenti righe selezionavano (ERRONEAMENTE) sia "U1.x" sia "U1.neg.x" (se "x" e "neg.x" erano segmented covariates) #nameU<- grep(paste("\\.",x.name,"$", sep=""), obj.seg$nameUV$U, value = TRUE) #nameV<- grep(paste("\\.",x.name,"$", sep=""), obj.seg$nameUV$V, value = TRUE) nameU<-obj.seg$nameUV$U[f.U(obj.seg$nameUV$U,x.name)] nameV<-obj.seg$nameUV$V[f.U(obj.seg$nameUV$V,x.name)] diffSlope<-coef(obj.seg)[nameU] est.psi<-obj.seg$psi[nameV, "Est."] k<-length(est.psi) PSI <- matrix(rep(est.psi, rep(n, k)), ncol = k) newZ<-matrix(x.values, nrow=n,ncol=k, byrow = FALSE) dummy1<-pmax(newZ-PSI,0) if(psi.est){ V<-ifelse(newZ>PSI,-1,0) dummy2<- if(k==1) V*diffSlope else V%*%diag(diffSlope) #t(diffSlope*t(-I(newZ>PSI))) newd<-cbind(x.values,dummy1,dummy2) colnames(newd)<-c(x.name,nameU, nameV) } else { newd<-cbind(x.values,dummy1) colnames(newd)<-c(x.name,nameU) } if(!x.name%in%names(coef(obj.seg))) newd<-newd[,-1,drop=FALSE] return(newd) } #-------------- f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #------------- xvalues<-term nomeV <- ogg$nameUV$V nomeU <- ogg$nameUV$U nomeZ <- ogg$nameUV$Z n.seg<-length(nomeZ) if(is.null(xvalues)){ if(n.seg>1) stop("there are multiple segmented covariates. Please specify one.") xvalues<-ogg$model[nomeZ] } if(is.character(xvalues)){ if(!xvalues %in% nomeZ) stop("'xvalues' is not a segmented covariate") xvalues<-ogg$model[xvalues] } nomeOK<-names(xvalues) if(length(nomeOK)>1) stop("Please specify one variable") if(!nomeOK %in% nomeZ) stop("'names(xvalues)' is not a segmented covariate") #if(n.seg>1 && !is.list(x.values)) stop("with multiple segmented covariates, please specify a named dataframe") #x.values<-data.frame(x.values) #names(x.values)<-nomeZ nomi <- names(coef(ogg)) nomiSenzaV <- nomiSenzaU <- nomi nomiSenzaU[match(nomeU, nomi)] <- "" nomiSenzaV[match(nomeV, nomi)] <- "" index <- vector(mode = "list", length = length(nomeZ)) for (i in 1:n.seg) { index[[i]] <- c(match(nomeZ[i], nomi), f.U(ogg$nameUV$U, nomeZ[i]) + (match(ogg$nameUV$U[1], nomi)-1), f.U(ogg$nameUV$V, nomeZ[i]) + (match(ogg$nameUV$V[1], nomi)-1)) #grep(paste("\\.", nomeZ[i], "$", sep = ""), nomiSenzaV, value = FALSE), #grep(paste("\\.", nomeZ[i], "$", sep = ""), nomiSenzaU, value = FALSE)) } ste.fit<-fit <- vector(mode = "list", length = length(nomeZ)) for (i in 1:n.seg) { x.name <- nomeZ[i] X<-dummy.matrix(unlist(xvalues), x.name, ogg)#<--NB: xvalues non varia con i!!! perche' farlo calcolare comunque? ind <- as.numeric(na.omit(unlist(index[[i]]))) if(interc && "(Intercept)"%in%nomi) { ind<- c(match("(Intercept)",nomi),ind) X<-cbind(1,X) } cof <- coef(ogg)[ind] fit[[i]]<-drop(X%*%cof) ste.fit[[i]] <- if(!se.fit) 10 else sqrt(rowSums((X %*% vcov(ogg)[ind,ind]) * X)) #sqrt(diag(X%*%Var%*%t(X))) } names(fit)<- names(ste.fit)<- nomeZ r<-list(fit=fit[[nomeOK]], se.fit=ste.fit[[nomeOK]]) if (inherits(ogg, what = "glm", FALSE) && !link){ r[[2]] <- ogg$family$mu.eta(r[[1]])*r[[2]] r[[1]] <- ogg$family$linkinv(r[[1]]) } if(!se.fit) r<-r[1] return(r) } segmented/R/draw.history.R0000644000176200001440000000776013157724622015231 0ustar liggesusersdraw.history<-function(obj,term,...){ #show.history() se c'e' stato boot restart potrebbe produrre un grafico 2x1 di "dev vs it" and "no.of distinct vs it" #-- f.U<-function(nomiU, term=NULL){ #trasforma i nomi dei coeff U (o V) nei nomi delle variabili corrispondenti #and if 'term' is provided (i.e. it differs from NULL) the index of nomiU matching term are returned k<-length(nomiU) nomiUsenzaU<-strsplit(nomiU, "\\.") nomiU.ok<-vector(length=k) for(i in 1:k){ nomi.i<-nomiUsenzaU[[i]][-1] if(length(nomi.i)>1) nomi.i<-paste(nomi.i,collapse=".") nomiU.ok[i]<-nomi.i } if(!is.null(term)) nomiU.ok<-(1:k)[nomiU.ok%in%term] return(nomiU.ok) } #-- if(missing(term)){ if(length(obj$nameUV$Z)>1 ) {stop("please, specify `term'")} else {term<-obj$nameUV$Z} } range.ok<-obj$rangeZ[,term] #id.ok<-grep(paste("\\.",term,"$",sep=""), rownames(obj$psi),value=FALSE) id.ok<- f.U(rownames(obj$psi), term) est.psi<-obj$psi[id.ok,2] if(length(obj$psi.history)==5) { #boot (non-autom) par(mfrow=c(1,2)) plot(obj$psi.history$all.selected.ss, type="b", xlab="bootstrap replicates", ylab="RSS (selected values)", xaxt="n", pch=20) axis(1,at=1:length(obj$psi.history$all.selected.ss),cex.axis=.7) #unicita' delle soluzioni if(is.vector(obj$psi.history$all.selected.psi)){ psi.matr<-m<-matrix(obj$psi.history$all.selected.psi, ncol=1) } else { psi.matr<-m<-obj$psi.history$all.selected.psi[,id.ok,drop=FALSE] } for(i in 1:nrow(m)) m[i,]<-apply(psi.matr[1:i,,drop=FALSE],2,function(xx)length(unique(xx))) m<-t(t(m)+.1*(0:(ncol(m)-1))) matplot(1:nrow(m),m, pch=1:ncol(m), type="b", col=1:ncol(m), ylab="no. of distinct solutions",xlab="bootstrap replicates", xaxt="n") axis(1,at=1:nrow(m),cex.axis=.7) } else { if(all(diff(sapply(obj$psi.history, length))==0)){ #non-boot, non-autom A<-t(matrix(unlist(obj$psi.history),nrow=nrow(obj$psi),byrow=FALSE)) colnames(A)<-rownames(obj$psi) matplot(1:nrow(A),A[,id.ok],type="o",pch=1:length(est.psi),col=1, xlab="iterations", ylab=paste("breakpoint ","(",term,")",sep=""), ylim=range.ok, xaxt="n",...) axis(1,at=1:nrow(A),cex.axis=.7) #if(rug) points(rep(1)) abline(h=est.psi,lty=3) } else { #non-boot, Autom id.iter<-rep(1:length(obj$psi.history), times=sapply(obj$psi.history, length)) psi.history<-unlist(obj$psi.history) nomi<-unlist(sapply(obj$psi.history, names)) d<-data.frame(iter=id.iter, psi=psi.history, nomi=nomi) #associa i nomi delle componenti di $psi.history (che sono indici 1,2,..) con i nomi della variabile term ii<-unique(names(obj$psi.history[[length(obj$psi.history)]])[id.ok]) if(length(ii)>1) stop("some error in the names?..") with(d[d$nomi==ii,], plot(iter, psi, xlab="iterations", ylab=paste("breakpoint ","(",term,")",sep=""), xaxt="n",...)) axis(1,at=unique(d$iter),cex.axis=.7) #se vuoi proprio associare le stime tra le diverse iterazioni #(per poi unire nel grafico i punti con le linee. Ovviamente alcune linee saranno interrotte) # for(i in 1:length(obj$psi.history)) { # a<-obj$psi.history[[i]] # for(j in 1:length(est.psi)){ # psij<-est.psi[j] #a<- ..names match # r[i,j]<-a[which.min(abs(a-psij))] # a<-setdiff(a, r[i,j]) # } # } } } } #end_fn segmented/R/segmented.R0000644000176200001440000000020713157724622014534 0ustar liggesusers`segmented` <- function(obj, seg.Z, psi, control=seg.control(), model=TRUE, ...){ UseMethod("segmented") } segmented/R/seg.def.fit.r0000644000176200001440000001736413157724622014731 0ustar liggesusersseg.def.fit<-function(obj, Z, PSI, mfExt, opz, return.all.sol=FALSE){ #----------------- fn.costr<-function(n.psi,isLeft=1,isInterc=1){ #build the constraint matrix #isLeft: TRUE (or 1) if there is the left slope #isInterc: TRUE (or 1) if there is the intercept.. IU<- -diag(n.psi) sumU<- diag(n.psi) #n. of diff slopes sumU[row(sumU)>col(sumU)]<-1 if(isLeft) { sumU<-cbind(1, sumU) IU<-diag(c(1, -rep(1, n.psi))) } A<-rbind(IU,sumU) if(isInterc) { A<-rbind(0,A) A<-cbind(c(1, rep(0,nrow(A)-1)), A) } #add zeros for the V coeffs A<-cbind(A, matrix(0,nrow(A), n.psi)) A } #----------------- dpmax<-function(x,y,pow=1){ #deriv pmax if(pow==1) ifelse(x>y, -1, 0) else -pow*pmax(x-y,0)^(pow-1) } #----------- vincoli<- FALSE c1 <- apply((Z <= PSI), 2, all) c2 <- apply((Z >= PSI), 2, all) if(sum(c1 + c2) != 0 || is.na(sum(c1 + c2))) stop("psi out of the range") # digits<-opz$digits pow<-opz$pow nomiOK<-opz$nomiOK toll<-opz$toll h<-opz$h gap<-opz$gap stop.if.error<-opz$stop.if.error dev.new<-opz$dev0 visual<-opz$visual id.psi.group<-opz$id.psi.group it.max<-old.it.max<-opz$it.max rangeZ <- apply(Z, 2, range) psi<-PSI[1,] names(psi)<-id.psi.group #H<-1 it <- 1 epsilon <- 10 dev.values<-psi.values <- NULL id.psi.ok<-rep(TRUE, length(psi)) nomiU<- opz$nomiU nomiV<- opz$nomiV call.ok <- opz$call.ok call.noV <- opz$call.noV #browser() if(is.null(opz$constr)) opz$constr<-0 if((opz$constr %in% 1:2) && class(obj)=="rq"){ vincoli<-TRUE call.ok$method<-"fnc" call.ok$R<-quote(R) call.ok$r<-quote(r) call.noV$method<-"fnc" call.noV$R<-quote(R.noV) call.noV$r<-quote(r) } fn.obj<-opz$fn.obj toll<-opz$toll k<-ncol(Z) while (abs(epsilon) > toll) { #k<-ncol(Z) U <- pmax((Z - PSI), 0)^pow[1]#U <- pmax((Z - PSI), 0) V <- dpmax(Z,PSI,pow=pow[2])# ifelse((Z > PSI), -1, 0) for(i in 1:k) { mfExt[nomiU[i]] <- U[,i] mfExt[nomiV[i]] <- V[,i] } R<- fn.costr(ncol(U),1,1) R.noV<-R[,-((ncol(R)-1)+seq_len(ncol(U))),drop=FALSE] r<- rep(0, nrow(R)) obj <- suppressWarnings(eval(call.ok, envir=mfExt)) dev.old<-dev.new dev.new <- dev.new1 <- eval(parse(text=fn.obj), list(x=obj)) #control$f.obj should be something like "sum(x$residuals^2)" or "x$dev" if(length(dev.new)<=0) stop("error in the objective to be minimized, see 'fn.obj'") if(return.all.sol) { obj.noV <- suppressWarnings(eval(call.noV, envir=mfExt)) dev.new1 <- eval(parse(text=fn.obj), list(x=obj.noV)) #dev.new1 <- sum(mylm(x = cbind(XREG, U), y = y, w = w, offs = offs)$residuals^2) } dev.values[[length(dev.values) + 1]] <- dev.new1 if (visual) { flush.console() if (it == 1) cat(0, " ", formatC(dev.old, 3, format = "f"), "", "(No breakpoint(s))", "\n") spp <- if (it < 10) "" else NULL cat(it, spp, "", formatC(dev.new, 3, format = "f"), "",length(psi),"\n") #cat(paste("iter = ", it, spp," dev = ",formatC(dev.new,digits=3,format="f"), " n.psi = ",formatC(length(psi),digits=0,format="f"), sep=""), "\n") } epsilon <- (dev.new - dev.old)/(dev.old + .001) obj$epsilon <- epsilon it <- it + 1 obj$it <- it beta.c<-coef(obj)[nomiU] gamma.c<-coef(obj)[nomiV] if (it > it.max) break psi.values[[length(psi.values) + 1]] <- psi.old <- psi # if(it>=old.it.max && h<1) H<-h psi <- psi.old + h*gamma.c/beta.c if(!is.null(digits)) psi<-round(psi, digits) PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #check if psi is admissible.. a <- apply((Z <= PSI), 2, all) #prima era solo < b <- apply((Z >= PSI), 2, all) #prima era solo > if(stop.if.error) { isErr<- (sum(a + b) != 0 || is.na(sum(a + b))) if(isErr) { if(return.all.sol) return(list(dev.values, psi.values)) else stop("(Some) estimated psi out of its range") } } else { id.psi.ok<-!is.na((a+b)<=0)&(a+b)<=0 Z <- Z[,id.psi.ok,drop=FALSE] psi <- psi[id.psi.ok] PSI <- PSI[,id.psi.ok,drop=FALSE] ToDeletenomiU<-nomiU[!id.psi.ok] #salva i nomi delle U per i psi ammissibili ToDeletenomiV<-nomiV[!id.psi.ok] #salva i nomi delle V per i psi ammissibili if(length(ToDeletenomiU)>0 || length(ToDeletenomiV)>0) {for(nn in c(ToDeletenomiU, ToDeletenomiV)) {mfExt[[nn]]<-NULL}} nomiOK<-nomiOK[id.psi.ok] #salva i nomi delle U per i psi ammissibili nomiU<-nomiU[id.psi.ok] #salva i nomi delle U per i psi ammissibili nomiV<-nomiV[id.psi.ok] #salva i nomi delle V per i psi ammissibili id.psi.group<-id.psi.group[id.psi.ok] names(psi)<-id.psi.group if(ncol(PSI)<=0) return(0) k<-ncol(Z) #aggiorna la call, altrimenti il modello avra' sempre lo stesso numero di termini anche se alcuni psi vengono rimossi!!! Fo <- update.formula(opz$formula.orig, as.formula(paste(".~.+", paste(c(nomiU, nomiV), collapse = "+")))) Fo.noV <- update.formula(opz$formula.orig, as.formula(paste(".~.+", paste(nomiU, collapse = "+")))) call.ok <- update(obj, formula = Fo, evaluate=FALSE, data = mfExt) call.noV <- update(obj, formula = Fo.noV, evaluate=FALSE, data = mfExt) } #end else #obj$psi <- psi } #end while psi<-unlist(tapply(psi, id.psi.group, sort)) names(psi)<-id.psi.group PSI <- matrix(rep(psi, rep(nrow(Z), length(psi))), ncol = length(psi)) #aggiunto da qua.. U <- pmax((Z - PSI), 0) V <- ifelse((Z > PSI), -1, 0) for(i in 1:k) { mfExt[nomiU[i]] <- U[,i] mfExt[nomiV[i]] <- V[,i] } ##LA DOMANDA E': PERCHE' QUI STIMA UN MODELLO SENZA V SE POI VIENE RISTIMATO in segmented.default (o segmented.lm o segmented.glm?) ##RE: il valore di SS.new serve per il boot restart. #Invece la domanda e': non si puo' restituire direttamente obj.new senza bisogno di sostituire i valori in obj ? obj.new <- suppressWarnings(eval(call.noV, envir=mfExt)) SS.new <- eval(parse(text=fn.obj), list(x=obj.new)) #sum(obj.new$residuals^2) if(!gap){ obj<-obj.new #names.coef<-names(obj$coefficients) #obj$coefficients<-c(obj.new$coefficients, rep(0,ncol(V))) #names(obj$coefficients)<-names.coef #obj$residuals<-obj.new$residuals #obj$fitted.values<-obj.new$fitted.values #obj$linear.predictors<-obj.new$linear.predictors #obj$deviance<-obj.new$deviance #obj$weights<-obj.new$weights #obj$aic<-obj.new$aic #+ 2*ncol(V) #ho fatto la modifica in segmented.glm(): "objF$aic<-obj$aic + 2*k" } else { obj <- suppressWarnings(eval(call.ok, envir=mfExt)) } obj$epsilon <- epsilon obj$it <- it #fino a qua.. obj<-list(obj=obj,it=it,psi=psi,psi.values=psi.values,U=U,V=V,rangeZ=rangeZ, epsilon=epsilon,nomiOK=nomiOK, SumSquares.no.gap=SS.new, id.psi.group=id.psi.group, nomiV=nomiV, nomiU=nomiU, mfExt=mfExt) #inserire id.psi.ok? #browser() if(vincoli) { obj$R<-R obj$R.noV<-R.noV obj$r<-r } return(obj) } segmented/R/seg.Ar.fit.boot.r0000644000176200001440000001232713157724622015471 0ustar liggesusersseg.Ar.fit.boot<-function(obj, XREG, Z, PSI, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE){ #random se TRUE prende valori random quando e' errore: comunque devi modificare qualcosa (magari con it.max) # per fare restituire la dev in corrispondenza del punto psi-random #nonParm. se TRUE implemneta il case resampling. Quello semiparam dipende dal non-errore di extract.psi<-function(lista){ #serve per estrarre il miglior psi.. dev.values<-lista[[1]] psi.values<-lista[[2]] dev.ok<-min(dev.values) id.dev.ok<-which.min(dev.values) if(is.list(psi.values)) psi.values<-matrix(unlist(psi.values), nrow=length(dev.values), byrow=TRUE) if(!is.matrix(psi.values)) psi.values<-matrix(psi.values) psi.ok<-psi.values[id.dev.ok,] r<-list(SumSquares.no.gap=dev.ok, psi=psi.ok) r } #------------- visualBoot<-opz$visualBoot opz.boot<-opz opz.boot$pow=c(1.1,1.2) opz1<-opz opz1$it.max <-1 n<-nrow(Z) o0<-try(seg.Ar.fit(obj, XREG, Z, PSI, opz), silent=TRUE) rangeZ <- apply(Z, 2, range) #serve sempre if(!is.list(o0)) { o0<- seg.Ar.fit(obj, XREG, Z, PSI, opz, return.all.sol=TRUE) o0<-extract.psi(o0) if(!nonParam) {warning("using nonparametric boot");nonParam<-TRUE} } if(is.list(o0)){ est.psi00<-est.psi0<-o0$psi ss00<-o0$SumSquares.no.gap if(!nonParam) fitted.ok<-fitted(o0) } else { if(!nonParam) stop("the first fit failed and I cannot extract fitted values for the semipar boot") if(random) { est.psi00<-est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0<-try(seg.Ar.fit(obj, Z, PSI1, opz1), silent=TRUE) ss00<-o0$SumSquares.no.gap } else { est.psi00<-est.psi0<-apply(PSI,2,mean) ss00<-opz$dev0 } } all.est.psi.boot<-all.selected.psi<-all.est.psi<-matrix(, nrow=n.boot, ncol=length(est.psi0)) all.ss<-all.selected.ss<-rep(NA, n.boot) if(is.null(size.boot)) size.boot<-n # na<- ,,apply(...,2,function(x)mean(is.na(x))) Z.orig<-Z if(visualBoot) cat(0, " ", formatC(opz$dev0, 3, format = "f"),"", "(No breakpoint(s))", "\n") count.random<-0 for(k in seq(n.boot)){ PSI <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) if(jt) Z<-apply(Z.orig,2,jitter) if(nonParam){ id<-sample(n, size=size.boot, replace=TRUE) o.boot<-try(seg.Ar.fit(obj, XREG[id,,drop=FALSE], Z[id,,drop=FALSE], PSI[id,,drop=FALSE], opz.boot), silent=TRUE) } else { yy<-fitted.ok+sample(residuals(o0),size=n, replace=TRUE) ##----> o.boot<-try(seg.lm.fit(yy, XREG, Z.orig, PSI, weights, offs, opz.boot), silent=TRUE) #in realta' la risposta dovrebbe essere "yy" da cambiare in mfExt o.boot<- try(seg.Ar.fit(obj, XREG, Z.orig, PSI, opz.boot), silent=TRUE) } if(is.list(o.boot)){ all.est.psi.boot[k,]<-est.psi.boot<-o.boot$psi } else { est.psi.boot<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) } PSI <- matrix(rep(est.psi.boot, rep(nrow(Z), length(est.psi.boot))), ncol = length(est.psi.boot)) opz$h<-max(opz$h*.9, .2) opz$it.max<-opz$it.max+1 o <- try(seg.Ar.fit(obj, XREG, Z.orig, PSI, opz, return.all.sol=TRUE), silent=TRUE) if(!is.list(o) && random){ est.psi0<-apply(rangeZ,2,function(r)runif(1,r[1],r[2])) PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o <- try(seg.Ar.fit(obj, XREG, Z, PSI1, opz1), silent=TRUE) count.random<-count.random+1 } if(is.list(o)){ if(!"coef"%in%names(o$obj)) o<-extract.psi(o) all.est.psi[k,]<-o$psi all.ss[k]<-o$SumSquares.no.gap if(o$SumSquares.no.gap<=ifelse(is.list(o0), o0$SumSquares.no.gap, 10^12)) o0<-o est.psi0<-o0$psi all.selected.psi[k,] <- est.psi0 all.selected.ss[k]<-o0$SumSquares.no.gap #min(c(o$SumSquares.no.gap, o0$SumSquares.no.gap)) } if(visualBoot) { flush.console() spp <- if (k < 10) "" else NULL cat(k, spp, "", formatC(o0$SumSquares.no.gap, 3, format = "f"), "\n") } } #end n.boot all.selected.psi<-rbind(est.psi00,all.selected.psi) all.selected.ss<-c(ss00, all.selected.ss) ris<-list(all.selected.psi=drop(all.selected.psi),all.selected.ss=all.selected.ss, all.psi=all.est.psi, all.ss=all.ss) if(is.null(o0$obj)){ PSI1 <- matrix(rep(est.psi0, rep(nrow(Z), length(est.psi0))), ncol = length(est.psi0)) o0 <- try(seg.Ar.fit(obj, XREG, Z, PSI1, opz1), silent=TRUE) } if(!is.list(o0)) return(0) o0$boot.restart<-ris return(o0) }segmented/R/davies.test.r0000644000176200001440000002671113157724622015062 0ustar liggesusers#se n=1000: value out of range in 'gammafn' #warning se "lm" con "glm"??? `davies.test` <- function (obj, seg.Z, k = 10, alternative = c("two.sided", "less", "greater"), type=c("lrt","wald"), values=NULL, dispersion=NULL) { # extract.t.value.U<-function(x){ # #estrae il t-value dell'ultimo coeff in un oggetto restituito da lm.fit # #non serve... in realta' viene usata extract.t.value.U.glm() # #x<-x$obj # R<-qr.R(x$qr) # p<-ncol(R) # n<-length(x$fitted.values) # invR<-backsolve(R,diag(p)) # hat.sigma2<-sum(x$residuals^2)/(n-p) # #solve(crossprod(qr.X(x$qr))) # V<-tcrossprod(invR)*hat.sigma2 # tt<-x$coefficients[p]/sqrt(V[p,p]) # tt} #------------------------------------------------------------------------------- daviesLM<-function(y, z, xreg, weights, offs, values, k, alternative){ #Davies test with sigma unknown #-------------- #> gammaA<-function(x){ # x^(x-.5)*exp(-x)*sqrt(2*pi)*(1+1/(12*x)+1/(288*x^2)-139/(51840*x^3) -571/(2488320*x^4)) # } #exp(lgamma()) fn="pmax(x-p,0)" y<-y-offs n<-length(y) n1<-length(values) RIS<-matrix(,n1,2) X.psi<-matrix(,n,length(fn)) df.res<- n - ncol(xreg) - length(fn) for(i in 1:n1){ for(j in 1:length(fn)) X.psi[,j]<-eval(parse(text=fn[[j]]), list(x=z, p=values[i])) xx1.new<-cbind(X.psi,xreg) #lrt #mu1.new<-xx1.new%*%solve(crossprod(xx1.new), crossprod(xx1.new,y)) #rss1<-sum((y-mu1.new)^2) #sigma2<-if(missing(sigma)) rss1/(n-ncol(xx1.new)) else sigma^2 #RIS[i]<-((rss0-rss1)/ncol(X.psi))/sigma2 #Wald invXtX1<-try(solve(crossprod(sqrt(weights)*xx1.new)), silent=TRUE) if(class(invXtX1)!="try-error"){ hat.b<-drop(invXtX1%*%crossprod(weights*xx1.new,y)) mu1.new<-xx1.new%*%hat.b devE<-sum((weights*(y-mu1.new)^2)) hat.sigma<- sqrt(devE/df.res) RIS[i,1]<-hat.b[1]/(hat.sigma*sqrt(invXtX1[1, 1])) Z<-hat.b[1]/(sqrt(invXtX1[1, 1])) D2<- Z^2 + devE RIS[i,2]<-Z^2/D2 #beta } } valori<-values[!is.na(RIS[,1])] RIS<- RIS[!is.na(RIS[,1]),] V<-sum(abs(diff(asin(RIS[,2]^.5)))) onesided <- TRUE if (alternative == "less") { M <- min(RIS[,1]) best<-valori[which.min(RIS[,1])] p.naiv <- pt(M, df=df.res, lower.tail = TRUE) } else if (alternative == "greater") { M <- max(RIS[,1]) best<-valori[which.max(RIS[,1])] p.naiv <- pt(M, df=df.res, lower.tail = FALSE) } else { M <- max(abs(RIS[,1])) best<-valori[which.max(abs(RIS[,1]))] p.naiv <- pt(M, df=df.res, lower.tail = FALSE) onesided <- FALSE } u<-M^2/((n-ncol(xx1.new))+ M^2) approxx<-V*(((1-u)^((df.res-1)/2))*gamma(df.res/2+.5))/(2*gamma(df.res/2)*pi^.5) p.adj <- p.naiv + approxx p.adj <- ifelse(onesided, 1, 2) * p.adj p.adj<-list(p.adj=p.adj, valori=valori, ris.valori=RIS[,1], best=best) return(p.adj) # M<-max(abs(RIS[,1])) # u<-M^2/((n-ncol(xx1.new))+ M^2) # approxx<-V*(((1-u)^((df.res-1)/2))*gamma(df.res/2+.5))/(2*gamma(df.res/2)*pi^.5) # p.naiv<-pt(-abs(M), df=df.res) #naive p-value # p.adj<-2*(p.naiv+approxx) #adjusted p-value (upper bound) # p.adj<-min(p.adj, 1) # p.adj<-list(p.adj=p.adj, valori=values, ris.valori=RIS[,1], approxx=approxx, p.naiv=p.naiv) # return(p.adj) } #-------------------------------- daviesGLM<-function(y, z, xreg, weights, offs, values=NULL, k, list.glm, alternative){ #Davies test for GLM (via LRT or Wald) est.dispGLM<-function(object){ df.r <- object$df.residual dispersion <- if(object$family$family%in%c("poisson","binomial")) 1 else object$dev/df.r dispersion } extract.t.value.U.glm<-function(object,dispersion,isGLM=TRUE){ #estrae il t-value dell'ultimo coeff in un oggetto restituito da lm.wfit/glm.fit est.disp <- FALSE df.r <- object$df.residual if (is.null(dispersion)) dispersion <- if(isGLM&&(object$family$family%in%c("poisson","binomial"))) 1 else if (df.r > 0) { est.disp <- TRUE if (any(object$weights == 0)) warning("observations with zero weight not used for calculating dispersion") sum((object$weights * object$residuals^2)[object$weights > 0])/df.r } else { est.disp <- TRUE NaN } dispersion<-max(c(dispersion, 1e-10)) p <- object$rank p1 <- 1L:p Qr <- object$qr coef.p <- object$coefficients[Qr$pivot[p1]] covmat.unscaled <- chol2inv(Qr$qr[p1, p1, drop = FALSE]) dimnames(covmat.unscaled) <- list(names(coef.p), names(coef.p)) covmat <- dispersion * covmat.unscaled tvalue <- coef.p[1]/sqrt(covmat[1,1]) #<0.4.0-0 era coef.p[p]/sqrt(covmat[p,p]) tvalue }#end extract.t.value.U.glm #-------------- fn<-"pmax(x-p,0)" dev0<-list.glm$dev0 eta0<-list.glm$eta0 family=list.glm$family type<-list.glm$type dispersion<-list.glm$dispersion n<-length(y) r<-length(fn) n1<-length(values) RIS<-rep(NA, n1) X.psi<-matrix(,n,length(fn)) for(i in 1:n1){ for(j in 1:length(fn)) X.psi[,j]<-eval(parse(text=fn[[j]]), list(x=z, p=values[i])) xreg1<-cbind(X.psi,xreg) o1<-glm.fit(x = xreg1, y = y, weights = weights, offset = offs, family=family, etastart=eta0) dev<-o1$dev if (is.list(o1) && ncol(xreg1)==o1$rank) { RIS[i]<- if(type=="lrt") sqrt((dev0-dev)/est.dispGLM(o1))*sign(o1$coef[1]) else extract.t.value.U.glm(o1,dispersion) } } valori<-values[!is.na(RIS)] ris.valori<-RIS[!is.na(RIS)] V<-sum(abs(diff(ris.valori))) #-----Questo e' se il test di riferimento e' una \chi^2_r. (Dovresti considerare il LRT non segnato) #V<-sum(abs(diff(sqrt(RIS))))#nota sqrt #M<- max(RIS) #approxx<-(V*(M^((r-1)/2))*exp(-M/2)*2^(-r/2))/gamma(r/2) #p.naiv<-1-pchisq(M,df=r) #naive p-value #p.adj<-min(p.naiv+approxx,1) #adjusted p-value (upper bound) onesided <- TRUE if (alternative == "less") { M <- min(ris.valori) best<-valori[which.min(ris.valori)] p.naiv <- pnorm(M, lower.tail = TRUE) } else if (alternative == "greater") { M <- max(ris.valori) best<-valori[which.max(ris.valori)] p.naiv <- pnorm(M, lower.tail = FALSE) } else { M <- max(abs(ris.valori)) best<-valori[which.max(abs(ris.valori))] p.naiv <- pnorm(M, lower.tail = FALSE) onesided <- FALSE } approxx<-V*exp(-(M^2)/2)/sqrt(8*pi) p.adj <- p.naiv + approxx p.adj <- ifelse(onesided, 1, 2) * p.adj p.adj<-list(p.adj=p.adj, valori=valori, ris.valori=ris.valori, best=best) return(p.adj) } #------------------------------------------------------------------------------- if(!inherits(obj, "lm")) stop("A 'lm', 'glm', or 'segmented' model is requested") if(class(seg.Z)!="formula") stop("'seg.Z' should be an one-sided formula") if(k<=1) stop("k>1 requested! k>=10 is recommended") if(k<10) warnings("k>=10 is recommended") alternative <- match.arg(alternative) type <- match.arg(type) #if(length(all.vars(seg.Z))>1) warning("multiple segmented variables ignored in 'seg.Z'",call.=FALSE) if(length(all.vars(seg.Z))>1) stop("Only a single segmented variable can be specified in 'seg.Z' ") isGLM<-"glm"%in%class(obj) Call<-mf<-obj$call mf$formula<-formula(obj) m <- match(c("formula", "data", "subset", "weights", "na.action","offset"), names(mf), 0L) mf <- mf[c(1, m)] mf$drop.unused.levels <- TRUE mf[[1L]] <- as.name("model.frame") mf$formula<-update.formula(mf$formula,paste(seg.Z,collapse=".+")) formulaOrig<-formula(obj) if(class(obj)[1]=="segmented"){ if(!is.null(eval(obj$call$obj)$call$data)) mf$data <- eval(obj$call$obj)$call$data mf$formula<-update.formula(mf$formula,paste("~.-",paste(obj$nameUV$V, collapse="-"))) for(i in 1:length(obj$nameUV$U)) assign(obj$nameUV$U[i], obj$model[,obj$nameUV$U[i]], envir=parent.frame()) formulaOrig<-update.formula(formulaOrig, paste("~.-",paste(obj$nameUV$V, collapse="-"))) } mf <- eval(mf, parent.frame()) weights <- as.vector(model.weights(mf)) offs <- as.vector(model.offset(mf)) if(!is.null(Call$weights)){ #"(weights)"%in%names(mf) names(mf)[which(names(mf)=="(weights)")]<-all.vars(Call$weights) #as.character(Call$weights) #aggiungere??? # mf["(weights)"]<-weights } mt <- attr(mf, "terms") interc<-attr(mt,"intercept") y <- model.response(mf, "any") XREG <- if (!is.empty.model(mt)) model.matrix(mt, mf, contrasts) n <- nrow(XREG) if (is.null(weights)) weights <- rep(1, n) if (is.null(offs)) offs <- rep(0, n) name.Z <- all.vars(seg.Z) Z<-XREG[,match(name.Z, colnames(XREG))] if(!name.Z %in% names(coef(obj))) XREG<-XREG[,-match(name.Z, colnames(XREG)),drop=FALSE] list.glm<-list(dev0=obj$dev, eta0=obj$linear.predictor, family=family(obj), type=type, dispersion=dispersion) if(is.null(values)) values<-seq(sort(Z)[2], sort(Z)[(n - 1)], length = k) #values<-seq(min(z), max(z), length=k+2) #values<-values[-c(1,length(values))] if(class(obj)=="lm" || identical(class(obj),c("segmented","lm")) ) { if(n<=300) { rr<-daviesLM(y=y, z=Z, xreg=XREG, weights=weights, offs=offs, values=values, k=k, alternative=alternative) } else { list.glm$family<-gaussian() list.glm$type<-"wald" rr<-daviesGLM(y=y, z=Z, xreg=XREG, weights=weights, offs=offs, values=values, k=k, list.glm=list.glm, alternative=alternative) } } if(identical(class(obj),c("glm","lm")) || identical(class(obj),c("segmented","glm","lm"))) rr<-daviesGLM(y=y, z=Z, xreg=XREG, weights=weights, offs=offs, values=values, k=k, list.glm=list.glm, alternative=alternative) best<-rr$best p.adj<-rr$p.adj valori<-rr$valori ris.valori<-rr$ris.valori if(is.null(obj$family$family)) { famiglia<-"gaussian" legame<-"identity"} else { famiglia<-obj$family$family legame<-obj$family$link } out <- list(method = "Davies' test for a change in the slope", # data.name=paste("Model = ",famiglia,", link =", legame, # "\nformula =", as.expression(formulaOrig), # "\nsegmented variable =", name.Z), data.name=paste("formula =", as.expression(formulaOrig), ", method =", obj$call[[1]] , "\nmodel =",famiglia,", link =", legame, if(isGLM) paste(", statist =", type) else NULL , "\nsegmented variable =", name.Z), statistic = c("'best' at" = best), parameter = c(n.points = length(valori)), p.value = min(p.adj,1), alternative = alternative, process=cbind(psi.values=valori, stat.values=ris.valori)) class(out) <- "htest" return(out) } segmented/MD50000644000176200001440000000563713160062250012543 0ustar liggesusers69f0f65bfbcc946d9586bac2c49d2543 *DESCRIPTION 0ffe8f412060912587d634afef274654 *NAMESPACE a3d677cb26ef4ce49683d183048d9ac6 *NEWS cf61d7b238288be0a75ea4c8eb35f7fb *R/broken.line.r 249a33f124608ea43e541b016a3d8110 *R/confint.segmented.R 82cb289649e5db67189c39cb67cccc1f *R/davies.test.r a42345d81421b74cdb8c8159fb2db83d *R/draw.history.R 811b0c28b99a4fe9b03a3aa9edf69ea7 *R/intercept.r 5e225c969dc3f328224c702ef14971e1 *R/lines.segmented.R 3ddeff1d297cd6bf639b93a7d411ccf4 *R/plot.segmented.R 5f850c48c5c18919b2524b68be9c93c1 *R/points.segmented.r 764fb065475ee2f14297cd2a92c687ad *R/predict.segmented.r 65e331633c5afec47f0c98ccc871f058 *R/print.segmented.R 50650e747dedabc28e07a8e1ab09fe16 *R/print.summary.segmented.R c34eb7ace2718dded6772bb386b37fa3 *R/pscore.test.r 3133f13e7a7ff6dc277ff3bbf8321c79 *R/seg.Ar.fit.boot.r c66b36ae348b6fac6c3117ca41318f95 *R/seg.Ar.fit.r f127c92f7aaeaf603a70bccec59e5d02 *R/seg.control.R 388cbb0ef6faf923beb3b13df9c4098e *R/seg.def.fit.boot.r 33419bf83d6f7cf8540b7f93a7b59eb4 *R/seg.def.fit.r 92ed650a2dba26a38b3c51997f26a22c *R/seg.glm.fit.boot.r 4e13424552caf91acad5028bc8968365 *R/seg.glm.fit.r 195d6dd356b9bc87584210e07d877ce4 *R/seg.lm.fit.boot.r 93c700743d40f3ddbd0ceb10f83c4b2e *R/seg.lm.fit.r dc05717bcef9538f666068c5af1b64e5 *R/segmented.Arima.r 6784ac55ef0763e1f1923bab5a59835d *R/segmented.R 223899d10d8f492a9f2e54f90a2a83c8 *R/segmented.default.r 1a0fbd6d9fd490fb790da3806c68ba67 *R/segmented.glm.R 13b7af2baabab3d658e8bb924f0af600 *R/segmented.lm.R e4f667cabee25328e891045bb8786c82 *R/slope.R ce83cd92e7ae3b7783257bea4a6fa3cb *R/summary.segmented.R fa2b6ae145168ca63999617dd351bde1 *R/vcov.segmented.R c4a5d60070a20fc1472a4e0efc2e8f55 *data/down.rda 2eedec2d09709515e3e5891351381172 *data/plant.rda f7f474367bcb4ef9a00c80aba9f2596b *data/stagnant.rda 3158df5caf091ede33d58eed15bdb230 *inst/CITATION a890933c26d412e555795aff354c326c *man/broken.line.Rd e38b11eb6601e9544619dfac019d96a6 *man/confint.segmented.Rd 16464e04983759a33507e1bca4567b63 *man/davies.test.Rd a1fd6bbde564db5be2a9dde1ecedfb2d *man/down.Rd bb4e4f96d8eee8acb68b5773ab2a5c5a *man/draw.history.Rd 5b8fa6f1223e85dfe081169cf96b19a2 *man/intercept.Rd 7a543b5123d9c64b5a4da6aba4e1d3ea *man/lines.segmented.Rd 925cd1c6b4a05d3fca632a7f93999d00 *man/plant.Rd 138e297b73d53673fabd9025142aa524 *man/plot.segmented.Rd b4da8d785ea0e901bf670dada5dd9a56 *man/points.segmented.Rd 1ba7c089bae8411532d0f40ff4091c7b *man/predict.segmented.Rd c5ff81f292c40cdc317fe2ddfc99c4cd *man/print.segmented.Rd ea04af66300b8bf636d6a08f2ef573d9 *man/pscore.test.rd 1af889f836c98e6f631ecd058f675df6 *man/seg.control.Rd 55f09cc33b69788ffaaad8aad83faa3a *man/seg.lm.fit.Rd fb7088872dfa30aeff1a55f778e10cc4 *man/segmented-package.Rd 563ffd883ecca7937eec4f2c2277f0ab *man/segmented.Rd 33cff252508df1d97eec9ecf9ec1f12f *man/slope.Rd 1c8095f4472b4cabf3dc3cae5132260f *man/stagnant.Rd 706dec93d7feda2b0e6ef5f878529db2 *man/summary.segmented.Rd 99a8c632f6534f724d479cd4bca71824 *man/vcov.segmented.Rd segmented/DESCRIPTION0000644000176200001440000000126413160062250013731 0ustar liggesusersPackage: segmented Type: Package Title: Regression Models with Break-Points / Change-Points Estimation Version: 0.5-2.2 Date: 2017-09-18 Authors@R: c(person(given = c("Vito","M.","R."), family = "Muggeo", role = c("aut", "cre"), email = "vito.muggeo@unipa.it")) Author: Vito M. R. Muggeo [aut, cre] Maintainer: Vito M. R. Muggeo Description: Given a regression model, segmented `updates' the model by adding one or more segmented (i.e., piece-wise linear) relationships. Several variables with multiple breakpoints are allowed. License: GPL NeedsCompilation: no Packaged: 2017-09-18 11:10:10 UTC; enea Repository: CRAN Date/Publication: 2017-09-19 00:28:24 UTC segmented/man/0000755000176200001440000000000013157724622013011 5ustar liggesuserssegmented/man/stagnant.Rd0000644000176200001440000000161013157724622015115 0ustar liggesusers\name{stagnant} \alias{stagnant} \docType{data} \title{Stagnant band height data} \description{ The \code{stagnant} data frame has 28 rows and 2 columns. } \usage{data(stagnant)} \format{ A data frame with 28 observations on the following 2 variables. \describe{ \item{\code{x}}{log of flow rate in g/cm sec.} \item{\code{y}}{log of band height in cm} } } \details{ Bacon and Watts report that such data were obtained by R.A. Cook during his investigation of the behaviour of stagnant surface layer height in a controlled flow of water. } \source{ Bacon D.W., Watts D.G. (1971) Estimating the transistion between two intersecting straight lines. \emph{Biometrika} \bold{58}: 525 -- 534. Originally from the PhD thesis by R.A. Cook } %\references{ % PhD thesis by R.A. Cook %} \examples{ data(stagnant) ## plot(stagnant) } \keyword{datasets} segmented/man/predict.segmented.Rd0000644000176200001440000000440213157724622016704 0ustar liggesusers\name{predict.segmented} \alias{predict.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Predict method for segmented model fits } \description{ Returns predictions and optionally associated quantities (standard errors or confidence intervals) from a fitted segmented model object. } \usage{ \method{predict}{segmented}(object, newdata, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{object}{ a fitted segmented model coming from \code{segmented.lm} or \code{segmented.glm}. } \item{newdata}{ An optional data frame in which to look for variables with which to predict. If omitted, the fitted values are used. } \item{\dots}{ further arguments passed to \code{predict.lm} or \code{predict.glm}. Usually these are \code{se.fit}, or \code{interval} or \code{type}. } } \details{ Basically \code{predict.segmented} builds the right design matrix accounting for breakpoint and passes it to \code{predict.lm} or \code{predict.glm} depending on the actual model fit \code{object}. } \value{ \code{predict.segmented} produces a vector of predictions with possibly associated standard errors or confidence intervals. See \code{predict.lm} or \code{predict.glm}. } %\references{ %% ~put references to the literature/web site here ~ %} \author{ Vito Muggeo } \note{ If \code{type="terms"}, \code{predict.segmented} returns predictions for each component of the segmented term. Namely if `my.x' is the segmented variable, predictions for `my.x', `U1.my.x' and `psi1.my.x' are returned. These are meaningless individually, however their sum provides the predictions for the segmented term. } %% ~Make other sections like Warning with \section{Warning }{....} ~ \seealso{ \code{\link{plot.segmented}}, \code{\link{broken.line}}, \code{\link{predict.lm}}, \code{\link{predict.glm}} } \examples{ n=10 x=seq(-3,3,l=n) set.seed(1515) y <- (x<0)*x/2 + 1 + rnorm(x,sd=0.15) segm <- segmented(lm(y ~ x), ~ x, psi=0.5) predict(segm,se.fit = TRUE)$se.fit #wrong (smaller) st.errors (assuming known the breakpoint) olm<-lm(y~x+pmax(x-segm$psi[,2],0)) predict(olm,se.fit = TRUE)$se.fit } % \dontrun{..} % KEYWORDS - R documentation directory. \keyword{models} \keyword{regression} segmented/man/vcov.segmented.Rd0000644000176200001440000000306713157724622016235 0ustar liggesusers\name{vcov.segmented} \alias{vcov.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Variance-Covariance Matrix for a Fitted Segmented Model} \description{ Returns the variance-covariance matrix of the parameters (including breakpoints) of a fitted segmented model object.} \usage{ \method{vcov}{segmented}(object, var.diff = FALSE, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{object}{a fitted model object of class "segmented", returned by any \code{segmented} method.} \item{var.diff}{logical. If \code{var.diff=TRUE} and there is a single segmented variable, the covariance matrix is computed using a sandwich-type formula. See Details in \code{\link{summary.segmented}}.} \item{\dots}{additional arguments. } } \details{ The returned covariance matrix is based on an approximation of the nonlinear segmented term. Therefore covariances corresponding to breakpoints are reliable only in large samples and/or clear cut segmented relationships. } \value{ The full matrix of the estimated covariances between the parameter estimates, including the breakpoints. } %\references{} \author{Vito M. R. Muggeo, \email{vito.muggeo@unipa.it}} \note{\code{var.diff=TRUE} works when there is a single segmented variable.} \seealso{\code{\link{summary.segmented}}} \examples{ ##continues example from summary.segmented() # vcov(oseg) # vcov(oseg,var.diff=TRUE) } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{regression} segmented/man/points.segmented.Rd0000644000176200001440000000432713157724622016574 0ustar liggesusers\name{points.segmented} \alias{points.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Points method for segmented objects } \description{ Takes a fitted \code{segmented} object returned by \code{segmented()} and adds on the current plot the joinpoints of the fitted broken-line relationships. } \usage{ \method{points}{segmented}(x, term, interc = TRUE, link = TRUE, rev.sgn=FALSE, transf=I, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{x}{ an object of class \code{segmented}. } \item{term}{ the segmented variable of interest. It may be unspecified when there is a single segmented variable. } \item{interc}{ If \code{TRUE} the computed joinpoints include the model intercept (if it exists). } \item{link}{ when \code{TRUE} (default), the fitted joinpoints are plotted on the link scale } \item{rev.sgn}{ when \code{TRUE}, the fitted joinpoints are plotted on the `minus' scale of the current \code{term} variable. This is useful when a null-constraint has been set on the last slope. } \item{transf}{ A possible function to convert the fitted values before plotting. } \item{\dots}{ other graphics parameters to pass on to \code{points()} function. } } \details{ We call 'joinpoint' the plane point having as coordinates the breakpoint (on the x scale) and the fitted value of the segmented relationship at that breakpoint (on the y scale). \code{points.segmented()} simply adds the fitted joinpoints on the current plot. This could be useful to emphasize the changes of the piecewise linear relationship. } %\value{ %% ~Describe the value returned %% If it is a LIST, use %% \item{comp1 }{Description of 'comp1'} %% \item{comp2 }{Description of 'comp2'} %% ... %} %\references{ %% ~put references to the literature/web site here ~ %} %\author{ %% ~~who you are~~ %} %\note{ %% ~~further notes~~ %} %% ~Make other sections like Warning with \section{Warning }{....} ~ \seealso{ \code{\link{plot.segmented}} to plot the fitted segmented lines. } \examples{ \dontrun{ #see examples in ?plot.segmented } } \keyword{ nonlinear } \keyword{ regression }% __ONLY ONE__ keyword per line segmented/man/plot.segmented.Rd0000644000176200001440000001326313157724622016235 0ustar liggesusers\name{plot.segmented} \alias{plot.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Plot method for segmented objects } \description{ Takes a fitted \code{segmented} object returned by \code{segmented()} and plots (or adds) the fitted broken-line for the selected segmented term. } \usage{ \method{plot}{segmented}(x, term, add=FALSE, res=FALSE, conf.level=0, interc=TRUE, link=TRUE, res.col=1, rev.sgn=FALSE, const=0, shade=FALSE, rug=TRUE, dens.rug=FALSE, dens.col = grey(0.8), transf=I, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{x}{ a fitted \code{segmented} object. } \item{term}{ the segmented variable having the piece-wise relationship to be plotted. If there is a single segmented variable in the fitted model \code{x}, \code{term} can be omitted.} \item{add}{ when \code{TRUE} the fitted lines are added to the current device.} \item{res}{ when \code{TRUE} the fitted lines are plotted along with corresponding partial residuals. See Details.} \item{conf.level}{ If greater than zero, it means the confidence level at which the pointwise confidence itervals have to be plotted.} \item{interc}{ If \code{TRUE} the computed segmented components include the model intercept (if it exists).} \item{link}{ when \code{TRUE} (default), the fitted lines are plotted on the link scale, otherwise they are tranformed on the response scale before plotting. Ignored for linear segmented fits. } \item{res.col}{when \code{res=TRUE} it means the color of the points representing the partial residuals.} \item{rev.sgn}{ when \code{TRUE} it is assumed that current \code{term} is `minus' the actual segmented variable, therefore the sign is reversed before plotting. This is useful when a null-constraint has been set on the last slope.} \item{const}{ constant to add to each fitted segmented relationship (on the scale of the linear predictor) before plotting.} \item{shade}{if \code{TRUE} and \code{conf.level>0} it produces shaded regions (in grey color) for the pointwise confidence intervals embracing the fitted segmented line. } \item{rug}{when \code{TRUE} (default) then the covariate values are displayed as a rug plot at the foot of the plot.} \item{dens.rug}{when \code{TRUE} then smooth covariate distribution is plotted on the x-axis.} \item{dens.col}{if \code{dens.rug=TRUE}, it means the colour to be used to plot the density.} % \item{show.gap}{ when \code{FALSE} the (possible) gaps between the fitted lines at the estimated breakpoints % are hidden. When bootstrap restarting has been employed (default in \code{segmented}), \code{show.gap} is meaningless % as the gap coefficients are always set to zero in the fitted model.} \item{transf}{ A possible function to convert the fitted values before plotting. It is only effective if the fitted values refer to a linear or a generalized linear model (on the link scale) \emph{and} \code{res=FALSE}.} \item{\dots}{ other graphics parameters to pass to plotting commands: `col', `lwd' and `lty' (that can be vectors, see the example below) for the fitted piecewise lines; `ylab', `xlab', `main', `sub', `xlim' and `ylim' when a new plot is produced (i.e. when \code{add=FALSE}); `pch' and `cex' for the partial residuals (when \code{res=TRUE}). } } \details{ Produces (or adds to the current device) the fitted segmented relationship between the response and the selected \code{term}. If the fitted model includes just a single `segmented' variable, \code{term} may be omitted. %Due to the parameterization of the segmented terms, sometimes %the fitted lines may not appear to join at the estimated breakpoints. If this is the case, the apparent %`gap' would indicate some lack-of-fit. However, since version 0.2-9.0, the gap coefficients are set to zero by default %(see argument \code{gap} in in \code{\link{seg.control}}). The partial residuals are computed as `fitted + residuals', where `fitted' are the fitted values of the segmented relationship relevant to the covariate specified in \code{term}. Notice that for GLMs the residuals are the response residuals if \code{link=FALSE} and the working residuals if \code{link=TRUE}. %weighted by the IWLS weights [fino alla versione 0.5-2.0 i workRes were weighted by the IWLS weights] } \value{ None. } %\references{ } \author{ Vito M. R. Muggeo } \note{ For models with offset, partial residuals on the response scale are not defined. Thus \code{plot.segmented} does not work when \code{link=FALSE}, \code{res=TRUE}, and the fitted model includes an offset.} % % ~Make other sections like Warning with \section{Warning }{....} ~ %} \seealso{ \code{\link{lines.segmented}} to add the estimated breakpoints on the current plot. \code{\link{points.segmented}} to add the joinpoints of the segmented relationship. \code{\link{predict.segmented}} to compute standard errors and confidence intervals for predictions from a "segmented" fit. } \examples{ set.seed(1234) z<-runif(100) y<-rpois(100,exp(2+1.8*pmax(z-.6,0))) o<-glm(y~z,family=poisson) o.seg<-segmented(o, ~z) #single segmented covariate and one breakpoint:'psi' can be omitted par(mfrow=c(2,1)) plot(o.seg, conf.level=0.95, shade=TRUE) points(o.seg, link=FALSE, col=2) ## new plot plot(z,y) ## add the fitted lines using different colors and styles.. plot(o.seg,add=TRUE,link=FALSE,lwd=2,col=2:3, lty=c(1,3)) lines(o.seg,col=2,pch=19,bottom=FALSE,lwd=2) points(o.seg,col=4, link=FALSE) } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } \keyword{ nonlinear } \keyword{ hplot }segmented/man/draw.history.Rd0000644000176200001440000000322313157724622015735 0ustar liggesusers\name{draw.history} \alias{draw.history} %- Also NEED an '\alias' for EACH other topic documented here. \title{ History for the breakpoint estimates } \description{ Displays breakpoint iteration values for segmented fits. } \usage{ draw.history(obj, term, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{obj}{ a segmented fit returned by any "segmented" method. } \item{term}{ a character to mean the `segmented' variable whose breakpoint values throughout iterations have to be displayed. } \item{\dots}{ graphic parameters to be passed to \code{matplot()}. } } \details{ For a given \code{term} in a segmented fit, \code{draw.history()} displays the different breakpoint values obtained during the estimating process, since the starting values up to the final ones. When bootstrap restarting is employed, \code{draw.history()} produces two plots, the values of objective function and the number of distinct solutions against the bootstrap replicates. } \value{ None. } %\references{ } \author{ Vito M.R. Muggeo } %\note{ ~~further notes~~ % ~Make other sections like Warning with \section{Warning }{....} ~ %} %\seealso{ ~~objects to See Also as \code{\link{help}}, ~~~ } \examples{ data(stagnant) os<-segmented(lm(y~x,data=stagnant),seg.Z=~x,psi=-.8) draw.history(os) #diagnostics with boot restarting os<-segmented(lm(y~x,data=stagnant),seg.Z=~x,psi=-.8, control=seg.control(n.boot=0)) draw.history(os) #diagnostics without boot restarting } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } \keyword{ nonlinear } segmented/man/down.Rd0000644000176200001440000000203513157724622014247 0ustar liggesusers\name{down} \alias{down} \docType{data} \title{ Down syndrome in babies} \description{ The \code{down} data frame has 30 rows and 3 columns. Variable \code{cases} means the number of babies with Down syndrome out of total number of births \code{births} for mothers with mean age \code{age}. } \usage{data(down)} \format{ A data frame with 30 observations on the following 3 variables. \describe{ \item{\code{age}}{the mothers' mean age.} \item{\code{births}}{count of total births.} \item{\code{cases}}{count of babies with Down syndrome.} } } %\details{ % ~~ If necessary, more details than the description above ~~ %} \source{ Davison, A.C. and Hinkley, D. V. (1997) \emph{Bootstrap Methods and their Application}. Cambridge University Press. } \references{ Geyer, C. J. (1991) Constrained maximum likelihood exemplified by isotonic convex logistic regression. \emph{Journal of the American Statistical Association} \bold{86}, 717--724. } \examples{ data(down) } \keyword{datasets} segmented/man/seg.control.Rd0000644000176200001440000001774213157724622015550 0ustar liggesusers\name{seg.control} \alias{seg.control} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Auxiliary for controlling segmented model fitting } \description{ Auxiliary function as user interface for 'segmented' fitting. Typically only used when calling any 'segmented' method (\code{segmented.lm} or \code{segmented.glm}). } \usage{ seg.control(toll = 1e-04, it.max = 10, display = FALSE, stop.if.error = TRUE, K = 10, quant = FALSE, last = TRUE, maxit.glm = 25, h = 1, n.boot=20, size.boot=NULL, gap=FALSE, jt=FALSE, nonParam=TRUE, random=TRUE, powers=c(1,1), seed=NULL, fn.obj=NULL, digits=NULL) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{toll}{ positive convergence tolerance. } \item{it.max}{ integer giving the maximal number of iterations. } \item{display}{ logical indicating if the value of the \emph{working} objective function should be printed at each iteration. The \emph{working} objective function is the objective function of the working model including the gap coefficients (and therefore it should not be compared with the value at convergence). If bootstrap restarting is employed, the value of the \emph{real} objective function (without gap coefficients) after every bootstrap iteration is printed. This value should decrease throughout the iterations.} \item{stop.if.error}{ logical indicating if non-admissible break-points should be removed during the estimating algorithm. Set it to \code{FALSE} if you want to perform a sort of `automatic' breakpoint selection, provided that several starting values are provided for the breakpoints. See argument \code{psi} in \code{\link{segmented.lm}} or \code{\link{segmented.glm}}. The idea of removing `non-admissible' break-points during the iterative process is discussed in Muggeo and Adelfio (2011) and it is not compatible with the bootstrap restart algorithm. This approach, indeed, should be considered as a preliminary and tentative approach to deal with an unknown number of breakpoints. } \item{K}{ the number of quantiles (or equally-spaced values) to supply as starting values for the breakpoints when the \code{psi} argument of \code{segmented} is set to \code{NA}. \code{K} is ignored when \code{psi} is different from \code{NA}. } \item{quant}{logical, indicating how the starting values should be selected. If \code{FALSE} equally-spaced values are used, otherwise the quantiles. Ignored when \code{psi} is different from \code{NA}.} \item{last}{ logical indicating if output should include only the last fitted model.} \item{maxit.glm}{ integer giving the maximum number of inner IWLS iterations (see details). } \item{h}{ positive factor (from zero to one) modifying the increments in breakpoint updates during the estimation process (see details). } \item{n.boot}{ number of bootstrap samples used in the bootstrap restarting algorithm. If 0 the standard algorithm, i.e. without bootstrap restart, is used. Default to 20 that appears to be sufficient in most of problems. However when multiple breakpoints have to be estimated it is suggested to increase \code{n.boot}, e.g. \code{n.boot=50}.} \item{size.boot}{the size of the bootstrap samples. If \code{NULL}, it is taken equal to the actual sample size.} \item{gap}{logical, if \code{FALSE} the gap coefficients are \emph{always} constrained to zero at the convergence.} \item{jt}{logical. If \code{TRUE} the values of the segmented variable(s) are jittered before fitting the model to the bootstrap resamples.} \item{nonParam}{ if \code{TRUE} nonparametric bootstrap (i.e. case-resampling) is used, otherwise residual-based. Currently working only for LM fits. It is not clear what residuals should be used for GLMs.} \item{random}{ if \code{TRUE}, when the algorithm fails to obtain a solution, random values are employed to obtain candidate values. } \item{powers}{ The powers of the pseudo covariates employed by the algorithm. These are possibly altered during the iterative process to stabilize the estimation procedure. Usually of no interest for the user. } \item{seed}{ The seed to be passed on to \code{set.seed()} when \code{n.boot>0}. Setting the seed can be useful to replicate the results when the bootstrap restart algorithm is employed. In fact a segmented fit includes \code{seed} representing the integer vector saved just before the bootstrap resampling. Re-use it if you want to replicate the bootstrap restarting algorithm with the \emph{same} samples. } \item{fn.obj}{ A \emph{character string} to be used (optionally) only when \code{segmented.default} is used. It represents the function (with argument \code{'x'}) to be applied to the fit object to extract the objective function to be \emph{minimized}. Thus for \code{"lm"} fits (although unnecessary) it should be \code{fn.obj="sum(x$residuals^2)"}, for \code{"coxph"} fits it should be \code{fn.obj="-x$loglik[2]"}. If \code{NULL} the `minus log likelihood' extracted from the object, namely \code{"-logLik(x)"}, is used. See \code{\link{segmented.default}}. } \item{digits}{optional. If specified it means the desidered number of decimal points of the breakpoint to be used during the iterative algorithm.} } \details{ Fitting a `segmented' GLM model is attained via fitting iteratively standard GLMs. The number of (outer) iterations is governed by \code{it.max}, while the (maximum) number of (inner) iterations to fit the GLM at each fixed value of psi is fixed via \code{maxit.glm}. Usually three-four inner iterations may be sufficient. When the starting value for the breakpoints is set to \code{NA} for any segmented variable specified in \code{seg.Z}, \code{K} values (quantiles or equally-spaced) are selected as starting values for the breakpoints. In this case, it may be useful to set also \code{stop.if.error=FALSE} to automate the procedure, see Muggeo and Adelfio (2011). The maximum number of iterations (\code{it.max}) should be also increased when the `automatic' procedure is used. If \code{last=TRUE}, the object resulting from \code{segmented.lm} (or \code{segmented.glm}) is a list of fitted GLM; the i-th model is the segmented model with the values of the breakpoints at the i-th iteration. Sometimes to stabilize the procedure, it can be useful to set \code{h<1} to reduce the increments in the breakpoint updates. At each iteration the updated estimate is usually given by \code{psi.new=psi.old+increm}. By setting \code{h<1} (actually \code{min(abs(h),1)} is considered) causes the following updates of the breakpoint estimate: \code{psi.new=psi.old+h*increm}. Since version 0.2-9.0 \code{segmented} implements the bootstrap restarting algorithm described in Wood (2001). The bootstrap restarting is expected to escape the local optima of the objective function when the segmented relationship is flat. Notice bootstrap restart runs \code{n.boot} iterations regardless of \code{toll} that only affects convergence within the inner loop. } \value{ A list with the arguments as components. } \references{ Muggeo, V.M.R., Adelfio, G. (2011) Efficient change point detection in genomic sequences of continuous measurements. \emph{Bioinformatics} \bold{27}, 161--166. Wood, S. N. (2001) Minimizing model fitting objectives that contain spurious local minima by bootstrap restarting. \emph{Biometrics} \bold{57}, 240--244. } \author{ Vito Muggeo } %\note{ ~~further notes~~ % ~Make other sections like Warning with \section{Warning }{....} ~ %} %\seealso{ ~~objects to See Also as \code{\link{help}}, ~~~ } \examples{ #decrease the maximum number inner iterations and display the #evolution of the (outer) iterations seg.control(display = TRUE, maxit.glm=4) } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } segmented/man/confint.segmented.Rd0000644000176200001440000000535113157724622016716 0ustar liggesusers\name{confint.segmented} \alias{confint.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Confidence intervals for breakpoints} \description{ Computes confidence intervals for the breakpoints in a fitted `segmented' model. } \usage{ \method{confint}{segmented}(object, parm, level=0.95, rev.sgn=FALSE, var.diff=FALSE, digits=max(3, getOption("digits") - 3), ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{object}{a fitted \code{segmented} object. } \item{parm}{the segmented variable of interest. If missing all the segmented variables are considered. } \item{level}{the confidence level required (default to 0.95).} \item{rev.sgn}{vector of logicals. The length should be equal to the length of \code{parm}; recycled otherwise. when \code{TRUE} it is assumed that the current \code{parm} is `minus' the actual segmented variable, therefore the sign is reversed before printing. This is useful when a null-constraint has been set on the last slope.} \item{var.diff}{logical. If \code{var.diff=TRUE} and there is a single segmented variable, the standard error is based on sandwich-type formula of the covariance matrix. See Details in \code{\link{summary.segmented}}.} \item{digits}{controls the number of digits to print when printing the output. } \item{\dots}{additional parameters } } \details{ Currently \code{confint.segmented} computes confidence limits for the breakpoints using the standard error coming from the Delta method for the ratio of two random variables. This value is an approximation (slightly) better than the one reported in the `psi' component of the list returned by any \code{segmented} method. The resulting confidence intervals are based on the asymptotic Normal distribution of the breakpoint estimator which is reliable just for clear-cut kink relationships. See Details in \code{\link{segmented}}. } \value{ A list of matrices. Each matrix includes point estimate and confidence limits of the breakpoint(s) for each segmented variable in the model. } %\references{ } \author{ Vito M.R. Muggeo } %\note{ ~~further notes~~ % % ~Make other sections like Warning with \section{Warning }{....} ~ %} \seealso{ \code{\link{segmented}} and \code{\link{lines.segmented}} to plot the estimated breakpoints with corresponding confidence intervals. } \examples{ set.seed(10) x<-1:100 z<-runif(100) y<-2+1.5*pmax(x-35,0)-1.5*pmax(x-70,0)+10*pmax(z-.5,0)+rnorm(100,0,2) out.lm<-lm(y~x) o<-segmented(out.lm,seg.Z=~x+z,psi=list(x=c(30,60),z=.4)) confint(o) } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } \keyword{ nonlinear } segmented/man/segmented.Rd0000644000176200001440000002374113157724622015262 0ustar liggesusers\name{segmented} \alias{segmented} \alias{segmented.lm} \alias{segmented.glm} \alias{segmented.default} \alias{segmented.Arima} %\alias{print.segmented} %\alias{summary.segmented} %\alias{print.summary.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Segmented relationships in regression models } \description{ Fits regression models with segmented relationships between the response and one or more explanatory variables. Break-point estimates are provided. } \usage{ segmented(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) \method{segmented}{default}(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) \method{segmented}{lm}(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) \method{segmented}{glm}(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) \method{segmented}{Arima}(obj, seg.Z, psi, control = seg.control(), model = TRUE, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{obj}{ standard `linear' model of class "lm" or "glm". Since version 0.5.0-0 any regression fit may be supplied (see 'Details').} \item{seg.Z}{ a formula with no response variable, such as \code{seg.Z=~x1+x2}, indicating the (continuous) explanatory variables having segmented relationships with the response. Currently, formulas involving functions, such as \code{seg.Z=~log(x1)} or even \code{seg.Z=~sqrt(x1)}, or selection operators, such as \code{seg.Z=~d[,"x1"]} or \code{seg.Z=~d$x1}, are \emph{not} allowed. It can be missing when \code{obj} ("lm" or "glm" fit) includes only one covariate which is taken as segmented variable. } \item{psi}{ named list of vectors. The names have to match the variables in the \code{seg.Z} argument. Each vector includes starting values for the break-point(s) for the corresponding variable in \code{seg.Z}. If \code{seg.Z} includes only a variable, \code{psi} may be a numeric vector or even missing (and the median of the segmented variable is used as a starting value). A \code{NA} value means that `\code{K}' quantiles (or equally spaced values) are used as starting values; \code{K} is fixed via the \code{\link{seg.control}} auxiliary function. } \item{control}{ a list of parameters for controlling the fitting process. See the documentation for \code{\link{seg.control}} for details. } \item{model}{ logical value indicating if the model.frame should be returned.} \item{\dots}{ optional arguments. } } \details{ Given a linear regression model (usually of class "lm" or "glm"), segmented tries to estimate a new model having broken-line relationships with the variables specified in \code{seg.Z}. A segmented (or broken-line) relationship is defined by the slope parameters and the break-points where the linear relation changes. The number of breakpoints of each segmented relationship is fixed via the \code{psi} argument, where initial values for the break-points must be specified. The model is estimated simultaneously yielding point estimates and relevant approximate standard errors of all the model parameters, including the break-points. Since version 0.2-9.0 \code{segmented} implements the bootstrap restarting algorithm described in Wood (2001). The bootstrap restarting is expected to escape the local optima of the objective function when the segmented relationship is flat and the log likelihood can have multiple local optima. Since version 0.5-0.0 the default method \code{segmented.default} has been added to estimate segmented relationships in general (besides "lm" and "glm" fits) regression models, such as Cox regression or quantile regression (for a single percentile). The objective function to be minimized is the (minus) value extracted by the \code{logLik} function or it may be passed on via the \code{fn.obj} argument in \code{seg.control}. See example below. While the default method is expected to work with any regression fit (where the usual \code{coef()}, \code{update()}, and \code{logLik()} returns appropriate results), it is not recommended for "lm" or "glm" fits (as \code{segmented.default} is slower than the specific methods \code{segmented.lm} and \code{segmented.glm}), although final results are the same. However the object returned by \code{segmented.default} is \emph{not} of class "segmented", as currently the segmented methods are not guaranteed to work for `generic' (i.e., besides "lm" and "glm") regression fits. The user could try each "segmented" method on the returned object by calling it explicitly (e.g. via \code{plot.segmented()} or \code{confint.segmented()}). } \value{ The returned object depends on the \code{last} component returned by \code{seg.control}. If last=TRUE, the default, segmented returns an object of class "segmented" which inherits from the class "lm" or "glm" depending on the class of \code{obj}. Otherwise a list is returned, where the last component is the fitted model at the final iteration, see \code{\link{seg.control}}. \cr An object of class "segmented" is a list containing the components of the original object \code{obj} with additionally the followings: \item{psi}{estimated break-points and relevant (approximate) standard errors} \item{it}{number of iterations employed} \item{epsilon}{difference in the objective function when the algorithm stops} \item{model}{the model frame} \item{psi.history}{a list or a vector including the breakpoint estimates at each step} \item{seed}{the integer vector containing the seed just before the bootstrap resampling. Returned only if bootstrap restart is employed} \item{..}{Other components are not of direct interest of the user} } \references{ Muggeo, V.M.R. (2003) Estimating regression models with unknown break-points. \emph{Statistics in Medicine} \bold{22}, 3055--3071. Muggeo, V.M.R. (2008) Segmented: an R package to fit regression models with broken-line relationships. \emph{R News} \bold{8/1}, 20--25. } \author{ Vito M. R. Muggeo, \email{vito.muggeo@unipa.it} } \note{ \enumerate{ \item The algorithm will start if the \code{it.max} argument returned by \code{seg.control} is greater than zero. If \code{it.max=0} \code{segmented} will estimate a new linear model with break-point(s) fixed at the values reported in \code{psi}. \item In the returned fit object, `U.' is put before the name of the segmented variable to mean the difference-in-slopes coefficient. \item Methods specific to the class \code{"segmented"} are \itemize{ \item \code{print.segmented} \item \code{summary.segmented} \item \code{print.summary.segmented} \item \code{plot.segmented} \item \code{lines.segmented} \item \code{confint.segmented} \item \code{vcov.segmented} \item \code{predict.segmented} \item \code{points.segmented} } Others are inherited from the class \code{"lm"} or \code{"glm"} depending on the class of \code{obj}. } } \section{ Warning }{It is well-known that the log-likelihood function for the break-point may be not concave, especially for poor clear-cut kink-relationships. In these circumstances the initial guess for the break-point, i.e. the \code{psi} argument, must be provided with care. For instance visual inspection of a, possibly smoothed, scatter-plot is usually a good way to obtain some idea on breakpoint location. However bootstrap restarting, implemented since version 0.2-9.0, is relatively more robust to starting values specified in \code{psi}. Alternatively an automatic procedure may be implemented by specifying \code{psi=NA} and \code{stop.if.error=FALSE} in \code{\link{seg.control}}: experience suggests to increase the number of iterations via \code{it.max} in \code{seg.control()}. This automatic procedure, however, is expected to overestimate the number of breakpoints. } \seealso{ \code{\link{lm}}, \code{\link{glm}} } \examples{ set.seed(12) xx<-1:100 zz<-runif(100) yy<-2+1.5*pmax(xx-35,0)-1.5*pmax(xx-70,0)+15*pmax(zz-.5,0)+rnorm(100,0,2) dati<-data.frame(x=xx,y=yy,z=zz) out.lm<-lm(y~x,data=dati) #simple example: 1 segmented variable, 1 breakpoint: you do not need to specify # the starting value for psi o<-segmented(out.lm,seg.Z=~z) #1 segmented variable, 2 breakpoints: you have to specify starting values (vector) for psi: o<-segmented(out.lm,seg.Z=~x,psi=c(30,60), control=seg.control(display=FALSE)) slope(o) #2 segmented variables: starting values requested via a named list out.lm<-lm(y~z,data=dati) o1<-update(o,seg.Z=~x+z,psi=list(x=c(30,60),z=.3)) #the default method leads to the same results (but it is slower) #o1<-segmented.default(out.lm,seg.Z=~x+z,psi=list(x=c(30,60),z=.3)) #o1<-segmented.default(out.lm,seg.Z=~x+z,psi=list(x=c(30,60),z=.3), # control=seg.control(fn.obj="sum(x$residuals^2)")) #automatic procedure to estimate breakpoints in the covariate x # Notice: bootstrap restart is not allowed! o<-segmented.lm(out.lm,seg.Z=~x+z,psi=list(x=NA,z=.3), control=seg.control(stop.if.error=FALSE,n.boot=0, it.max=20)) #assess the progress of the breakpoint estimates throughout the iterations \dontrun{ par(mfrow=c(2,1)) draw.history(o, "x") draw.history(o, "z") } #try to increase the number of iterations and re-assess the #convergence diagnostics #An example using the default method: # Cox regression with a segmented relationship \dontrun{ library(survival) data(stanford2) o<-coxph(Surv(time, status)~age, data=stanford2) os<-segmented(o, ~age, psi=40) #estimate the breakpoint in the age effect summary(os) #actually it means summary.coxph(os) plot(os) #it does not work plot.segmented(os) #call explicitly plot.segmented() to plot the fitted piecewise lines } } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{regression} \keyword{nonlinear } segmented/man/intercept.Rd0000644000176200001440000000501713157724622015300 0ustar liggesusers\name{intercept} \alias{intercept} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Intercept estimates from segmented relationships } \description{ Computes the intercepts of each `segmented' relationship in the fitted model. } \usage{ intercept(ogg, parm, rev.sgn = FALSE, var.diff=FALSE, digits = max(3, getOption("digits") - 3)) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{ogg}{ an object of class "segmented", returned by any \code{segmented} method. } \item{parm}{ the segmented variable whose intercepts have to be computed. If missing all the segmented variables in the model are considered. } % \item{gap}{ % logical. should the intercepts account for the (possible) gaps? %} \item{rev.sgn}{vector of logicals. The length should be equal to the length of \code{parm}, but it is recycled otherwise. When \code{TRUE} it is assumed that the current \code{parm} is `minus' the actual segmented variable, therefore the order is reversed before printing. This is useful when a null-constraint has been set on the last slope. } \item{var.diff}{Currently ignored as only point estimates are computed. %logical. If \code{var.diff=TRUE} and there is a single segmented variable, the computed standard errors % are based on a sandwich-type formula of the covariance matrix. See Details in \code{\link{summary.segmented}}. } \item{digits}{controls number of digits in output.} } \details{ A broken-line relationship means that a regression equation exists in the intervals `\eqn{min(x)}{min(x)} to \eqn{\psi_1}{psi1}', `\eqn{\psi_1}{psi1} to \eqn{\psi_2}{psi2}', and so on. \code{intercept} computes point estimates of the intercepts of the different regression equations for each segmented relationship in the fitted model. } \value{ \code{intercept} returns a list of one-column matrices. Each matrix represents a segmented relationship. } %\references{ %% ~put references to the literature/web site here ~ %} \author{Vito M. R. Muggeo, \email{vito.muggeo@unipa.it}} %\note{ %% ~~further notes~~ %} %% ~Make other sections like Warning with \section{Warning }{....} ~ \seealso{ See also \code{\link{slope}} to compute the slopes of the different regression equations for each segmented relationship in the fitted model. } \examples{ ## see ?slope \dontrun{ intercept(out.seg) } } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } segmented/man/segmented-package.Rd0000644000176200001440000000641513157724622016652 0ustar liggesusers\name{segmented-package} \alias{segmented-package} %\alias{segmented} \docType{package} \title{ Segmented relationships in regression models with breakpoints / changepoints estimation } \description{ Estimation and Inference of Regression Models with piecewise linear relationships having a fixed number of break-points. } \details{ \tabular{ll}{ Package: \tab segmented\cr Type: \tab Package\cr Version: \tab 0.5-2.2\cr Date: \tab 2017-09-18\cr License: \tab GPL\cr } Package \code{segmented} is aimed to estimate linear and generalized linear models (and virtually any regression model) having one or more segmented relationships in the linear predictor. Estimates of the slopes and breakpoints are provided along with standard errors. The package includes testing/estimating functions and methods to print, summarize and plot the results. \cr The algorithm used by \code{segmented} is \emph{not} grid-search. It is an iterative procedure (Muggeo, 2003) that needs starting values \emph{only} for the breakpoint parameters and therefore it is quite efficient even with several breakpoints to be estimated. Moreover since version 0.2-9.0, \code{segmented} implements the bootstrap restarting (Wood, 2001) to make the algorithm less sensitive to starting values. \cr Since version 0.5-0.0 a default method \code{segmented.dafault} has been added. It may be employed to include segmented relationships in \emph{general} regression models where specific methods do not exist. Examples include quantile and Cox regressions. See examples in \code{\link{segmented.default}}.\cr A tentative approach to deal with unknown number of breakpoints is also provided, see option \code{stop.if.error} in \code{\link{seg.control}}. } \author{ Vito M.R. Muggeo } \references{ Muggeo, V.M.R. (2016) Testing with a nuisance parameter present only under the alternative: a score-based approach with application to segmented modelling. \emph{J of Statistical Computation and Simulation} \bold{86}, 3059--3067. Davies, R.B. (1987) Hypothesis testing when a nuisance parameter is present only under the alternative. \emph{Biometrika} \bold{74}, 33--43. Seber, G.A.F. and Wild, C.J. (1989) \emph{Nonlinear Regression}. Wiley, New York. Bacon D.W., Watts D.G. (1971) Estimating the transistion between two intersecting straight lines. \emph{Biometrika} \bold{58}: 525 -- 534. Muggeo, V.M.R. (2003) Estimating regression models with unknown break-points. \emph{Statistics in Medicine} \bold{22}, 3055--3071. Muggeo, V.M.R. (2008) Segmented: an R package to fit regression models with broken-line relationships. \emph{R News} \bold{8/1}, 20--25. Muggeo, V.M.R., Adelfio, G. (2011) Efficient change point detection in genomic sequences of continuous measurements. \emph{Bioinformatics} \bold{27}, 161--166. Wood, S. N. (2001) Minimizing model fitting objectives that contain spurious local minima by bootstrap restarting. \emph{Biometrics} \bold{57}, 240--244. Muggeo, V.M.R. (2010) Comment on `Estimating average annual per cent change in trend analysis' by Clegg et al., Statistics in Medicine; 28, 3670-3682. \emph{Statistics in Medicine}, \bold{29}, 1958--1960. } \keyword{ regression } \keyword{ nonlinear } segmented/man/print.segmented.Rd0000644000176200001440000000127013157724622016406 0ustar liggesusers\name{print.segmented} \alias{print.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Print method for the segmented class } \description{ Printing the most important feautures of a segmented model. } \usage{ \method{print}{segmented}(x, digits = max(3, getOption("digits") - 3), ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{x}{ object of class \code{segmented} } \item{digits}{ number of digits to be printed } \item{\dots}{ arguments passed to other functions } } \author{ Vito M.R. Muggeo } \seealso{ \code{\link{summary.segmented}}, \code{\link{print.summary.segmented}} } \keyword{ models } segmented/man/plant.Rd0000644000176200001440000000226613157724622014424 0ustar liggesusers\name{plant} \alias{plant} \docType{data} \title{ Plan organ dataset} \description{ The \code{plant} data frame has 103 rows and 3 columns. } \usage{data(plant)} \format{ A data frame with 103 observations on the following 3 variables: \describe{ \item{\code{y}}{measurements of the plant organ.} \item{\code{time}}{times where measurements took place.} \item{\code{group}}{three attributes of the plant organ, \code{RKV}, \code{RKW}, \code{RWC}.} } } \details{ Three attributes of a plant organ measured over time where biological reasoning indicates likelihood of multiple breakpoints. The data are scaled to the maximum value for each attribute and all attributes are measured at each time. } \source{ The data have been kindly provided by Dr Zongjian Yang at School of Land, Crop and Food Sciences, The University of Queensland, Brisbane, Australia. } %\references{ % ~~ possibly secondary sources and usages ~~ %} \examples{ \dontrun{ data(plant) attach(plant) %lattice::xyplot(y~time,groups=group,pch=19,col=2:4,auto.key=list(space="right")) lattice::xyplot(y~time,groups=group,auto.key=list(space="right")) } } \keyword{datasets} segmented/man/seg.lm.fit.Rd0000644000176200001440000001016313157724622015247 0ustar liggesusers\name{seg.lm.fit} \alias{seg.lm.fit} \alias{seg.glm.fit} \alias{seg.def.fit} \alias{seg.Ar.fit} \alias{seg.lm.fit.boot} \alias{seg.glm.fit.boot} \alias{seg.def.fit.boot} \alias{seg.Ar.fit.boot} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Fitter Functions for Segmented Linear Models } \description{ \code{seg.lm.fit} is called by \code{segmented.lm} to fit segmented linear (gaussian) models. Likewise, \code{seg.glm.fit} is called by \code{segmented.glm} to fit generalized segmented linear models, and \code{seg.def.fit} is called by \code{segmented.default} to fit segmented relationships in general regression models (e.g., quantile regression and Cox regression). \code{seg.lm.fit.boot}, \code{seg.glm.fit.boot}, and \code{seg.def.fit.boot} are employed to perform bootstrap restart. These functions should usually not be used directly by the user. } \usage{ seg.lm.fit(y, XREG, Z, PSI, w, offs, opz, return.all.sol=FALSE) seg.lm.fit.boot(y, XREG, Z, PSI, w, offs, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE) seg.glm.fit(y, XREG, Z, PSI, w, offs, opz, return.all.sol=FALSE) seg.glm.fit.boot(y, XREG, Z, PSI, w, offs, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE) seg.def.fit(obj, Z, PSI, mfExt, opz, return.all.sol=FALSE) seg.def.fit.boot(obj, Z, PSI, mfExt, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE) seg.Ar.fit(obj, XREG, Z, PSI, opz, return.all.sol=FALSE) seg.Ar.fit.boot(obj, XREG, Z, PSI, opz, n.boot=10, size.boot=NULL, jt=FALSE, nonParam=TRUE, random=FALSE) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{y}{ vector of observations of length \code{n}. } \item{XREG}{ design matrix for standard linear terms. } \item{Z}{ appropriate matrix including the segmented variables whose breakpoints have to be estimated. } \item{PSI}{ appropriate matrix including the starting values of the breakpoints to be estimated. } \item{w}{ possibe weights vector. } \item{offs}{ possibe offset vector. } \item{opz}{ a list including information useful for model fitting. } \item{n.boot}{ the number of bootstrap samples employed in the bootstrap restart algorithm. } \item{size.boot}{ the size of the bootstrap resamples. If \code{NULL} (default), it is taken equal to the sample size. values smaller than the sample size are expected to increase perturbation in the bootstrap resamples. } \item{jt}{ logical. If \code{TRUE} the values of the segmented variable(s) are jittered before fitting the model to the bootstrap resamples. } \item{nonParam}{ if \code{TRUE} nonparametric bootstrap (i.e. case-resampling) is used, otherwise residual-based. } \item{random}{ if \code{TRUE}, when the algorithm fails to obtain a solution, random values are used as candidate values. } \item{return.all.sol}{ if \code{TRUE}, when the algorithm fails to obtain a solution, the values visited by the algorithm with corresponding deviances are returned. } \item{obj}{ the starting regression model where the segmented relationships have to be added. } \item{mfExt}{ the model frame. } } \details{ The functions call iteratively \code{lm.wfit} (or \code{glm.fit}) with proper design matrix depending on \code{XREG}, \code{Z} and \code{PSI}. \code{seg.lm.fit.boot} (and \code{seg.glm.fit.boot}) implements the bootstrap restarting idea discussed in Wood (2001). } \value{ A list of fit information. } \references{ Wood, S. N. (2001) Minimizing model fitting objectives that contain spurious local minima by bootstrap restarting. \emph{Biometrics} \bold{57}, 240--244. } \author{ Vito Muggeo } \note{ These functions should usually not be used directly by the user. } %% ~Make other sections like Warning with \section{Warning }{....} ~ \seealso{ \code{\link{segmented.lm}}, \code{\link{segmented.glm}} } \examples{ ##See ?segmented } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{regression} \keyword{nonlinear } segmented/man/slope.Rd0000644000176200001440000000744113157724622014430 0ustar liggesusers\name{slope} \alias{slope} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Slope estimates from segmented relationships } \description{ Computes the slopes of each `segmented' relationship in the fitted model. } \usage{ slope(ogg, parm, conf.level = 0.95, rev.sgn=FALSE, var.diff=FALSE, APC=FALSE, digits = max(3, getOption("digits") - 3)) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{ogg}{ an object of class "segmented", returned by any \code{segmented} method. } \item{parm}{ the segmented variable whose slopes have to be computed. If missing all the segmented variables are considered. } \item{conf.level}{ the confidence level required. } \item{rev.sgn}{vector of logicals. The length should be equal to the length of \code{parm}, but it is recycled otherwise. When \code{TRUE} it is assumed that the current \code{parm} is `minus' the actual segmented variable, therefore the sign is reversed before printing. This is useful when a null-constraint has been set on the last slope.} \item{var.diff}{logical. If \code{var.diff=TRUE} and there is a single segmented variable, the computed standard errors are based on a sandwich-type formula of the covariance matrix. See Details in \code{\link{summary.segmented}}.} \item{APC}{logical. If \code{APC=TRUE} the `annual percent changes', i.e. \eqn{100\times(\exp(\beta)-1)}{100*(exp(b)-1)}, are computed for each interval (\eqn{\beta}{b} is the slope). Only point estimates and confidence intervals are returned. } \item{digits}{controls number of digits printed in output.} } \details{ To fit broken-line relationships, \code{segmented} uses a parameterization whose coefficients are not the slopes. Therefore given an object \code{"segmented"}, \code{slope} computes point estimates, standard errors, t-values and confidence intervals of the slopes of each segmented relationship in the fitted model. } \value{ \code{slope} returns a list of matrices. Each matrix represents a segmented relationship and its number of rows equal to the number of segments, while five columns summarize the results. } \references{ Muggeo, V.M.R. (2003) Estimating regression models with unknown break-points. \emph{Statistics in Medicine} \bold{22}, 3055--3071. Muggeo, V.M.R. (2010) Comment on `Estimating average annual per cent change in trend analysis' by Clegg et al., Statistics in Medicine; 28, 3670-3682. \emph{Statistics in Medicine}, \bold{29}, 1958--1960. } \author{Vito M. R. Muggeo, \email{vito.muggeo@unipa.it} } \note{The returned summary is based on limiting Gaussian distribution for the model parameters involved in the computations. Sometimes, even with large sample sizes such approximations are questionable (e.g., with small difference-in-slope parameters) and the results returned by \code{slope} might be unreliable. Therefore is responsability of the user to gauge the applicability of such asymptotic approximations. Anyway, the t values may be not assumed for testing purposes and they should be used just as guidelines to assess the estimate uncertainty. } \seealso{See also \code{\link{davies.test}} and \code{\link{pscore.test}} to test for a nonzero difference-in-slope parameter. } \examples{ set.seed(16) x<-1:100 y<-2+1.5*pmax(x-35,0)-1.5*pmax(x-70,0)+rnorm(100,0,3) out<-glm(y~1) out.seg<-segmented(out,seg.Z=~x,psi=list(x=c(20,80))) ## the slopes of the three segments.... slope(out.seg) rm(x,y,out,out.seg) # ## an heteroscedastic example.. set.seed(123) n<-100 x<-1:n/n y<- -x+1.5*pmax(x-.5,0)+rnorm(n,0,1)*ifelse(x<=.5,.4,.1) o<-lm(y~x) oseg<-segmented(o,seg.Z=~x,psi=.6) slope(oseg) slope(oseg,var.diff=TRUE) #better CI } \keyword{ regression } \keyword{ htest } segmented/man/lines.segmented.Rd0000644000176200001440000000564113157724622016372 0ustar liggesusers\name{lines.segmented} \alias{lines.segmented} %- Also NEED an '\alias' for EACH other topic documented here. \title{ Bars for interval estimate of the breakpoints } \description{ Draws bars relevant to breakpoint estimates (point estimate and confidence limits) on the current device } \usage{ \method{lines}{segmented}(x, term, bottom = TRUE, shift=TRUE, conf.level = 0.95, k = 50, pch = 18, rev.sgn = FALSE, ...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{x}{ an object of class \code{segmented}. } \item{term}{ the segmented variable of the breakpoints being drawn. It may be unspecified when there is a single segmented variable.} \item{bottom}{ logical, indicating if the bars should be plotted at the bottom (\code{TRUE}) or at the top (\code{FALSE}).} \item{shift}{ logical, indicating if the bars should be `shifted' on the y-axis before plotting. Useful for multiple breakpoints with overlapped confidence intervals.} \item{conf.level}{ the confidence level of the confidence intervals for the breakpoints. } \item{k}{ a positive integer regulating the vertical position of the drawn bars. See Details. } \item{pch}{ either an integer specifying a symbol or a single character to be used in plotting the point estimates of the breakpoints. See \code{\link{points}}. } \item{rev.sgn}{ should the signs of the breakpoint estimates be changed before plotting? see Details. } \item{\dots}{ further arguments passed to \code{\link{segments}}, for instance `col' that can be a vector. } } \details{ \code{lines.segmented} simply draws on the current device the point estimates and relevant confidence limits of the estimated breakpoints from a "segmented" object. The y coordinate where the bars are drawn is computed as \code{usr[3]+h} if \code{bottom=TRUE} or \code{usr[4]-h} when \code{bottom=FALSE}, where \code{h=(usr[4]-usr[3])/abs(k)} and \code{usr} are the extremes of the user coordinates of the plotting region. Therefore for larger values of \code{k} the bars are plotted on the edges. The argument \code{rev.sgn} allows to change the sign of the breakpoints before plotting. This may be useful when a null-right-slope constraint is set. } %\value{ % ~Describe the value returned % If it is a LIST, use % \item{comp1 }{Description of 'comp1'} % \item{comp2 }{Description of 'comp2'} % ... %} %\references{ ~put references to the literature/web site here ~ } %\author{ ~~who you are~~ } %\note{ ~~further notes~~ % ~Make other sections like Warning with \section{Warning }{....} ~ %} \seealso{ \code{\link{plot.segmented}} to plot the fitted segmented lines, and \code{\link{points.segmented}} to add the fitted joinpoints. } \examples{ ## See ?plot.segmented } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } \keyword{ nonlinear } segmented/man/davies.test.Rd0000644000176200001440000001407313157724622015536 0ustar liggesusers\name{davies.test} \alias{davies.test} \title{ Testing for a change in the slope } \description{ Given a generalized linear model, the Davies' test can be employed to test for a non-constant regression parameter in the linear predictor. } \usage{ davies.test(obj, seg.Z, k = 10, alternative = c("two.sided", "less", "greater"), type=c("lrt","wald"), values=NULL, dispersion=NULL) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{obj}{ a fitted model typically returned by \code{glm} or \code{lm}. Even an object returned by \code{segmented} can be set (e.g. if interest lies in testing for an additional breakpoint).} \item{seg.Z}{ a formula with no response variable, such as \code{seg.Z=~x1}, indicating the (continuous) segmented variable being tested. Only a single variable may be tested and an error is printed when \code{seg.Z} includes two or more terms. } \item{k}{ number of points where the test should be evaluated. See Details. } \item{alternative}{ a character string specifying the alternative hypothesis. } \item{type}{ the test statistic to be used (only for GLM, default to lrt. Ignored if \code{obj} is a simple linear model.} \item{values}{ optional. The evaluation points where the Davies approximation is computed. See Details for default values.} \item{dispersion}{ the dispersion parameter for the family to be used to compute the test statistic. When \code{NULL} (the default), it is inferred from \code{obj}. Namely it is taken as \code{1} for the Binomial and Poisson families, and otherwise estimated by the residual Chi-squared statistic (calculated from cases with non-zero weights) divided by the residual degrees of freedom. } } \details{ \code{davies.test} tests for a non-zero difference-in-slope parameter of a segmented relationship. Namely, the null hypothesis is \eqn{H_0:\beta=0}{H_0:beta=0}, where \eqn{\beta}{beta} is the difference-in-slopes, i.e. the coefficient of the segmented function \eqn{\beta(x-\psi)_+}{beta*(x-psi)_+}. The hypothesis of interest \eqn{\beta=0}{beta=0} means no breakpoint. Roughtly speaking, the procedure computes \code{k} `naive' (i.e. assuming fixed and known the breakpoint) test statistics for the difference-in-slope, seeks the `best' value and corresponding naive p-value (according to the alternative hypothesis), and then corrects the selected (minimum) p-value by means of the \code{k} values of the test statistic. If \code{obj} is a LM, the Davies (2002) test is implemented. This approach works even for small samples. If \code{obj} represents a GLM fit, relevant methods are described in Davies (1987), and the Wald or the Likelihood ratio test statistics can be used, see argument \code{type}. This is an asymptotic test. The \code{k} evaluation points are \code{k} equally spaced values between the second and the second-last values of the variable reported in \code{seg.Z}. \code{k} should not be small; I find no important difference for \code{k} larger than 10, so default is \code{k=10}. } \value{ A list with class '\code{htest}' containing the following components: \item{method}{title (character)} \item{data.name}{the regression model and the segmented variable being tested} \item{statistic }{the point within the range of the covariate in \code{seg.Z} at which the maximum (or the minimum if \code{alternative="less"}) occurs} \item{parameter }{number of evaluation points} \item{p.value }{the adjusted p-value} \item{process}{a two-column matrix including the evaluation points and corresponding values of the test statistic} } \references{ Davies, R.B. (1987) Hypothesis testing when a nuisance parameter is present only under the alternative. \emph{Biometrika} \bold{74}, 33--43. Davies, R.B. (2002) Hypothesis testing when a nuisance parameter is present only under the alternative: linear model case. \emph{Biometrika} \bold{89}, 484--489. } \author{ Vito M.R. Muggeo } \note{ Strictly speaking, the Davies test is not confined to the segmented regression; the procedure can be applied when a nuisance parameter vanishes under the null hypothesis. The test is slightly conservative, as the computed p-value is actually an upper bound. Results should change slightly with respect to previous versions where the evaluation points were computed as \code{k} equally spaced values between the second and the second last observed values of the segmented variable. } \section{Warning }{ The Davies test is \emph{not} aimed at obtaining the estimate of the breakpoint. The Davies test is based on \code{k} evaluation points, thus the value returned in the \code{statistic} component (and printed as "'best' at") is the best among the \code{k} points, and typically it will differ from the maximum likelihood estimate returned by \code{segmented}. Use \code{\link{segmented}} if you are interested in the point estimate. To test for a breakpoint in \emph{linear} models with small samples, it is suggested to use \code{davies.test()} with objects of class "lm". If \code{obj} is a \code{"glm"} object with gaussian family, \code{davies.test()} will use an approximate test resulting in smaller p-values when the sample is small. However if the sample size is large (n>300), the exact Davies (2002) upper bound cannot be computed (as it relies on \code{gamma()} function) and the \emph{approximate} upper bound of Davies (1987) is returned. } %%\section{Warning }{Currently \code{davies.test} does not work if the fitted model \code{ogg} %% does not include the segmented variable \code{term} being tested.} \seealso{See also \code{\link{pscore.test}} which is more power, especially when the signal-to-noise ratio is low. } \examples{ \dontrun{ set.seed(20) z<-runif(100) x<-rnorm(100,2) y<-2+10*pmax(z-.5,0)+rnorm(100,0,3) o<-lm(y~z+x) davies.test(o,~z) davies.test(o,~x) o<-glm(y~z+x) davies.test(o,~z) #it works but the p-value is too small.. } } \keyword{ htest } segmented/man/summary.segmented.Rd0000644000176200001440000000723413157724622016755 0ustar liggesusers\name{summary.segmented} \alias{summary.segmented} \alias{print.summary.segmented} \title{ Summarizing model fits for segmented regression } \description{ summary method for class \code{segmented}. } \usage{ \method{summary}{segmented}(object, short = FALSE, var.diff = FALSE, ...) \method{print}{summary.segmented}(x, short=x$short, var.diff=x$var.diff, digits = max(3, getOption("digits") - 3), signif.stars = getOption("show.signif.stars"),...) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{object}{ Object of class "segmented". } \item{short}{ logical indicating if the `short' summary should be printed. } \item{var.diff}{ logical indicating if different error variances should be computed in each interval of the segmented variable, see Details. } \item{x}{a \code{summary.segmented} object produced by \code{summary.segmented()}.} \item{digits}{controls number of digits printed in output.} \item{signif.stars}{logical, should stars be printed on summary tables of coefficients?} \item{\dots}{ further arguments. } } \details{ If short=TRUE only coefficients of the segmented relationships are printed. If var.diff=TRUE and there is only one segmented variable, different error variances are computed in the intervals defined by the estimated breakpoints of the segmented variable. For the jth interval with nj observations the error variance is estimated via \eqn{RSS_j/(n_j-p)}{RSSj/(nj-p)}, where \eqn{RSS_j} is the residual sum of squares in interval jth, and \eqn{p} are the model parameters. Note \code{var.diff=TRUE} does \emph{not} affect the parameter estimation which is performed via ordinary (and not weighted) least squares. However if \code{var.diff=TRUE} the variance-covariance matrix of the estimates is computed via the sandwich formula, \deqn{(X^TX)^{-1}X^TVX(X^TX)^{-1}}{(X'X)^{-1}X'VX(X'X)^{-1}} where V is the diagonal matrix including the different error variance estimates. Standard errors are the square root of the main diagonal of this matrix. } \value{ A list (similar to one returned by \code{segmented.lm} or \code{segmented.glm}) with additional components: \item{psi }{estimated break-points and relevant (approximate) standard errors} \item{Ttable }{estimates and standard errors of the model parameters. This is similar to the matrix \code{coefficients} returned by \code{summary.lm} or \code{summary.glm}, but without the rows corresponding to the breakpoints. Even the p-values relevant to the difference-in-slope parameters have been replaced by NA, since they are meaningless in this case, see \code{\link{davies.test}}.} \item{gap}{estimated coefficients, standard errors and t-values for the `gap' variables} \item{cov.var.diff}{if \code{var.diff=TRUE}, the covaraince matrix accounting for heteroscedastic errors.} \item{sigma.new}{if \code{var.diff=TRUE}, the square root of the estimated error variances in each interval.} \item{df.new}{if \code{var.diff=TRUE}, the residual degrees of freedom in each interval.} } %\references{ ~put references to the literature/web site here ~ } \author{ Vito M.R. Muggeo } \seealso{ \code{\link{print.segmented}}, \code{\link{davies.test}} } \examples{ ##continues example from segmented() # summary(segmented.model,short=TRUE) ## an heteroscedastic example.. # set.seed(123) # n<-100 # x<-1:n/n # y<- -x+1.5*pmax(x-.5,0)+rnorm(n,0,1)*ifelse(x<=.5,.4,.1) # o<-lm(y~x) # oseg<-segmented(o,seg.Z=~x,psi=.6) # summary(oseg,var.diff=TRUE)$sigma.new } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } segmented/man/pscore.test.rd0000644000176200001440000001123313157724622015611 0ustar liggesusers\name{pscore.test} \alias{pscore.test} \title{ Testing for existence of one breakpoint} \description{ Given a (generalized) linear model, the (pseudo) Score statistic tests for the existence of one breakpoint. } \usage{ pscore.test(obj, seg.Z, k = 10, alternative = c("two.sided", "less", "greater"), values=NULL, dispersion=NULL, df.t=NULL, more.break=FALSE) } %- maybe also 'usage' for other objects documented here. \arguments{ \item{obj}{ a fitted model typically returned by \code{glm} or \code{lm}. Even an object returned by \code{segmented} can be set. Offset and weights are allowed.} \item{seg.Z}{ a formula with no response variable, such as \code{seg.Z=~x1}, indicating the (continuous) segmented variable being tested. Only a single variable may be tested and an error is printed when \code{seg.Z} includes two or more terms. \code{seg.Z} can be omitted if i)\code{obj} is a segmented fit with a single segmented covariate (and that variable is taken), or ii)if it is a "lm" or "glm" fit with a single covariate (and that variable is taken).} \item{k}{ optional. Number of points used to compute the pseudo Score statistic. See Details. } \item{alternative}{ a character string specifying the alternative hypothesis. } \item{values}{ optional. The evaluation points where the Score test is computed. See Details for default values.} \item{dispersion}{ optional. the dispersion parameter for the family to be used to compute the test statistic. When \code{NULL} (the default), it is inferred from \code{obj}. Namely it is taken as \code{1} for the Binomial and Poisson families, and otherwise estimated by the residual Chi-squared statistic in the model \code{obj} (calculated from cases with non-zero weights divided by the residual degrees of freedom).} \item{df.t}{ optional. The degress-of-freedom used to compute the p-value. When \code{NULL}, the df extracted from \code{obj} are used.} \item{more.break}{ optional logical. If \code{obj} is a segmented fit, \code{more.break=FALSE} tests for the actual breakpoint in \code{obj}, while \code{more.break=TRUE} tests for an \emph{additional} breakpoint. Ignored when \code{obj} is not a segmented fit.} } \details{ \code{pscore.test} tests for a non-zero difference-in-slope parameter of a segmented relationship. Namely, the null hypothesis is \eqn{H_0:\beta=0}{H_0:beta=0}, where \eqn{\beta}{beta} is the difference-in-slopes, i.e. the coefficient of the segmented function \eqn{\beta(x-\psi)_+}{beta*(x-psi)_+}. The hypothesis of interest \eqn{\beta=0}{beta=0} means no breakpoint. Simulation studies have shown that such Score test is more powerful than the Davies test (see reference) when the alternative hypothesis is `one changepoint'. The \code{dispersion} value, if unspecified, is taken from \code{obj}. If \code{obj} represents the fit under the null hypothesis (no changepoint), the dispersion parameter estimate will be usually larger, leading to a (potentially severe) loss of power. The \code{k} evaluation points are \code{k} equally spaced values in the range of the segmented covariate. \code{k} should not be small. Specific values can be set via \code{values}. However I have found no important difference due to number and location of the evaluation points, thus default is \code{k=10} equally-spaced points. } \value{ A list with class '\code{htest}' containing the following components: \item{method}{title (character)} \item{data.name}{the regression model and the segmented variable being tested} \item{statistic }{the point within the range of the covariate in \code{seg.Z} at which the maximum (or the minimum if \code{alternative="less"}) occurs} \item{parameter }{number of evaluation points} \item{p.value }{the p-value} \item{process}{the alternative hypothesis set} } \references{ Muggeo, V.M.R. (2016) Testing with a nuisance parameter present only under the alternative: a score-based approach with application to segmented modelling. \emph{J of Statistical Computation and Simulation}, \bold{86}, 3059--3067. } \author{ Vito M.R. Muggeo } \seealso{See also \code{\link{davies.test}}. } \examples{ \dontrun{ set.seed(20) z<-runif(100) x<-rnorm(100,2) y<-2+10*pmax(z-.5,0)+rnorm(100,0,3) o<-lm(y~z+x) #testing for one changepoint #use the simple null fit pscore.test(o,~z) #compare with davies.test(o,~z).. #use the segmented fit os<-segmented(o, ~z) pscore.test(os,~z) #smaller p-value, as it uses the dispersion under the alternative (from 'os') #test for the 2nd breakpoint in the variable z pscore.test(os,~z, more.break=TRUE) } } \keyword{ htest } segmented/man/broken.line.Rd0000644000176200001440000000511313157724622015506 0ustar liggesusers\name{broken.line} \alias{broken.line} \title{ Fitted values for segmented relationships} \description{ Given a segmented model (typically returned by a \code{segmented} method), \code{broken.line} computes the fitted values (and relevant standard errors) for each `segmented' relationship. } \usage{ broken.line(ogg, term = NULL, link = TRUE, interc=TRUE, se.fit=TRUE) } \arguments{ \item{ogg}{ A fitted object of class segmented (returned by any \code{segmented} method). } \item{term}{ Three options. A list (whose name should be one of the segmented covariates) including values for which segmented predictions should be computed. A character meaning the name of any segmented covariate in the model. \code{NULL} if the model includes a single segmented covariate. } \item{link}{ Should the predictions be computed on the scale of the link function? Default to \code{TRUE}. } \item{interc}{ Should the model intercept be added? (provided it exists).} \item{se.fit}{ If \code{TRUE} also standard errors for predictions are returned.} } \details{ If \code{term=NULL} or \code{term} is a valid segmented covariate name, predictions for each segmented variable are the relevant fitted values from the model. If \code{term} is a (correctly named) list with numerical values, predictions corresponding to such specified values are computed. If \code{link=FALSE} and \code{ogg} inherits from the class "glm", predictions and standard errors are returned on the response scale. The standard errors come from the Delta method. Argument \code{link} is ignored whether \code{ogg} does not inherit from the class "glm". } \value{ A 2-component (if \code{se.fit=TRUE}) list representing predictions and standard errors for the segmented covariate values. } %\references{ ~put references to the literature/web site here ~ } \author{ Vito M. R. Muggeo } %\note{ %This function will be probably removed in the next versions. See \code{predict.segmented} instead. %} % ~Make other sections like Warning with \section{Warning }{....} ~ %} \seealso{ \code{\link{segmented}}, \code{\link{predict.segmented}}, \code{\link{plot.segmented}}} \examples{ set.seed(1234) z<-runif(100) y<-rpois(100,exp(2+1.8*pmax(z-.6,0))) o<-glm(y~z,family=poisson) o.seg<-segmented(o,seg.Z=~z,psi=.5) \dontrun{plot(z,y)} \dontrun{points(z,broken.line(o.seg,link=FALSE)$fit,col=2)} #just to illustrate, use plot.segmented } % Add one or more standard keywords, see file 'KEYWORDS' in the % R documentation directory. \keyword{ regression } \keyword{ nonlinear }